TECHNICAL WORKING GROUP ON PESTICIDES (TWG)

(Q)UANTITATIVE STRUCTURE ACTIVITY RELATIONSHIP

NAFTA [(Q)SAR] GUIDANCE DOCUMENT

Page 1 of 186
 North American Free Trade Agreement (NAFTA)
Technical Working Group on Pesticides (TWG)

(Quantitative) Structure Activity Relationship [(Q)SAR]

Guidance Document

November, 2012
Contributors
Mary Manibusan, US EPA
Joel Paterson, PMRA
Dr. Ray Kent, US EPA
Dr. Jonathan Chen, US EPA

Dr. Jenny Tao, US EPA

Dr. Edward Scollon, US EPA
Christine Olinger, US EPA
Dr. Patricia Schmieder, US EPA
Dr. Chris Russom, US EPA
Dr. Kelly Mayo, US EPA
Dr. Yin-tak Woo, US EPA
Dr. Thomas Steeger, US EPA
Dr. Edwin Matthews, US FDA
Dr. Sunil Kulkarni, Health Canada

External Peer Reviewers

Kirk Arvidson, US FDA
Mark Bonnell, Environment Canada
Bob Diderich, OECD
Terry Schultz, OECD
Andrew Worth, European Commission – Joint Research Centre

Page 2 of 186
 PREFACE
Integrated Approaches to Testing and Assessment (IATA) and
(Q)SAR
Pesticide regulatory agencies have traditionally relied on extensive in vivo and in
vitro testing to support regulatory decisions on human health and environmental
risks. While this approach has provided strong support for risk management
decisions, there is a clear recognition that it can often require a large number of
laboratory animal studies which can consume significant amounts of resources in
terms of time for testing and evaluation. Even with the significant amounts of
information from standard in vivo and in vitro testing, pesticide regulators are
often faced with questions and issues relating to modes of action for toxicity,
novel toxicities, susceptible populations, and other factors that can be
challenging to address using traditional approaches.

Recognizing the limitations of current testing approaches and the rapid

development of new biochemical and cellular assay systems and computational
predictive methods, pesticide and other regulatory agencies have initiated the
long-term investigation of Integrated Approaches to Testing and Assessment
(IATA). IATA integrate existing knowledge bases on classes of chemicals with
the results of biochemical and cellular assays, computational predictive methods,
exposure studies, and other sources of information to identify requirements for
targeted testing or develop assessment conclusions. In some cases, the
application of IATA could lead to the refinement, reduction, and/or replacement of
selected conventional tests (e.g., animal toxicity tests). IATA also have the
potential to further enhance the understanding of mode/mechanism of action 1
including the consideration of relevant adverse outcome pathways (AOPs) that
provide biological linkages between molecular initiating events to adverse

1
In this context, mode of action for toxicity is the description of key events and processes,
starting with interaction of an agent with the cell through functional and anatomical changes,
resulting in cancer or other health endpoints. Mechanism of action for toxicity is the detailed
molecular description of key events in the induction of cancer or other health endpoints and
represents a more detailed understanding and description of events than is meant by mode of
action. Mode of action for toxicity can also be differentiated from the pesticidal mode of action
which is the specific biochemical or physical effect(s) by which the pesticide kills, inactivates or
otherwise controls pests. Mechanism and mode of action for toxicity are important components of
adverse outcome pathways (AOPs).

Page 3 of 186
outcomes in individual organisms and populations that are the bases for risk
assessments.

The subject of this guidance document, (Quantitative) Structure Activity

Relationships [(Q)SAR], is an important set of predictive tools that can be
considered when applying IATA to pesticide assessments. (Q)SAR represents a
variety of techniques for predicting activities and properties of untested chemicals
based on their structural similarity to chemicals with known activities and
properties. 2 (Q)SAR methods have a long history of use both for the industrial
design and regulatory assessment of pharmaceuticals, pesticides, and other
chemicals. While historical and current applications of (Q)SAR methods have
focused on the prediction of physical-chemical properties and apical endpoints
(e.g., toxicity, ecotoxicity), as IATA are developed and applied to pesticides, a
greater emphasis will be placed on using (Q)SAR to predict key events along the
cascade of obligatory steps toward the adverse outcome in modes of
toxicological action and AOPs (e.g., receptor binding potential, enzyme
activation/inhibition, DNA/protein binding).

The development and application of IATA and (Q)SAR methods to pesticide

assessments is consistent with the United States Environmental Protection
Agency (US EPA) commissioned National Research Council (NRC) report,
Toxicity Testing in the 21st Century: A Vision and a Strategy (NRC, 2007). The
NRC’s vision emphasizes moving away from checklists of conventional toxicity
studies towards integrated approaches using existing knowledge of chemicals
and the results of alternative testing methods, including computational tools such
as (Q)SAR, to identify toxicity pathways and streamline data requirements for
more efficient, and effective, targeted toxicity testing. (Q)SAR has also been
highlighted as an important IATA tool in the report, Integrating Emerging
Technologies into Chemical Safety Assessment, sponsored by Health Canada
and prepared by the Expert Panel on the Integrated Testing of Pesticides of the
Canadian Council of Academies (CCA) (CCA, 2012). The CCA report provides
an update on the status of IATA and IATA tools, and a vision for the evolution of
IATA in the regulatory context.

2
(Q)SAR is the study of the correlation between chemical structure and associated biological
activity, with the ultimate goal of predicting the activity of untested chemicals based on structurally
related compounds with known activity. The parentheses around the “Q” in (Q)SAR indicates that
the term refers to both qualitative predictive tools (i.e., structure-activity relationships (SARs)) and
quantitative predictive methods (quantitative structure-activity relationships (QSARs)). Although
the term (Q)SAR is often used to refer to predictive models, especially computer-based models, it
should be noted that (Q)SAR is actually inclusive of a wide variety of computerized and non-
computerized tools and approaches.

Page 4 of 186
Moving IATA from a long-term vision into mainstream practice for pesticide
assessments will require the development and application of biochemical and
cellular assays, along with the further development and broader application of
existing tools such as (Q)SAR. Towards that end, the United States
Environmental Protection Agency Office of Pesticide Programs (US EPA OPP)
has partnered with the Pest Management Regulatory Agency (PMRA) of Health
Canada to develop common approaches to IATA for the human health and
ecological risk assessment of pesticides. The formalized framework for this
partnership is a North American Free Trade Agreement (NAFTA) Joint Project on
“21st Century Toxicology: Integrated Approaches to Testing and Assessment”.
While this project is intended to cover a broad array of computational toxicity
tools, a key current activity is the development of this NAFTA (Q)SAR guidance
document for pesticide risk assessors.

The primary purpose of this guidance document is to help pesticide evaluators to

evaluate (Q)SAR predictions and to identify the important issues that may be
involved when incorporating predictions in the risk assessment process. The
document is not intended to reproduce or replace the ever-expanding volume of
journal articles, reports, documents, and textbooks on the development and
application of (Q)SAR, but to provide an introduction to the evaluation of (Q)SAR
tools and their application to pesticide regulatory risk assessments. While the
focus of this document is on the application of (Q)SAR to pesticide risk
assessments, the principles and issues described in this document are general
and may also be used for other types of chemical assessments. Regardless of
the scenario to which (Q)SAR is being applied, the peer review process is critical
and relevant to the consistent application of this tool. To that end, appropriate
(Q)SAR experts should be consulted and peer review procedures used to ensure
scientific excellence and rigor.

Page 5 of 186
TABLE OF CONTENTS
PREFACE ................................................................................................................ 3

GLOSSARY ......................................................................................................... 11

1. EXECUTIVE SUMMARY ..................................................................... 18

2. INTRODUCTION ..................................................................................... 22

2.0 Current Applications of (Q)SAR in Pesticide Risk Assessments .. 22

2.0.1 United States Environmental Protection Agency,

Office of Pesticide Programs (US EPA OPP) .............................22

2.0.1.1 Application of (Q)SAR to Pesticide Metabolites

and Degradates .............................................................22

2.0.1.2 Application of (Q)SAR to Antimicrobial Agents ...............23

2.0.1.3 Application of (Q)SAR to Ecological Risks

from Pesticides ..............................................................24

2.0.2 Pest Management Regulatory Agency (PMRA),

Health Canada...........................................................................25

2.1 Other Regulatory Applications of (Q)SAR....................................... 25

2.1.1 US EPA, Office of Pollution Prevention and Toxics (OPPT) .......26

2.1.2 US EPA, Office of Research and Development .........................26

2.1.3 US FDA, Office of Food Additive Safety.....................................27

2.1.4 Health Canada and Environment Canada..................................27

2.1.5 Organization for Economic Cooperation and Development

(OECD)......................................................................................27
2.1.6 European Commission Joint Research Centre ..........................28

2.2 Purpose of the NAFTA (Q)SAR Guidance Document ................... 29

Page 6 of 186
3. BACKGROUND INFORMATION ON (Q)SAR ............................ 32

3.0 Introduction ........................................................................................ 32

3.1 Definition of (Q)SAR .......................................................................... 32

3.1.1 Defining Similarity ......................................................................33

3.2 Types of (Q)SAR Approaches .......................................................... 35

3.2.1 Analogs .....................................................................................36

3.2.2 Chemical Categories .................................................................36

3.2.3 (Q)SAR Models .........................................................................37

3.3 Importance of Data Quality in (Q)SAR Model Development ......... 39

3.4 Importance of Mode/Mechanism of Action

in (Q)SAR Model Development ........................................................ 40

3.5 Examples of (Q)SAR Tools and Their Applications ....................... 41

3.6 Summary ............................................................................................ 43

4. PROBLEM FORMULATION AND (Q)SAR .................................. 44

4.0 Introduction ........................................................................................ 44

4.1 Assessment Context that (Q)SAR is being Applied to ................... 45

4.2 Characteristics of the Pesticide that is the Subject

of the Prediction ................................................................................. 46

4.2.1 Chemical Identifiers and Mixtures ..............................................47

4.2.2 Transformation, Degradation, and Metabolism ..........................47

4.2.3 Isomers and Structural Representations for (Q)SAR..................48

4.3 Characteristics of the (Q)SAR Tool and the Prediction ................. 50

4.4 Empirical Data Including Information on Mode of Action ............... 53

4.5 Summary ............................................................................................ 56

Page 7 of 186
5. EVALUATING THE ADEQUACY OF (Q)SAR
PREDICTIONS ........................................................................................ 57
5.0 Introduction ........................................................................................ 57

5.1 Scientific Validity of the (Q)SAR Tool .............................................. 59

5.1.1 OECD (Q)SAR Validation Principles ..........................................59

5.1.1.1 Principle 1 — Defined Endpoint......................................61

5.1.1.2 Principle 2 — Unambiguous Algorithm ...........................62

5.1.1.3 Principle 3 — Defined Domain of Applicability ................63

5.1.1.4 Principle 4 — Appropriate Measures of Goodness-of-fit,

Robustness, and Predictivity ..........................................64

5.1.1.5 Principle 5 — Defined Mechanism of Action,

if Possible ......................................................................67

5.2 Applicability of the (Q)SAR Tool to the Pesticide ........................... 68

5.3 Relevance of the (Q)SAR Tool to the Assessment Context .......... 69

5.4 Reliability of the (Q)SAR Prediction ................................................. 70

5.4.1 Relationship of the Pesticide to the Domain of Applicability

of the (Q)SAR Tool ....................................................................70

5.4.2 Strengths and Limitations of the (Q)SAR Tool ...........................72

5.4.3 Prediction Results and How They are Interpreted ......................73

5.4.4 Predictive Performance of the (Q)SAR Tool

for Similar Chemicals .................................................................75

5.4.5 Other Available Information .......................................................75

5.5 Documentation of (Q)SAR Tool and Prediction .............................. 77

5.5.1 General Types of Information ....................................................77

5.5.2 Documentation of (Q)SAR Predictions — EC QMRF

and QPRF .................................................................................78

5.6 Summary ............................................................................................ 79

Page 8 of 186
6. COMBINING INFORMATION FROM MULTIPLE
PREDICTIONS ........................................................................................ 80
6.0 Introduction ........................................................................................ 80

6.1 Approaches to Combining Multiple Predictions .............................. 81

6.2 Advantages and Disadvantages of Combining Predictions .......... 82

6.2.1 Advantages of Combining Predictions .......................................82

6.2.2 Disadvantages of Combining Predictions...................................84

6.3 Selecting (Q)SAR Tools for Multiple Predictions ............................ 85

6.4 Evaluation of Multiple Predictions .................................................... 86

6.5 Summary ............................................................................................ 88

7. INTEGRATION OF (Q)SAR PREDICTIONS

INTO HAZARD ASSESSMENTS ..................................................... 89
7.0 Introduction ........................................................................................ 89

7.1 Incorporating (Q)SAR in Hazard Characterizations: Overview ..... 90

7.2 Problem Formulation and Adequacy Determination ...................... 91

7.3 Evaluating Empirical Data versus (Q)SAR Predictions.................. 92

7.4 Mode of Action Consideration .......................................................... 95

7.5 Overall Weight of Evidence .............................................................. 95

7.6 Hazard Characterization and Risk Communication ....................... 96

7.7 Summary ............................................................................................ 97

8. CONCLUSIONS AND FUTURE VISION FOR (Q)SAR AND

PESTICIDES ............................................................................................ 99
8.0 Toxicity Testing in the 21st Century:
Shift in the Risk Assessment Paradigm .......................................... 99

8.1 Weight of Evidence Approach: Biological Plausibility .................. 100

Page 9 of 186
 8.2 Adverse Outcome Pathway: Conceptual Framework .................. 101

8.3 Expert Scientific Judgment and Peer Review ............................... 103

REFERENCES .................................................................................................. 105

Appendix I Where to Go to Learn More About (Q)SAR .......... 122

Appendix II European Commission QSAR Model

Reporting Format (QMRF) and QSAR
Prediction Reporting Format (QPRF) ..................... 129

Appendix III Examples of the use of (Q)SAR

in the Assessment of Pesticides
and Other Chemicals ...................................................... 133

LIST OF TABLES
Table 3–1: Pesticidal Model of Action and Associated Chemical Class for
a Select Group of Insecticides .............................................................. 34

Table 5–1: Recommended General Types of Information to Include when

Documenting (Q)SAR Predictions ........................................................ 78

Table 6–1: Consensus Modeling Interpretation Criteria ....................................... 81

LIST OF FIGURES
Figure 2–0: (Q)SAR Guidance Document Schematic........................................... 31

Figure 3–1: Example of Structural Analogs ........................................................... 33

Figure 3–2: Measures that can be Described in Similarity Indices ...................... 35

Figure 3-3: A Schematic of a Chemical Category Matrix Table ........................... 37

Figure 4–1: Fenvalerate Racemic Mixture ............................................................. 49

Figure 5–1: Evaluating the Adequacy of a (Q)SAR Prediction for a Pesticide ... 59

Figure 7–1: Weight of Evidence Analysis: Integration of (Q)SAR Predictions

and Empirical Data ................................................................................. 91

Figure 8–1: Adverse Outcome Pathway Diagram ............................................... 103

Page 10 of 186
GLOSSARY
This Glossary section is intended to provide additional explanation for common
scientific terms which are presented in order to enhance communication between
(Q)SAR experts and users of (Q)SAR models, particularly in the field of
pesticides.

There are two portions in the glossary — abbreviations (acronyms) and terms
with more detailed explanations.

ABBREVIATIONS

A/I ratio Ratio of active to inactive chemicals

ACC American Chemistry Council

ADME Absorption, distribution, metabolism, and elimination

AOP Adverse Outcome Pathway

BCF Bioconcentration Factor

CAS Chemical Abstract Service

CCA Council of Canadian Academies

CEPA Canadian Environmental Protection Act

CFR Code of Federal Regulations

DER Data evaluation record

DSL Domestic Substances List (Canada)

EC European Commission

ECHA European Chemicals Agency

EDSP Endocrine Disruptor Screening Program (US EPA)

EEC European Economic Community

EFSA European Food Safety Authority

ER Estrogen Receptor

EU European Union

FAO Food and Agriculture Organization (of the United Nations)

FDA (US) Food and Drug Administration

FQPA Food Quality Protection Act

Page 11 of 186
HPV High Production Volume Chemicals Program (US EPA)

IATA Integrated Approaches to Testing and Assessment

ILSI International Life Sciences Institute

TM
InChI IUPAC International Chemical Identifier

IPCS International Program on Chemical Safety

IUPAC International Union of Pure and Applied Chemistry

JRC Joint Research Centre (European Commission)

kNN k Nearest Neighbor

Kow Octanol-water partition coefficient

Kp Permeability coefficient through the skin for a chemical in water

LMO Leave many out

LOAEL Lowest Observed Adverse Effect Level

Log P Logarithm to the base 10 of the 1-octanol/water partition coefficient, also Log Kow

LOO Leave one out

MED Mid-Continent Ecology Division (US EPA ORD)

MIE Molecular initiating event

MOA Mode of (toxicological) Action

NAFTA North American Free Trade Agreement

NAS (US) National Academy of Sciences

NHEERL National Health and Environmental Effects Research Laboratory (US EPA ORD)

NOAEL No Observed Adverse Effect Level

NRC (US) National Research Council

OECD Organization for Economic Co-operation and Development

OFAS Office of Food Additive Safety (US FDA)

OPP US EPA Office of Pesticide Programs

OPPT US EPA Office of Pollution Prevention and Toxics

ORD US EPA Office of Research and Development

PBT Persistent, bioaccumulative and toxic

PCKOC Organic Carbon Partition Coefficient model components within the EPI Suite
(US EPA)

PMN Premanufacturing notification

Page 12 of 186
PMRA Pest Management Regulatory Agency (Health Canada)
2
Q Cross-validated correlation coefficient
2
Q ext External correlation coefficient

(Q)SAR Quantitative structure-activity relationship or structure-activity relationship

QSAR Quantitative structure-activity relationship

QMRF QSAR Model Reporting Format (European Commission)

QPRF QSAR Prediction Reporting Format (European Commission)

QSPR Quantitative structure-property relationship

2
R Coefficient of determination

REACH Registration, Evaluation, Authorization (and Restriction) of Chemicals legislation

(European Union)

SAR Structure-activity relationship

SDF Structure Data Format

SEE Standard error of the estimate

Spress Cross-validated standard error of prediction

TSCA US Toxic Substances Control Act

TTC Threshold of Toxicological Concern

WHO World Health Organization

TERMS

Adverse outcome A conceptual construct that portrays existing knowledge concerning the
pathway (AOP) linkage between a direct molecular initiating event and an adverse
outcome at a biological level of organization relevant to risk assessment.

Algorithm A sequence of instructions for carrying out a defined task. Typically the
instructions are mathematical equations or computer code.

Analog A chemical compound that has a similar structure and similar chemical
properties to those of another compound, but differs from it by one or a
few atoms or functional groups.

Apical endpoint Observable effects of exposure to a toxic chemical in a test animal. The
effects reflect relatively gross changes in animals after substantial
durations of exposure.

Chemical category A group of chemicals with similar physicochemical, human health, or

ecotoxicological properties usually resulting from structural similarity.

Congeneric series A group of chemicals with a common base structure (e.g. aliphatic
alcohols) but differing in the arrangement of common substituents. The
polychlorinated biphenyls are considered a congeneric series.

Page 13 of 186
Cross-validation A statistical technique for assessing the predictive ability of a QSAR by
the removal of different proportions of the chemicals from the training
set, developing a QSAR on the remaining chemicals and using that
QSAR to predict the activity of those removed. This procedure is
repeated a number of times, so that a number of statistics can be derived
from the comparison of predicted data with the known data.

Data mining A collective term that refers to all procedures (informatic and statistical)
that are applied to large heterogeneous data sets, in order to develop a
data matrix amenable to statistical methods.

Descriptor A quantifiable physical, chemical, or structural property specific to a

chemical that can be correlated with an endpoint under investigation.
There are three main categories of descriptors: hydrophobic, steric, and
electronic. Steric descriptors are those relating to molecular size or
shape. Electronic descriptors are those concerning molecular
interactions such as hydrogen bonding and dipole forces and they
include quantum mechanical and quantum chemical descriptors such as
atomic charge. Hydrophobic descriptors such as Log P are those relating
to the tendencies of chemical to partition between hydrophilic (aqueous)
and hydrophobic/lipophilic (lipid) phases.

Domain of Applicability The domain of applicability of a (Q)SAR model is the chemical structure
and response space in which the model makes predictions with a given
reliability. It can be thought of as a theoretical region in multi-dimensional
space in which the model is expected to make reliable predictions. It
depends on the nature of the chemicals in the training set, and the
method used to develop the model and helps the user of the model to
judge whether the prediction for a new chemical is reliable or not.

EC50 Half Maximal Effective Concentration. Statistically derived concentration

of a substance expected to induce a response halfway between baseline
and maximum effect.

Endpoint The measure of a biological effect, e.g., LC50 or EC50. A large number of
endpoints are used in regulatory assessments of chemicals. These
include lethality, carcinogenicity, immunological responses, organ
effects, developmental and reproductive effects, etc. In (Q)SAR analysis,
it is important to develop models for individual toxic endpoints.

Expert system A formalized system (often computer based) that utilizes a

knowledgebase of structure-activity relationships accumulated from
human experts. The knowledgebase is applied using a set of expert rules
to derive predictions of biological activity for chemicals of interest based
on the presence of specific chemical structures.

Page 14 of 186
External validation A validation exercise in which the chemical structures selected for
inclusion in the test set are different from those included in the training
set, but which should be representative of the same chemical domain.
The QSAR model developed by using the training set chemicals is then
applied to the test set chemicals in order to assess the predictive ability
of the model.

Functional group A molecular moiety that imparts certain characteristics to a molecule,

– – –
e.g., hydroxyl (OH ), amino (NH2 ), or nitro (NO2 ). When only a limited
number of functional groups are present, they may be the primary basis
for the specific chemical, physical or biological characteristics of a
chemical. However, for complex chemicals with many functional groups,
the simple interactions associated with individual functional groups may
not be reliable predictors of chemical behavior unless one functional
group predominates for the particular activity.

Genetic algorithm A statistical method that selects the best combination of descriptors to
describe a given property, modeled on the principle of the survival of the
fittest (best) in the breeding of organisms.

In silico An expression that means “performed on computer or via computer

simulation.”

LC50 Median Lethal Concentration. Statistically derived concentration of a

substance expected to cause death in 50% of test animals, usually
expressed as the weight of substance per weight or volume of water, air
or feed, e.g., mg/l, mg/kg or ppm.

LD50 Median Lethal Dose. Statistically derived single dose causing death in
50% of test animals when administered by the route indicated (oral,
dermal, inhalation), expressed as a weight of substance per unit weight
of animal, e.g., mg/kg.

Lipinski’s rule of 5 A rule of thumb developed by Christopher Lipinski for evaluating whether
the properties of a chemical are likely to make it an orally active drug in
humans. The rule states that, in general, an orally active drug has no
more than one violation of the following criteria: not more than 5
hydrogen bond donors, not more than 10 hydrogen bond acceptors, a
molecular weight under 500, and an octanol-water partition coefficient
less than 5.

Mechanism of Action The detailed molecular description of key events in the induction of
(Toxicity) cancer or other health endpoints. Mechanism of action for toxicity
represents a more detailed understanding and description of events than
is meant by mode of action. Mechanism of action of toxicity is an
important component of an adverse outcome pathway (AOP).

Page 15 of 186
Mode of Action The mode of action of a pesticide refers to the specific biochemical or
(Pesticide) physical effect(s) by which the pesticide kills, inactivates, or otherwise
controls pests.

Mode of Action The description of key events and processes, starting with interaction of
(Toxicity) an agent with the cell through functional and anatomical changes,
resulting in cancer or other health endpoints. Mode of action for toxicity is
an important component of an adverse outcome pathway (AOP).

OECD QSAR Toolbox The OECD QSAR Toolbox is a software application intended to be used
by governments, chemical industry, and other stakeholders in filling gaps
in (eco)toxicity data needed for assessing the hazards of chemicals. The
Toolbox incorporates information and tools from various sources in a
logical workflow. Crucial to this workflow is the grouping of chemicals into
chemical categories ( http://www.oecd.org/document/54/
0,3746,en_2649_34379_42923638_1_1_1_1,00.html ).

Outlier A data point that is far removed from other members of the dataset.
Typically, the outlier of a (Q)SAR model has a cross-validated
standardized residual greater than three standard deviation units.

Partition coefficient The ratio of equilibrium concentrations of a chemical distributed between

two immiscible solvents. Frequently octanol and water are used to mimic
a chemical distributing between lipid and aqueous phases in an
organism, normally expressed as a logarithm to base 10, i.e., Log Kow, or
Log P, a descriptor of hydrophobicity.

Point of Departure Commonly abbreviated POD, the point of departure is the dose-response
point that marks the beginning of a low dose extrapolation. This point is
often the lower bound on an observed incidence or on an estimated
incidence from a dose-response model.

Predictivity A measure of a model’s ability to make reliable predictions for chemical

structures not included in the training set of the model.

Read-across Endpoint information for one or more chemicals (the source chemical(s))
is used to predict the same endpoint for another chemical (the target
chemical), which is considered to be “similar” in some way (usually on
the basis of structural similarity or similar mode or mechanisms of
action). Sometimes, it is also referred to as “data bridging.” In principle,
read-across can be used to estimate physicochemical properties,
toxicity, environmental fate, and ecotoxicity. For any of these endpoints,
it may be performed in a qualitative or quantitative manner.

QSAR Quantitative structure-activity relationship — a quantitative relationship

between an endpoint (biological activity, e.g., toxicity) and one or more
descriptors associated with the endpoint/activity.

Page 16 of 186
SAR Structure-activity relationship — a qualitative relationship (i.e., an
association) between a molecular (sub)structure and the presence or
absence of a biological activity, or the capacity to modulate a biological
activity imparted by another substructure.
SMILES Simplified Molecular Input Line Entry System — a computer-compatible,
standardized, two-dimensional description of chemical structure. The
SMILES string is written by following a small number of rules. In brief,
each non-hydrogen atom (hydrogen is only explicitly included in special
circumstances) is denoted by its symbol; double and triple bonds are
shown by “=” and “#” symbols, respectively; branches are shown in
parentheses; and rings are opened and closed by the use of numbers.
For example, CCO represents ethanol, and c1ccccc1N represents
aniline (the digits indicate the beginning and ending of ring, and lower
case “c” indicates aromatic carbon).
Structural alert A molecular (sub)structure associated with the presence of a specific
(usually adverse) biological activity.
Substructure A portion of the overall structure of a chemical that may be associated or
correlated with a biological activity or property of the chemical.
TD50 The statistically derived median toxic dose of a drug or toxin at which
toxicity occurs in 50% of the test population.
Test set A set of chemicals, not included in the training set used to develop a
QSAR, that is used to validate (assess the predictive ability of) the
QSAR. It is sometimes called an “independent” or “external” test set or
validation set. For the purpose of (Q)SAR validation, it is important that
the test set has the same domain of applicability as the training set, and
contains a sufficient number of chemical structures.
Toxicity pathway A cellular response pathway that, when sufficiently perturbed, is
expected to result in adverse health effects (NRC, 2007). Toxicity
pathways are important components of adverse outcome pathways
(AOPs).
Training set A set of chemicals used to derive a QSAR. The data in a training set are
typically organized in the form of a matrix of chemicals and their
measured properties or effects observed in a toxicity test. A
homogeneous training set is a set of chemicals which belong to a
common chemical class or share a common chemical functionality or a
common mechanism of action. A heterogeneous training set is a set of
chemicals which belong to multiple chemical classes, or which do not
share a common chemical functionality or common mechanism of action.
Validation The testing of a (Q)SAR tool to assess its reliability and relevance. The
OECD Guidance Document on the Validation of (Quantitative)Structure-
Activity Relationship (Q)SAR Models (OECD Series on Testing and
Assessment No. 69) defines validation as the process by which the
reliability and relevance of a particular approach, method, process or
assessment is established for a defined purpose.

Page 17 of 186
1. EXECUTIVE SUMMARY

While pesticide regulatory agencies have traditionally relied on extensive in vivo

and in vitro testing to support regulatory decisions on human health and
environmental risks, these and other agencies have initiated the long-term
investigation of Integrated Approaches to Testing and Assessment (IATA). The
application of IATA could lead to the refinements, reduction, and/or replacement
of conventional tests through the integration of existing knowledge bases on
chemicals, biochemical and cellular assays, computational predictive methods,
exposure studies, and other sources of information to identify targeted testing
requirements or develop assessment conclusions. (Quantitative) Structure
Activity Relationships represent an important set of predictive tools to be
considered when applying IATA to pesticide risk assessments.

Moving IATA from a long-term vision into mainstream practice for pesticide
assessments will require the development and application of new predictive tools
and the further development and broader application of existing tools such as
(Q)SAR. In recognition of these requirements and the need to develop common
approaches to IATA for the risk assessment of pesticides, the United States
Environmental Protection Agency (US EPA) Office of Pesticide Programs (OPP)
and the Pest Management Regulatory Agency (PMRA) of Health Canada have
established a North American Free Trade Agreement (NAFTA) Joint Project on
“21st Century Toxicology: Integrated Approaches to Testing and Assessment”.

(Q)SAR is the study of the correlation between chemical structure and

associated (biological) activity, with the ultimate goal of predicting the activity of
untested chemicals based on structurally related compounds with known activity.
The parentheses around the “Q” in (Q)SAR indicates that the term refers to both
qualitative predictive tools (i.e., structure-activity relationships (SARs)) and
quantitative predictive methods (i.e., quantitative structure-activity relationships
(QSARs)). Although the term (Q)SAR is often used to refer to predictive models,
especially computer-based models, (Q)SAR is actually inclusive of a wide variety
of computerized and non-computerized tools and approaches.

The development of this NAFTA (Q)SAR Guidance Document is a key activity

under the NAFTA Joint Project. The purpose of this guidance document is to help
pesticide evaluators to evaluate (Q)SAR related information and to identify the
important issues that may be involved when incorporating (Q)SAR information
into the risk assessment process. This document does not reproduce or replace
the ever-expanding volume of journal articles, reports, documents, and textbooks

Page 18 of 186
that provide guidance on the development and application of (Q)SAR, but
provides an introduction to the evaluation of (Q)SAR tools and their application to
pesticide regulatory risk assessments. While the focus of this document is on the
application of (Q)SAR to pesticide risk assessments, the principles and issues
are general enough to be applied to other types of chemicals. Regardless of the
type of risk assessment scenario, (Q)SAR experts should be consulted and peer
review procedures used to ensure scientific excellence and rigor.

The document is organized into eight sections including this executive summary.

Section 2 provides an introduction to some current applications of (Q)SAR to

pesticide risk assessments with an emphasis on the use of (Q)SAR by the US
EPA OPP and the PMRA. It also includes a brief discussion of other regulatory
applications of (Q)SAR at the US EPA, the US FDA, Health Canada and
Environment Canada, the OECD, and the European Commission. Finally the
overall purpose of the guidance document is discussed and a schematic is
provided as guide to the contents of the document.

The purpose of section 3 is to provide some brief background information on the

definition of (Q)SAR, types of (Q)SAR tools and approaches, and some key
issues associated with the development of (Q)SAR tools. In particular, the
importance of data quality and mode/mechanism of action for toxicity information
in the development of (Q)SAR models is highlighted. Also, while (computerized)
(Q)SAR models are frequently cited in examples in this document, section 3
illustrates that (Q)SAR actually consists of a range of tools and approaches.

Section 4 focuses on the preliminary analysis of (Q)SAR predictions as one of

the several potential sources of information to be integrated at the problem
formulation stage of a pesticide assessment. Problem formulation for (Q)SAR
essentially involves answering questions on the assessment context for (Q)SAR,
the characteristics of the pesticide, the characteristics of the (Q)SAR tool and
prediction, and what empirical data are available including any information on
mode/mechanism of action for toxicity.

The topic of section 5, evaluating whether a (Q)SAR prediction is adequate or “fit

for purpose”, is an important component of applying a prediction to a pesticide
assessment. Four key factors originally outlined by the European Commission:
the scientific validity of the model, the applicability of the model to the query
chemical, the reliability of the (Q)SAR result, and the relevance of the (Q)SAR
model for the regulatory purpose are used to guide pesticide evaluators through
the information to be considered when evaluating whether predictions from

Page 19 of 186
(Q)SAR tools are adequate for consideration in pesticide assessments. Section 5
also includes a discussion of the documentation of (Q)SAR tools and predictions.

Section 6 briefly discusses approaches to combining information from multiple

(Q)SAR predictions, advantages and disadvantages of combining predictions,
selecting (Q)SAR tools for multiple predictions, and the evaluation of multiple
predictions. Because different (Q)SAR tools may have different prediction
paradigms and different strengths and limitations, combining predictions has the
potential to increase the confidence in the overall prediction. It is also noted that
combining predictions from multiple (Q)SAR tools does not eliminate the need to
ensure that each prediction is adequate or fit for purpose and it is not always
necessary to combine predictions.

The National Academy of Sciences risk assessment paradigm (i.e., hazard

identification, dose response assessment, exposure assessment, and risk
characterization) provides the context for section 7 which emphasizes guidance
on the integration of (Q)SAR tools into the hazard identification component of the
risk assessment process for pesticides. Section 7 builds upon previous sections
and includes a consideration of the findings at the problem formulation stage,
evaluating empirical data versus (Q)SAR predictions, a consideration of mode of
action data, the overall weight of evidence, and hazard identification, and risk
communication.

Section 8 provides conclusions and perspectives on the future vision for (Q)SAR
and pesticides. It is noted that the conclusions of the NAS report on Toxicity
Testing in the 21st Century: A Vision and a Strategy with respect to increased
reliance on existing knowledge-bases for chemical classes and alternative testing
methods is especially relevant for pesticide regulatory authorities and will require
research on new testing technologies and integrated approaches to testing and
assessment (IATA) for more efficient and effective reviews that don’t compromise
public health and the environment. (Q)SAR tools are one example of an
alternative method that could be applicable to IATA and the increasing use of
these tools by pesticide authorities make it important to communicate a
systematic and transparent approach to using (Q)SAR in pesticide assessments.
This guidance document is consistent with the current hazard/risk assessment
paradigm with an overall emphasis on not using (Q)SAR in isolation. In addition
to the validity and relevance of the individual (Q)SAR tools and predictions, the
defensibility of predictions depends on biological consistency and plausibility
across all scientific lines of evidence in a holistic weight of evidence approach.
Future applications will involve anchoring (Q)SAR predictions with what is known
about chemical classes/categories, biological mode of action, toxicity pathways

Page 20 of 186
and population effects. Eventually, (Q)SAR predictions will be built into larger
conceptual frameworks or Adverse Outcome Pathways (AOPs) that delineate the
documented, biologically plausible, measurable, and testable processes by which
a chemical induces molecular perturbations and subsequent biological responses
that are relevant for risk assessment.

In addition to the eight sections discussed above, the document also includes an
appendix of the web pages for a number of national and international
organizations that may be useful to evaluators seeking additional information on
general (Q)SAR concepts, and the development, validation, and evaluation of
(Q)SAR tools and predictions (Appendix I), an appendix summarizing the content
of the European Commission’s (Q)SAR model and prediction reporting formats,
examples of detailed information templates that could be considered when
(Q)SAR predictions are used as critical sources of data in pesticide assessments
(Appendix II), and an appendix of several examples of the application of (Q)SAR
tools and methods to pesticides and other chemicals (Appendix III).

Page 21 of 186
2. INTRODUCTION

INTRODUCTION

Topics Discussed in this Section:

 Current applications of (Q)SAR in pesticide risk assessments

 Other regulatory applications of (Q)SAR

 Purpose of the NAFTA (Q)SAR Guidance Document

2.0 Current Applications of (Q)SAR in Pesticide Risk

Assessments
In general, pesticide regulatory programs have extensive testing requirements as
part of the registration process and as a result, they have not had to rely heavily
on predictive methods such as (Q)SAR. However, this is changing over time as
pesticide agencies have begun to investigate alternative testing methods such as
(Q)SAR to help enhance the efficiency of their assessment processes. This is
particularly the case for the investigation and application of Integrated
Approaches to Testing and Assessment (IATA) to pesticide risk assessments.
IATA have the potential to integrate existing data on pesticides with the results of
alternative methods (e.g., biochemical/cellular assays, (Q)SAR) leading to the
refinement, reduction, and/or replacement of conventional test requirements.

Provided below is a brief overview of some of the current applications of (Q)SAR

by pesticide regulatory agencies.

2.0.1 United States Environmental Protection Agency, Office of Pesticide

Programs (US EPA OPP)
2.0.1.1 Application of (Q)SAR to Pesticide Metabolites and Degradates

The United States Environmental Protection Agency, Office of Pesticide

Programs (US EPA OPP) generally considers that the toxicity and ecotoxicity
studies required to support the evaluation of pesticides adequately address the
mammalian and environmental hazard profiles. The agency does not typically
require separate toxicity testing of pesticide plant or livestock metabolites or
environmental degradates even though there may be much greater human and
non-human exposure to the metabolites and degradates than the parent
pesticides. Also, historically, the US EPA OPP has typically included only major

Page 22 of 186
(>10%) pesticide metabolites and degradates in human dietary and
environmental risk assessments.

The advent of the Food Quality Protection Act (FQPA) has necessitated an
increased refinement of pesticide risk assessments including a closer scrutiny of
all metabolites and degradates. In order to determine whether these metabolites
and degradates should be included in human dietary and environmental risk
assessments in the absence of detailed toxicity data, the US EPA OPP has relied
upon various types of structural similarity evaluations. Also, more recently, the
agency has explored the use of (Q)SAR models to predict the potential toxicity of
pesticide metabolites/ degradates in order to provide scientific rationales and
support for requiring additional toxicity testing, to substantiate the use of
metabolites/degradates in estimates of total toxic residues, or to exclude
metabolites/degradates from further testing based on a lack of toxicity concerns.
Similarly, the US EPA OPP has made sporadic use of bridging techniques and
structure activity relationships to identify whether additional ecotoxicity testing of
environmental degradates should be required and whether these residues should
be included in environmental exposure estimates for pesticides.

Since empirical data are typically available on the parent pesticide, one of the
key factors considered when determining whether (Q)SAR model predictions for
the toxicity or ecotoxicity of metabolites and degradates are reliable enough to be
used is how well predictions from the same model for the parent pesticide
compare to the empirical data for the parent pesticide.

2.0.1.2 Application of (Q)SAR to Antimicrobial Agents

The US EPA OPP has several on-going initiatives and projects related to the
application of (Q)SAR to antimicrobial pesticide agents and one of the most
important initiatives is the proposed revised testing requirements for antimicrobial
agents (40CFR158 subpart W). In this proposed rule, EPA has indicated that it
will consider any submission using appropriate SAR analyses and QSAR
modeling to supplement or fulfill data requirements for antimicrobial pesticide
chemicals.

Approaches with concepts similar to structure activity relationships (SARs) are

also being utilized in a pilot project on non-animal eye irritation tests for
antimicrobial products with cleaning claims. The purpose of this project is to
assess the predictive performance of registrant submitted non-animal eye
irritation studies for antimicrobial agents by having registrants include any
available Draize rabbit test results for structurally related compounds in their
submissions.

Page 23 of 186
There is also an ongoing threshold of toxicological concern (TTC) project that is
designed to determine a level of concern for various chemical classes of
antimicrobial pesticides. Human exposures below the TTC would not be
considered to be of concern and no additional toxicological data would be
required. SAR will be used to characterize the toxicity of all chemicals within
specific classes of antimicrobial chemicals. This is an American Chemistry
Council (ACC) Biocide Panel sponsored project conducted through the
International Life Sciences Institute (ILSI) with the US EPA participating on the
ILSI Steering Committee and Expert Working Group.

2.0.1.3 Application of (Q)SAR to Ecological Risks from Pesticides

The US EPA OPP estimates chemical properties, environmental fate parameters,
and ecological toxicity values for pesticides, inert compounds, and degradates
using the EPI Suite and the Assessment Tools for Ecological Risk (ASTER)
software on a case-by-case basis when measured values are not available from
studies submitted to the Agency or from the open literature. In EPI Suite, the
organic carbon partition coefficient (PCKOC) model is used to estimate soil
mobility, the KOWWIN model is used to estimate the octanol-water partition
coefficients, and the BCFWIN model is used to estimate bioconcentration factors.
The Ecological Structure Activity Relationships (ECOSAR) component of EPI
Suite and ASTER are used to estimate pesticide ecotoxicity values. These
estimates may be used to support human dietary and ecological risk
assessments although their use, at this time, is not uniform across the US EPA
OPP since formal guidance has not yet been developed. Also, in evaluations
against measured values, the organic carbon partition coefficient (PCKOC)
model component of EPI Suite has acceptable predictive performance for organic
pesticides, but does not perform well for ionic compounds, organometallics, and
highly fluorinated pesticides. In addition, while ASTER contains models for five
aquatic species (i.e., fathead minnow, bluegill sunfish, water fleas, rainbow trout,
and channel catfish), it does not support models for terrestrial species. Similar to
ASTER, ECOSAR only predicts toxicity for aquatic species and cannot be used
to profile inorganic or organometallic chemicals. In the case of environmental
degradates, since empirical data are typically available on the parent compound,
(Q)SAR-generated toxicity estimates for the parent compound are compared to
the available empirical data in order to decide on whether it is appropriate to use
(Q)SAR models to estimate the potential ecological toxicity of the environmental
degradates. When determining which (Q)SAR model to use, consideration is also
given as to which model(s) has the best predictive performance. (Q)SARs have
been used by the US EPA OPP to address data gaps in ecological risk
assessments on an ad hoc basis.

Page 24 of 186
For ecological risk assessments, OPP has made increasing use of bridging
techniques and structure activity relationships (SARs) to identify whether
additional testing of degradates/transformation products should be required and
whether these residues should be included in modeling exposure estimates.

2.0.2 Pest Management Regulatory Agency (PMRA), Health Canada

The Pest Management Regulatory Agency (PMRA) of Health Canada takes into
account the same kinds of considerations as the US EPA OPP when addressing
the potential toxicity of pesticide metabolites/degradates of chemical pesticides. If
metabolites or degradates of a pesticide are identified in plants or soil, but not in
rat metabolism studies, the agency will require the submission of available
toxicity data on those metabolites. Also, toxicity data on metabolites/degradates
are sometimes voluntarily submitted to the PMRA by applicants. In terms of the
application of (Q)SAR, the PMRA can include a request for (Q)SAR predictions
on metabolites/degradates when requiring the submission of existing data and
can also generate (Q)SAR predictions to help identify potential concerns.

2.1 Other Regulatory Applications of (Q)SAR

Unlike pesticide regulatory agencies, programs involved in the regulatory
assessment of industrial chemicals, food additives, and other chemicals often
only have a limited amount of data available to support their assessments.
Consequently, many of these programs have a longer history with the
development and use of (Q)SAR tools and approaches.

Several examples of non-pesticidal regulatory applications of (Q)SAR at the US

EPA, the US FDA, Health Canada and Environment Canada, the OECD, and the
European Commission are summarized below. This is not intended as an
exhaustive listing of the uses of (Q)SAR by regulatory agencies, but it should
give the reader some context on the development and application of (Q)SAR
tools and approaches by a number of prominent national and international
agencies. For further information on current developments and applications of
(Q)SAR by various national and international agencies, the reader is directed to
the various websites listed in Appendix I.

Page 25 of 186
2.1.1 US EPA, Office of Pollution Prevention and Toxics (OPPT)

(Q)SAR methods have been used for identification of potential mutagenic,

carcinogenic and other potential health and ecotoxicological hazards and
subsequent regulation of new industrial chemicals (premanufacturing notification,
or PMN, chemicals) for more than two decades by the US EPA’s Office of
Pollution Prevention and Toxics (OPPT) under the Toxic Substances Control Act
(TSCA) which regulates all industrial chemicals in US commerce. Under TSCA,
OPPT is charged with assessing, and if necessary, regulating all phases of the
life cycle of industrial chemicals including manufacturing, processing, use and
disposal (OECD 2007b).

OPPT has also developed a number of publicly available (Q)SAR tools that are
used in regulating substances under TSCA. Examples include the EPI Suite
program which includes several models for estimating physical-chemical
properties and environmental fate parameters. EPI Suite also contains the
ECOSAR model for predicting ecotoxicity. Other tools developed by OPPT
include Oncologic, an expert system for predicting carcinogenicity, and an analog
identification tool (AIM) for identifying structural analogs.

2.1.2 US EPA, Office of Research and Development

The US EPA Office of Research and Development (ORD), National Health and
Environmental Effects Laboratory (NHEERL), Mid-Continent Ecology Division
(MED) has been developing (Q)SAR models and related databases since the
1980s. Examples include a database of ecotoxicity information (ECOTOX) as
well as ASTER3, a collection of databases and (Q)SAR models for toxicity to
aquatic species. ASTER also includes models to estimate physical-chemical
properties, bioconcentration, and environmental fate.

Research at ORD on receptor based toxicity mechanisms in aquatic species has

led to the development of a QSAR based expert system for predicting the
estrogen receptor binding potential of data poor pesticidal inerts and
antimicrobial pesticide active ingredients. The system is designed to prioritize
chemicals for further testing in the US EPA Endocrine Disruptor Screening
Program (EDSP) and it has been incorporated into the OECD QSAR Toolbox.

3
ASTER is a US EPA intranet application only accessible to US EPA staff and contractors.
http://www.epa.gov/med/Prods_Pubs/aster.htm

Page 26 of 186
2.1.3 US FDA, Office of Food Additive Safety

The US Food and Drug Administration’s (FDA) Office of Food Additive Safety
(OFAS) has utilized (Q)SAR analysis in the pre-market review of food contact
substances for many years and has recently implemented the use of multiple
commercial and publicly available (Q)SAR software models (Lo Piparo et al.,
2011; Arvidson et al., 2010; Bailey et al., 2005). OFAS is also investigating the
potential application of metabolism prediction software to the review of food
contact substances. OFAS uses (Q)SAR analysis as a decision support tool in
conjunction with open literature data and submitted test results, and (Q)SAR may
also be used to identify the need for additional toxicity testing during pre-
submission consultations for food contact substances.

2.1.4 Health Canada and Environment Canada

Health Canada and Environment Canada have extensive experience with the
use of (Q)SAR to address selected data requirements for new substances under
the Canadian Environmental Protection Act (CEPA). Adequately validated
(Q)SAR predictions may be submitted by notifiers or in some cases generated by
government evaluators to address physical-chemical properties,
persistence/bioaccumulation, human health effects, ecotoxicity endpoints and
other endpoints included in the New Substances notification requirements under
CEPA. For instance, predictions are sometimes used for assessing substances
with low production volumes and in cases where the substance cannot be
isolated in pure enough form to provide meaningful test results. (Q)SAR data are
generally combined with empirical data and expert judgment in a weight of
evidence approach. (Q)SAR was also utilized by both departments for the
categorization (prioritization) of existing substances on the Domestic Substances
List (DSL) for further assessment. Environment Canada used (Q)SAR predictions
to assist with determinations of persistence, bioaccumulation and inherent
toxicity to non-human organisms from existing substances while Health Canada
used (Q)SAR to generate physical-chemical data to support determinations of
greatest potential for human exposure and as part of the hazard tools used to
prioritize chemicals for inherent toxicity to humans when data for specific
endpoints were not available. (Q)SAR can also be used by both departments as
supporting information in screening level risk assessments for DSL substances
when experimental data are not available.

2.1.5 Organization for Economic Cooperation and Development (OECD)

Starting in the 1990s, the Organization for Economic Cooperation and

Development (OECD) began the investigation of various (Q)SAR methodologies

Page 27 of 186
with the aim of facilitating the application of (Q)SAR approaches in regulatory
settings and their regulatory acceptance. One of the most important products
from the OECD (Q)SAR project has been the principles for the validation of
(Q)SAR models (OECD, 2004). Comprehensive guidance has also been
produced on the development and application of grouping methods for chemicals
including chemical categories, read-across and trend-analysis (OECD, 2007a).

The OECD has also done extensive work on software for identifying structural
characteristics and mode/mechanism of action data on chemicals, systematically
grouping them into chemical categories and applying read-across, trend analysis
and (Q)SARs to fill data gaps. The end result of these efforts, the OECD QSAR
Toolbox, is intended for use by government agencies and stakeholders for
addressing gaps in the toxicity and ecotoxicity databases used in the hazard and
risk assessment of chemicals and is freely available (OECD, 2011a).

2.1.6 European Commission Joint Research Centre

The European Union's Registration, Evaluation, and Authorization of Chemicals

(REACH) legislation is designed to improve the protection of human health and
the environment while maintaining competitiveness and increasing innovation in
the European chemicals industry. Under the REACH legislation there is a strong
emphasis on the use of alternative testing methods to refine, reduce or replace
conventional animal testing. The European Commission Joint Research Centre
(JRC) Computational Toxicology Group is involved in projects to promote the
availability for regulatory application of validated computational methods for
assessing environmental distribution and fate, and the effects on human health
and the environment in support of the REACH legislation, the European
Cosmetics Directive and the assessment of food safety (Mostrag-Szlichtyng et
al., 2010). The group conducts research on the development of freely available
(Q)SAR tools (e.g., Toxtree, DART, Toxmatch), regulatory applications of
(Q)SARs and grouping approaches, the use of computational methods to assess
the properties of nanomaterials, and the consideration of molecular interactions
in the assessment of toxicity. The JRC has also developed templates for
documenting the application of the OECD (Q)SAR validation principles to
(Q)SAR models including the (Q)SAR Model Reporting Format (QMRF) (EC
2008a) and (Q)SAR Prediction Reporting Format (QPRF) (EC 2008b), and is
leading the development of a reporting format for describing key
events/intermediate effects in AOPs (OHT 201) in collaboration with the OECD
and the European Chemicals Agency (ECHA). The Joint Research Centre has
also established a database of (Q)SAR models.

Page 28 of 186
2.2 Purpose of the NAFTA (Q)SAR Guidance Document
The purpose of this guidance document is to help pesticide evaluators to
evaluate the (Q)SAR-related information and to identify the important issues that
may be involved when incorporating (Q)SAR information into the risk assessment
process. It is recognized that there is an ever-expanding volume of journal
articles, national and international reports and guidance documents, and
academic textbooks on the subject of (Q)SAR. This document does not
reproduce or replace these journal articles, reports, guidance documents, and
textbooks on (Q)SAR, but provides an introduction to the evaluation of (Q)SAR
tools and their application to pesticide regulatory risk assessments.

(Q)SAR predictions can be considered as one of the many potential sources of

data for the weight of evidence approaches used in the risk assessment of
pesticides. Similar to other sources of data considered, the defensibility of the
use of (Q)SAR predictions can be related to the consistency of the predictions
generated from the various (Q)SAR tools used as well as the consistency
between the predictions and the results of other lines of evidence considered in
the weight of evidence approaches.

While many of the illustrative examples in this document involve the application
of (Q)SAR to the prediction of toxicity in pesticide hazard assessments, the
general principles discussed can also be applied to (Q)SAR predictions for
ecotoxicity, physical chemical parameters, and other activities and properties of
relevance to pesticide assessments. Similarly, although many issues are raised
in the context of the prediction of apical endpoints, pesticide evaluators should
recognize that most of these issues will also apply when (Q)SAR is eventually
used in IATA to predict key events related to mechanism/mode of action for
toxicity and AOPs such as receptor binding, gene activation, enzyme
inhibition/activation, etc.

Although this document focuses primarily on (Q)SAR in the context of pesticide

risk assessments, the principles and issues discussed are general enough to
also be broadly applicable to the use of (Q)SAR in risk assessments for other
types of chemicals. Regardless of the type of assessment that (Q)SAR is being
applied to, it is recommended that experts in the (Q)SAR field be consulted and
that adequate peer review procedures be in place to ensure overall scientific
excellence and rigor.

The overall structure of this guidance document is presented in schematic form in

Figure 2–0. The document is organized to navigate the pesticide evaluator
through sections that provide an introduction to (Q)SAR and (Q)SAR tools, and

Page 29 of 186
information on problem formulation and (Q)SAR, evaluating the adequacy of
(Q)SAR predictions, combining information from multiple predictions, and
incorporating predictions into weight of evidence assessments. Each section can
also be considered as stand-alone guidance on its particular subject area.

As mentioned previously, Appendix I provides a listing of the websites of a

number of national and international agencies involved in the development and
application of (Q)SAR tools and approaches. These websites could be a useful
starting point for those who are interested in learning more about (Q)SAR and
obtaining more guidance on its use beyond what is presented in this document.
Appendix II summarizes the key features of the European Commission’s (Q)SAR
model and reporting formats, and Appendix III provides several case study
examples of the application of (Q)SAR to pesticides and other chemicals.

Page 30 of 186
 Figure 2–0: (Q)SAR Guidance Document Schematic

INTRODUCTION AND BACKGROUND INFORMATION

INTRODUCTION
(Section 2)

• Current applications of (Q)SAR in pesticide risk assessments

• Other regulatory applications of (Q)SAR
• Purpose of the NAFTA (Q)SAR Guidance Document

BACKGROUND INFORMATION ON (Q)SAR

(Section 3)

• Definition of (Q)SAR
• Types of (Q)SAR tools and approaches
• Importance of data quality in (Q)SAR model development
• Importance of mode/mechanism of action in (Q)SAR development
• Examples of (Q)SAR tools and their applications

APPLYING (Q)SAR PREDICTIONS TO PESTICIDES

PROBLEM FORMULATION AND (Q)SAR COMBINING INFORMATION FROM
(Section 4) MULTIPLE PREDICTIONS
(Section 6)
• Assessment context that (Q)SAR is being applied to
• Characteristics of the pesticide that is the subject of the • Approaches to combining multiple predictions
prediction • Advantages and disadvantages of combining predictions
• Characteristics of the (Q)SAR tool and the prediction • Selecting (Q)SAR tools for multiple predictions
• Empirical data including information on mode of action • Evaluation of multiple predictions

EVALUATING THE ADEQUACY OF (Q)SAR

PREDICTIONS
(Section 5)

• Scientific validity of a (Q)SAR tool

• Applicability of the (Q)SAR tool to the pesticide
• Relevance of the (Q)SAR tool to the assessment context
• Reliability of the (Q)SAR prediction
• Documentation of (Q)SAR tools and predictions

INCORPORATING (Q)SAR INTO

WEIGHT OF EVIDENCE ASSESSMENTS
(Section 7)

• Incorporating (Q)SAR in hazard characterizations: Overview

• Problem formulation and Adequacy Determination
• Evaluating empirical data versus (Q)SAR predictions
• Mode of action considerations
• Overall weight of evidence
• Hazard characterization and risk communication

CONCLUSIONS
CONCLUSIONS AND FUTURE VISION FOR (Q)SAR
(Section 8)

• Toxicity Testing in the 21st Century: Shift in the Risk Assessment Paradigm
• Weight of Evidence Approach: Biological Plausibility
• Adverse Outcome Pathway: Conceptual Framework
• Expert Scientific Judgement and Peer Review

Page 31 of 186
3. BACKGROUND INFORMATION ON (Q)SAR

BACKGROUND INFORMATION ON (Q)SAR

Topics Discussed in this Section:

 Definition of (Q)SAR

 Types of (Q)SAR tools and approaches

 Importance of data quality in (Q)SAR model development

 Importance of mode/mechanism of action in (Q)SAR model

development

 Examples of (Q)SAR tools and their applications

3.0 Introduction
The purpose of this section is to provide some brief background information on
the definition of (Q)SAR, types of (Q)SAR tools and approaches and some key
issues associated with the development of (Q)SAR tools. In particular, the
importance of data quality and mode/mechanism of action in the development of
(Q)SAR models is highlighted. Also, while (computerized) (Q)SAR models are
frequently cited in examples elsewhere in this document, this section illustrates
that (Q)SAR actually consists of a range of tools and approaches.

3.1 Definition of (Q)SAR

(Q)SAR is the study of the correlation between chemical structure and
associated biological activity, with the ultimate goal of predicting the activity of
untested chemicals based on structurally related compounds with known activity
(Cronin, 2010). Structure-activity relationships (SARs) are qualitative
relationships, often in the form of structural alerts that incorporate molecular
substructures or fragments related to the presence or absence of activity
(Dearden et al., 2009). Quantitative structure-activity relationships (QSARs)
attempt to quantify the relationship between an aspect of chemical structure and
an activity or property imparted by that structure. Chemical structure is often
described by descriptors (e.g., electrophilicity, hydrogen bonding, molecular
fragments) or physical-chemical properties (e.g., Log P) which are then used to
develop a mathematical correlation between a group of structures and a defined
activity or endpoint. The mathematical correlations usually take the form of

Page 32 of 186
statistical algorithms developed through a variety of techniques (e.g., univariate
regression, multiple linear regression, partial least squares analysis).

3.1.1 Defining Similarity

Structurally similar chemicals or structural analogs usually have similar chemical

structures but with one or more atoms or groups of atoms replaced with other
atoms or groups of atoms. Figure 3-1 lists the chemical structures of two
pyrethroid insecticides, deltamethrin and cypermethrin. These two structural
analogs share a cyclopropane carboxylic acid substructure that is common to
most pyrethoid structures.

Figure 3–1: Example of Structural Analogs

Deltamethrin Cypermethrin

Listed below are some common criteria used to identify structurally similar
substances. Many of these have been proposed by the OECD and the US EPA
as a basis for building chemical categories (OECD, 2007a; US EPA, 1999).

• a common functional group or sub-structure (e.g., phenols, aldehydes)

• a common precursor or break-down product, which can result from

structurally-similar chemicals; this approach can be used to examine related
chemicals such as acids/esters/salts. (e.g., short-chained alkyl-methacrylate
esters which are metabolized to methacrylic acid)

• an incremental or constant change in a chemical structure (e.g., increased

carbon chain lengths; typically used for physicochemical properties such as
boiling point)

• common constituents or chemical classes, such as similar carbon range

numbers, often used with “substances of unknown or variable composition,
complex reaction products or biological material” (UVCBs)

• functionally similar chemicals or functional analogs that have similar biological

activities (e.g., toxicity endpoints, pesticidal mode of action) or physical-
chemical properties (e.g., Log P, solubility, vapour pressure). Note that

Page 33 of 186
 functional analogs are not necessarily structural analogs and vice versa
(Saliner et al., 2005; Russom et al., 1997).

Table 3–1 lists examples of pesticidal modes of action for several examples of
insecticides. Most pesticidal modes of action include more than one chemical
class. Consequently, ‘similarity’ can be based on at least three aspects of a
pesticide, (i) pesticidal mode of action (e.g., acetylcholinesterase inhibition),
(ii) pesticide classification (e.g., insecticide) and (iii) chemical
classification/common functional group (e.g., carbamate, organophosphate, etc.).
Therefore, a weight of evidence approach can be important when defining
similarity for the purpose of developing (Q)SAR tools and approaches. As
discussed in example 5, Appendix III, information on the pesticidal mode of
action and structural similarity can be combined with pharmacokinetic and
empirical animal study results in a weight of evidence approach in pesticide
assessments.

Table 3–1: Pesticidal Mode of Action and Associated Chemical Class for a
Select Group of insecticides (adapted from Insecticide Resistance
Action Committee (IRAC); http://eclassification.irac-online.org/)

Pesticidal Mode of Action Chemical Class

Acetylcholine esterase inhibitor Carbamates

Organophosphates

GABA-gated chloride channel antagonists Cyclodiene organochlorines

Phenylpyrazoles (Fiproles)

Sodium channel modulators Pyrethroids

Organochlorines

Nicotinic acetylcholine receptor agonists Neonicotinoids

Juvenile hormone mimics Juvenile hormone analog

Carbamates

Pyridine insect growth regulator

A common mathematical approach to defining structural similarity is the use of

algorithms or similarity indices that calculate similarity based on pattern

Page 34 of 186
matching. These estimation tools rank chemicals based on (structural)
characteristics or features of each chemical that are similar (match/overlap), and
features that are dissimilar (mismatch/difference) (Saliner et al., 2005; Monev,
2004.). Figure 3–2 provides a schematic of the measures that can be described
in similarity indices. Similarity indices can utilize two- or three-dimensional
structural information and examples include correlation-type indices (e.g.,
Tanimoto Index (also known as Jaccard coefficient), Hodgkin Ricards Index,
Cosine-similarity index), dissimilarity measures (e.g., Euclidean distance index,
Hamming distance), and composite measures of similarity and dissimilarity (e.g.,
Hamann measure, Yule measure). For an overview of these approaches see
Saliner et al., 2005; Monev, 2004; and Urbano-Cuadrado et al., 2008. It is also
important to know that a high degree of similarity based on mathematical
similarity indices does not necessary indicate there are similarities in the mode of
action (MOA) for the concerned effects.

Figure 3-2: Measures that can be Described in Similarity Indices

Comparing Chemicals A and B

• a = number of features present in A and absent in B
• b = number of features present in B and absent in A
• c = number of features common to both A and B
• d = number of features absent from both A and B

Chemical A Chemical B

a c b d

3.2 Types of (Q)SAR Approaches

Although the term (Q)SAR is often used to refer to predictive models, especially
computer-based models, it should be noted that (Q)SAR is actually inclusive of a
wide variety of tools and approaches such as analogs, chemical categories and
computer-based or non-computer based SAR/QSAR models. A brief overview of
these tools and approaches is provided below.

Page 35 of 186
3.2.1 Analogs
Analog approaches have traditionally involved predicting an endpoint or property
of one chemical based on the available data for the same endpoint or property of
a similarly structured chemical (OECD, 2007a). An example of an analog
technique is bridging or extrapolating the results of toxicological studies on a
parent pesticide compound to a metabolite or transformation product of that
same parent pesticide. When using an analog approach for bridging from a
parent pesticide to metabolites or transformation products, it is important to have
sufficient evidence to link a particular substructure or substructures to the toxicity
endpoint of interest, and that the substructure is conserved from the parent
pesticide to the metabolite or transformation product.

3.2.2 Chemical Categories

A chemical category is defined as a group of substances with physical-chemical,
human health, or ecotoxicological attributes that are similar or follow a pattern as
a result of structural similarity (OECD, 2007a). As discussed in section 3.1.1, the
US EPA and the OECD have identified a number of ways of identifying similar
chemicals for the purpose of building categories. Both agencies have also
developed a consistent approach for defining chemical categories (OECD,
2007a; OECD, 2009; US EPA, 1999). Chemicals within a category are not
required to be similar with respect to all properties, and a substance can belong
to more than one chemical category. In most instances, chemical category
approaches are based on a weight of evidence, considering multiple lines of
information from many tested chemicals and inferring information for an untested
substance.

When using the chemical category approach, it is common to construct a matrix

table as depicted in Figure 3–3. The matrix consists of chemical category
members in each of the columns and corresponding sets of properties and/or
activities in each of the rows. The solid dots are properties/activities for which
reliable data exist and the hollow dots are data gaps. Data gap filling in
categories can be done using techniques such as read-across, interpolation,
extrapolation, and trend analysis (see examples in Figure 3–3). Read-across is
estimating the activity/property for one untested chemical from a tested chemical
or chemicals. Read-across can be qualitative or quantitative. Interpolation is the
estimation of a property/activity for a data poor category member based on
existing data from other category members on both sides of the data poor
chemical in the matrix. Extrapolation is estimating an activity/property for a
chemical that is near or at the boundary of the category based on data for other
category members. Extrapolation is more prone to error than interpolation,

Page 36 of 186
especially when the boundary of the category is difficult to define. The
observation of a quantitative trend (increasing, decreasing, or constant) in the
experimental data for a given endpoint across chemicals in a category can also
be used as the basis for interpolation or extrapolation (i.e., trend analysis). In
addition, it is possible to develop a QSAR within a chemical category by plotting
the activities versus the properties of chemicals with empirical data. By using a
combination of tools, i.e., read-across, trend analysis and (Q)SAR, the matrix of
properties/activities for chemicals under consideration can be rendered less
uncertain through the greater use of existing data (OECD, 2007a).

Figure 3–3: A Schematic of a Chemical Category Matrix Table

(modified from van Leeuwen et al., 2009)

Chemical 1 Chemical 2 Chemical 3 Chemical 4

Property 1 SAR / read-across

Property 2 Interpolation

Property 3 Extrapolation

Property 4

Trend analysis / QSAR

Activity 1

Activity 2
Empirical data
Activity 3 Missing data

Activity 4

Chemical category approaches have been used for assessing chemicals with
data gaps by the US EPA’s OPPT, in the US EPA HPV Challenge Program,
under the REACH legislation, and in OECD SIDS program (van Leeuwen et al.,
2009). Additional examples of categories can be found in Enoch, 2010; Enoch et
al., 2009; US EPA, 1999; and Worth and Patlewicz, 2007.

3.2.3 (Q)SAR Models

(Q)SAR models generally refer to computerized systems developed to predict

activities or properties of chemicals using SAR or (Q)SAR methods. There are
numerous commercial (i.e., available for a fee) and non-commercial (i.e.,

Page 37 of 186
freeware) models available for predicting human health related and
environmental activities, physical-chemical properties, and other parameters.

SAR models generally follow a process of identifying active and inactive

chemicals based on the presence or absence of specific structural features. For
example, SAR/expert systems use decision logic to categorize potential activity
of untested chemicals based on expert knowledge gathered from the analysis of
data on tested chemicals. Some systems use a series of questions that the user
responds to or the system automatically responds to. The questions may be
based on databases of structural alerts or chemical parameters known to be
associated with biological activity and can capture multiple types of interactions
within a specific biological system.

(Q)SAR models usually consist of computerized mathematical correlations (i.e.,

algorithms) that relate descriptors of chemical structure or physical-chemical
properties to an activity or property to be predicted. The descriptors or physical-
chemical properties for a chemical of interest may be input by the user or
generated by the model and then used in the algorithm to make a prediction. For
example, in a QSAR model developed to use the octanol and water partition
coefficient (Kow) to estimate the permeability coefficient through the skin for a
chemical in water (Kp), the Kow is the descriptor which can be a measured value
or estimated by the model (US EPA, 2007). QSAR models can produce
qualitative predictions of activity/inactivity or quantitative (continuous) values
related to biological activity (e.g., receptor binding affinity, acute oral LD50 in rats,
etc.) or other parameters (e.g., bioconcentration factor). QSAR models generally
rely on data for many chemicals (i.e., training sets) for the development of the
algorithms used to predict the activity of a single chemical lacking data.

In addition to classifying (Q)SAR models as relying on SAR versus QSAR

approaches, they can also be considered in terms of statistical versus
mechanistic approaches and global versus local approaches. In general,
statistically-based (Q)SAR models rely on a statistical association between
structure and activity, can be developed objectively with little mechanism of
action expertise, are useful for detecting structural features/molecular descriptors
predictive of toxicity, but may be noisier and tend to perform poorer for endpoints
with multiple mechanisms. Mechanistically-based models can focus on key
features that provide more clear-cut relationships and mechanistic backing but
generally require considerable expert knowledge of the relationship between
mechanism of action and descriptors of chemical structure, may be subjective,
and could have high levels of uncertainty if the mechanism is unclear or

Page 38 of 186
presumptive. Ideally, (Q)SAR models should strive to achieve statistical
association but have a mechanistic foundation.

Local QSAR models are generally developed for individual classes of chemicals.
Their training sets usually consist of highly structurally homogeneous or
congeneric chemicals or classes of chemicals with similar known biological
activity/function (e.g., peroxisome proliferators). Local models require fewer
training set chemicals and tend to perform better, presumably because they are
more likely to focus on a single mechanism of action. However, they are often
limited in scope to a small subset of narrowly defined chemicals. Global models
are generally derived using training sets of structurally heterogeneous or non-
congeneric chemicals. Due to the diversity of the training set chemicals, these
models often cover a range of different mechanisms of action, usually resulting in
poorer predictive performance than local models, unless the training sets are
subdivided based on mechanism of action. Global predictive models tend to be
more adept in discovering new insights, but may be more likely to yield incorrect
results if the predicted chemical structure is not well-represented in the training
set. Several publications have investigated the ability of global and local (Q)SAR
approaches to fulfill regulatory requirements (e.g., EC , 2010; Yuan et al., 2007;
and Worth et al., 2011).

3.3 Importance of Data Quality in (Q)SAR Model Development

Developing (Q)SAR models depends on experimental data, molecular
representation (2-D or 3-D structures), availability of chemical descriptor or
parameter data (measured or calculated) associated with structure, and fitting
relationships (e.g., algorithms) to the data (Bradbury et al., 2003; Perkins et al.,
2003; Tong et al., 2003; Walker et al., 2003). Among these factors, experimental
data are generally the most important determinants of the accuracy of the
predictions from (Q)SAR models as the confidence in a model can be no greater
than the understanding of, and confidence in, the underlying data.

In (Q)SAR model development, usually a set of chemicals with reliable data are
collected for a particular biological/chemical activity. Typically the original test
data are randomly separated into a training set and a validation set, with the
training data set used to develop a model and the validation data set used to test
the assumptions that the model works for chemicals not involved in the
development of the original model (Leonard and Roy, 2006).

Model training sets can be assembled prospectively or retrospectively. In the

prospective approach bioassays are developed and optimized for testing the type
of chemicals for which the (Q)SAR predictions are needed. Mechanistic

Page 39 of 186
information on the training set chemicals can also be obtained prospectively. In
the more commonly used retrospective approach data are collected from readily
available sources (e.g., the open literature). This often results in noisier
predictions because of the lack of control or consistency in study protocols,
interpretation criteria, etc., although this can be compensated for by evaluating
the available data and selecting only consistent higher quality studies (i.e.,
chemical identity/form confirmed, concentrations/purity measured, standard test
protocols, assays optimized for the type of chemicals, etc.). Also, it is important
to verify that the identities of the chemicals in the training set correspond to their
structural representations used in the predictions. Sometimes information on the
metabolism of the chemicals tested and mechanisms of action can also be
obtained retrospectively to help enhance the interpretability of the predictions.

The importance of data quality in (Q)SAR model development is also discussed

further in section 5.1.1 of this document.

3.4 Importance of Mode/Mechanism of Action in (Q)SAR Model

Development
An understanding of a chemical’s mode/mechanism of action is highly sought
when developing (Q)SARs. Mode/mechanism of action considerations can help
in the selection of appropriate molecular descriptors or physicochemical
properties that are associated with activity, determination of whether the training
set is applicable to the chemicals to be predicted, separation of the training set
into more mechanistically homogenous groups to help improve predictive
performance, and the interpretation of model outliers. An understanding of
mode/mechanism of action can also provide support for predictions, help in the
assessment of the human significance of predictions of toxicity in laboratory
animals, and help to identify and prioritize additional testing to fill data gaps.

One example of the utility of mode/mechanism of action data in (Q)SAR model

development is a study by Russom et al. (1997). A diverse dataset of more than
600 chemicals was divided into mechanistic groups prior to developing (Q)SARs
for fathead minnow acute toxicity. Combining all of the chemicals into a single
training set would have resulted in a much poorer correlation of the LC50 values
to the chemical parameter, log Kow, which in turn, would have resulted in much
poorer predictive performance.

An understanding of mechanistic considerations is especially important for

complex biological systems which may have metabolism and chemical kinetics
adding to the complexity. In general, the less complex the biological system, the
greater the confidence that the structure of the chemical is directly related to the

Page 40 of 186
observed activity and that the relationship can be reliably modeled. For example,
in vitro systems (e.g., Ames mutagenicity) are often less complex and more
reliably modeled than many in vivo systems (e.g., carcinogenicity, teratogenicity).
However, this is not always the case as in vivo fish acute toxicity LC50 values are
well predicted for several modes of action because chemical concentration in
water is a good surrogate for chemical activity in the blood (MacKay et al., 1983).
Also, if in vitro systems include metabolic components, their complexity for
(Q)SAR development will increase.

The importance of mode/mechanism of action information in (Q)SAR model

development is also discussed further in section 5.1.1.5 of this document.

3.5 Examples of (Q)SAR Tools and their Applications

This section is not intended to provide an exhaustive overview of computational
tools available via government, open access, or commercial sources, but rather
an overview of the types of tools that currently exist.

Several reviews have been written on the types of tools available (EC, 1995a,b;
Hulzebos et al., 1999; Jensen et al., 2008; Pavan et al., 2005a,b; Rorije and
Hulzebos, 2005; Tsakovska et al., 2005, 2008), but it should be kept in mind that
the inventories of available tools is constantly changing with emerging research
in this area.

With the development of Simplified Molecular Input Line Entry System (SMILES)
notation (Weininger, 1988) as a means to identify structure information in a
computer readable format, and the advancement of desk top computing in the
1970’s, (Q)SAR tools have become more readily accessible to risk assessors
(Benfenati, 2007). Although initially (Q)SAR approaches were primarily used in
the drug and pesticide discovery and development fields, these techniques
became especially important to regulatory risk assessment after the promulgation
of the Toxic Substances Control Act (TSCA) (Zeeman et al., 1995). The use of
QSARs in assessing potential toxic effects of organic chemicals on ecologically
relevant species and humans evolved as computational efficiency and
toxicological understanding advanced, and in many cases has proved to be
scientifically-credible for use in estimating toxicity for substances with little or no
available empirical data (OECD, 2007b).

(Q)SAR models also exist for specific endpoints such as skin sensitization
(Patlewicz et al., 2008), eye irritation (Tsakovska et al., 2005), acute toxicity and
repeated-dose endpoints for mammalian species (Tsakovska et al., 2008),
bioaccumulation (Arnot and Gobas, 2004), mutagenicity and carcinogenicity

Page 41 of 186
(Benfenati et al., 2009; Benigni et al., 2007a,b), estimating physical chemical
properties (EC, 1995a,b; Deardon and Worth, 2007), toxicity to aquatic species
(EC, 1995a,b; Netzeva et al., 2007; Pavan et al., 2005a,b), and reproductive
toxicity (Jensen et al., 2008).

Software applications are available for assisting in the identification of chemical

similarity (Gallegos-Saliner et al., 2008; Patlewicz et al., 2005). In addition, the
availability of state-of-the-art software programs for use in the development of
QSAR models from any data set allows one to generate models at their desk top
for endpoints/applicability domains not covered by existing models (see the
series of FDA papers as an example: Matthews et al., 2009 a, b; Matthews et al.,
2007 a,b; Ursem et al., 2009). Key to this type of analysis is a high quality,
structurally-annotated data set for use in the development of models (Judson et
al., 2009; Richard et al. 2006, 2008; Williams et al., 2009). Another important
aspect of many risk assessments is metabolism/degradation products, and
(Q)SAR tools to simulate metabolism have been developed to assist in
identifying these products (Dimitrov et al., 2005a,b; Mekenyan et al., 2006;
Ringeissen, et al., 2010).

(Q)SAR approaches can be used to better inform testing strategies via

screening, prioritization, and ranking of large chemical inventories based on
receptor binding (Jensen et al., 2008; Klopman and Chakravarti, 2003 a,b;
Schmieder et al., 2004), human health endpoints (Demchuk et al., 2008;
Klopman et al., 2003; Ruggeri, 2009), and environmental toxicity, fate, and
persistence (Brown and Wania, 2008; Daginnus et al., 2009; Walker et al., 2004).
These rankings can be used for a variety of risk assessment purposes including
developing chemical categories, identification of PBT (persistent,
bioaccumulative and toxic) substances, and risk characterization (Pavan and
Worth, 2008). Similarly, (Q)SAR tools have been investigated in combination with
physical-chemical data and read-across to improve the application of TTC
methods (Bassan et al., 2011; Worth et al., 2011).

Under REACH, information on models that meet the OECD validation principles
and are proposed for use in filling data gaps are currently being gathered. A
searchable catalog of all models including background information required to
validate the models, authors/source of model, related publications, endpoint
estimated and related experimental protocol, algorithm with training set and
validation set, including all input variables for the models can be found at the
following website: http://qsardb.jrc.ec.europa.eu/qmrf/index.jsp. Some actual
example cases are listed in Appendix III.

Page 42 of 186
3.6 Summary
(Q)SAR tools and approaches involve the study of correlations between chemical
structure and associated biological activity, physical-chemical properties or other
properties, with the ultimate goal of predicting the activity or properties of
untested chemicals using available data from structurally-related compounds.
While frequently associated with computerized models, (Q)SAR tools actually
encompass a wide range of approaches such as analogs, chemical categories,
and computer or non-computer based SAR and (Q)SAR. The development of
reliable (Q)SAR models depends upon a number of factors, among which,
experimental data are probably the most important. In particular, data quality and
a good understanding of the available information on mode/mechanism of action
can contribute to the confidence in (Q)SAR model predictions. Types of
endpoints or properties from a pesticide context that can be predicted using
(Q)SAR and related methods include in vivo ecotoxicity and human health-
related toxicity endpoints, specialized in vitro endpoints, metabolism, physical-
chemical parameters, and environmental fate parameters. While this document is
not intended to recommend or endorse individual (Q)SAR tools, it is recognized
that there are currently a variety of computerized and non-computerized,
commercial and non-commercial (Q)SAR tools for predicting the endpoints or
properties described above. Sections 2 and 3 of this document were designed to
provide a brief introduction and background information on (Q)SAR tools and
approaches. The subsequent sections of this document (4, 5, 6, and 7) focus on
issues associated with applying (Q)SAR predictions to pesticides including
problem formulation and (Q)SAR (section 4), evaluating the adequacy of (Q)SAR
predictions (section 5), combining information from multiple predictions (section
6) and incorporating (Q)SAR into weight of evidence assessments (section 7).

Page 43 of 186
4. Problem Formulation and (Q)SAR

PROBLEM FORMULATION AND (Q)SAR

Topics Discussed in this Section:

 Assessment context that (Q)SAR is being applied to

 Characteristics of the pesticide that is the subject of the prediction

 Characteristics of the (Q)SAR tool and the prediction

 Available empirical data including information on mode of action

4.0 Introduction
Problem formulation is an important initial step for framing the specific
question(s) to be addressed in assessments of human health and environmental
risks from pesticides. In its Guidelines for Ecological Risk Assessment, the US
EPA has indicated that problem formulation involves the on-going integration of
the available information that eventually leads to three products: assessment
endpoints, a conceptual model of the risk to be investigated, and an analysis plan
(US EPA, 1998).

Since guidance on the general problem formulation process for the risk
assessment of chemicals such as pesticides has been outlined in other published
documents (e.g., US EPA, 1998; Doull et al., 2007), the details of that guidance
will not be discussed here. Instead, this section will focus on the preliminary
analysis of (Q)SAR predictions as one of the several potential sources of
information to be integrated at the problem formulation stage. Preliminary
analysis of (Q)SAR prediction for a pesticide at the problem formulation
essentially involves answering the following questions:

 What is the assessment context that the (Q)SAR prediction is being

applied to?

 What are the characteristics of the pesticide that is the subject of the
prediction?

 What are the characteristics of the (Q)SAR tool and the prediction?

 What empirical data are available including any information on mode of

action?

Page 44 of 186
Answering these questions at the problem formulation stage may enable an
evaluator to immediately determine that a prediction is not suitable or relevant for
addressing the specific pesticide risk assessment question. Alternatively, these
questions may lead to a more in-depth evaluation of whether the (Q)SAR
prediction is adequate or “fit for purpose” (see section 5) and eventually to the
consideration of how the results of a fit for purpose prediction could be
incorporated into an overall weight of evidence decision (see section 7).

4.1 Assessment Context that (Q)SAR is being Applied to

Identifying the assessment context for a (Q)SAR prediction involves
understanding why the prediction is being considered for the assessment of a
pesticide and the specific endpoint or property that the prediction is intended to
address. Both of these points will assist the evaluator in determining whether a
(Q)SAR prediction should be considered in a pesticide assessment and if yes,
what will be an acceptable level of reliability and uncertainty associated with the
use of (Q)SAR.

(Q)SAR predictions are generally used to try to gain some insights into the
toxicity, ecotoxicity, behavior in the environment or other aspects of a pesticide in
the absence of empirical data. Consideration of a (Q)SAR prediction for the
premarket assessment of a pesticide would likely involve one of the following
scenarios: 1) submission of a (Q)SAR prediction by a registrant to address a data
requirement or as supporting evidence for a data requirement for pesticide, a
metabolite or a transformation product, or 2) use of a prediction by an evaluator
to identify or support a data requirement for a pesticide, metabolite or
transformation product.

In the first scenario, an applicant would likely submit a (Q)SAR prediction or

predictions as a replacement for or as supporting evidence to waive a
requirement for a specific type of empirical data (e.g., to address a requirement
for acute irritation toxicity data). In most cases, using a (Q)SAR prediction as a
stand alone replacement for a data requirement is not likely to be acceptable
depending on the nature of the endpoint and the specific policies of the pesticide
regulatory agency. Combining a (Q)SAR prediction with other types of data to
support a waiver request may be more acceptable depending on what other
types of data are available, the reliability and level of uncertainty for the (Q)SAR
prediction, the overall scientific defensibility of the rationale, and regulatory
agency policies.

The second scenario could involve using a (Q)SAR prediction to justify a

requirement for a study not normally included in regulatory data requirements for

Page 45 of 186
pesticides or to justify a requirement for a study on a metabolite or transformation
product for which no data have been submitted. This would also include cases in
which (Q)SAR predictions are used as supporting information when questioning
the reliability of experimental data, leading to a requirement for the submission of
more reliable studies. Criteria for what constitutes a reliable (Q)SAR prediction
and acceptable levels of uncertainty would likely be less stringent for scenarios in
which (Q)SAR predictions are used to drive data requirements compared to
cases where (Q)SAR is used to support waiving data requirements.

While these premarket scenarios are likely to be the most frequent applications
of (Q)SAR, there may also be instances where (Q)SAR tools could be used post-
market such as the toxicity characterization of a novel impurity (e.g.,
leachable/extractable) not originally characterized during the pre-market approval
process.

Endpoints or properties that can be predicted by (Q)SAR and could be relevant

to pesticide assessments include toxicity (e.g., carcinogenicity, developmental
toxicity), metabolism, ecotoxicity (e.g., fat head minnow LC50, longer-term toxicity
in terrestrial species), other biological activities (e.g., estrogen receptor binding),
and physical-chemical properties (e.g., Log Kow, partition coefficients,
bioaccumulation factor). The type of endpoint and whether it is a critical data
point for a pesticide assessment (e.g., used for a point of departure analysis) will
have an influence on how reliable a (Q)SAR prediction should be (see section
5.4). For example, it may be possible to accept a less reliable prediction for an
acute toxicity endpoint used as supporting information for labeling requirements
compared to a predicted NOAEL for chronic toxicity that is to be considered in a
point of departure analysis. Furthermore, the use (Q)SAR predictions to address
critical endpoints in pesticide risk assessments would likely require much more
detailed analyses of whether the predictions are fit for purpose compared to
predictions generated for non-critical endpoints.

4.2 Characteristics of the Pesticide that is the Subject of the

Prediction
Understanding the characteristics of the pesticide that is the subject of the
(Q)SAR prediction is critical to determining whether the pesticide is correctly
identified; whether the prediction is to be made for an active ingredient, a
metabolite or a transformation product; whether an accurate structural
representation is available; or whether it is even possible to generate a prediction
for the pesticide in question.

Page 46 of 186
4.2.1 Chemical Identifiers and Mixtures
Examples of the types of pesticides for which a (Q)SAR prediction may be
required include discrete substances; individual isomers or mixtures of isomers;
crystalline structures (e.g., minerals); substances with unknown or variable
composition, complex reaction products and biological materials; polymers; other
mixtures or formulations; and complex salts and metal-containing compounds.
Therefore, accurate information on the identity, composition and structure of a
pesticide is critical to determining whether a prediction was based on a correct
structure. Confusion can result when common or trade names are applied to
multiple isomers of a pesticide, salt forms, acid/base forms or polymeric and
monomeric forms. The use of more precise chemical nomenclature (e.g.,
International Union of Pure and Applied Chemistry (IUPAC)) can assist with more
accurate identification (IUPAC, 2010). While Chemical Abstract Service (CAS)
numbers (American Chemical Society, 2010) are frequently used as unique
identifiers for pesticides, in some cases they may actually represent isomer
mixtures, polymers, and unknown or variable composition substances rather than
discrete, single chemicals, so it may be necessary to review the CAS number to
clearly determine which structure(s) it actually represents.

In general, mixtures cannot be run through (Q)SAR models, nor can synergistic
or antagonistic effects of chemicals in mixtures be accounted for because models
typically use single, discrete chemical structures as input. For mixtures of
discrete organic chemicals, one option may be to make separate predictions for
each chemical and compare and contrast the results. Alternatively, if one
component of a mixture is predominant, in some cases that component may be
used to represent the entire mixture. However, for pesticides with variable
compositions, (i.e., oligomers, natural fats, or mixtures that change composition
depending on reaction conditions) evaluators should be aware that (Q)SAR
predictions generated using a representative structure may not accurately reflect
the true nature of the material used in the pesticide application.

4.2.2 Transformation, Degradation, and Metabolism

A number of pesticides are reactive chemicals that can be readily transformed in
the environment or in the body (e.g., hydrolyzable acid halides, isocyanates,
etc.). Transformation products may have dramatically different toxicity profiles
than the original pesticides and need to be considered when identifying the
correct structures for (Q)SAR predictions. Similarly, information on potential
environmental degradates and metabolic by-products of pesticides in livestock,
food plants or the body should also be considered as the toxicity may not reflect
the parent pesticide, but rather a reactive intermediate, degradate or metabolite.

Page 47 of 186
Information on degradates and metabolites may be available from empirical
pharmaco/toxicokinetic studies or from in silico models of potential metabolic or
degradation pathways. If degradates and metabolites are identified, it will also be
important to consider their stability and in the case of (Q)SAR predictions of
metabolites, the likelihood that they could occur in vivo. Individual pesticide
regulatory agencies have specific criteria they use to identify probable, stable
metabolites.

When (Q)SAR models are developed, available information on the metabolism of

the training set compounds should be taken into consideration whenever
possible. Training set compounds that require metabolic transformation prior to
inducing a specific endpoint are likely to generate models that are unreliable if
those models are constructed based on the parent structures alone. Failure to
consider the structure of a metabolite could lead to an inaccurate assessment of
the chemical features or properties associated with the predicted endpoint, errors
in analog selection, problems with characterization of similarity based on mode of
action, errors in inter-species extrapolation when metabolic differences exist
between species, and ultimately poor predictive performance. Another aspect of
metabolism that may need to be considered during (Q)SAR model development
is differences in metabolism from different routes of exposure. While the
industrial chemical, bis-(chloromethyl) ether is one of the most potent human and
animal respiratory carcinogens, it is not expected to be carcinogenic via the oral
route because it hydrolyses in seconds upon contact with aqueous solution (Woo
and Lai, 2010; ATSDR, 1989). Consequently, (Q)SAR predictions for direct
acting reactive pesticides that are used to support data requirements for
inhalation toxicity should be treated with caution if they come from (Q)SAR tools
whose training sets only include analog substances tested via the oral route.
Finally, the results of in vitro tests can also be impacted by metabolic
transformations, so information on the degree to which metabolic capability is
incorporated into in vitro toxicity tests should be considered when constructing
models for those tests.

4.2.3 Isomers and Structural Representations for (Q)SAR

A pesticide’s three-dimensional molecular structure or shape and its molecular
conformation can influence properties such as absorption, distribution, and
excretion, as well as enzyme or receptor binding, and the resulting differences
can readily impact toxicity profiles. Consequently, the isomeric form of a pesticide
is another important piece of information to consider when determining whether a
prediction is based on the correct structure and whether the (Q)SAR tool is
applicable to the structure in question.

Page 48 of 186
Examples of isomeric forms to consider include stereoisomers that differ in their
spatial orientation of atoms. The pyrethroid insecticide fenvalerate is a racemic
mixture of stereoisomers (i.e., R/S enantiomers) of a chiral active ingredient,
although the S-isomer in the mixture (esfenvalerate) has the greatest insecticidal
activity (WHO/FAO, 1996) (see Figure 4–1). Because these different isomers
have the same molecular formula, molecular weight, and physical-chemical
properties, it can be difficult for some (Q)SAR models to distinguish them,
especially models that do not take stereoisomerism into account.

Many regulatory agencies make the conservative assumption that stereoisomers

will have similar mammalian toxicity and ecotoxicity, unless data are available to
demonstrate the contrary. In addition to isomeric forms, the position of flexible
groups in a molecule can also be important as relatively free rotation of attached
groups about single bonds can influence the conformation of a molecule and
determine the overall molecular size, especially in complex molecules with
multiple rotation points. More advanced (Q)SAR techniques may employ three-
dimensional molecular descriptors to account for rotation of flexible groups and
other characteristics, but calculating these descriptors can be complex and time
consuming.

Figure 4–1: Fenvalerate Racemic Mixture

Fenvalerate Esfenvalerate
(RS)-alpha-Cyano-3-phenoxybenzyl (RS)-2-(4- (S)-alpha-Cyano-3-phenoxybenzyl (S)-2-(4-
chlorophenyl)-3-methylbutyrate chlorophenyl)-3-methylbutyrate

CAS No. 51630-58-1 CAS No. 66230-04-4

In addition to understanding a pesticide's three dimensional structure and

conformation, the method of entering structures into a (Q)SAR model should also
be taken into consideration. Some of the more common structural entry options
that have been historically employed for single structure entries include the
SMILES (simplified molecular input line entry system), International Chemical
Identifier (InChITM) codes, the MDL Mol file (MOL), and various drawing applets

Page 49 of 186
and molecular editors (Daylight Chemical Information Systems, 2008; IUPAC,
2010b; Dalby et al., 1992). For multiple (batch) chemical entries, the Structure
Data Format (SDF) file and SMILES (SMI) file formats are commonly used
(Dalby et al., 1992). These structural entry methods have strengths and
limitations, and in some cases, it may be necessary to verify the accuracy of the
structural representations from these methods to ensure that correct structures
are used for predictions.

4.3 Characteristics of the (Q)SAR Tool and the Prediction

Prior to considering and weighting the results of an empirical study in the
assessment of a pesticide, it is necessary to obtain and evaluate the details of
the study protocol and how the study was conducted, as well as the results of the
study and how they were interpreted. Similar concepts apply to the use of
(Q)SAR predictions in pesticide assessments, as the characteristics of the tools
used to make the predictions and the predictions themselves need to be
understood and evaluated before weighting the predictions in an assessment.

Many of the concepts discussed in this section overlap with the evaluation of the
scientific validity of a (Q)SAR tool as discussed in section 5.1. However, at the
problem formulation stage, it is intended that the evaluator will gain a basic
understanding of these issues. This may enable an immediate decision that the
(Q)SAR prediction is not adequate for the assessment context or it could lead to
a more detailed evaluation as discussed in section 5.1, especially with respect to
the application of the OECD (Q)SAR validation principles (section 5.1.1).

A starting point for characterizing a (Q)SAR tool at the problem formulation stage
is a sufficient understanding of the general methodology behind the tool. Is the
tool based on simple analog extrapolations, read-across or trend-analysis
approaches using chemical categories, a structural alert/rule based SAR/expert
system, a statistical (e.g., regression based) QSAR derived from a specific
database of chemicals and their descriptors or some other method? Each of
these methods has strengths and limitations that can influence how they should
be interpreted and the reliability of predictions from them. For instance,
SAR/expert systems based on structural alerts may be supported by expert
reviews of relevant research, and can include a mechanistic rationale to support
predictions. However, in some cases these systems do not include structural
alerts associated with inactivity, may have limited databases of alerts, and may
not have a clearly defined domain of applicability. Statistical QSAR models based
on training sets of active/inactive chemicals and descriptors of chemical structure
may provide insights into associations between specific structures and activity
that were not previously investigated, help to identify structures that modify or

Page 50 of 186
eliminate specific activities, and may be capable of generating quantitative
predictions (e.g., probabilities or specific numerical values) rather than
dichotomous active/inactive (yes/no) predictions. However, in some cases QSAR
models may overemphasize statistical associations in the absence of
mechanistic rationales, their domains may be restricted by the structural diversity
in their training sets, and their training sets may include chemicals with a variety
of different mechanisms which can result in poor predictive performance and/or
considerable uncertainty in their predictions. A number of reviews of the
strengths and limitations of (Q)SAR models are available in the scientific
literature (e.g., Hulzebos et al., 2001; Greene, 2002).

Gaining an understanding of the empirical data from which the (Q)SAR tool was
derived is another important starting point for determining whether a (Q)SAR
prediction is likely to be relevant to a pesticide assessment. It may be possible to
quickly discount (Q)SAR tools derived from studies based on outdated protocols
not conducted according to GLP standards, based on endpoints that are vague
or inconsistent, interpreted according to non-standard criteria, involving
chemicals significantly structurally dissimilar to the pesticide of interest, and/or
obtained from non-peer reviewed sources. On the other hand, (Q)SAR tools
based on higher quality empirical data may be subjected to a more detailed
evaluation and potentially included in a weight of evidence assessment.

As discussed in section 4.1, information on the endpoint on which a (Q)SAR tool

is based can be one of the important factors to consider at the problem
formulation stage for determining whether a (Q)SAR prediction will be relevant to
the specific pesticide assessment context. In particular, because many pesticide
assessment questions involve quantitative toxicity values (e.g., LD50, EC50,
NOAEL, etc.) for identifying labeling requirements, and calculating margins of
exposure, reference doses, etc., it is important to determine whether a (Q)SAR
tool can generate quantitative or qualitative predictions, and if quantitative, the
type of value predicted. A qualitative yes/no prediction for chronic toxicity will not
be particularly useful if a prediction of a NOAEL is required to derive a regulatory
point of departure for a pesticide. Alternatively, a (Q)SAR model that only
predicts quantitative LOAELs for short-term endpoints may also have limited
applicability. As mentioned above, predicted endpoints that are somewhat vague
such as general developmental toxicity potential may not be specific enough to
address questions about endpoints such as post implantation loss,
developmental delays, fetal dysmorphogenesis, etc. An overall point to consider
is whether there is likely to be sufficient, high quality empirical data available on
an endpoint of interest so that (Q)SAR tools could be developed that are relevant
to a particular assessment context.

Page 51 of 186
Investigating other details of the (Q)SAR tool used may also assist in determining
the relevance of a (Q)SAR prediction to a pesticide assessment during problem
formulation. For example, details on a (Q)SAR model such as the specific name
of the model, version number, date it was developed, and contact information for
the developer can be important for determining the relevance of a (Q)SAR
prediction. Model developers can make significant changes from one version to
another such as increasing the number and diversity of the chemicals in the
training set, modifying the library of descriptors or structural alerts, and modifying
model algorithms. As a result, predictions from a newer version of a model may
not be comparable to predictions from previous versions. Model developers can
even discontinue support for older versions making it difficult to obtain additional
information on training sets, interpretation criteria, etc.

Information on the prediction output should also be considered including the

actual prediction and information on the structural or other features of the test
pesticide that influenced the prediction. For dichotomous endpoints, predictions
may take the form of a positive/negative or active/inactive result, but often a
dichotomous result will be expressed as a numerical probability (i.e., 0 – 1) by
QSAR models or as a semi-quantitative probability (e.g., probable, likely, not
likely) by SAR/expert systems which can then be interpreted as positive or
negative according to various interpretation criteria. Information on the predicted
probability, and the interpretation criteria and the rationale for their use may
assist an evaluator to determine whether a prediction is relevant at the problem
formulation stage and/or can be considered when determining the reliability of a
prediction during a more detailed evaluation (see section 5.4). Numerical
endpoints (e.g., NOAEL, LD50, BCF) predicted from QSAR models may be taken
at face value, but in some cases specific criteria may be recommended by the
model developer or by the regulatory agency if the predictions are used to
support labeling requirements or hazard classifications (e.g., specific
classification/label statement when predicted value is within an order of
magnitude of value X). It should also be remembered that some numeric
predictions will need to be converted before application to a weight of evidence
assessment (e.g., conversion of a predicted LOAEL in units of mmoles/kg bw/day
to mg/kg bw/day).

Other information that can be important to consider at the problem formulation

stage includes the structural or other features of the test pesticide that have
influenced the (Q)SAR prediction including structural fragments, specialized
descriptors of structure (e.g., molecular size, shape and electronic parameters),
and physical-chemical properties (e.g., molecular weight, Log Kow, boiling point)
that are used as variables in QSAR model algorithms, and structural alerts that

Page 52 of 186
are used by SAR/expert systems to identify potentially active compounds.
Information on how these features influenced the overall prediction either
quantitatively or qualitatively can impact on the level of reliability assigned to a
(Q)SAR prediction (see section 5.4). In particular, it can be important to
investigate whether the structural fragments, descriptors, and/or physical-
chemical properties that drive a prediction are consistent with available
information on mechanism of action or not.

When using QSAR models, it can be important to review the identities of the
compounds similar to the test pesticide that influenced the prediction. This would
likely be obtained from an analysis of the training set compounds that formed the
basis for the model algorithm. Similarly, for SAR/expert systems, the compounds
that were utilized to support the development of any structural alerts identified in
the test pesticide could be reviewed. The compounds that make up a category or
group used in a read-across or trend analysis approach can also be considered
as compounds that are similar to the test pesticide and that directly influenced
the (Q)SAR prediction from that approach. Regardless of the type of tool used,
the identities of the compounds that influenced the prediction, how their similarity
to the test pesticide was assessed and the degree of similarity, how they
influenced the prediction, the nature of the empirical data for them that is related
to the predicted endpoint, how well they are predicted by the (Q)SAR tool (i.e.,
internal validation), and whether a mode and/or mechanism of action has been
established are all important considerations when determining the reliability of
the (Q)SAR prediction for a test pesticide (see section 5.4).

4.4 Empirical Data Including Information on Mode of Action

Although this section is intended to focus on the preliminary analysis of (Q)SAR
predictions as one of the sources of information in a problem formulation for
pesticide risk assessment, it must be remembered that the empirical database for
a pesticide can impact on the determination of relevance of a (Q)SAR tool for a
particular regulatory application (see section 5.3) and the reliability of predictions
obtained from that tool (see section 5.4). Empirical data that may influence the
use of (Q)SAR predictions includes not only the results of conventional toxicity
tests, but also information on mode (and mechanism) of action.

As indicated previously, it is likely that in most pesticide assessment scenarios,

(Q)SAR will not be used in a stand alone manner, but will represent only one of
multiple lines of evidence considered. Therefore, understanding what relevant
empirical data are available, the strengths and limitations of these data, and any
gaps that need to be addressed will facilitate the determination of whether there
are any (Q)SAR tools relevant for those gaps. Integrating existing empirical data

Page 53 of 186
on a pesticide with relevant and reliable (Q)SAR predictions could also help build
defensible rationales for requiring additional empirical studies on specific
endpoints, mode of action, etc. (e.g., targeted testing).

The integration of the empirical database with (Q)SAR predictions at the problem
formulation stage could also be important for more detailed evaluations of the
reliability of the (Q)SAR predictions at a later stage in the assessment (see
section 5.4). Questions to consider include whether a predicted endpoint for a
pesticide is consistent with and supported by empirical data for related endpoints
for the same pesticide or whether the prediction contradicts these data. Empirical
data for similar compounds, metabolites and degradation products can be
particularly important to consider when assessing the reliability of a (Q)SAR
prediction. Knowledge of the toxicity database for a parent pesticide compound
could impact on the level of confidence assigned to a (Q)SAR prediction for a
metabolite. In some cases, the consistency of the results of (Q)SAR predictions
for a parent pesticide versus a metabolite may be useful in determining the
confidence in the prediction for the metabolite, especially if the parent compound
contains structural alerts known to be associated with specific mechanisms of
toxicity and those alerts are preserved or activated following metabolic
transformation (e.g., substructures associated with DNA/protein binding). Also,
an evaluation of the existing empirical data for a pesticide may provide
justification for using more or less conservative criteria to interpret a (Q)SAR
prediction for that same pesticide.

As mentioned previously, information on mode of action for toxicity is one type of

empirical data that could impact on the consideration of (Q)SAR predictions at
the problem formulation stage of a pesticide assessment. A consideration of
mode of action can also include the pesticidal mode of action against the target
species, and any postulated modes of action of toxicity in non-target species
(e.g., humans) which could be used to support the results of existing (Q)SAR
predictions or rationales for generating additional (Q)SAR predictions and/or
obtaining additional empirical data.

If a pesticidal mode of action is not species specific (e.g., acetylcholinesterase

inhibition), information on this mode of action may support the need to investigate
related endpoints (e.g., neurotoxicity, developmental neurotoxicity) using (Q)SAR
predictions for various taxa, which could, in turn, lead to requirements for
additional in vivo studies of those endpoints in the relevant organisms. A
common pesticidal mode of action may also be a means of identifying groups of
similar pesticides from which to build categories and support data bridging
through read-across or other types of predictions.

Page 54 of 186
For a postulated (eco)toxicological mode of action, the extent to which the initial
chemical-biological system relationship is understood and how well the cascade
of key events leading to the adverse outcome is understood (i.e., mode of action,
mechanism of action, adverse outcome pathway) in taxa under consideration
could directly influence the level of confidence in (Q)SAR predictions for
endpoints associated with this mode of action. For instance, when a
(eco)toxicological mode of action has already been established for a structurally
similar compound, or for a chemical class in which the pesticide in question
resides, this mode of action could be used at the problem formulation stage to
focus (Q)SAR predictions on particular endpoints and taxa, bridge from the
structurally similar compound to inform dose selection for any study required for
the pesticide in question, provide support for waiving the need for specific studies
based on the current pesticide dataset, and/or help to rule out the relevance of
the observed or predicted effect to humans or other species, so that additional
studies are unlikely to be required.

Although information on postulated modes of toxicological action can provide

support for (Q)SAR predictions at the problem formulation stage and during
weight of evidence assessments, it should be noted that, mode of action
determinations are generally data rich decisions that must be made on a case-
by-case basis. The International Program on Chemical Safety (IPCS) has
developed an extensive framework for mode of action analysis based on the
Bradford Hill criteria which can be used for cancer and non-cancer endpoints in
the context of human health, and for ecological endpoints (Boobis et al., 2008). In
most situations, information on toxicological mode of action will not be readily
available for a majority of pesticides.

In some cases, comprehensive toxicological mode of action data for pesticide

may not be available, but it may be possible to use information on the chemical
structure of a pesticide and/or selected (Q)SAR tools (e.g., OECD QSAR
Toolbox) to identify potential (chemical) mechanisms of action of pesticides (e.g.,
mechanisms of protein or DNA binding) to assist in identifying analogs, grouping
chemicals into categories and supporting read-across extrapolations (OECD,
2011b; 2009; 2007a).

Page 55 of 186
4.5 Summary
The initial step in framing the questions to be addressed in the human health or
environmental assessment of a pesticide is problem formulation. Although the
questions to be addressed in pesticide risk assessments have traditionally been
framed in terms of the available empirical data, (Q)SAR predictions are another
source of information that can be considered during the problem formulation
process. The assessment context in which (Q)SAR is being applied, the
characteristics of the pesticide that is the subject of the prediction, the
characteristics of the (Q)SAR tool and the prediction, and the available empirical
data including mode of action data that could impact on the application of
(Q)SAR are all important factors to consider when integrating (Q)SAR into a
problem formulation. This type of preliminary analysis of the (Q)SAR information
on a pesticide could lead to an immediate conclusion that (Q)SAR is not suitable
for the particular pesticide assessment question or it could set the stage for a
more in depth evaluation of whether a (Q)SAR prediction is fit for purpose for
integrating into a weight of evidence decision (see section 5).

Page 56 of 186
5. Evaluating the Adequacy of (Q)SAR Predictions

EVALUATING THE ADEQUACY OF (Q)SAR PREDICTIONS

Topics Discussed in this Section:

 Scientific validity of a (Q)SAR tool

 Applicability of the (Q)SAR tool to the pesticide

 Relevance of the (Q)SAR tool to the assessment context

 Reliability of the (Q)SAR prediction

 Documentation of (Q)SAR tools and predictions

5.0 Introduction

Evaluating whether a (Q)SAR prediction is adequate or “fit for purpose” is an

important component of applying the prediction to a pesticide assessment. The
European Commission Joint Research Centre (JRC) has noted that whether a
prediction from a (Q)SAR model is adequate or not depends upon four key
factors: the scientific validity of the model, the applicability of the model to the
query chemical, the reliability of the (Q)SAR result, and the relevance of the
(Q)SAR model for the regulatory purpose. The validity of the model was to be
established through the application of the OECD QSAR validation principles
(OECD, 2004), the applicability of the model relates to whether the chemical of
interest lies within the model domain of applicability, reliability is based on the
application of a valid (Q)SAR to a chemical within its domain of applicability, and
relevance involves considering whether a predicted endpoint can be directly
applied to a particular regulatory purpose (EC, 2008b). Similarly, the REACH
guidance for applying (Q)SARs provides a flexible framework for using (Q)SAR
models in lieu of experimental data that is based on four main conditions: the
scientific validity of the model used, the applicability of the model to the chemical
of interest, the relevancy of the prediction for the regulatory purpose, and
whether appropriate documentation on the (Q)SAR and the prediction is provided
(ECHA, 2008; ECHA, 2010; Worth et al., 2011).
In this section of the NAFTA (Q)SAR Guidance Document, the key factors noted
by the JRC for assessing the adequacy of (Q)SAR models and the REACH

Page 57 of 186
framework have have been adapted to guide pesticide evaluators through the
information to be considered when evaluating whether predictions from (Q)SAR
tools are adequate for use in pesticide assessments. A schematic for the
resulting modified framework is shown in Figure 5–1. Evaluating the adequacy of
(Q)SAR predictions relies on a lot of the same information initially considered at
the problem formulation stage (see section 4), but with a more focussed
consideration of validity, applicability, relevance, and reliability. This type of
evaluation can be done in advance of or at least independently of the process of
combining the prediction with other information in a weight of evidence
assessment (see section 7). Since clear and complete documentation of (Q)SAR
tools and predictions is important both to the evaluation of the adequacy of
predictions and their incorporation into weight of evidence assessments, this
section also includes a discussion of documentation.

The guidance provided here is not meant to be prescriptive, but is intended to

allow for case-by-case flexibility and the incorporation of expert scientific
judgment. As such, it is recognized that the level of detail and effort employed in
evaluating the adequacy of predictions and documenting them will vary
depending on a number of factors including the assessment context in which
(Q)SAR is being applied.

Although the evaluation of the adequacy of a (Q)SAR prediction may be a new

concept to many pesticide evaluators, the process can be thought of as parallel
to evaluating the adequacy of empirical studies. When evaluating traditional
animal toxicity studies, evaluators can generally rely on the existence of validated
test guidelines that are applicable to most pesticides, whereas for (Q)SAR
predictions, additional effort needs to be invested to assess the validity of the
(Q)SAR tool and its applicability to the pesticide in question.

Much of the discussion in this section focuses on (Q)SAR models. However, it

should be recognized that the key issues to consider when evaluating the
adequacy of (Q)SAR predictions are applicable to all types of (Q)SAR tools.

Page 58 of 186
Figure 5–1: Evaluating the Adequacy of a (Q)SAR Prediction for a Pesticide
(modified from ECHA, 2008 and Worth et al., 2011)

RELIABLITY OF
THE (Q)SAR
PREDICTION APPLICABILITY OF
SCIENTIFIC VALIDITY OF
THE (Q)SAR TOOL TO
THE (Q)SAR TOOL
THE PESTICIDE

ADEQUACY OF
THE (Q)SAR
PREDICTION

RELEVANCE OF THE (Q)SAR TOOL

TO THE PESTICIDE
ASSESSMENT CONTEXT

5.1 Scientific Validity of the (Q)SAR Tool

The OECD has defined (Q)SAR validation as “the process by which the reliability
and relevance of a particular approach, method, process or assessment is
established for a defined purpose” (OECD, 2007c). In the context of (Q)SAR
model validation, the OECD considers that reliability focuses on the predictive
accuracy of the (Q)SAR tool for a range of different chemicals and relevance
refers to specific toxicological pathways and mechanisms that culminate in the
test endpoint. In particular, it is assumed that a (Q)SAR tool that has a
mechanistic basis for the predicted endpoint tends to be more relevant and
reliable for groups of chemicals acting via the mechanism in question (OECD,
2007c).

5.1.1 OECD (Q)SAR Validation Principles

The OECD previously noted that one of the critical challenges to the regulatory
acceptance of (Q)SAR predictions was the lack of an internationally harmonized
framework for assessing (Q)SARs. In particular, there was a need for an
internationally-agreed-upon set of principles for (Q)SAR validation to provide a

Page 59 of 186
scientific basis for making decisions on the acceptability of (Q)SAR predictions,
and to improve the transparency and consistency of (Q)SAR reporting leading to
a greater mutual acceptance of predictions (OECD, 2007c).

In response, the OECD developed the Principles for the Validation, for
Regulatory Purposes, of (Q)SAR Models which can be used as guidance for the
types of information to review when determining if a (Q)SAR model is acceptable
or not for use in a regulatory or decision-making framework. The principles
include “1) a defined endpoint, 2) an unambiguous algorithm, 3) a defined
domain of applicability, 4) appropriate measures of goodness-of-fit, robustness
and predictivity, and 5) a mechanistic interpretation, if possible.” (OECD, 2004).
The OECD also drafted and finalized a separate guidance document (Guidance
Document on the Validation of (Quantitative) Structure-Activity Relationships
[(Q)SAR] Models) that includes a discussion of the principles and information on
how to validate (Q)SARs for different applications (OECD, 2007c).

The five OECD (Q)SAR validation principles are presented in sections 5.1.1.1–
5.1.1.5 along with a summary of some of the key issues identified in the OECD
guidance document and other sources that should be considered in the context
of evaluating (Q)SAR tools for application to specific purposes in pesticide risk
assessments. For further details on the principles and their application,
evaluators should consult the OECD guidance document (OECD, 2007c).
Evaluators may also be interested in consulting a recent paper by Dearden et al.
(2009) which outlined 21 types of errors related to the OECD (Q)SAR validation
principles which were identified in various (Q)SAR analyses published in the
scientific literature.

Application of the principles is an important step in determining the adequacy of

(Q)SAR predictions for use in pesticide assessments. However, the OECD has
noted that because of the designs of many of the currently available (Q)SAR
models, it may not be possible to completely address all of the principles in every
case. Consequently, evaluators will need to be flexible and take into account the
available information on (Q)SAR tools and predictions, and individual regulatory
program requirements when applying the principles (OECD, 2007c). Also,
because of the range of (Q)SAR tools that could be used to make predictions for
pesticides and the varying levels of complexity of these tools, use of the OECD
(Q)SAR validation principles will require the application of expert scientific
judgment, in some cases from a multidisciplinary team.

Example No. 1 in Appendix III provides a summarized version of a case study of

reliability and validation testing of a set of (Q)SAR models for predicting acute
toxicity to fish species.

Page 60 of 186
5.1.1.1 Principle 1 — Defined Endpoint
The purpose of this principle is to make sure that the endpoint being predicted by
a given (Q)SAR tool is transparent. According to the OECD a “defined endpoint”
can be considered as “any physicochemical, biological or environmental effect
that can be measured and therefore modeled.” (OECD 2007c).

Unlike empirical data derived from standardized guideline based studies

designed to meet regulatory requirements for pesticides, studies for chemicals in
(Q)SAR model training sets may be based on non-standardized, non-uniform,
experimental protocols and conditions. The variability induced by these
differences can affect predictive performance and may be a limitation for some
(Q)SAR models. However, this variability does not necessarily invalidate the data
or models derived from them, but the characteristics of the data and their
potential impacts on model predictions must be taken into account.

No (Q)SAR model can be better than the data upon which it is based. Optimally,
all of the training set data for a particular (Q)SAR model should correspond to the
specific regulatory endpoint of interest, have been generated using the same
experimental protocol (ideally a standardized guideline type protocol), and be
interpreted using evaluation criteria that correspond to those of the specific
pesticide regulatory program. While this type of approach would help to ensure
the reliability and relevance of (Q)SAR predictions, (Q)SAR model developers
often have to rely on studies conducted under different protocols and conditions
in order to ensure sufficient numbers and diversity of chemical structures in the
model training sets (OECD, 2007c).

Variability can also be induced by the nature of the regulatory endpoint.

Regulatory test guideline type endpoints such as developmental toxicity may
actually encompass a range of subendpoints (e.g., teratogenicity, fetal growth
retardation, fetal death). Attempting to model poorly defined endpoints may result
in the use of model training sets containing a variety of chemical structures
producing different subendpoints via different mechanisms of action in a variety
of study types. Failing to take this variability into account can result in poor
correlations between model parameters and predicted endpoints resulting in poor
predictive performance. Alternatively, building a model for a more defined
endpoint such as a 96-hour LC50 in fish using a more mechanistically
homogeneous training set will likely produce better correlations and predictive
performance. Finally, in some cases, there may be uncertainties associated with
the model endpoint and training set data because information on study protocols
and evaluation criteria may not be readily available for some (Q)SAR models,
particularly certain commercial models. Pesticide evaluators should take these

Page 61 of 186
potential sources of variability and uncertainty into account when evaluating the
validity of a (Q)SAR tool.

5.1.1.2 Principle 2 — Unambiguous Algorithm

The second principle states that a (Q)SAR model should be associated with an
unambiguous algorithm. This means that the specifics of the relationship
between the chemical structures and the predicted endpoint or property (e.g., an
equation) should be clear and transparent. For a mathematically-based QSAR
model, the algorithm may take the form of a regression equation that relates
descriptors of the chemical structures to the predicted endpoint. Although it is
recognized that the unambiguous algorithm principle may be best applied to
statistical QSAR models, the OECD has extended the principles to other model
types, such as structural alert based SAR/expert system, where the algorithm
would take the form of expert-derived rules (OECD, 2007c).

Ideally, a (Q)SAR algorithm should be clear enough that an independent (Q)SAR

analyst should be able to explain how predictions were generated and reproduce
the results, if required. Although some (Q)SAR models that do not have
transparent algorithms may have equal or better predictive performance than
more transparent models, the lack of transparency of the former may negatively
impact their regulatory acceptance. While transparency is critical, the OECD has
stated that there is a difference between having a transparent algorithm and
being able to interpret the algorithm as a cause-and-effect relationship. The
descriptor values and equation for a QSAR model may be readily available, but a
mechanistic/causal link between the descriptors and the predicted endpoint may
not have been identified (OECD, 2007c).

In practice, the degree of transparency varies depending on the type of (Q)SAR

tool considered, as some non-commercial models have fully transparent
algorithms, while most commercial model developers consider specific algorithms
and how they were derived to be proprietary information. Also, what constitutes a
sufficient level of transparency for regulatory purposes will likely depend upon the
assessment context such that only limited information may suffice when (Q)SAR
predictions are used to prioritize inventories of chemicals for further
testing/assessment, whereas much more detail would likely be required for the
algorithm of a (Q)SAR model used to derive a quantitative prediction for a
regulatory point of departure estimate. In some cases, it may be possible to
compensate for less than complete transparency by analyzing a model’s
predictive performance for a set of chemicals similar to the test chemical, but not
used in the model training set, and for which empirical data are available (ECHA,
2008).

Page 62 of 186
5.1.1.3 Principle 3 — Defined Domain of Applicability
Netzeva et al. (2005) have defined the applicability domain of a (Q)SAR model as
“the response and chemical structure space in which the model makes
predictions with a given reliability.” This means the range of chemical structures,
physicochemical properties, mechanisms, and responses over which the (Q)SAR
tool can generate reliable predictions for the intended regulatory purpose. The
domain of applicability is dependent upon the set of chemicals on which the tool
is based (e.g., (Q)SAR model training set).

While it is possible to make predictions for chemicals outside of the applicability

domain of a (Q)SAR model, such predictions are extrapolations that are assumed
to be less reliable than predictions for chemicals within the domain of applicability
(i.e., interpolations). Also, because there are multiple ways of defining domain of
applicability (e.g., structures, physico-chemical properties, mechanisms), there
may be variations in the reliability of predictions even for chemicals within the
domain of applicability of a (Q)SAR tool. For instance, a prediction for a test
chemical that is structurally similar to chemicals in the training set of a (Q)SAR
model may still be unreliable if the test chemical has a different mechanism of
action compared to the chemicals in the training set (OECD, 2007c).

There is a balance between the overall range of the domain of applicability and
the predictivity of a (Q)SAR tool. Models with large training sets and diverse
domains of applicability may be capable of generating predictions for a wider
variety of chemical structures than smaller more structurally and mechanistically
homogeneous models, but there is a greater chance that many of those
predictions will be unreliable (OECD, 2007c; ECHA, 2008). Using information on
mechanisms, mode of action, and/or adverse outcome pathways to group
chemicals can improve predictive performance for large heterogeneous training
sets.

A number of existing commercial and non-commercial QSAR models have built-

in methods for determining whether a compound lies within the domain of
applicability. Examples include the univariate analysis (whether the training set
substructures include the substructures in the query chemical ) and multivariate
analysis (whether the query chemical’s descriptors are in the optimum prediction
space) in the TOPKAT program, the CAESAR models’ warnings for descriptor
values outside the range of the training set compounds and classes or groups of
compounds known to be less than optimally predicted, and the ASTER system’s
notification when a chemical is outside the predictive capability of a model
(Accelrys Inc., 2004; Benfenati, 2010; US EPA, 2011).

Page 63 of 186
The OECD guidance document on the validation of (Q)SAR models summarizes
a variety of different methods for defining domain of applicability including the use
of structural features that enhance (toxicophores) or modulate toxicity to define
the mechanistic domain, characterizing the descriptor or interpolation space by
graphing and distance (geometric) analysis, using Williams plots to visualize
outliers in descriptor and response space, comparing the structural and physical-
chemical similarity of the test chemical to the training set by fragment based
approaches, and other methods (OECD, 2007c). A number of reviews of different
methods for defining domain of applicability have also been published (Nikolova
and Jaworska, 2003; Dimitrov et al., 2005a; Jaworska et al., 2005; Netzeva et al.,
2005).

It should be noted that there is no single approach, or set of accepted

approaches, to assessing domain applicability. Consequently, whatever
approach is adopted should be transparently presented and documented.

In the context of (Q)SAR predictions for pesticide active ingredients, the need to
assess the domain of applicability cannot be over-emphasized. A long-standing
limitation of many commercial and non-commercial (Q)SAR models has been
domains of applicability that are not sufficiently representative of the structures
and mechanisms of action associated with pesticide active ingredients. This is in
part related to the nature of pesticide data (i.e., confidential unpublished studies
accessible only by regulatory agencies). Fortunately, this is changing over time
as resources such as the US EPA ToxRef database should it make it possible to
build models and other tools with domains of applicability that are more
encompassing of pesticide active ingredients.

5.1.1.4 Principle 4 — Appropriate Measures of Goodness-of-fit,

Robustness, and Predictivity
According to principle 4, a (Q)SAR should be associated with “appropriate
measures of goodness-of-fit, robustness and predictivity” which are obtained
through statistical validation of a (Q)SAR tool. For a QSAR model, goodness-of-
fit refers to how well the model accounts for the variability in the endpoint or
property measured for the training set chemicals. Robustness is a measure of
how much change will be induced in the coefficients, etc. in the model algorithm if
the training set chemicals are changed. Predictivity involves determining how well
the model can make predictions, generally for an external test set of data not
included in the training set (Eriksson et al., 2003).

Goodness-of-fit for regression-based QSAR models is usually expressed as a

multiple correlation coefficient (R2 value; range: 0 – 1) which is the amount of

Page 64 of 186
variation in the predicted values that can be explained by the regression
equation, and the standard error of the estimate (s) which measures the
dispersion of the predicted values around the regression line. Well-fitted models
have R2 values close to 1 and low s values. Poorly-fitted models are not likely to
be too useful for regulatory applications. However, it should be noted that
deceptively high R2 and low s values can be obtained by including a large
number of variables or descriptors in the regression equation (i.e., over-fitting a
model). Generally, better predictive performance can be obtained when the ratio
of the number of chemicals in the training set to the number of descriptors in the
regression equation (i.e., the Topliss ratio) is 5:1 or more. Note that R2 and s
values alone are not enough to assess model validity as they do not provide
information on the predictive performance for chemicals not included in the
training set of a (Q)SAR model (OECD, 2007c; ECHA, 2008).

For (Q)SAR tools that make dichotomous classifications (i.e., active/inactive,

positive/negative), goodness-of-fit is usually expressed as Cooper statistics such
as sensitivity (fraction of true positive chemicals predicted as positive), specificity
(fraction of true negative chemicals predicted as negative), accuracy (fraction of
true positive and negative chemicals correctly predicted as positive and negative,
respectively), and positive and negative predictivities (probabilities that chemicals
predicted as positive and negative are actually positive and negative,
respectively). Some (Q)SAR models can be biased towards high specificity or
sensitivity depending on the specific application they are designed for. Because
the Cooper statistics are interrelated, designing a model for high specificity can
result in decreased sensitivity (i.e., high false negative prediction rate) and vice
versa. This can be an important consideration as there is generally a greater
emphasis on correctly predicting positive chemicals (i.e., high sensitivity) for
pesticides and other environmental chemicals. Cooper statistics can also be
influenced by the distribution of positive and negative chemicals in the test set
such that the predictive performance for the largest class of chemicals in the test
set (i.e., positive or negative) will impact on the accuracy of the model and the
proportion of positive chemicals in the test set will influence the positive and
negative predictivities. The OECD has recommended that the Cooper statistics
be significantly greater than 50% for classification models used in a stand-alone
manner, but there is no absolute value for differentiating good from poor
predictive performance and a (Q)SAR tool with poor performance for one Cooper
statistic may still be useful depending upon the application (OECD, 2007c;
ECHA, 2008). Also, as discussed in section 6, it may be possible to combine
predictions from multiple (Q)SAR tools to enhance overall predictive
performance.

Page 65 of 186
Predictivity can be assessed by external validation, either through the use of a
test set of chemicals separate from the (Q)SAR model training set or by
separating a set of chemicals into a training set and a test set at the design stage
(Gramatica, 2007). External validation is usually measured by an external
correlation coefficient (Q2ext). External test sets should be of sufficient size and
representative of the types of chemicals to be predicted using the (Q)SAR model.

In some cases, model developers may also present the results of internal
validation techniques such as leave-one-out (LOO) and leave-many-out (LMO)
methods. For these methods, one or more chemicals is removed from the
training set, the model is re-built, the removed chemicals are predicted, the
process is repeated, and the average predictivity across the various versions of
the model is estimated as a cross-validated regression coefficient (Q2). One of
the reasons internal validation statistics are presented is that there may be
limited data from which to construct an independent external test set because
(Q)SAR model developers generally want to maximize the number of training set
chemicals, leaving few chemicals for external validation testing.

Q2 or Q2ext values of >0.5 and >0.9 are considered to represent good and
excellent performance, respectively, but it should be noted that predictivity is
dependent on the statistical method used and the composition of the test set.

Also, as stated previously, predictions outside the domain of the training set are
likely to be less reliable than predictions within the domain of applicability, so that
validation principle 4 is closely linked to validation principle 3 (OECD, 2007c;
ECHA, 2008).

It should be noted that not all elements of principle 4 are applicable to all (Q)SAR
tools, so the assessment of goodness-of-fit, robustness, and predictivity may
have to be made on a case by case basis. Rule-based SAR/Expert Systems that
use databases of structural alerts are one example in which there is generally no
training set and as such LOO, LMO, and other methods will not be applicable.
Also, when considering Cooper statistics for external validation testing, the
determination of specificity and negative predictivity may be difficult if the expert
system is only based on structural alerts for activity.

Page 66 of 186
5.1.1.5 Principle 5 — Defined Mechanism of Action, if Possible

The fifth validation principle states that a (Q)SAR model should be associated
with a mechanistic interpretation wherever possible. Although it is recognized that
mechanistic information is not always available for (Q)SAR models, whenever it
is available, it should be investigated and reported. A transparent mechanistic
interpretation can assist in the determination of whether the domain of
applicability of a model is suitable for predictions for the chemical of interest, help
with the interpretation of outliers, guide hypothesis testing, and provide support
for the biological plausibility (i.e., toxicological interpretation) and reliability of the
predictions from a model. However, the absence of a clearly identified
mechanistic basis for a model does not necessarily mean that the model is not
potentially useful for a given regulatory application (OECD, 2007c).

For QSAR models, a mechanistic interpretation represents the physical, chemical

and/or biological basis for the model descriptors and their relationship with the
endpoint or property to be predicted (ECHA, 2008). A mechanistic interpretation
can be associated with a QSAR model through the selection of mechanistically
relevant descriptors at the time of model development (i.e., a priori) or through
the investigation and delineation of the mechanistic basis for the descriptors in an
existing model (i.e., a posteriori) (OECD, 2007c).

As indicated previously using mechanistic similarity to group chemicals in the

training set of a (Q)SAR tool can provide a solid basis for QSAR model
development and interpretation. Consequently, when evaluating the mechanistic
basis of a (Q)SAR tool, it is important that the rationales for grouping training set
chemicals be presented, particularly with respect to any mechanistic hypotheses
that were applied (e.g., skin sensitization associated with protein binding
potential) and how the mechanistic hypotheses were translated into structural
inclusion/exclusion rules (e.g., grouping thiol compounds with potential for protein
binding via disulfide formation).

For knowledge-based SAR/expert systems and other related tools, the

mechanistic interpretation can be related to observed empirical data, expert
knowledge, and expert derived rules on the chemical reactivity and/or biological
activity of various chemical substructures (OECD, 2007c).

Page 67 of 186
5.2 Applicability of the (Q)SAR Tool to the Pesticide
Whether a (Q)SAR tool can be considered as applicable to a pesticide depends
upon the characteristics of the pesticide (see section 4.2) and the domain of
applicability of the (Q)SAR tool (see section 5.1.1.3).

In terms of the characteristics of a pesticide, accurate information on identity,

composition, and structure is necessary when determining whether a (Q)SAR
tool could be applicable. Many (Q)SAR models are limited to making predictions
for discrete organic chemicals and are incompatible with pesticides that are
mixtures, salts, or polymers. These incompatibilities may necessitate the use of
surrogate compounds such as monomers, uncharged acid forms, and single
mixture components to make predictions. Although surrogates may in some
cases be a useful approach to making predictions for pesticides that are
incompatible with available models, their use should be supported by rationales
that account for the potential impacts of molecular size and weight, ionization
state, variations in mixture composition, synergism/ antagonism between mixture
components, and other factors.

Similarly, if the isomeric form of a pesticide could have an impact on the endpoint
or property to be predicted, the (Q)SAR tool will need to be capable of
differentiating between isomers to be applicable. A QSAR model that uses 2-D
structural descriptors and only accepts 2-D structural representations of
chemicals to be predicted will not be very useful for predicting differences in
toxicity between stereoisomeric forms of a pesticide. A better approach would be
to use a QSAR model capable of recognizing structural representations of
isomers, that includes isomer specific descriptors, and whose training set is
sufficiently diversified with respect to data on different isomeric forms.

As discussed in section 4.2, for pesticides that can be transformed in the

environment or through metabolism in the body, the toxicity, ecotoxicity, physical-
chemical properties and other properties of the transformation products,
degradates or metabolites may differ from those of the parent pesticide. Whether
or not a (Q)SAR tool could be applicable to a pesticide that can be metabolized
or transformed would involve identifying metabolites or transformation products
(i.e., from empirical data or model predictions) and determining whether
predictions can be generated for them or not.

Section 5.1.1.3 outlines the concept of defining domain of applicability during the
evaluation of the validity of a (Q)SAR tool. While it is possible for some (Q)SAR
tools to make predictions for pesticides outside their domains of applicability,
those predictions are likely to be less reliable at best or in some cases the

Page 68 of 186
pesticides will be so far outside the domain of applicability that the (Q)SAR tools
should not be considered as applicable. As discussed, there are a number of
commercial and non-commercial (Q)SAR models that include automated
methods for assessing whether a chemical lies within their domain of applicability
based on limits on descriptor values, the presence of unrecognized structural
features, and other parameters. Also, a variety of different methods of defining
domain of applicability have been published (OECD 2007c; Nikolova and
Jaworska, 2003; Dimitrov et al., 2005a; Jaworska et al., 2005; Netzeva et al.,
2005).

5.3 Relevance of the (Q)SAR Tool to the Assessment Context

As noted by the JRC, the relevance of a (Q)SAR model involves considering
whether a predicted endpoint can be directly applied to a particular regulatory
purpose (EC, 2008b). This is based on the endpoint or property that the tool is
capable of predicting and the specific type of prediction information that the tool
can generate for a particular assessment context. The information obtained at the
problem formulation stage on the assessment context that (Q)SAR is being
applied to (section 4.1) and the characteristics of the (Q)SAR tool and the
prediction (section 4.3) can provide a useful starting point for assessing the
relevance of the (Q)SAR tool.

In order for a (Q)SAR tool to be relevant, the endpoint or property that it predicts
must correspond to the endpoint or property for which a data requirement exists
in a given pesticide assessment context. A (Q)SAR model, capable of generating
reliable predictions for the mutagenicity of chemicals in Salmonella typhimurium
TA1538 may provide useful information on the in vitro mutagenicity of a pesticide,
but it will not provide specific information to address a data requirement for an in
vivo clastogenicity study. Similarly, a positive prediction for general pre-natal
developmental toxicity for a pesticide may not be sufficient to address a question
about whether a pesticide can induce specific skeletal malformations.

Whether the endpoint or property predicted by a (Q)SAR tool could address a

specific pesticide data requirement involves a clear understanding of the data
from which the tool was derived. This corresponds to the OECD validation
principle of defined endpoint which was discussed in section 5.1.1 — i.e.,
understanding study protocols, data interpretation criteria, and other study
elements.

The type of information that a (Q)SAR tool can generate can also impact on its
relevance to a pesticide assessment question. In particular, (Q)SAR models are
usually designed to generate qualitative or quantitative predictions for particular

Page 69 of 186
endpoints. A model that can provide a qualitative (e.g., yes/no, positive/negative)
estimate of the toxicity of a pesticide to freshwater fish may provide some useful
information, but will be of limited relevance if a prediction of an acute LC50 in trout
is required for a particular assessment context.

5.4 Reliability of the (Q)SAR Prediction

In addition to considering the validity of a (Q)SAR tool for a particular pesticide
assessment context, the applicability of the tool to the pesticide, and the
relevance of the tool to the assessment context, it is also necessary to evaluate
the level of reliability (or confidence) in the individual prediction itself. Evaluating
the reliability of a prediction takes into account information gleaned from the
problem formulation process (see section 4) and information obtained when the
(Q)SAR tool is evaluated using the OECD validation principles (see section 5.1).

The pesticide assessment context is an overlying consideration when evaluating

the reliability of a (Q)SAR prediction. Moving from a less comprehensive to a
more comprehensive assessment context will likely require a higher level of
reliability from any (Q)SAR predictions used in the assessment. When rapidly
prioritizing chemicals for further assessment, it may be possible to take
predictions from validated (Q)SAR tools almost on face value. However, prior to
relying on a (Q)SAR prediction as a critical piece of information in a human
health or environmental risk assessment for a pesticide, the relationship of the
pesticide to the domain of applicability of the (Q)SAR tool, the strengths and
limitations of the tool, the prediction results and how they are interpreted, the
predictive performance of the tool for similar chemicals, and the potential impact
of other available information all have to be evaluated in more detail in order to
judge the reliability of the (Q)SAR prediction.

5.4.1 Relationship of the Pesticide to the Domain of Applicability of

the (Q)SAR Tool
The importance of considering the domain of applicability of the (Q)SAR tool has
already been mentioned with respect to the OECD validation principles (see
section 5.1.1.3) and the applicability of a (Q)SAR tool to a pesticide (see section
5.2), respectively. Section 5.1.1.3 also references a number of methods for
assessing domain of applicability.

Evaluating the relationship of the pesticide to the domain of applicability of the

(Q)SAR tool essentially involves determining whether the pesticide lies within the
domain of applicability or outside of it. As indicated previously, predictions for
pesticides outside of the domain of applicability of a (Q)SAR tool are not

Page 70 of 186
necessarily inaccurate, but are generally considered less reliable than predictions
for compounds falling with the domain of applicability.

As mentioned in section 5.1.1.3, domain of applicability may be defined in

different ways (e.g., descriptor, structural fragment, mechanistic, and metabolic
domains). Whether a pesticide is within the domain of a descriptor based QSAR
model is usually based on comparing the pesticide descriptor values to the range
of values for the chemicals in the training set. Structural fragment domain
analyses would involve ensuring that the pesticide doesn’t contain fragments that
are not present in the training set of the model. For the mechanism of action or
metabolic domain, the key question is whether the pesticide is likely to act via the
same mode/mechanism of action and/or be metabolized in the same manner as
other chemicals for which the (Q)SAR tool is applicable (EC, 2010). The OECD
has noted that because there are different ways of defining domain of
applicability, a prediction for a pesticide that is within the domain of applicability
of a (Q)SAR tool based on structural and physicochemical parameters may still
not be reliable if it has a unique mechanism of action not covered by the
mechanistic domain(s) of applicability of the (Q)SAR tool (OECD, 2007c).

The age of the QSAR model and its training set may also have impacts on the
consideration of the domain of applicability of the model and the reliability of the
prediction. An older, global type QSAR model may make a negative prediction for
a pesticide because its training set is populated with a limited number of
chemicals that contain the key structural elements in the pesticide and that all
tested negative in historical empirical studies. However, a more up-to-date
model, whose training set has been tested in more modern empirical studies, has
been segregated into groups according to mechanism of action, and contains a
larger number of compounds from the same chemical class as the pesticide of
interest, many of which have positive empirical test results, may generate a
positive prediction that is more reliable even though the pesticide falls within the
domains of applicability of both models. Consequently the use of the most up-to-
date versions of models and training sets is recommended and could be
particularly important when combining information from multiple predictions (see
section 6).

Finally, as discussed above, assessing the domain of applicability may be

particularly important for pesticides as (Q)SAR tool developers have not always
had access to proprietary pesticide empirical studies for incorporation into
training sets.

Page 71 of 186
5.4.2 Strengths and Limitations of the (Q)SAR Tool

The strengths and limitations of a (Q)SAR tool can impact on the evaluation of
the reliability of the predictions from that tool (Hulzebos et al., 2001; Greene,
2002). One source of strengths and limitations is the general methodologies on
which various (Q)SAR tools are based (e.g., analog approaches, chemical
categories, SAR and QSAR models, etc.) (see section 4.3). An example already
cited in this document is the lack of structural alerts linked to inactivity or negative
test results in some SAR/expert systems. If no structural alerts are identified for a
pesticide using this type of system and this is considered as equivalent to a
prediction of inactivity (negative), the prediction may be less reliable than a
positive prediction from the same system or a negative prediction from another
type of (Q)SAR tool that uses descriptors, alerts or other parameters directly
related to inactivity, depending on the assessment context. Similarly, the
overemphasis on statistical associations and lack of a mechanistic basis for
predictions may make some statistical QSAR models less reliable.

Built-in biases are another source of strengths and limitations of (Q)SAR tools
that could influence the reliability of predictions. For instance, some QSAR
models for pharmaceutical applications have training sets with distributions of
positive and negative compounds designed to generate higher specificity versus
sensitivity scores (Section 5.1.1.4). This type of bias needs to be taken into
account when models of this type are applied to pesticides as they may generate
a higher proportion of false negative predictions. The European Chemicals
Agency noted a potential source of bias for biodegradation models in their
guidance for the implementation of the REACH legislation. Because QSAR
models for biodegradation are often biased towards non-ready biodegradability,
predictions of biodegradability may be less reliable than predictions of non-ready
biodegradability (ECHA, 2008).

The sources of data for training set compounds, and the sources of data or
methods of calculation for descriptors (see section 5.1.1.1) can be another type
of strength or limitation of (Q)SAR model that could impact on the reliability of
predictions. While empirical datasets for registered pesticides usually consist of
peer reviewed guideline type studies, many model training sets are based on
open literature studies of varying quality. Also, as noted by Doull et al. (2007), for
some chemical classes, potential training set data may not be available from the
published literature. Similarly, the sources of the descriptor values and/or
methods used to estimate them may need to be scrutinized when evaluating the
reliability of a QSAR model prediction. Whether calculated descriptors, especially
obscure types, are reproducible or whether methods used to estimate descriptors

Page 72 of 186
for older versions of QSAR models have been supplanted by newer methods
could impact on the acceptability of predictions. These considerations also apply
to chemical category/read-across approaches. The methods used to identify
similar compounds, and the sources used for the endpoint related, physical-
chemical property, mechanistic and other data used to support chemical category
development and read-across predictions may need to be carefully considered
when determining the reliability of those predictions (OECD, 2007a).

5.4.3 Prediction Results and How They are Interpreted

Along with the basic qualitative (e.g., positive, negative, marginal) or quantitative
(e.g., LC50, LOAEL, TD50, etc.) prediction results, additional information is
available from most (Q)SAR tools which can be used to assist in evaluating the
reliability of predictions. Many QSAR models provide information on the structural
fragments, descriptors or physical-chemical parameters used as variables in their
algorithms. Examination of the values of these variables and their coefficients in
the model algorithm can indicate whether they positively or negatively influenced
a (Q)SAR prediction and the magnitude of their impact. Combining information on
the influence of structural fragments or descriptors on a model prediction with
knowledge of their relationship to the mechanism of action for the predicted
endpoint can provide powerful evidence to support or question the reliability of
the prediction. For instance, an increased level of reliability could be assigned to
a QSAR model prediction for a toxicity endpoint directly related to a receptor
binding process if the model algorithm contains descriptors of molecular size and
shape known to be related to receptor binding affinity and the values of those
descriptors for the pesticide in question are similar to those for chemicals known
to bind to the receptor and produce the effect in question.

Other information generated by some (Q)SAR models includes calculated values

for molecular weight of the test chemical, and properties such as Log Kow and
bioavailability (e.g., based on Lipinski’s rule of 5) which can help with the
consideration of whether a prediction for an endpoint could reliably represent
what might occur following an in vivo exposure to a chemical. Similarly, some
SAR/expert systems have expert rules that can take into account physical-
chemical factors which can impact on absorption/bioavailability by discounting
the presence of a structural alert associated with toxicity in a pesticide if the
physical-chemical parameters of the pesticide are outside the range normally
associated with the endpoint in question.

Understanding how the results of (Q)SAR predictions have been interpreted is

another consideration in the evaluation of their reliability. Algorithms for QSAR
models are generally derived from training sets of empirical study results for

Page 73 of 186
chemicals. How those empirical study results are interpreted can influence the
nature of the algorithm, the predictive performance of the model, and ultimately
the reliability of predictions from that model. For a carcinogenicity (Q)SAR model
developed from a training set of rodent bioassays, the bioassay results may have
been interpreted as positive based on a specific percentage increase in tumor
incidence over controls, a statistically significant increase in incidence over
controls, a statistically significant trend over several dose groups, and/or other
criteria. The reliability of a prediction from such a model could be influenced by
whether the study interpretation criteria were consistent among the training set
chemicals and whether the criteria correspond to regulatory agency specific
interpretation criteria.

The criteria for interpreting predictions that have been developed by the
originator of the (Q)SAR tool and the rationale for them should also be taken into
account when evaluating the reliability of predictions. Statistical-based QSAR
models often generate probabilities (i.e., 0 – 1.0) for dichotomous (e.g.,
positive/negative) endpoints and the model developers recommend specific
criteria for interpreting the predicted probabilities (e.g., TOPKAT criteria: ≥0.7 and
≤1.0 = positive; ≥0.0 and <0.3 = negative; ≥0.3 and <0.7 = inconclusive; Accelrys
Inc., 2004). Criteria of this nature are usually developed based on internal and/or
external validation testing to optimize the predictive performance of the model.

Although originators of (Q)SAR tools may recommend prediction interpretation

criteria, users of the tools may make modifications to those criteria. A pesticide
applicant or regulatory evaluator may decide that interpretation criteria put forth
by a (Q)SAR model developer are too conservative or not conservative enough
based on previous experience with the model, the results of validation studies or
other information. In some cases, comparisons of predictions for parent
compounds versus metabolites, data from related empirical studies, kinetic
and/or mechanism of action data, and other information may be used to modify,
override or contradict the interpretation criteria recommended by a model
developer. Also, regulatory agencies may develop standing policies on how
selected (Q)SAR tools should be interpreted that may differ from those of the tool
developers. Regardless of which criteria are used, they should be transparent so
that they can be considered in the evaluation of the reliability of the predictions.

Page 74 of 186
5.4.4 Predictive Performance of the (Q)SAR Tool for Similar
Chemicals
Testing the predictive performance of a (Q)SAR tool on chemicals that are similar
to the pesticide in question and have empirical data available for them can
provide another source of information for evaluating the reliability of predictions.
Chemicals from the same chemical class as the pesticide in question, as well as
isomers, salts, and other forms could be considered for testing the predictive
performance of the (Q)SAR tool. For example, a starting point for testing the
predictive performance of a (Q)SAR tool for a sodium salt of an organic acid
would be to generate a prediction for a de-salted acid form of the compound for
which empirical data are available. Which chemicals to use would depend on the
type and quality of empirical data available for them, the parameter used to
assess similarity (e.g., physical-chemical parameters, structure, metabolism) and
the degree of similarity.

As discussed in other sections of this document, one of the advantages of

applying (Q)SAR to pesticide metabolites, transformation products or
manufacturing impurities can be the abundance of high quality guideline study
derived data on a structurally related parent pesticide. In some cases, parent
pesticides with existing empirical databases can be used as external validation
sets or “positive and negative controls” for the associated metabolites,
transformation products or manufacturing impurities. When an endpoint for a
parent pesticide is well predicted by a (Q)SAR tool and the structures or
descriptors with the greatest influence on predictions are similar for the parent
pesticide and its metabolite, it may be possible to assign a greater reliability to
the prediction for the metabolite from the same tool.

5.4.5 Other Available Information

Although there may be some instances where (Q)SAR predictions will be used in
a stand-alone manner, in most pesticide assessments, predictions will be only
one of many lines of evidence to be considered. Therefore, when evaluating a
(Q)SAR prediction for a pesticide, it is important to consider the impact of other
available information on the evaluation of the reliability of the (Q)SAR prediction.
Available empirical data on a pesticide could contradict the (Q)SAR prediction or
support it. Also, even if the prediction is supported by the empirical data, the
degree of support may justify a more or less conservative interpretation of the
prediction. For example, a prediction of high acute inhalation toxicity for a
pesticide may be generated by a QSAR model, but contradictory empirical data
on the physical-chemical properties of the pesticide may indicate low volatility or
low potential for aerosolization which may lead the evaluator to question the

Page 75 of 186
reliability of the prediction and seek additional data or predictions for structurally
similar chemicals. On the other hand, there may be cases where precursor
effects in a target organ consistently reported in short-term studies may be used
to support a (Q)SAR prediction of a related longer-term effect (e.g.,
carcinogenicity) in the same target organ and species by the same route of
exposure.

Just as a defined mechanism of action can be an important consideration when

evaluating the validity of a (Q)SAR tool (section 5.1.1.5), empirical data on mode
or mechanism of action can be an important consideration when evaluating the
reliability of an individual (Q)SAR prediction. Even if a (Q)SAR tool is not based
on a defined mode or mechanism of action, high quality, empirical toxicological
mode or mechanism of action data for a pesticide can represent a very powerful
line of supporting or contradictory evidence for a prediction from that tool,
including supporting or contradicting the relevance to humans and/or
environmental organisms of the predicted toxicity endpoint. For instance,
because of the relationship between the protein binding capability of chemicals
and their skin sensitization potential, it could be assumed that a category
composed of chemicals that are not only structurally similar to a test pesticide,
but are also known to bind to proteins via the same mechanism as the test
pesticide will likely result in a read-across skin sensitization prediction that is
more reliable compared to predictions where information on protein binding
mechanism is not available (Dimitrov et al., 2005b). One example of the impact of
mode of action data on the human relevance of a (Q)SAR prediction could
involve the interpretation of a positive prediction for renal tumors in male rats that
has been statistically validated and has a domain of applicability that
encompasses the pesticide in question. The interpretation of such a prediction
may have to be tempered if data are available from short-term or specialized
mode of action studies that indicate the accumulation of alpha-2u-globulin in the
kidneys of male rats administered the pesticide, a potential mode of action of
questionable relevance to humans.

As outlined in section 4.4, empirical data on the pesticidal mode of action of

pesticides could provide support and enhance the reliability of (Q)SAR
predictions for related endpoints in humans and other non-target species. Also, a
common pesticidal mode of action could be one of several lines of evidence
supporting the formation of a category of related chemicals. A category that is
based not only on structural and physical-chemical similarities, but also similarity
of pesticidal mode of action amongst category members is likely to result in more
reliable read-across and trend analysis predictions.

Page 76 of 186
5.5 Documentation of (Q)SAR Tool and Prediction
In order for an evaluator to critically review the adequacy of a (Q)SAR prediction
for a pesticide, the (Q)SAR tool and the prediction must be documented with a
sufficient level of transparency. This is similar to the concept of sufficient
documentation for empirical studies as delineated in empirical study guidelines
and in guidance for producing robust study summaries for regulatory purposes.
What constitutes a sufficient level of transparency will depend on the assessment
context, specific data reporting requirements or policies of the regulatory agency
and the type of (Q)SAR tool. When predictions of toxicity, ecotoxicity,
environmental fate, etc. are used in a prioritization or screening context, it may
not be necessary to provide full details on the adequacy of the (Q)SAR
predictions. However, for a (Q)SAR prediction to be accepted as a critical data
point in a pesticide assessment would likely require less uncertainty, and thus
more extensive documentation analogous to pesticide data evaluation records
(DERs) used to capture critical information from conventional toxicity, exposure,
and other study types.

5.5.1 General Types of Information

At the present time, a standardized template for reporting information on (Q)SAR
predictions included in pesticide assessments has not yet been developed.
However, in lieu of such a template, some recommendations can be made
regarding the general types of information to report (see Table 5–1). In general,
for any (Q)SAR prediction, sufficient information must be provided to clearly
identify the chemical for which the prediction is being made and the model or
other (Q)SAR tool used to generate the prediction. A description of the results of
the prediction and how they were interpreted should also be presented as well as
a discussion of the validity of the tool in the context in which it is being used.
These recommendations should only be considered as a starting point for what to
include when documenting (Q)SAR predictions for pesticides. As indicated
above, the level of detail to be included when reporting on predictions will depend
on the assessment context, specific data reporting requirements or policies of the
regulatory agency and the type of (Q)SAR tool.

While the recommendations in Table 5–1 are fairly general, they include a
number of information elements that may be more suitable for (Q)SAR models
than other tools such as analog and category approaches. Specific guidance on
reporting formats for analog and category approaches has been developed by
the OECD (OECD, 2007a).

Page 77 of 186
Table 5–1: Recommended General Types of Information to Include when
Documenting (Q)SAR Predictions
Information on the chemical
• Chemical (systematic) and common names
• CAS number
• Structural formula
• Form of the chemical (including relevant stereochemistry)
• Structural entry format

Information on the (Q)SAR model or tool

• Type of model or tool
• Name of the software platform
• Name of model/submodel or tool
• Version number/date of model/submodel or tool
• Characterization of the training set

Validity of the (Q)SAR model or tool

• Information on the predicted endpoint
• Information on the algorithm
• Domain of applicability
• Internal/external validation statistics
• Mechanistic information (if available)

Results and interpretation of the prediction

• Qualitative prediction (i.e., yes/no, +/-)
• Quantitative prediction (e.g., LOAEL = 50 mg/kg bw)
• Predictive probability
• Criteria for interpreting predictions

5.5.2 Documentation of (Q)SAR Predictions — EC QMRF and QPRF

The European Commission (EC) has developed the QSAR Model Reporting
Format (QMRF) and the QSAR Prediction Reporting Format (QPRF) as detailed
documentation templates for providing sufficient information to facilitate
regulatory consideration of (Q)SAR models and predictions. The QMRF has been
designed to provide information related to the OECD principles for the validation,
for regulatory purposes, of (Q)SAR models and the QPRF has been developed to
provide information to assist in the consideration of the adequacy of a (Q)SAR
prediction for a defined regulatory purpose (EC, 2008a; 2008b). In terms of
parallels with the assessment of empirical studies, the QMRF is somewhat
analogous to an empirical study test guideline and the QPRF is analogous to a
study DER. However, the QMRF and QPRF are very detailed and were not
specifically designed for use in a pesticide context. They are discussed here only
as useful examples to consider when specific (Q)SAR reporting templates for
pesticides are being developed. One of the potential projects under consideration
as a follow-up to this document is the development of pesticide specific (Q)SAR
DERs. Further background information on the QMRF and QPRF, tables

Page 78 of 186
summarizing the information fields in each, and the EC website link for the
templates and additional guidance on their use are included in Appendix II.

5.6 Summary
Evaluating the adequacy of a (Q)SAR prediction is an essential step for
determining whether the prediction is useful source of data for a pesticide
assessment. Adequate or fit for purpose predictions can be incorporated into
weight of evidence assessments, whereas inadequate predictions necessitate a
reliance on other sources of data alone. Whether a prediction is adequate or fit
for purpose should always be determined within a specific assessment context
such that a prediction that is adequate to support a request that specific empirical
studies be conducted may not be adequate enough to replace those studies in
an assessment. In this section, a framework has been presented for evaluating
the adequacy of predictions based on a consideration of the validity of the
(Q)SAR tool, the applicability of the (Q)SAR tool to the pesticide of interest, the
relevance of the (Q)SAR tool to the pesticide assessment context, and the
reliability of the prediction (Figure 5–1). This framework relies on information
obtained during problem formulation for (Q)SAR (see section 4) and it is flexible
enough to be useful for a variety of different assessment contexts. The next
section of this document (section 6) deals with combining information from
multiple predictions. While there are specific issues associated with combining
multiple predictions, in general, the framework for evaluating the adequacy of
(Q)SAR predictions outlined in section 5 can also be applied to multiple
predictions. For either single or multiple predictions that have been determined to
be adequate, the next step is the incorporation into a weight of evidence
assessment which is the subject of section 7 of this document.

Page 79 of 186
6. Combining Information from Multiple Predictions

COMBINING INFORMATION FROM MULTIPLE PREDICTIONS

Topics Discussed in this Section:

 Approaches to combining multiple predictions

 Advantages and disadvantages of combining predictions

 Selecting (Q)SAR tools for multiple predictions

 Evaluation of multiple predictions

6.0 Introduction
Combining information from multiple (Q)SAR predictions can be thought of as
analogous to combining the results of multiple in vivo and in vitro studies to
strengthen a weight of evidence argument for a toxicity or ecotoxicity endpoint.
Because different (Q)SAR tools may have different prediction paradigms and
different strengths and limitations, combining predictions has the potential to
increase the confidence in the overall prediction. However, it should be noted that
combining predictions from multiple (Q)SAR tools does not eliminate the need to
ensure that each prediction is adequate or fit for purpose (see section 5). In
particular, it is important that each (Q)SAR tool used be scientifically valid,
applicable to the pesticide of interest, and relevant to the assessment context.

Also, while there are advantages to combining predictions, it is not always

necessary to do so. In some instances, relying on a single prediction from a
(Q)SAR tool that is valid for the stated purpose, applicable to the pesticide in
question, and relevant to the assessment context may be much more acceptable
than trying to combine predictions from tools with significant limitations.

This section briefly discusses approaches to combining information from multiple

predictions, some advantages and disadvantages of combining predictions,
selecting (Q)SAR tools for multiple predictions, and the evaluation of multiple
predictions. Although much of the discussion focuses on combining predictions
from (Q)SAR models, many of the concepts mentioned can also be applied when
predictions from different types of (Q)SAR tools are combined.

Page 80 of 186
6.1 Approaches to Combining Multiple Predictions
Combining the output of multiple (Q)SAR models into an overall prediction has
been referred to as consensus modeling, battery approaches, or weight of
evidence approaches to (Q)SAR modeling (Abshear et al., 2006; OECD, 2007;
Hewitt et al., 2007; Matthews et al., 2008, 2009a; Ellison et al., 2010; Hewitt et
al., 2010). An example of a fairly simple approach to consensus predictions is the
set of interpretation criteria in the KnowITAll computational system described by
Abshear et al. (2006) which are summarized in Table 6–1 below. These criteria
range from single hit/unanimity requirements for true/false predictions based on
the worst case scenario and vice versa for the best case scenario, to the
assessment of counts of true and false predictions for the majority and
percentage agreement scenarios. For toxicity predictions, a value of true can be
considered as equivalent to a positive prediction and value of false as equivalent
to a negative prediction.

Table 6-1: Consensus Modeling Interpretation Criteria

(Abshear et al., 2006)

Scenario Definition
1. If any model returns a value of true, return a value of true
Worst Case
2. Only if all models return a value of false, return a value of false
1. If any model returns a value of false, return a value of false
Best Case
2. Only if all models return a value of true, return a value of true
1. If the majority of the models return true, the consensus will be true
Majority Rules
2. If the majority of the models return false, the consensus will be false
1. If a specified percentage of the models returns a true value, the consensus
Percentage
will be true
Agreement 2. Otherwise, the consensus will be false

In contrast, a slightly more complex method of combining predictions is the

weight of evidence approach of Ellison et al. (2010) in which predictions from the
OECD QSAR Toolbox, Derek for Windows, CAESAR and SMARTS rules were
combined for skin sensitization with positive predictions given a weighting of
+0.5, negative predictions given a weighting of -0.5, and the absence of structural
alerts in a compound given a weighting of -0.25. The sum of the weightings was
then interpreted in a weight of evidence argument as positive:
≥ +0.5, negative: ≤ -0.5, and inconclusive: > -0.5 and < +0.5.

Additional factors that could be considered when weighting individual predictions

in a consensus approach include characteristics of the models or tools that are
being combined such as the presence/absence of a mechanistic basis for the
predictions, the characteristics of the model training set (e.g., data sources and

Page 81 of 186
domain of applicability), and the known predictive performance of the models or
tools. Finally, when predictions are made for a quantitative (i.e., continuous)
endpoint (e.g., LC50, LOAEL, TD50, EC50, etc.), it may, in some cases, be
possible to average the predicted numerical values or combine them using other
statistical methods. Consensus predictions can vary in complexity depending on
the tools considered and the methods of counting, scoring or weighting the
individual predictions.

6.2 Advantages and Disadvantages of Combining Predictions

6.2.1 Advantages of Combining Predictions

A number of researchers have demonstrated improvement in measures of
statistical fit and predictive performance when multiple (Q)SAR predictions for
human health related and environmental toxicity endpoints are combined. One
example is the study of Matthews et al. (2009a) involving the generation of
consensus predictions for drug-induced adverse liver and urinary tract effects
using training sets configured for four (Q)SAR programs (i.e., MC4PC,
MDL-QSAR, BioEpisteme and Predictive Data Miner). Consensus predictions
where a positive prediction from at least one program was considered as an
overall positive result resulted in an increase in sensitivity from 39% for one
program to 56% for two, and to 68% for four programs. Increased sensitivity was
at the cost of specificity which decreased from 86% for one program, to 78% for
two, and to 67% for three. Consensus predictions requiring agreement between
two, three and four programs increased the specificity by 4, 9, and 12%,
respectively, compared to single program predictions. In this case, sensitivity was
increased for consensus predictions from two or more programs, but reduced by
11 and 26% for predictions requiring agreement between three and four or more
programs, respectively. Similar improvements in predictive performance were
obtained for carcinogenicity predictions using models built from the four (Q)SAR
programs above in an earlier study (Matthews et al., 2008). Matthews et al.
(2009a) concluded that no one (Q)SAR model can provide both high specificity
and high sensitivity. Combining models that have good specificity individually in a
consensus approach can enhance the overall sensitivity, which can be an
important consideration for models used in the assessment of pesticides.

Another example of improved predictive performance from consensus modeling,

is the work of Lewis et al. (2002) in which a predictive concordance of 100% was
obtained from combined COMPACT and Hazard Expert predictions for a small
group (14) of carcinogens. The concordances were 71% for COMPACT alone

Page 82 of 186
and 57% for Hazard Expert alone. 4 However, not all studies have demonstrated
improved model statistics for consensus versus individual models. Hewitt et al.
(2007) used genetic algorithms to construct a range of models for four different
data sets (silastic membrane flux, toxicity of phenols to the ciliated protozoan
Tetrahymena pyriformis, acute toxicity in fathead minnow and flash point). There
was no consistent improvement in statistical fit or predictivity (i.e., R2, Q2 root
mean square error) for average predictions from a consensus of the 10 best
models (i.e., models with the highest R2 and Q2 values); or a consensus of a
diverse set of models that best covered the available model space compared to
the single regression model with the best R2 and Q2 values.

The potential for multiple (Q)SAR models to provide complementary or

confirmatory information compared to individual predictions is another advantage
of consensus approaches. Matthews et al. (2009a) defined complementary
models as two or more models that predict different sets of active and inactive
chemicals when used on the same test set and noted that combining
complementary models can enhance predictive performance (e.g., sensitivity).
Combining complementary models could also enhance or expand the overall
coverage or domain of applicability. For instance, if within a set of 100 pesticides,
40 contain key structures that are in the domain of applicability of (Q)SAR model
A, whereas the domain of applicability of (Q)SAR model B covers half of those
same 40 pesticides plus key structures found in the 60 pesticides not covered by
model A, then combining predictions could provide complementary coverage of
all 100 pesticides. Combining predictions from multiple models for similar or
related endpoints could also provide complementary information that increases
the reliability of the overall assessment. A single positive prediction for
mutagenicity in a (Q)SAR model for one specific strain of Salmonella
typhimurium provides limited information on the potential microbial (prokaryotic)
mutagenicity of a compound. Combining multiple positive predictions from
models for several Salmonella strains, models for in vitro mutagenicity in
mammalian cell systems, and models for in vitro and in vivo chromosomal
aberration assays could provide complementary information that when combined
with evidence of carcinogenicity from in vivo bioassays may provide an initial
indication of a potential genotoxic mode of action for a carcinogenic compound.

Models that are based on different predictive paradigms (e.g., molecular

fragment versus molecular descriptor paradigms) that predict the same
chemicals to be active and inactive from a single test set provide confirmatory

4
Although the COMPACT model is not currently available, the example was included to illustrate
some of the advantages of combining predictions from different (Q)SAR models.

Page 83 of 186
information (Matthews et al., 2009a). This type of confirmatory information can
increase the overall confidence in predictions (Contrera et al., 2007), especially
when additional mechanistic insights are provided by one or more of the models.
Individual models can emphasize a set of structural features in a molecule while
placing reduced or no emphasis on other features (Gramatica et al., 2007).
Consequently, combining multiple (Q)SAR models that are based on different
methodologies can help to relate the activity of a compound to different aspects
of its structure, confirming the impact of key structural features on activity or
providing additional insights into the key parts of a compound's structure that
influence activity (Contrera et al., 2007). Confirmatory predictions also increase
the likelihood that the structures in the active compounds are causally related to
the activity in question and that the compounds come from clusters with the same
mechanism of action (Matthews et al., 2008; 2009a). However, combining
predictions from models based on the same methodology (e.g., several statistical
(Q)SAR models based on similar descriptors) and developed from the same
training set would not be expected to provide much additional information.

6.2.2 Disadvantages of Combining Predictions

Combining information from multiple predictions has the potential to greatly
increase the complexity of the predictive process in terms of selecting models to
be combined, approaches to combining models, and interpreting the combined
predictions. In the study of Hewitt et al. (2007) described above, the authors
noted the complexity of generating multiple models from a range of descriptors
for each endpoint, and in assessing which models and how many should be used
from the global model space to construct consensus models. Also, when a large
number of descriptors is considered, consensus models may be based on
descriptors that are difficult to interpret mechanistically.

Conflicting predictions from individual models for the same endpoint may
represent another source of complexity in consensus approaches. While the
evaluation of the adequacy of the individual (Q)SAR predictions (see section 5)
may help to resolve some differences, conflicting predictions may still occur if
different (Q)SAR paradigms are employed and/or the models are based on
different training data sets. While simplified approaches for reconciling conflicting
predictions may be adopted, such as the criteria summarized by Abshear et al.
(2006), quantitative or semi-quantitative weighting of models may be necessary
to account for differences in model domains of applicability, scoring criteria for
training set data, performance measures, mechanistic bases and other
characteristics. When multiple factors are influencing the weighting, multiple
predictions can be very complex to interpret, especially when there is a need to

Page 84 of 186
additionally weight the (Q)SAR predictions against other data available on a
chemical.

Resources needed to develop optimized consensus approaches could be

another disadvantage of combining multiple predictions. While it may be possible
to predefine weightings of (Q)SAR tools based on their known characteristics,
finalized weightings of tools may require multiple rounds of testing and analysis
(EC, 2010).

6.3 Selecting (Q)SAR Tools for Multiple Predictions

Many of the considerations in selecting models for multiple predictions
encompass similar issues to those discussed in the problem formulation section
(section 4) of this guidance document — i.e., the assessment context,
characteristics of the pesticide, characteristics of the (Q)SAR tool, and available
empirical data including information on mode of action.

The assessment context, including whether the prediction is used to support a

data waiver or to identify data requirements, what type of endpoint or property is
being predicted, and whether the predicted endpoint or property is critical to a
pesticide assessment will influence the level of confidence or reliability required
in the combined predictions which, in turn, can be factored into the selection of
appropriate (Q)SAR tools. Similarly, the amount and quality of the available
supporting empirical data including data on mode of action can also impact on
the required level of confidence in the combined predictions and the choice of
tools used to make those predictions.

As discussed in section 4, certain (Q)SAR models may not be compatible with

the molecular structure of some pesticides (e.g., ionic compounds, complex
mixtures, polymers, etc.) and others, though compatible, may have limitations
such as insufficient structural diversity in their domains of applicability or the
inability to generate quantitative predictions. As a result, the characteristics of the
pesticide and of the (Q)SAR tool can also influence the selection of appropriate
(Q)SAR tools for the generation of multiple predictions for a pesticide.

While selection of models for specific scenarios will most likely be made on a
case by case basis taking into account the factors described above and expert
judgment, one example of an approach that has been used previously is the
selection of a molecular fragment-based method (Q)SAR model and a descriptor-
based model, both having high predictive performance for a toxicological
endpoint of interest. Such a combination could have a good chance of improving
the combined domain of applicability of the models, and positive predictions

Page 85 of 186
could provide strong evidence that the fragments/descriptors associated with the
toxicological activity are highly significant and well separated from the structural
features of inactive molecules.

Another example of an approach to selecting (Q)SAR tools for multiple

predictions would be combining complementary models with different prediction
paradigms, such as molecular fragment-based (Q)SAR methods that detect small
molecular fragment alerts, large molecular fragments, and very large molecular
fragments. Relative confidence in the predictions could be increased if these
different methods identify a common region of a molecule that is correlated with
activity. Similarly, approaches based on combinations of predictions from
molecular descriptor-based methods that use different pools of descriptors and/or
different statistical methods to detect activity could also be considered.
Confidence in the predictions is increased when these different methods identify
a common physicochemical structural feature of a molecule that is correlated with
activity. The work of Matthews et al. (2008, 2009a) provides some examples of
combining predictions from (Q)SAR models with different prediction paradigms.

6.4 Evaluation of Multiple Predictions

Evaluation of multiple predictions involves similar considerations to those
discussed in section 5 of this guidance document relating to evaluating the
adequacy of single predictions — i.e., scientific validity of the (Q)SAR tools,
applicability of the (Q)SAR tools to the pesticide, reliability of the combined
predictions, relevance of the (Q)SAR tools to the assessment context, and
documentation of the tools and predictions. Also, as mentioned previously,
combining predictions from multiple (Q)SAR tools does not eliminate the need to
ensure that each prediction is adequate or fit for purpose.

In addition to the concepts discussed in section 5, another issue to consider

when evaluating the results of multiple or consensus predictions is the overall
objective of combining the predictions. Have predictions been combined to
confirm or to increase the confidence in the predictions from one (Q)SAR model
based on the results of other models that use different predictive paradigms? As
mentioned above, identical predictions from multiple (Q)SAR models developed
from the same training set using very similar predictive methodologies may not
provide much additional information, whereas models based on different
paradigms that identify the same chemicals to be active and inactive may
increase the confidence in the predictions (Contrera et al., 2007). Alternatively, is
there an interest in improving the overall predictive performance for a given
endpoint or parameter by combining complementary models that individually
predict different chemicals to be active and inactive? Such an approach could

Page 86 of 186
potentially enhance predictive performance compared to the individual models
(Matthews et al., 2009a). Knowing the objective of combining predictions and the
characteristics of the (Q)SAR tools used can help the pesticide evaluator
determine whether it is appropriate to combine predictions or not.

The adequacy of the training sets in the models is another important factor in
determining whether it is appropriate to combine predictions (Matthews, 2009b).
If confirmatory predictions are required from multiple models then the models
should have comparable training sets in terms of coverage of the chemicals to be
predicted. Also, the scoring systems or criteria used to characterize the data on
the training set chemicals should be comparable. For example, two models
based on similar training sets of Ames test results, but constructed using different
scoring systems for what constitutes a positive versus a negative assay result
may not give reliable predictions when combined. Whether the training sets have
been designed to be balanced or heavily weighted towards active or inactive
chemicals should also be considered. If the training set for one model has a
relatively low ratio of active to inactive chemicals (A/I ratio) and a high sensitivity
prediction is desired, then it may be better to combine the predictions from this
model with predictions from models that have higher A/I ratios in order to
enhance the chance of correctly predicting positive chemicals.

If it is considered appropriate to combine predictions, then how the predictions

are combined and interpreted (see section 6.1) is another important question. For
example, a pesticide applicant may put a higher priority or heavier weight on
models with a high level of specificity when addressing submission data
requirements in an effort to support waiver rationales for multiple in vivo studies.
Conversely, for a metabolite or residue of potential concern for which little
empirical data are available, pesticide evaluators may assign a greater weight to
models with high sensitivity in order to ensure potential endpoints of concern are
flagged for additional data requirements. Interpretation criteria for multiple
predictions may also be designed to compensate for the strengths and limitations
of the individual models being combined such as giving precedence or higher
weighting to predictions from tools with more extensive domains of applicability
encompassing the test compound, superior overall predictive performance,
clearer mechanistic bases, greater transparency, etc. Regardless of the
approach used to combine and interpret multiple predictions, the interpretation
criteria should be transparent and the rationale for their use should be included
when documenting predictions. Documentation for combining predictions may
consist of separate records for each prediction (e.g., QPRFs) with an
accompanying rationale for combining predictions, or, alternatively a single

Page 87 of 186
document may be used to capture the individual predictions and the rationale for
combining predicitions.

Finally, if it is determined that combining (Q)SAR predictions is appropriate for a

particular assessment context, sufficient information on training sets, scoring
systems, interpretation criteria, etc., should be available on each model used in
the combined prediction so that their strengths and limitations are transparent
and if necessary, their predictions can be weighted before they are combined.

6.5 Summary
In general, the aims of combining multiple predictions include enhancing
predictive performance, expanding domain of applicability, obtaining
complementary or confirmatory information, and ultimately increasing confidence
in (Q)SAR predictions. Although there can be advantages to combining multiple
predictions, it is not always necessary, as a single prediction from a validated,
applicable and relevant (Q)SAR tool is likely to be more acceptable than
combined predictions from tools with significant limitations. The selection of
appropriate tools for generating multiple predictions for a pesticide will involve a
trade-off of the desired advantages against the potential disadvantages and a
consideration of the assessment context, characteristics of the pesticide,
characteristics of the (Q)SAR tools, and the available empirical data. Combining
information from multiple predictions does not represent a new data stream for
consideration in pesticide assessments. Rather it is a variation in the (Q)SAR
data stream that needs to undergo the same types of problem formulation,
evaluation of adequacy, and weight of evidence considerations as single (Q)SAR
predictions prior to being incorporated into pesticide assessments.

Page 88 of 186
7. Integration of (Q)SAR Predictions into Hazard
Assessments

INTEGRATION OF (Q)SAR PREDICTIONS INTO HAZARD ASSESSMENTS

Topics Discussed in this Section:

 Incorporating (Q)SAR in hazard characterizations: Overview

 Problem formulation and Adequacy Determination

 Evaluating empirical data versus (Q)SAR predictions

 Mode of action considerations

 Overall weight of evidence

 Hazard characterization and risk communication

7.0 Introduction
Traditional pesticide risk assessments in regulatory agencies have routinely been
based on the results of laboratory animal testing and estimates of exposure
according to the following four key steps from the National Academy of Sciences
(NAS) risk assessment paradigm (NRC, 1983):
• Hazard Identification
• Dose Response Assessment
• Exposure Assessment
• Risk Characterization

Twenty four years later, the NAS presented a vision for toxicity testing and risk
assessment in the document Toxicity Testing in the 21st Century A Vision and a
Strategy; this document recommended the use of predictive tools such as
(Q)SAR (NRC, 2007). In the outline of this vision, the NAS described some of the
risk assessment-related applications of (Q)SAR including the prediction of
toxicity, ADME properties, environmental fate, and ecologic effects for chemicals.
Consistent with the NAS vision and the existing risk assessment paradigm, the
emphasis in this section is on the integration of (Q)SAR tools into the hazard
identification (biological endpoint) component of the risk assessment process for
pesticides. This section builds upon concepts discussed in section 4 (Problem
Formulation for (Q)SAR) and section 5 (Evaluating the Adequacy of (Q)SAR

Page 89 of 186
Predictions), and discusses the process of integrating the overall toxicity
database (including (Q)SAR predictions) to arrive at conclusions regarding
hazard with consideration of confidence and level of uncertainty. As with
traditional pesticide risk assessments, the characterization of the hazards and
associated uncertainties are communicated to risk managers for consideration in
regulatory decision making.

7.1 Incorporating (Q)SAR in Hazard Characterizations:

Overview
In general, the hazard identification and characterization process should not be
greatly different for situations where (Q)SAR predictions are an additional source
of data compared to traditional hazard assessments in which empirical data
alone are considered. As discussed in section 4, there are two situations where
(Q)SAR predictions are likely to be considered by a pesticide regulatory authority:
a pesticide applicant submits a (Q)SAR prediction to address a data requirement,
either to fulfill or support the requirement or a waiver for the requirement or the
evaluator of a pesticide risk assessment uses a (Q)SAR prediction to support the
case for an additional or refined data requirement. For the first situation where a
pesticide applicant has submitted a (Q)SAR prediction, the primary tasks for the
evaluator are to determine the purpose of the (Q)SAR submission, determine
whether the (Q)SAR prediction fulfills the intended purpose and to factor the
(Q)SAR prediction into the overall weight of evidence for the pesticide chemical
in the appropriate context. If the (Q)SAR prediction is not clearly deficient, the
prediction is evaluated according to the principles and procedures discussed in
sections 4 and 5 to determine whether the prediction accomplishes its intended
purpose as defined in the problem formulation. At the same time, the empirical
data are evaluated with respect to the validity and acceptability of each study. If
mode of action data have been submitted, these data are also evaluated for
individual validity and acceptability.

For the second situation where an evaluator uses a (Q)SAR prediction, usually
there is no empirical information involved and it is still important to assess the
validity and reliability of the chosen (Q)SAR method or model for the evaluator’s
purpose.

The basic steps in the integration of (Q)SAR predictions and empirical data into a
weight of evidence analysis are listed in Figure 7–1 and described in more detail
in subsequent sections.

Page 90 of 186
Figure 7–1: Weight of Evidence Analysis: Integration of (Q)SAR Predictions
and Empirical Data

1. Problem formulation: What is the goal of the assessment? And the role of
(Q)SAR in that assessment?

2. Determination of Adequacy: What is the adequacy of the (Q)SAR

prediction? Is it fit for the purpose intended in the problem formulation step?

3. Weight of Scientific Evidence: Integration of the existing empirical data

and (Q)SAR predictions. Weighing the scientific data including mode of
action information, if available.*

4. Data base sufficient: Is the data base sufficient for risk assessment?
What data are missing? What is the level of data base uncertainty?

5. Hazard characterization: Telling a clear and transparent hazard story and

presenting the determination of confidence level and level of uncertainty,
and risk communication.

* For most pesticides, a complete understanding of the mode of toxicological action may be
absent. To the extent that a toxicological mode of action is postulated for an analog of the
pesticide of interest, it would be important to consider this information to help build confidence in
a predicted endpoint.

7.2 Problem Formulation and Adequacy Determination

As discussed in section 4, problem formulation in the context of (Q)SAR
prediction for pesticides involves asking and answering a number of key
questions during the course of the review with respect to assessment context,
characteristics of the chemical subject to the (Q)SAR prediction, characteristics
of the selected (Q)SAR tool(s) and the prediction, and identification of empirical
data including mode of action data that are relevant to the (Q)SAR prediction.

The assessment context is essentially a determination of the objective of the

(Q)SAR analysis, the specific endpoint that is predicted, the role that the (Q)SAR
prediction plays in the risk assessment and the acceptable level of reliability that
the (Q)SAR prediction must have if the prediction is to be accepted for its
proposed role.

Pesticide characterization involves identifying whether the prediction is for a

pesticide active ingredient, an impurity, a metabolite or transformation product, a
pesticide inert/formulant ingredient, or an analog of the pesticide of interest. It

Page 91 of 186
also involves ensuring that the correct structure was the subject of the prediction
and whether it is appropriate to use (Q)SAR predictions for that structure.

The characterization of the (Q)SAR tool and the prediction includes a

consideration of the general methodology behind the tool, the empirical data on
which the tool is based, the endpoint predicted by the tool, other details on the
tool, and details on the prediction. As discussed in section 5, this information is
used in the evaluation of the adequacy of the prediction that includes the
scientific validity of the (Q)SAR tool (i.e., OECD (Q)SAR validation principles),
applicability of the tool to the pesticide, reliability of the prediction, and the
relevance of the tool to the assessment question (context). The characterization
of the (Q)SAR tool and the evaluation of the adequacy of the prediction can aid in
the interpretation of the prediction relative to any available empirical data.

Empirical data related to the (Q)SAR prediction and available mode of action
data are important to consider if they either support the (Q)SAR prediction or
contradict it. These data will be discussed further in the following sections.

7.3 Evaluating Empirical Data versus (Q)SAR Predictions

For many pesticide active ingredients the empirical database will consist of a
suite of guideline toxicity studies that are typically required by regulatory
agencies for the registration and reregistration of pesticide chemicals. These
studies address systemic and local effects (e.g., acute toxicity, skin irritation,
developmental toxicity, genotoxicity, carcinogenicity), multiple routes of exposure
(e.g., oral, inhalation, dermal), multiple durations of exposure (e.g., acute, short-
term, long-term), and are conducted across multiple taxa (e.g., freshwater fish,
aquatic invertebrates, rats, mice, rabbits).

Toxicity studies tend to be very detailed. The description of the conduct of the
study (i.e., materials and methods) is typically very extensive. The results
sections of the studies are also reported in even greater detail and many of the in
vivo studies cover a variety of biological endpoints. The studies also have a
conclusion section in which the data submitter will often propose a point of
departure for the study depending on how the submitter interprets the outcome of
the study. The reviewer or evaluator of each study prepares a written evaluation
of the study (“Data Evaluation Record” or “DER” at EPA and PMRA), and in this
DER the evaluator summarizes the key points of the study. Perhaps most
importantly, the evaluator records his or her own conclusions about the adequacy
of the study and the appropriate points of departure for the endpoints supported
by the study. In the normal review of empirical data for pesticide risk assessment,
each study is evaluated individually for scientific rigor and those studies that are

Page 92 of 186
considered acceptable are integrated to “tell a story” or “paint a picture” of the
hazard profile of the pesticide chemical.

The review of a (Q)SAR prediction is similar to the review of empirical data. In the
typical scenario, the pesticide applicant will submit a documented (Q)SAR
prediction including the purpose of the prediction, the rationale for selection of a
model (or models), information on how the query structure was entered into the
model, a discussion of the OECD validation principles as applied to the model,
why the training set chemicals are applicable to the query structure, why the
(Q)SAR prediction satisfies a data requirement (or supports a waiver from the
data requirement) and a discussion of limitations and uncertainties associated
with the prediction. The evaluator reviews the submitted (Q)SAR prediction and
all of the supporting documentation and prepares a written record of the review
much like the DER for empirical data (see section 5.5).

Once the review of the empirical data and the (Q)SAR prediction have taken
place, and provided that the review indicates that the prediction is scientifically
valid, the evaluator is in a position to determine whether the prediction is reliable,
i.e., whether it fills a data gap or supports a waiver from a data requirement.

Evaluating empirical data with (Q)SAR predictions enables a determination of

whether relevant empirical data on structurally related chemicals will support,
detract from, or influence the weighting of the (Q)SAR prediction in the
assessment. For example, if there are only limited empirical data available from
short-term animal studies on a pesticide impurity of potential concern and the
objective is to determine whether the impurity has carcinogenic potential and, if
so, to obtain some information on how it might elicit carcinogenicity, then a
weight of evidence approach could be considered. Such an approach could
combine information from the available short-term animal dataset on the impurity,
genotoxicity/carcinogenicity studies for structurally similar compounds if
applicable, and carcinogenicity and genotoxicity predictions for the impurity in
question. In this example, the short-term animal study data may provide
information on precursor effects that may support the carcinogenicity of the
impurity. Similarly, if the studies on the structurally similar compounds are in
agreement with the (Q)SAR predictions for the impurity, this would enhance the
confidence in the predictions.

On occasion the (Q)SAR prediction may be in apparent conflict with empirical

data. In such situations, it is useful to re-examine both the empirical database
and the data supporting the prediction for a possible resolution of the apparent
conflict. A theoretical example of using a closer examination of the empirical
database to resolve a conflict with a (Q)SAR prediction would involve a pesticide

Page 93 of 186
that demonstrates clear systemic toxicity in short-term dermal toxicity tests. If
read-across extrapolations from analog chemicals indicate low potential for
dermal absorption, then there is an apparent conflict. However, if a closer
examination of the physical-chemical properties of the pesticide and skin irritation
testing data reveals that the pesticide is likely to be poorly absorbed at low
concentrations, but at high concentrations, it is corrosive, destroying the barrier
properties of the skin, facilitating access to the systemic circulation then it may be
possible to resolve the differences between the (Q)SAR predictions and the
empirical data.

Depending on the basis for a conflict between a (Q)SAR prediction and the
results of empirical studies, it may be necessary to fully examine the adequacy of
the (Q)SAR prediction (see section 5), and also the adequacy of the empirical
data. (Q)SAR tools are reductionist methods that may not fully account for the
impact of physical-chemical properties and pharmacokinetics/dynamics, may
over-emphasize the contribution of a particular structural alert or property or may
miss a toxicologically relevant alert because of database/training set limitations.
In general, (Q)SAR predictions should not be used to override the results of well-
conducted, guideline type studies for the same endpoint. However, in cases
where the empirical studies are of questionable reliabilty because they are non-
guideline studies, conducted according to older protocols, restricted to examining
specific research endpoints or have other limitations, it may be necessary to give
greater weight to reliable and relevant (Q)SAR predictions and/or develop
recommendations for further testing to help resolve the conflict. Weighting of
predictions and empirical data is discussed further in section 7.5.

As discussed in section 6 (Q)SAR evidence may actually consist of multiple

predictions from multiple (Q)SAR tools (e.g., read-across from a chemical
category, QSAR model prediction, SAR prediction, etc.) each based on different
prediction paradigms with different strengths and limitations. There are a number
of important advantages of combining predictions from multiple (Q)SAR models
(e.g., improved predictive performance, enhanced domain of applicability,
complementary information, increased confidence in the predictions). Because of
these advantages, it may be possible to assign a greater level of confidence and
a greater weighting to combined or consensus predictions when comparing them
to the available empirical data and the other streams of evidence.

There are three likely outcomes to the weight of evidence evaluation: 1) The
(Q)SAR prediction adequately addresses the data requirement, i.e., the data
requirement is satisfied or the study need not be done; 2) The (Q)SAR prediction
is not relevant to the data requirement, i.e., the (Q)SAR prediction is scientifically

Page 94 of 186
adequate but does not address the specific data requirement; and 3) The
(Q)SAR prediction is scientifically acceptable and addresses the data
requirement, but the data requirement is critical, requiring a high degree of
certainty and confidence that the prediction as submitted is unable to meet. In
this last instance, the result of reviewing the empirical data and the (Q)SAR
prediction together may point the way to follow up action such as additional
information from the submitter on the identity of, and relevant data for the training
set chemicals; or perhaps targeted mode of action testing on the query structure
and specific training set chemicals that could help fulfill the data requirement.

7.4 Mode of Action Consideration

If the weight of evidence includes a known or suspected mode of toxicological
action (MOA), this understanding could substantially strengthen the overall
database and provide additional support to determine whether the (Q)SAR
prediction(s) is biologically plausible and consistent with what is known about the
chemical of interest. The mode of action for the chemical of interest provides the
overall biological basis for the phenotypically expressed adverse effects reported
among the traditional in vitro and in vivo toxicology studies, and can also support
a (Q)SAR prediction for the specific endpoint of interest (e.g., cancer,
genotoxicity). Some of the most robust (Q)SARs are ones with the greatest
confidence that all the chemicals being combined in a model are producing
toxicity through a single molecular initiating event (mechanism).

7.5 Overall Weight of Evidence

Evaluating empirical data versus (Q)SAR predictions is part of the overall
concept of integrating multiple lines of evidence into a weight of evidence. While
(Q)SAR predictions have strengths, limitations, and uncertainties that can differ
from those associated with empirical studies, (Q)SAR predictions can be
considered as one potential line of evidence within a weight of evidence
assessment of a pesticide.

A detailed discussion of weight of evidence approaches in pesticide risk

assessment is outside the scope of this document, but some general principles
can be highlighted that are applicable in scenarios where (Q)SAR predictions and
empirical data are being integrated. Similar to weight of evidence for empirical
data alone, it is likely that qualitative, scientific expert judgement based
approaches will be the most frequently used ones for integrating (Q)SAR
predictions and empirical data. These approaches can involve considering
information from each individual prediction, collectively examining multiple
predictions within a (Q)SAR line of evidence (see section 6), considering multiple

Page 95 of 186
data points within each individual empirical study, combining information from
similar empirical studies within one line of evidence, and finally, integrating
(Q)SAR, empirical and other lines of evidence together to arrive at an
assessment conclusion (Health Canada, 2011).

Just as with empirical data based weight of evidence, approaches that integrate
(Q)SAR and empirical lines of evidence usually include a qualitative weighting or
ranking of the importance of the different lines of evidence for the overall
assessment conclusion. Such a weighting involves a consideration of the
adequacy (i.e., see section 5) and uncertainties associated with the different lines
of evidence. Regardless of the weighting/ranking approach adopted,
transparency is critical and can be addressed via comprehensive narrative
rationales outlining the approaches followed in considering each line of evidence
and integrating the lines of evidence together. It is particularly important to outline
the approaches taken when there are conflicts between the (Q)SAR and the
empirical data lines of evidence. For ease of interpretation, tabular presentations
of the (Q)SAR and empirical data lines of evidence can also be considered.

While qualitative approaches are likely to be used most often, it is also possible
to consider quantitative scoring systems or mathematical algorithms that may be
more systematic for weighting (Q)SAR and empirical data lines of evidence than
qualitative, expert judgement-based approaches. Examples of such systems
usually involve numerical weightings for each line of evidence, multiplying the
scores for each line of evidence by its weighting, and summing up the weighted
scores into an overall result. Just as transparency is critical for qualitative
weighting/ranking approaches, it is especially critical to clearly outline the
rationale behind any quantitative weight of evidence scoring systems.

7.6 Hazard Characterization and Risk Communication

The hazard characterization should include a clear, straightforward hazard
narrative, piecing together all the components of the problem formulation, hazard
identification, weight of evidence and addressing the level of uncertainty in the
database and confidence in the overall assessment. As with all characterizations
of hazard, it is imperative to consider all available hazard data whether these are
in vivo or in vitro toxicity data or, in the absence of measured values, relevant
data from (Q)SAR predictions or other predictive methods.

Subsequent to developing a clear, straightforward hazard narrative,

communicating this narrative to risk assessors and risk managers becomes
critical. The successful communication of risk is not only dependent on relaying
the adverse health outcome(s), susceptible subpopulations, dose response and

Page 96 of 186
exposure assessment, but also the quality of the data, level of uncertainty, and
confidence in the overall assessment. Any risk mitigation decision that is based
on risk assessment conclusions, must be made with a clear understanding of the
level of uncertainty surrounding the risk assessment conclusions and what level
of confidence should be placed on those conclusions to support a regulatory
decision. If, for example, the level of uncertainty in the database is high because
most of the non-cancer endpoints are predicted and the level of confidence in the
overall risk assessment is weak, the risk manager should be cognizant of this low
level of confidence before selecting an adequate risk mitigation option. In short,
the regulatory option selected should be consistent with the level of uncertainty
identified for the predicted and empirical datasets so as not to over or under-
inflate the confidence in these datasets. If that level of database uncertainty can
be addressed by additional research, the decision on when the data will be
required to be submitted and when they can be considered in future risk
assessment and management decisions may be dependent on the potential
health outcome. In any scenario, it is critical that the risk manager has all the
relevant information from the risk assessor in order to develop appropriate risk
management options and make a good regulatory decision based on sound
science.

7.7 Summary
The integration of (Q)SAR predictions into the risk assessment involves many
steps which are similar to the risk assessment paradigm: hazard identification,
dose response assessment, exposure assessment and risk characterization. The
only difference between a risk assessment based on traditional empirical data
from that which involves in silico predictions is the judgement of adequacy of the
(Q)SAR predictions and determination of database completeness.

The steps involved with integrating (Q)SAR predictions rely on starting with a
solid problem formulation to establish what the (Q)SAR prediction is intended to
inform for the assessment and what type of assessment to be performed will
indicate the amount of uncertainty that is deemed acceptable. For example, a
screening level assessment would allow for more uncertainty than a risk
assessment. The determination that the (Q)SAR prediction is valid and reliable
for the purpose described in the problem formulation step is critical to proceed
forward in the subsequent steps. Without the determination of scientific
adequacy, the (Q)SAR prediction would be rendered unacceptable and therefore,
could not be considered in the risk assessment.

The next step of integrating the (Q)SAR prediction with extant scientific data on
that compound or a structural analog is critical to a scientific weight of evidence

Page 97 of 186
analysis. Issues of reproducibility of observations, consistency of effects across
species, strain, time of exposure and routes of exposure, as well as the
determination of biological plausibility and incorporation of mode of action
information is considered in this analysis. After developing the weight of scientific
evidence, the risk assessor proceeds forward to determine the completeness of
the database to support the risk assessment.

If the database is considered deficient, and missing critical studies, the weight of
evidence, including any mode of action information, should be informative in
determining the type of study(s) needed to fulfill the database deficiency; this
may be in vitro and/or short term studies depending on the data deficiency and
as the concept of the adverse outcome pathway becomes elucidated for the
particular toxicity endpoint, the determination of what study will be needed will be
more clearly defined. In short, the combination of the (Q)SAR prediction,
empirical data, mode of action, and/or adverse outcome pathway in a weight of
evidence approach will inform the risk assessment on database deficiency and
identification of critical research needed to address this level of uncertainty.

Page 98 of 186
8. Conclusions and Future Vision for (QSAR) and
Pesticides
CONCLUSIONS AND FUTURE VISION FOR (Q)SAR AND
PESTICIDES
Topics Discussed in this Section:

 Toxicity Testing in 21st Century: Shift in the Risk Assessment Paradigm

 Weight of Evidence Approach: Biological Plausibility

 Adverse Outcome Pathway: Conceptual Framework

 Expert Scientific Judgment and Peer Review

8.0 Toxicity Testing in the 21st Century: Shift in the Risk

Assessment Paradigm
The NAS report Toxicity Testing in the 21st Century: A Vision and a Strategy
emphasizes the need for moving away from prescriptive assessments based on
checklists of traditional animal toxicity studies towards an integrated approach
that relies on the existing knowledge-base for a class of chemicals and the
results of alternative testing methods to identify toxicity pathways and to focus
data requirements on more targeted toxicity testing (NRC, 2007). This is
especially relevant for pesticide regulatory authorities that receive and evaluate
large volumes of animal and ecological toxicity data submitted in support of new
chemical registrations. To continue to meet the demands of new pesticide
registrations requires new technologies that allow for faster, more efficient and
effective technical reviews; this change, however, must not come at the price of
public health and environmental protection.

While there is the need for more efficient review processes, there is also the
impetus of research on newer technologies, a recognition of the accelerated pace
of scientific innovation. New technologies will allow pesticide regulatory
authorities to build upon existing knowledge of pesticide toxicity to develop
integrated approaches for testing and assessment (IATA) of pesticides. These
parallel regulatory and risk assessment changes and advancements in the state-
of-the-science propel the agencies forward and expedite the transition towards
global application of newer, swifter risk assessment and testing methodologies.

Page 99 of 186
Computational tools vary widely depending on the purpose for their use in risk
assessment. (Q)SAR tools represent an example of an alternative testing method
that could be a useful component of integrated approaches to testing and
assessment. (Q)SAR tools have had a long history of use by industry and
regulatory agencies for hazard determinations and other applications; there are
also many different types of commercial and non-commercial (Q)SAR tools that
are either currently available or rapidly under development. However, in spite of
the regulatory experience and the on-going developments in the field of (Q)SAR,
there are very few examples of formal (Q)SAR guidance documents that discuss
the unique issues and considerations associated with the application of (Q)SAR
to pesticide regulatory risk assessments.

Because pesticide regulatory authorities are transitioning down the path of

utilizing this type of predictive technology, it becomes increasingly important that
a systematic and transparent approach to the use of (Q)SAR predictions in
pesticide assessments be adopted and communicated to ensure that the
application of sound scientific judgment. To that end, this document seeks to
provide risk assessors and hazard evaluators with some general guidance on
how to review (Q)SAR predictions included in pesticide submissions. It is well
recognized, however, that there are a variety of different guidance documents for
(Q)SAR which have been published by agencies such as the OECD, the EU, the
US EPA and others. This particular NAFTA guidance document does not
endeavor to reproduce or to replace any of the other guidance documents, but
seeks to provide much needed ready-to-use, streamlined, plain language
guidance to pesticide evaluators on the application of (Q)SAR to pesticide
regulatory decision making.

8.1 Weight of Evidence Approach: Biological Plausibilty

This guidance document is consistent with the current hazard/risk paradigm in
terms of the approach to the evaluation of (Q)SAR predictions. The document
starts at the problem formulation stage as explained in section 4, through
evaluating the adequacy of (Q)SAR predictions as described in section 5,
combining information from multiple predictions in section 6 and integrating
(Q)SAR predictions into hazard assessments assessment in section 7. The
overall emphasis throughout this document is that (Q)SAR predictions should be
considered among the many other data streams in a weight of evidence
approach for determination of hazard/risk. Similar to the consideration of multiple
lines of evidence when identifying a toxicological mode of action, (Q)SAR
predictions should not be used in isolation when reaching human and
environmental hazard assessment conclusions for a pesticide. In addition to

Page 100 of 186

considerations of the reliability, validity and relevance of the individual (Q)SAR
tools and predictions, the defensibility of the predictions is dependent on
biological consistency and plausibility across all scientific lines of evidence in a
holistic weight of evidence approach. This weight of evidence analysis should be
described in the hazard characterization section of a pesticide assessment and
should include the level of confidence, range of uncertainty, data gaps and any
needed research for further refinement of the risk assessment.

8.2 Adverse Outcome Pathway: Conceptual Framework

As stated earlier, (Q)SAR is not a new predictive tool; it has been used for many
decades within various regulatory programs to provide predictions of apical
endpoints, biological and physiochemical properties. The limitations of using
(Q)SAR when the domain of applicability is not inclusive of the chemical under
review or when the database is not targeted for a particular endpoint of concern,
are well understood. These limitations and others are discussed in this
document, particularly in the context of the OECD validation principles, as
described in section 5. In spite of these limitations, (Q)SAR remains a predictive
tool worthy of continued use and development.

The historical use of (Q)SAR predictions were two-dimensional, stopping at the

binary prediction (e.g., yes/no carcinogenicity). The future use of (Q)SAR will
involve anchoring (Q)SAR predictions with what is known about that chemical
class/category, the biological mode of action, toxicity pathways and population
effects; (Q)SAR predictions will be built into larger conceptual frameworks called
adverse outcome pathways (AOPs) (Ankley et al., 2010). More simplistically, and
as illustrated in Figure 8–1 below, AOPs delineate the documented, biologically
plausible, measurable and testable processes by which chemicals induce
molecular perturbations and subsequent biological responses relevant for risk
assessment. The basic concept describes how molecular perturbations cause
effects at different levels of biological organizations including at the subcellular,
cellular, tissue, organ, and population levels (OECD 2011a). As indicated in
Figure 8–1, toxicity pathways, the initial perturbations of cell-signaling motifs,
genetic circuits, and cellular-response networks that might eventually result in
disease, are components of AOPs, but unlike AOPs, toxicity pathways are not
necessarily directly linked to apical effects (i.e., disease outcomes). The mode of
action (MOA) is also a component of the AOP; it is inclusive of the events from
the initial molecular perturbations to an adverse effect at the individual level, but
does not usually consider exposure or effects at the population level (OECD,
2012a). AOPs capture the continuum of metabolism, molecular perturbations,
cellular interactions, effects on the tissue and organ leading to individual effects

Page 101 of 186

and effects at the population level in a holistic approach. They allow regulatory
authorities to move away from an overdependence on single chemical (Schultz
and Diderich, 2011) in vivo animal testing and make greater use of
computational, molecular and in vitro tools as described and advocated for in the
2007 NAS report Toxicity Testing in the 21st Century: A Vision and a Strategy.

One way in which (Q)SARs could contribute to the AOP approach would be
through the identification of structural alerts associated with key events in an
AOP, particularly molecular initiating events (MIEs). The OECD has noted that a
close linkage between an MIE and an observed adverse outcome in vivo can be
used as a basis for developing a chemical category for the relationship between
chemical structure and the in vivo endpoint. Thus, rather than just relying on
intrinsic chemical activity, AOPs potentially provide a comprehensive mechanistic
basis for forming toxicologically meaningful categories for making predictions
using read-across or (Q)SAR models (OECD, 2012b). As noted previously the
European Commission Joint Research Centre is developing a reporting format
for describing key events/intermediate effects in AOPs in collaboration with the
OECD and ECHA (OECD, 2012a).

An example of an AOP discussed in a US EPA hosted workshop in December,

2010 involves the binding of a xenobiotic to an hepatic nuclear receptor as the
molecular initiating event for a variety of toxicity pathways, including pathways
leading to liver cancer (Hester et al., 2006). The identification of molecular
initiating events of this type allows the development of methods to screen for
chemical interactions with biological targets and in this case, receptors. This is
where (Q)SAR models become critical, in identifying chemical categories based
on chemical structures and their linkage to biological activities by understanding
the toxicity pathways relevant for risk assessment (OECD, 2011b).

Page 102 of 186

Figure 8–1: Adverse Outcome Pathway Diagram (Ankley et al., 2010)

In summary, (Q)SAR tools are important predictive technologies for today’s risk
assessment as well as those for tomorrow — provided that they are applied with
appropriate constraints and cautionary guidance and that there is a meaningful
attempt to build data bridges in a weight of evidence approach among (Q)SAR
and future emerging predictive technologies to better target efforts to more
efficiently and effectively maximize overall biological predictive capability (Benigni
et al., 2007b).

8.3 Expert Scientific Judgment and Peer Review

While this guidance document is intended to cover the main issues that
evaluators should consider when reviewing (Q)SAR predictions included in
pesticide submissions, the document is not intended to provide stand-alone,
step-by-step instructions for all potential applications of (Q)SAR tools to
pesticides. This guidance document should be supplemented with expert
scientific judgment and expert peer review to ensure consistency, reproducibility,
and scientific defensibility in the use of (Q)SAR in pesticide hazard assessments.

To this end, the NAFTA pesticide regulatory authorities are currently assembling
a (Q)SAR expert committee to provide advice to pesticide evaluators in complex
assessments that seek to integrate (Q)SAR predictions with empirical data in a
weight of evidence approach for hazard/risk determinations that may trigger
regulatory risk management decisions. One of the mandates of this expert

Page 103 of 186

committee will be to identify updates, modifications or additions to the guidance
document.

Finally, as mentioned previously, there are a variety of other available guidance

documents on regulatory applications of (Q)SAR (see Appendix I). Pesticide
evaluators are encouraged to consult these other documents for additional
information on specific topics or scenarios as required.

Page 104 of 186

References
Note that all website links in the reference list were accessed at the time this
document was finalized. Any any problems accessing these links after the
finalization of this document should be referred to the originating organizations
for the websites.
Abshear, T., G.M. Banik, M.L. D’Souza, K. Nedwed, and C. Peng. 2006. A model
validation and consensus building environment. SAR and QSAR Environ. Res.
17(3):311–321.

Accelrys Inc. 2004. TOPKAT, Version 6.2, User Guide, Accelrys Inc., San Diego,
CA, USA.

American Chemical Society. 2010. CAS RegistrySM and CAS Registry System.
CAS® A Division of the American Chemical Society.

Ankley, G.T., G.P. Daston, S.J. Degitz, N.D. Denslow, R.A. Hoke, S.W. Kennedy,
A.L. Miracle, E.J. Perkins, J. Snape, D.E. Tillitt, C.R. Tyler, and D. Versteeg.
2006. Toxicogenomics in regulatory ecotoxicology. Envron. Sci. Technol.
40(13):4055–4065.

Ankley, G.T., R.S. Bennett, R.J. Erickson, D.J. Hoff, M.W. Hornung, R.D.
Johnson, D.R. Mount, J.W. Nichols, C.L. Russom, P.K. Schmieder, J.A.
Serrrano, J.E. Tietge, and D.L. Villeneuve. 2010. Adverse outcome pathway: A
conceptual framework to support ecotoxicology research and risk assessment.
Environ. Toxicol. Chem. 29(3):730–741.

Arnot, J.A., F.A.P.C. Gobas. 2004. A Food Web Bioaccumulation Model for
Organic Chemicals in Aquatic Ecosystems. Environ. Toxicol. Chem.
23(10):2343–2355.

Aronson, D., R. Boethling, P. Howard, W. Stiteler. 2006. Estimating

biodegradation half-lives for use in chemical screening. Chemosphere 63:1953–
1960.

Arvidson, K.B., R. Chanderbhan, K. Muldoon-Jacobs, J. Mayer, and A.

Ogungbesan. 2010. Regulatory use of computational toxicology tools and
databases at the United States Food and Drug Administration's Office of Food
Additive Safety. Expert Opinion on Drug Metabolism & Toxicology 6, 793–796.

Asikainen, A.H., J. Ruuskanen, and K.A. Tuppurainen. 2004. Performance of

(consensus) kNN QSAR for predicting estrogenic activity in a large diverse set of
organic compounds. SAR QSAR Environ. Res. 15(1):19–32.

Page 105 of 186

ATSDR. 1989. Toxicological Profile for Bis(Chloromethyl)Ether. Agency for Toxic
Substances and Disease Research. U.S. Public Health Service in collaboration
with: U.S. Environmental Protection Agency. December 1989.
http://www.atsdr.cdc.gov/ToxProfiles/tp128.pdf

Bailey, A., N. Chanderbhan, N. Collazo-Braier, M. Cheeseman, and M. Twaroski.

2005. The use of structure-activity relationship analysis in the food contact
notification program. Reg. Toxicol. and Pharm. 42:225–235.

Bassan, A., E. Fioravanzo, M. Pavan, and M. Stocchero. 2011. Applicability of

physicochemical data, QSARs and read-across in Threshold of Toxicological
Concern assessment. Final report of a study carried out by Soluzioni
Informatiche (S-In, Vicenza, Italy) for the European Food Safety Authority
(EFSA). Accessible from: http://www.efsa.europa.eu/fr/supporting/pub/159e.htm

Benfenati, E. (editor). 2007. Quantitative Structure-Activity Relationships (QSAR)

for Pesticide Regulatory Purposes. Elsevier. Amsterdam.

Benfenati, E. 2010. The CAESAR project for in silico models for the REACH
legislation. Chem. Cent. J. 4(Suppl 1):I1.

Benfenati, E., R. Benigni, D.M. Demarini, C. Helma, D. Kirkland, T.M. Martin, P.

Mazzatorta, G. Ouedraogo-Arras, A.M. Richard, B. Schilter, W.G. Schoonen,
R.D. Snyder, and C. Yang. 2009. Predictive Models for Carcinogenicity and
Mutagenicity: Frameworks,State-of-the-Art, and Perspectives. J. Environ. Sci.
Health. C, 27(2):57–90.

Benigni, R., and C. Bossa. 2008. Predictivity and reliability of QSAR models: The
case of mutagens and carcinogens. Toxicol. Mechan. Methods 18:137–147.

Benigni, R., C. Bossa, T. Netzeva, and A. Worth. 2007a. Collection and

evaluation of (Q)SAR models for mutagenicity and carcinogenicity. European
Commission, Joint Research Centre, EUR 22772 EN, 126 p. Ispra, Italy.
http://ihcp.jrc.ec.europa.eu/our_labs/computational_toxicology/doc/EUR_22772_
EN.pdf

Benigni, R., T.I. Netzeva, E. Benfenati, C. Bossa, R. Franke, C. Helma, E.

Hulzebos, C. Marchant, A. Richard, Y.-T. Woo, and C. Yang. 2007b. The
expanding role of predictive toxicology: An update on the (Q)SAR models of
mutagens and carcinogens. J. Environ. Sci. Health C 25:53–97.

Page 106 of 186

Boobis, A.R., J.E. Doe, B. Heinrich-Hirsch, M.E. Meek, S. Munn, M. Ruchirawat,
J. Schlatter, J. Seed, and C. Vickers. 2008. IPCS framework for analyzing the
relevance of a noncancer mode of action for humans. Crit. Rev. Toxicol.
38(2):87–96.

Bradbury, S.P. 1994. Predicting modes of toxic action from chemical structure: an
overview. SAR QSAR Environ. Res. 2:89–104.

Bradbury, S.P., C.L. Russom, G.T. Ankley, T.W. Schultz, and J.D. Walker. 2003.
Overview of data and conceptual approaches for derivation of quantitative
structure-activity relationships for ecotoxicological effects of organic chemicals.
Environ. Toxicol. Chem. 22(8):1789–1798.

Bradbury, S.P., T.C.J. Feijtel, and C.J. van Leeuwen. 2004. Meeting the scientific
needs of ecological risk assessment in a regulatory context. Environ. Sci.
Technol. 38(23):463A–470A.

Brown, T.N. and F. Wania. 2008. Screening chemicals for the potential to be
persistent organic pollutants. A case study of arctic contaminants. Environ. Sci.
Technol. 42:5202–5209.

Clark, M. 2005. Generalized fragment-substructure based property prediction

method. J. Chem. Inf. Model. 45(1):30–38.

Contrera, J.F., N.L. Kruhlak, E.J. Matthews, and R.D. Benz. 2007. Comparison of
MC4PC and MDL-QSAR rodent carcinogenicity predictions and the
enhancement of predictive performance by combining QSAR models. Reg.
Toxicol. Pharm. 49:172–182.

CCA. 2012. Integrating Emerging Technologies into Chemical Safety

Assessment. The Expert Panel on the Integrated Testing of Pesticides. Council
of Canadian Academies.
http://www.scienceadvice.ca/en/assessments/completed/pesticides.aspx

Cronin M. 2010. Quantitative structure-activity relationships (QSARs) —

applications and methodology. Chapter 10 in Recent Advances in QSAR Studies:
Methods and Applications. T. Puzyn, J. Leszczynski, and M.T.D. Cronin (eds).
Springer, Heidelberg, Germany, pp. 3–11.

Cronin, M.T.D., J.D. Walker, J.S. Jaworska, M.H.I. Comber, C.D. Watt, and A.P.
Worth. 2003. Use of QSARs in International Decision-Making Frameworks to
Predict Ecologic Effects and Environmental Fate of Chemical Substances.
Environ Health Perspect 111:1376–1390.

Page 107 of 186

Daginnus, K., S. Gottardo, A. Mostrag-Szlichtyng, H. Wilkinson, P. Whitehouse,
A. Paya-Perez, and J.M. Zaldivar. 2009. A modeling approach for the
prioritization of chemicals under the water framework directive. European
Commission, Joint Research Centre, EUR 24292 EN, 54 p. Ispra, Italy

Dalby, A., J.G. Nourse, W.D. Hounshell, A.K.I. Gushurst, D.L. Grier, B.A. Leland,
and J. Laufer. 1992. Description of several chemical structure file formats used
by computer programs developed at molecular design limited. J. Chem. Inf.
Comput. Sci. 32:244–255.

Daylight Chemical Information Systems. 2008. Simplified Molecular Input Line

Entry System (SMILESTM) Daylight Chemical Information Systems, Inc.
http://www.daylight.com/smiles/index.html

Dearden, J.C., M.T.D. Cronin, and K.L.E. Kaiser. 2009. How not to develop a
quantitative structure-activity or structure-property relationship (QSAR/QSPR).
SAR and QSAR in Environmental Research 20:241–266.

Dearden, J. and A. Worth. 2007. In silico prediction of physicochemical

properties. European Commission, Joint Research Centre, EUR 23051 EN, 73 p.
Ispra, Italy

Demchuk, E., P. Ruiz, J.D. Wilson, F. Scinicariello, H.R. Pohl, M. Fay, M.M.
Mumtaz, H. Hansen, and C.T. De Rosa. 2008. Computational toxicology methods
in public health practice. Toxicol. Mech. Methods 18:119–135.

Dimitrov, S.D., G.D. Dimitrova, T.S. Pavlov, N. Dimitrova, G.Y. Patlewicz, J.

Niemela, and O.G. Mekenyan. 2005a. A stepwise approach for defining the
applicability domain of SAR and QSAR models. J. Chem. Inf. Model. 45:839–
849.

Dimitrov, S.D., L.K. Low, G.Y. Patlewicz, P.S. Kern, G.D. Dimitrova, M.H.I.
Comber, R.D. Philips, J. Niemela, P.T. Bailey, and O.G. Mekenyan. 2005b. Skin
sensitization: Modeling based on skin metabolism simulation and formation of
protein conjugates. Int. J. Toxicol. 24:189–204.

Doull, J., J.F. Borzelleca, R. Becker, G. Daston, J. DeSesso, A. Fan, P. Fenner-

Crisp, M. Holsapple, J. Holson, G.C. Llewellyn, J. MacGregor, J. Seed, I. Walls,
Y.-T. Woo, and S. Olin. 2007. Framework for use of toxicity screening tools in
context-based decision making. Food and Chem. Toxicol. 45:759–796.

Page 108 of 186

EC. 1995a. Overview of structure-activity relationships for environmental
endpoints. Part 1: General outline and procedures. European Commission.
Report of the EU-DG-XII Project QSAR for predicting fate and effects of chemical
in the environment. Contract No. EV5V-CT92-0211. 149 p. Ispra, Italy.
http://ihcp.jrc.ec.europa.eu/our_labs/computational_toxicology/information-
sources/qsar-document-area/QSAR_environ_background.pdf

EC. 1995b. Overview of structure-activity relationships for environmental

endpoints. Part 2: Description of selected models. Report of the EU-DG-XII
Project QSAR for predicting fate and effects of chemical in the environment.
European Commission. Contract No. EV5V-CT92-0211. 216 p. Ispra, Italy.
http://ihcp.jrc.ec.europa.eu/our_labs/computational_toxicology/information-
sources/qsar-document-area/QSAR_environ_models.pdf

EC. 2008a. QSAR Model Reporting Format (Version 1.2). European Commission
Directorate General Joint Research Centre. Institute for Health and Consumer
Protection. Toxicology and Chemical Substances Unit. Ispra 26/05/2008.
http://ihcp.jrc.ec.europa.eu/our_labs/computational_toxicology/qsar_tools/qrf/QM
RF_version_1.2.pdf

EC. 2008b. QSAR Prediction Reporting Format (QPRF) (version 1.1, May 2008).
European Commission Directorate General Joint Research Centre. Institute for
Health and Consumer Protection. Toxicology and Chemical Substances Unit.
Ispra 26/05/2008.
http://ihcp.jrc.ec.europa.eu/our_labs/computational_toxicology/qsar_tools/qrf/QP
RF_version_1.1.pdf

EC. 2010. Applicability of QSAR analysis to the evaluation of the toxicological

relevance of metabolites and degradates of pesticide active substances for
dietary risk assessment. Scientific report submitted to EFSA. Prepared by
Computational Toxicology Group, Institute for Health & Consumer Protection,
European Commission — Joint Research Centre, Ispra, Italy.
http://www.efsa.europa.eu/en/supporting/pub/50e.htm

ECHA, 2010. Practical Guide 5. How to report (Q)SAR. European Chemicals

Agency. http://echa.europa.eu/documents/10162/13655/pg_report_qsars_en.pdf.

ECHA. 2008. Guidance on information requirements and chemical safety

assessment. Chapter R.6: QSARs and grouping of chemicals. May 2008.
Guidance for the implementation of REACH. European Chemicals Agency.
http://echa.europa.eu/documents/10162/13632/information_requirements_r6_en.pdf

Ellison, C.M., J. C. Madden, P.N. Judson, and M.T.D. Cronin. 2010. Application
of in silico tools in a weight of evidence approach to aid toxicological assessment.
Poster 4.4. 14th International Workshop on Quantitative Structure-Activity

Page 109 of 186

Relationships in Environmental and Health Sciences. May 24–28, 2010.
Montreal, Quebec, Canada.

Enoch, S.J. 2010. Chapter 7. Chemical category formation and read-across for
the prediction of toxicity. In: Puzyn, T., J. Leszczynski, and M.T. Cronin (eds.)
Recent Advances in QSAR Studies, Springer, pp. 209–219.

Enoch, S.J., M.T.D. Cronin, J.C. Madden, and M. Hewitt. 2009. Formation of
structural categories to allow for read-across for teratogenicity. QSAR Comb. Sci.
28(6-7):696–708.

Eriksson, L., J. Jaworska, A.P. Worth, M.T.D. Cronin, R.M. McDowell, and P.
Gramatica. 2003. Methods for reliability and uncertainty assessment and for
applicability evaluations of classification- and regression-based QSARs. Environ.
Health Perspect. 111(10):1361–1375.

Ferguson, J. 1939. The use of chemical potentials as indices of toxicity. Proc R

Soc Lond Series B Biol Sci 127:387–404.

Gallegos-Saliner, A., A. Poater, N. Jeliazkova, G. Patlewicz, and A.P. Worth.

2008. Toxmatch — A chemical classification and activity prediction tool based on
similarity measures. Reg. Toxicol. Pharm. 52:77-84.

Geiss, K.T. and J.M. Frazier. 2001. QSAR modeling of oxidative stress in vitro
following hepatocyte exposures to halogenated methanes. Toxicology in Vitro
15(4-5), 557–563.

Gramatica, P. 2007. Principles of QSAR models validation: Internal and external.

QSAR & Comb. Sci. 26(5):694–701.

Gramatica, P. and E. Papa. 2007. Screening and ranking of POPs for global half-
life: QSAR approaches for prioritization based on molecular structure. Environ.
Sci. Technol. 41:2833–2839.

Gramatica, P., E. Giani, and E. Papa. 2007. Statistical external validation and
consensus modeling: A QSPR case study for KOC prediction. J. Mol. Graph.
Model. 25:755–766.

Greene, N. 2002. Computer systems for the prediction of toxicity: an update. Adv.
Drug Deliv. Rev. 54:417–431.

Hansch, C. and T. Fujita. 1964. A method for the correlation of biological activity
and chemical structure. J. Am. Chem. Soc. 86:1616–1626.

Hansch, C. and W.J. Dunn. 1972. Linear relationships between lipophilic

character and biological activity of drugs. J. Pharm. Sci. 61:1–19.

Page 110 of 186

Health Canada. 2011. Weight of Evidence: General Principles and Current
Applications at Health Canada. November 2011. Prepared by the Task Force on
Scientific Risk Assessment’s Weight of Evidence Working Group. Tao, T., Y.
Bhuller, Y. Bonvalot, M. Hill, A. Klein, G. Kozak, I. Plante, and N. Robert.

Hester, S.D., D.C. Wolf, S. Nesnow, and S.F. Thai. 2006 Transcriptional profiles
in liver from rats treated with tumorigenic and non-tumorigenic triazole conazole
fungicides: Propiconazole, triadimefon, and myclobutanil. Toxicol. Pathol.
34:879–94.

Hewitt, M., C.M. Ellison, E.J. Enoch, J.C. Madden, and M.T.D. Cronin. 2010.
Integrating (Q)SAR models, expert systems and read-across approaches for the
prediction of developmental toxicity. Repro. Tox. 30:147–160.

Hewitt, M., M.T.D. Cronin, J.C. Madden, P.H. Rowe, C. Johnson, A. Obi, and S.J.
Enoch. 2007. Consensus QSAR Models: Do the Benefits Outweigh the
Complexity? J. Chem. Inf. Model 47:1460–1468.

Hulzebos, E.M., P.A.H. Janssen, L. Maślankiewicz, M.C.M. Mijerink, J.J.A.

Muller, S.M.G. Pelgrom, L. Verdam, and T.G. Vermeire. 2001. The application of
structure-activity relationships in human hazard assessment: a first approach.
National Institute of Public Health and the Environment. RIVM report 601516 008.
http://rivm.openrepository.com/rivm/bitstream/10029/9562/1/601516008.pdf

Hulzebos, E.M., P.C.J.I. Schielen, and L. Wijkhuizen-Maslankiewicz. 1999.

(Q)SARs for human toxicological endpoints: a literature search. RIVM, National
Institute of Public health and the Environment Report No. 601516.001, 65 p.
Bilhoven, The Netherlands.
http://www.rivm.nl/bibliotheek/rapporten/601516001.pdf

IUPAC. 2010a. Nomenclature and Symbols. International Union of Pure and

Applied Chemistry. http://old.iupac.org/general/FAQs/ns.html

IUPAC. 2010b. The IUPAC International Chemical Identifier (InChITM).

International Union of Pure and Applied Chemistry.
http://www.iupac.org/home/publications/e-resources/inchi.html

Jaworska, J., N. Nikolova-Jelizkova, and T. Aldenberg. 2005. Review of methods

for QSAR applicability domain estimation by the training set. A.T.L.A. 33:445–
459.

Jaworska, J.S., M. Comber, C. Auer, and C.J. van Leeuwen. 2003. Summary of a
workshop on regulatory acceptance of (Q)SARs for human health and
environmental endpoints. Environ. Health Perspect. 111(10):1358–1360.

Page 111 of 186

Jensen, G.E., J.R. Niemela, E.B. Wedebye, and N.G. Nikolov. 2008. QSAR
models for reproductive toxicity and endocrine disruption in regulatory use — a
preliminary investigation. SAR QSAR Environ. Res. 19(7–8):631–641.

Judson,R., A. Richard, D.J. Dix, K.Houck, M. Martin, R. Kavlock, V. Dellarco, T.

Henry, T. Holderman, P. Sayre, S. Tan, T. Carpenter, and E. Smith. 2009. The
Toxicity Data Landscape for Environmental Chemicals. Environ. Health Perspect.
117(5):685–695.

Judson, R.S., K.A. Houck, R.J. Kavlock, T.B. Knudsen, M.T. Martin, H.M.
Mortensen, D.M. Reif, D.M. Rotroff, I. Shah, A.M. Richard, and D.J. Dix. 2010. In
vitro screening of environmental chemicals for targeted testing prioritization: The
ToxCast Project. Environ. Health Perspect. 118(4):485–492.

Karelson, M., V.S. Lobanov, and A.R. Katritzky. 1996. Quantum-chemical

descriptors in QSAR/QSPR studies. Chem. Rev. 96(3):1027–1043.

Klopman, G. and S.K. Chakravarti. 2003a. Structure-activity relationship study of

a diverse set of estrogen receptor ligands (I) using MultiCASE expert system.
Chemosphere 51:445–459.

Klopman, G. and S.K. Chakravarti. 2003b. Screening of high production volume

chemicals for estrogen receptor binding activity (II) by the MultiCASE expert
system. Chemosphere 51:461–468.

Klopman, G., S.K. Chakravarti, N. Harris, J. Ivanov, and R.D. Saiakhov. 2003. In-
silico screening of high production volume chemicals for mutagenicity using the
MCASE QSAR expert system. SAR QSAR Environ. Res. 14:165–180.

Lai, D.Y., Y.-T. Woo, M.F. Argus, and J.C. Arcos. 1996. Cancer risk reduction
through mechanism-based molecular design of chemicals. In: Designing Safer
Chemicals: Green Chemistry for Pollution Prevention (DeVito, S.C., and R.L.
Garrett, eds.). ACS Symposium Series 640. Washington, DC:American Chemical
Society, 1996; 62–73.

Leonard, J.T. and K. Roy. 2006. On selection of training and test sets for
development of predictive QSAR models. QSAR Comb. Sci. 25(3):235–251.

Lewis, D.F.V., M.G. Bird, and M.N. Jacobs. 2002. Human carcinogens: an
evaluation study via the COMPACT and Hazard Expert procedures. Hum. Exp.
Toxicol. 21:115–122.

Lipnick, R.L. 1995a. Structure-activity relationships. In: Rand, G.M., ed.,

Fundamentals of Aquatic Toxicology, 2nd ed. Taylor&Francis, Washington, DC,
pp 609–655.

Page 112 of 186

Lipnick, R.L., 1995b. Hans Horst Meyer and the lipoid theory of narcosis. Trends
Pharmacol. Sci. 10(7):265–269.

Lo Piparo, E., A. Worth, M. Manibusan, C. Yang, B. Schilter, P. Mazzatorta, M.N.

Jacobs, H. Steinkellner, and L. Mohimont. 2011. Use of computational tools in
the field of food safety. Reg. Toxicol. Pharm. 60:354–362.

MacKay, D., S. Paterson, and M. Joy. 1983. Application of fugacity models to the
estimation of chemical distribution and persistence in the environment. In:
Swann, R.L. and A. Eschenroeder, eds., Fate of Chemicals in the Environment.
American Chemical Society Symposium Series, Vol. 225, Chapter 9, pp 175–
196.

Matthews, E.J. 2009b. Meeting on the Development of a General Guidance

Document on the Application of QSAR Tools to Pesticides. July 10, 2009.
9:00am – 5:00pm Agenda/Meeting Notes.

Matthews, E.J. and J.F. Contrera. 1998. A new highly specific method for
predicting the carcinogenic potential of pharmaceuticals in rodents using
enhanced MCASE QSAR-ES software. Reg. Toxicol. Pharm. 28:242–264.

Matthews, E.J., C.J. Ursem, N.L. Kruhlak, R.D.Benz, D.A. Sabaté, C. Yang, G.
Klopman, and J.F. Contrera. 2009a. Identification of structure-activity
relationships for adverse effects of pharmaceuticals in humans: Part B. Use of
(Q)SAR systems for early detection of drug-induced hepatobiliary and urinary
tract toxicities. Reg. Toxicol. Pharm. 54:23–42.

Matthews, E.J., N.L. Kruhlak, R.D. Benz, and J.F. Contrera. 2008. Combined use
of MC4PC, MDL-QSAR, BioEpisteme, Leadscope PDM, and Derek for Windows
software to achieve high-performance, high-confidence, mode of action-based
predictions of chemical carcinogenesis in rodents. Toxicol. Mech. Method.
18:189–206.

Matthews, E.J., N.L. Kruhlak, R.D. Benz, and J.F. Contrera. 2007a. A
comprehensive model for reproductive and developmental toxicity hazard
identification: I. Development of a weight of evidence QSAR database. Reg.
Toxicol. Pharm. 47:115–135.

Matthews, E.J., N.L. Kruhlak, R.D.Benz, J. Ivanov, G. Klopman, and J.F.

Contrera. 2007b. A comprehensive model for reproductive and developmental
toxicity hazard identification: II. Construction of QSAR models to predict activities
of untested chemicals. Reg. Toxicol. Pharm. 47:136–155.

Mekenyan, O., S. Dimitrov, N. Dimitrova, G. Dimitrova, T. Pavlov, G. Chankov, S.

Kotov, K. Vasilev, R. Vasilev. 2006. Metabolic activation of chemicals: in-silico
simulation. SAR QSAR Environ. Res. 17(1):107–120.

Page 113 of 186

Monev, V. 2004. Introduction to similarity searching in chemistry. Match-Comm.
Math. Comp. Chem. 51:7–38.

Mostrag-Szlichtyng, A. and A. Worth. 2010b. Review of QSAR Models and

Software Tools for predicting Biokinetic Properties. JRC Technical Report EUR
24377 EN. Publications Office of the European Union, Luxembourg.
http://ihcp.jrc.ec.europa.eu/our_labs/computational_toxicology/publications/

Mostrag-Szlichtyng, A. and A.P. Worth. 2010a. In silico modelling of microbial

and human metabolism: a case study with the fungicide carbendazim. JRC
Scientific and Technical Report EUR 24523 EN. Publications Office of the
European Union, Luxembourg.
http://ihcp.jrc.ec.europa.eu/our_labs/computational_toxicology/publications/

Mostrag-Szlichtyng, A., J.-M. Zaldívar Comenges, and A.P. Worth. 2010.

Computational toxicology at the European Commission’s Joint Research Centre.
Expert Opinion on Drug Metabolism and Toxicology 6, 785–792.

National Research Council (NRC). 1983. Risk assessment in the federal

government. Managing the process. National Academy Press, Washington, D.C.

National Research Council (NRC). 2007. Toxicity testing in the 21st century: A
vision and a strategy. National Academy of Sciences, Washington, DC.
http://dels.nas.edu/resources/static-assets/materials-based-on-reports/reports-in-
brief/Toxicity_Testing_final.pdf

Netzeva, T., M. Pavan, and A. Worth. 2007. Review of data sources, QSARs,
and integrated testing strategies for aquatic toxicity. European Commission, Joint
Research Centre, EUR 22943 EN, 103 p. Ispra, Italy.
http://ihcp.jrc.ec.europa.eu/our_labs/computational_toxicology/doc/EUR_22943_
EN.pdf

Netzeva, T.I., A. Worth, T. Aldenberg, R. Benigni, M.T. Cronin, P. Gramatica, J.S.

Jaworska, S. Kahn, G. Klopman, C.A. Marchant, G. Myatt, N. Nikolova-
Jeliazkova, G.Y. Patlewicz, R. Perkins, D. Roberts, T. Schultz, D.W. Stanton, J.J.
van de Sandt, W. Tong, G. Veith, and C. Yang. 2005. Current Status of Methods
for Defining the Applicability Domain of (Quantitative) Structure-Activity
Relationships. The report and recommendations of ECVAM Workshop 52. Altern.
Lab. Anim. 33(2):155–173.

Nikolova, N. and J. Jaworska. 2003. Approaches to measure chemical similarity

— a review. QSAR & Comb. Sci. 22:1006–1026.

OECD. 2005. The OECD Guidance Document on the Validation and International
Acceptance of New or Updated Test Methods for Hazard Assessment (OECD

Page 114 of 186

series on testing and assessment No. 34). OECD Environment Directorate Joint
Meeting of the Chemicals Committee and the Working Party on Chemicals,
Pesticides and Biotechnology. ENV/JM/MONO(2005)14. 18-Aug-2005.
http://www.oecd.org/officialdocuments/displaydocumentpdf?cote=env/jm/mono(2
005)14&doclanguage=en

OECD. 2007a. Guidance on grouping of chemicals. OECD Environmental Health

and Safety Publications. Series on Testing and Assessment No. 80.
ENV/JM/MONO (2007)28: 99 p.
http://www.oecd.org/officialdocuments/displaydocumentpdf/?cote=env/jm/mono(2
007)28&doclanguage=en

OECD. 2007b. Report on the regulatory uses and applications in OED member
countries of (quantitative) structure-activity relationship [(Q)SAR] models in the
assessment of new and existing chemicals. OECD Environmental Health and
Safety Publications. Series on Testing and Assessment No. 58. ENV/JM/MONO
(2006)25: 79 p. http://www.oecd.org/dataoecd/55/22/38131728.pdf

OECD. 2007c. The OECD Guidance Document on the Validation of

(Quantitative) Structure-Activity Relationships [(Q)SAR] Models. OECD
Environmental Health and Safety Publications. Series on Testing and
Assessment No. 69. OECD Environment Directorate Joint Meeting of the
Chemicals Committee and the Working Party on Chemicals, Pesticides and
Biotechnology. ENV/JM/MONO(2007)2. 30-May-2007.
http://www.oecd.org/officialdocuments/displaydocumentpdf?cote=ENV/JM/MON
O(2007)2&doclanguage=en

OECD. 2009. Guidance document for using the OECD QSAR application
toolbox to develop chemical categories according to the OECD guidance on
grouping of chemicals. OECD Series on Testing and Assessment No. 102.
ENV/JM/MONO (2009)5: 118 p.
http://www.oecd.org/dataoecd/50/60/42294034.pdf

OECD. 2011a. OECD Quantitative Structure-Activity Relationships Project

[(Q)SARs].
http://www.oecd.org/document/23/0,2340,en_2649_34365_33957015_1_1_1_1,00.html

OECD. 2011b. Report of the Workshop on Using Mechanistic Information in

Forming Chemical Categories. No. 138. Paris 2011.

OECD.2004. OECD Principles for the validation, for regulatory purposes, of

(Quantitative) Structure-Activity Relationship Models. Organization for Economic
Cooperation and Development.
http://www.oecd.org/dataoecd/33/37/37849783.pdf

Overton, C.E. 1901. Studien über die Narkose. Gustav Fischer, Jena, Germany.

Page 115 of 186

Patlewicz, G., A. Gallegos Saliner, M. Pavan, A. Worth, R. Benigni, A. Aptula, A.
Bassan, C. Bossa, A. Galk-Filipsson, V. Gillet, N. Jeliazkova, A. McDougal, J.
Mestres, I. Munro, T. Netzeva, B. Safford, B. Simon-Hettich, I. Tsakovska, M.
Wallen, and C. Yang. 2005. Chemical Similarity and Threshold of Toxicological
Concern (TTC) Approaches. Report of an ECB Workshop held in Ispra,
November 2005 European Commission, Joint Research Centre, EUR 22657 EN,
43 p. Ispra, Italy.
http://ihcp.jrc.ec.europa.eu/our_labs/computational_toxicology/doc/EUR_22657_EN.pdf

Patlewicz, G., A.O. Aptulu, D.W. Roberts, and E. Uriarte. 2008. A minireview of
available skin sensitization (Q)SARs/Expert Systems. QSAR Combin. Sci.
27(1):60–76.

Pavan, M., A.P. Worth, and T.I. Netzeva. 2005a. Preliminary analysis of an
aquatic toxicity dataset and assessment of QSAR models for narcosis. European
Commission, Joint Research Centre, EUR 21749 EN, 42 p. Ispra, Italy.
http://ihcp.jrc.ec.europa.eu/our_labs/computational_toxicology/doc/Report_QSAR
_model_for_narcosis.pdf

Pavan, M., A.P. Worth, and T.I. Netzeva. 2005b. Comparative assessment of
QSAR models for aquatic toxicity. European Commission, Joint Research
Centre, EUR 21750 EN, 122 p. Ispra, Italy.
http://ihcp.jrc.ec.europa.eu/our_labs/computational_toxicology/doc/Report_Comp
arative_assessment_QSAR_models.pdf

Pavan, M., and A. Worth. 2008. A set of case studies to illustrate the applicability
of DART (Decision Analysis by Ranking Techniques) in the ranking of chemicals.
European Commission, Joint Research Centre, EUR 23481 EN, 80 p. Ispra, Italy.
http://ihcp.jrc.ec.europa.eu/our_labs/computational_toxicology/doc/EUR_23481_EN.pdf

Perkins, R., H. Fang, W. Tong, and W.J. Welsh. 2003. Quantitative structure-
activity relationship methods: Perspectives on drug discovery and toxicology.
Environ. Toxicol. Chem. 22(8):1666–1679.

Richard, A.M., C. Yang, and R. Judson. 2008. Toxicity data informatics:

Supporting a new paradigm for toxicity prediction. Toxicol. Mechan.Methods 18(2
& 3):103–118.

Richard, A.M., I. Swirsky Gold, and M.C. Nicklaus. 2006. Chemical structure
indexing of toxicity data on the internet: moving towards a flat world. Curr.
Opinion Drug Disc.Dev. 9(3):314–325.

Ringeissen, S., L. Marrot, R. Note, A. Labarussiat, S. Imbert, M. Todorov, O.

Mekenyan, and J. Meunier. 2010. Development of a mechanistic SAR model for
the detection of phototoxic chemicals and use in an integrated testing strategy.
Toxicol. in Vitro (2010), doi:10.1016/ j.tiv.2010.09.017.

Page 116 of 186

Rorije, E. and E. Hulzebos. 2005. Evaluation of (Q)SARs for the prediction of skin
irritation/corrosion potential. RIVM, National Institute of Public health and the
Environment. September 2005. 65 p. Bilhoven, The Netherlands. 45 p.
http://ihcp.jrc.ec.europa.eu/our_labs/computational_toxicology/information-
sources/qsar-document-area/Evaluation_of_Skin_Irritation_QSARs.pdf

Ruggeri, B. 2009. Chemical exposure: Scoring procedure and uncertainty

propagation in scenario selection for risk analysis. Chemosphere 77:330–338.

Russom, C.L., R.L. Breton, J.D. Walker, and S.P. Bradbury. 2003. An overview of
the use of quantitative structure-activity relationships for ranking and prioritizing
large chemical inventories for environmental risk assessments. Environ. Toxicol.
Chem. 22(8):1810–1821.

Russom, C.L., S.P. Bradbury, S.J. Broderius, D.E. Hammermeister, and R.A.
Drummond. 1997. Predicting modes of toxic action from chemical structure:
Acute toxicity in the fathead minnow (Pimephales promelas). Environ. Toxicol.
Chem. 16(5):948–967.

Saliner, G., G.Patlewicz, and A.P. Worth. 2005. A similarity based approach for
chemical category classification. European Commission, Joint Research Centre
Report number: EUR 21867 EN. Institute for Health and Consumer Protection,
Toxicology and Chemical Substances Unit, European Chemicals Bureau, I-21020
Ispra (VA) Italy.
http://ihcp.jrc.ec.europa.eu/our_labs/computational_toxicology/doc/Report_Chemi
cal_Similarity_for_Category_Classification.pdf

Schmieder, P., O. Mekenyan, S. Bradbury, and G. Veith. 2003c. QSAR

prioritization of chemical inventories for endocrine disruptor testing. Pure Appl.
Chem. 75:2389–2396.

Schmieder, P.K., G.T. Ankley, O. Mekenyan, J.D. Walker, and S. Bradbury.

2003b. Quantitative structure-activity relationship models for prediction of
estrogen receptor binding affinity of structurally diverse chemicals. Environ.
Toxicol. Chem. 22:1844–1854.

Schmieder, P.K., M.A. Tapper, J.S. Denny, R.C. Kolanczyk, B.R. Sheedy, T.R.
Henry, and G.D. Veith. 2004. Use of trout liver slices to enhance mechanistic
interpretation of estrogen receptor binding for cost-effective prioritization of
chemicals within large inventories. Environ. Sci. Technol. 38:6333–6342.

Schmieder, P.K., M.A. Tapper, R.C. Kolanczyk, D.E. Hammermeister, B.R.

Sheedy, and J.S. Denny. 2003a. Discriminating redox cycling and arylation
pathways of reactive chemical toxicity in trout hepatocytes. Toxicol. Sci. 72:66–
76.

Page 117 of 186

Schultz, T. and B. Diderich. 2011. Adverse outcome pathways and their role in
helping to formulate mechanistically relevant chemical categories. AltTox.org
Non-animal Methods for Toxicity Testing.
http://alttox.org/spotlight/050.html

Tong, W., W.J. Welsh, L. Shi, H. Fang, and R. Perkins. 2003. Structure-activity
relationship approaches and applications. Environ. Toxicol. Chem. 22(8):1680–
1695.

Tsakovska, I., I. Lessigaiarska, T. Netzeva, and A. Worth. 2008. A mini review of

mammalian toxicity (Q)SAR models. QSAR & Combin. Sci. 27(1):41–48.

Tsakovska, I., T. Netzeva, and A. Worth. 2005. Evaluation of (Q)SARs for the
prediction of eye irritation/corrosion potential. Physicochemical exclusion rules.
European Commission, Joint Research Centre, EUR 21897 EN, 42 p. Ispra, Italy.
http://ihcp.jrc.ec.europa.eu/our_labs/computational_toxicology/doc/Evaluation_of
_Eye_Irritation_QSARs.pdf

Urbano-Cuadrado, M., I.L. Ruiz, and M.A. Gomez-Nieto. 2008. Description and
application of similarity-based methods for fast and simple QSAR model
development. QSAR Comb. Sci. 27(4):457–468.

Ursem, C.J., N.L. Kruhlak, J.F. Contrera, P.M. MacLaughlin, R.D. Benz, and E.J.
Matthews. 2009. Identification of structure-activity relationships for adverse
effects of pharmaceuticals in humans. Part A: Use of FDA post-market reports to
create a database of hepatobiliary an urinary tract toxicities. Reg. Toxicol. Pharm.
54:1–22.

US EPA. 1998. Guidelines for Ecological Risk Assessment (Published on May

14, 1998, Federal Register 63(93):26846-26924). Risk Assessment Forum.
EPA/630/R-95/002F April 1998.

US EPA. 1999. Development of chemical categories in the HPV Challenge

Program. Office of Pollution Prevention and Toxics. 21 p.

US EPA, 2007. Dermal Exposure Assessment: A Summary of EPA Approaches.

National Center for Environmental Assessment, Office of Research and
Development, Washington, DC 20460.
http://cfpub.epa.gov/ncea/cfm/recordisplay.cfm?deid=183584

US EPA. 2010. TSCA New Chemicals Program (NCP) Chemical Categories.

Office of Pollution Prevention and Toxics, Washington, DC 20460.
http://www.epa.gov/oppt/newchems/pubs/npcchemicalcategories.pdf

Page 118 of 186

US EPA. 2011. Aster (Assessment Tools for the Evaluation of Risk). National
Health and Environmental Effects Research Laboratory (NHEERL).
http://www.epa.gov/med/Prods_Pubs/aster.htm

Valkó K. 2004. Application of high-performance liquid chromatography based

measurements of lipophilicity to model biological distribution. J. Chromatog. A
1037(1-2):299–310.

van der Jagt, K., S. Munn, J. Torslov, and J. de Bruijn. 2004. Alternative
approaches can reduce the use of test animals under REACH. European
Commission, Joint Research Centre, Institute for Health Consumer Protection.
EUR 21405 EN. 31 p.
http://publications.jrc.ec.europa.eu/repository/bitstream/111111111/8790/1/EUR
%2021405%20EN.pdf

van Leeuwen, C.J., G.Y. Patlewicz, and A.P. Worth. (2007). Intelligent testing
strategies. In: Risk Assessment of Chemicals: An Introduction. 2nd Edition. Van
Leeuwen, C.J. and T.G. Vermeire, (eds.).

van Leeuwen, K. T.W. Schultz, T. Henry, B. Diderich, and G.D. Veith. 2009.
Using chemical categories to fill data gaps in hazard assessment. SAR QSAR
Environ. Res. 20(3-4):207–220.

Veith, G.D., D.L. De Foe, and B.V. Bergstedt. 1979. Measuring and estimating
the bioconcentration factor of chemicals in fish. J. Fish Res. Board Can.
36:1040–1047.

Veith, G.D., D.J. Call, and L.T. Brooke. 1983. Structure-toxicity relationships for
the fathead minnow, Pimephales promelas: Narcotic industrial chemicals. Can. J.
Fish. Aquat. Sci. 40:743–748.

Vonk, A.J., R. Benigni, M. Hewitt, M. Nendza, H. Segner, D. van de Meent, D.

and M.T.D. Cronin. 2009. The Use of Mechanisms and Modes of Toxic Action in
Integrated Testing Strategies: The Report and Recommendations of a Workshop
held as part of the European Union OSIRIS Integrated Project. ATLA 37, 557–
571.

Walker, J.D., D. Knaebel, K. Mayo, J. Tunkel, and D.A. Gray. 2004. Use of
QSARs to promote more cost-effective use of chemical monitoring resources. 1.
Screening industrial chemicals and pesticides, direct food additives, indirect food
additives and pharmaceuticals for biodegradation, bioconcentration and aquatic
toxicity potential. Water Qual. Res. J. Canada 29(1):35–39.

Page 119 of 186

Walker, J.D., J. Jaworska, M.H.I. Comber, T.W. Schultz, and J.C. Dearden. 2003.
Guidelines for developing and using quantitative structure-activity relationships.
Environ. Toxicol. Chem. 22(8):1653–1665.

Walker, JD. 2003. Quantitative structure-activity relationship for pollution

prevention, toxicity screening, risk assessment, and web applications. Pensacola,
FL, USA: Society of Environmental Toxicology and Chemistry (SETAC). 282 p.

Weininger, D. 1988. SMILES, a chemical language and information system. 1.

Introduction to methodology and encoding rules. J. Chem. Inform. Comput. Sci.
28:31–36.

WHO/FAO. 1996. Fenvalerate. WHO/FAO Data Sheets on Pesticides No. 90.

World Health Organization. Food and Agriculture Organization of the United
Nations. WHO/PCS/DS/96.90.
http://www.inchem.org/documents/pds/pds/pest90_e.htm

Williams, C., M. Wolf, and A.M. Richard. 2009. DSSTox chemical-index files for
exposure-related experiments in ArrayExpress and Gene Expression Omnibus:
enabling toxico-chemogenomics data linkages. Bioinformatics 25(5):692–694.

Woo, Y.T., and D.Y. Lai. 2010. (Q)SAR analysis of genotoxic and nongenotoxic
carcinogens: a state-of-the-art overview. In: Cancer Risk Assessment: Chemical
Carcinogenesis, Hazard Evaluation, and Risk Quantification, edited by Hsu, C.H.
and T. Stedeford, Wiley. Chapter 20, 517–556.

Woo, Y.-T., D.Y. Lai, J.C. Arcos, and M.F. Argus. 1985. Chemical Induction of
Cancer, Structural Bases and Biological Mechanisms. Vol IIIB: Aliphatic and
Polyhalogenated Carcinogens. Orlando, FL: Academic Press.

Woo, Y.-T., D.Y. Lai, J.L. McLain, M.K. Manibusan, and V. Dellarco. 2002. Use of
Mechanism-Based Structure-Activity Relationships Analysis in Carcinogenic
Potential Ranking for Drinking Water Disinfection By-Products. Environ. Health
Perspect. 110(Suppl 1):75–87.

Woo, Y.-T., D.Y. Lai, M.F. Argus, and J.C. Arcos. 1998. Integrative approach of
combining mechanistically complementary short-term predictive tests as a basis
for assessing the carcinogenic potential of chemicals. Environ. Carcinog.
Ecotoxicol. Rev. (Part C of J. Environ. Sci. Health) C16(2):101–122.

Worth, A., and G. Patlewicz. 2007. A compendium of case studies that helped to
shape the REACH guidance on chemical categories and read across. European
Commission, Joint Research Centre, EUR 22481 EN. 188 p.
http://publications.jrc.ec.europa.eu/repository/bitstream/111111111/1234/1/7212
%20-%20Compendium%20of%20case%20studies_final_060707.pdf

Page 120 of 186

Worth, A., M. Fuart-Gatnik, S. Lapenna, and R. Serafimova. 2011. Applicability of
QSAR analysis in the evaluation of developmental and neurotoxicity effects for
the assessment of the toxicological relevance of metabolites and degradates of
pesticide active substances for dietary risk assessment. Report produced for
EFSA.
http://www.efsa.europa.eu/en/supporting/pub/169e.htm

Worth, A., S. Lapenna, E. Lo Piparo, A. Mostrag-Szlichtyng, and R. Serafimova.

2010. The Applicability of Software Tools for Genotoxicity and Carcinogenicity
Prediction: Case Studies relevant to the Assessment of Pesticides. JRC
Technical Report EUR 24640 EN. Publications Office of the European Union,
Luxembourg.
http://ihcp.jrc.ec.europa.eu/our_labs/computational_toxicology/publications/

Worth, A., S. Lapenna, E. Lo Piparo, A. Mostrag-Szlichtyng, and R. Serafimova.

2011. A Framework for assessing in silico Toxicity Predictions: Case Studies with
selected Pesticides. JRC report EUR 24705 EN. Publications Office of the
European Union, Luxembourg.
http://ihcp.jrc.ec.europa.eu/our_labs/computational_toxicology/publications/

Worth, A.P. 2010. The role of QSAR methodology in the regulatory assessment
of chemicals, Chapter 13 in Recent Advances in QSAR Studies: Methods and
Applications. Puzyn, T., J. Leszczynski, and M.T.D. Cronin (eds). Springer,
Heidelberg, Germany, pp. 367–382.

Worth, A.P. and A. Mostrag-Szlichtyng. 2011. Towards a Common Regulatory

Framework for Computational Toxicology: Current Status and Future
Perspectives, in New Horizons in Predictive Toxicology: Methods and
Applications. Wilson, A.G.E. (ed). The Royal Society of Chemistry, Cambridge,
UK, pp 38–69.

Yuan, H., Y. Wang, and Y. Cheng. 2007. Local and Global Quantitative
Structure−Activity Relationship Modeling and Prediction for the Baseline Toxicity.
J. Chem. Inf. Model., 2007, 47 (1), pp 159–169.

Zeeman, N., C.M. Auer, R.G. Clements, J.V. Nabholz, and R.S. Boethling. 1995.
U.S. EPA regulatory perspectives on the use of QSAR for new and existing
chemical evaluations. SAR/QSAR Environ. Res. 3:179–201.

Zvinavashe, E., A.J. Murk, and I.M.C.M. Rietjens. 2008. Promises and pitfalls of
quantitative structure-activity relationship approaches for predicting metabolism
and toxicity. Chem. Res. Toxicol. 21:2229–2236.

Page 121 of 186

Appendix I

Where To Go to Learn More About (Q)SAR

Listed below are world wide webpages (WWW) for a number of national and
international organizations involved in projects and the development of tools
related to (Q)SAR and the assessment of risks from chemicals. In most cases,
the descriptions of the activities and projects come directly from the webpages.
These links may be useful to pesticide evaluators who are seeking additional
information on general (Q)SAR concepts, and the development, validation, and
evaluation of (Q)SAR tools and predictions. Some of the links and information
below have been cited in various sections of this guidance document.

The list below is by no means exhaustive, and since the field of (Q)SAR is
constantly expanding, pesticide evaluators are advised to regularly monitor
various national and international agency websites and the open literature for
developments in the area of (Q)SAR of interest to them.

Danish Ministry of the Environment Environmental Protection

Agency (Danish EPA) (Q)SAR — Assessment of Chemical
Properties of Substances
http://www.mst.dk/English/Chemicals/assessment_of_chemicals/qsar_assessme
nt_chemical_properties_of_substances/

Descriptions of the key (Q)SAR activities from the Danish EPA webpage (i.e., the
(Q)SAR database and the Advisory list for self-classification of dangerous
substances) are provided below.

“The Danish EPA has for a number of years worked with the development and
use of (Q)SAR’s, also called ‘computer models’ for prediction of properties of
chemical substances. (Quantitative) Structure Activity Relationships — (Q)SAR
— are relations between structure properties of chemical substances and some
other property. The other property can be a physical-chemical property or a
biological activity, including the ability to cause toxic effects.”

“The Danish EPA has made a database, which comprise predictions from more
than 70 (Q)SAR models on endpoints for physico-chemical properties, fate, eco-
toxicity, absorption, metabolism and toxicity. The database is constantly growing
as new models are obtained and developed. More than half of all the estimates

Page 122 of 186

are for mammalian (human) toxicity endpoints and include commercial data sets
from TOPKAT and MULTICASE as well as many models developed in-house.”

“Lack of data on hazardous properties of chemicals makes it difficult for

companies to meet their obligations to self classify the chemicals they import or
produce. To address this issue, The Danish Environmental Protection Agency
(DK-EPA) publishes the advisory list for self classification of chemical substances
— with advisory classifications of more than 30,000 substances. Since the new
regulation for classification and labelling (the CLP-regulation) came into force, the
regulation is in a transitional phase until 2015 where both regulations still are
relevant in certain situations. Therefore both regulations are covered.”

European Commission Joint Research Centre (JRC) Institute for

Health and Consumer Protection (IHCP)
http://ihcp.jrc.ec.europa.eu/our_labs/computational_toxicology/

A description of the EU computational toxicology and modeling activities from the

IHCP website is listed below.

“One of the activities of the Institute is to support the implementation of EU

chemicals policy (including the safety assessment of industrial chemicals,
chemicals in consumer products, pesticides and biocides) through the
development, assessment and application of computational (in silico) methods.
These methods, sometimes referred to as ‘non-testing methods’, can be used to
reduce our reliance on experimental testing, and in particular animal testing. In
practice, these methods are used in Integrated Testing Strategies, along with
experimental data generated by alternative (non-animal) tests, such as in vitro
tests and high throughput screening (HTS) assays.”

“Quantitative Structure-Activity Relationship (QSAR) models can be used to

obtain information on the properties and activities of substances from chemical
structure alone, and can thus be used to fill data gaps in the safety assessment
of chemicals.”

“Another method, Physiologically Based Biokinetic (PBBK) Modelling, can be

used to extrapolate between in vitro and in vivo exposure conditions, thereby
helping to establish the relevance of data generated by in vitro toxicity tests.”

The IHCP website provides information and links to a number of documents and
several downloadable (Q)SAR tools:

Page 123 of 186

 • Background information on non-testing methods
• Information on QSAR reporting formats
• A range of other information sources, including chemical inventories and
documents on computational toxicology
• A list of IHCP publications on computational toxicology
• Webpages for free download or access to the following computational
tools:
o JRC QSAR Model Database
o Toxtree
o Dart
o Toxmatch
o Stat-4-tox

Organization for Economic Cooperation and Development

(OECD) OECD (Q)SAR Project
http://www.oecd.org/env/hazard/qsar

The following is a description of the OECD (Q)SAR project from the OECD
website:

“To facilitate practical application of (Q)SAR approaches in regulatory contexts by

governments and industry and to improve their regulatory acceptance, the OECD
(Q)SAR project has developed various outcomes, such as the principles for the
validation of (Q)SAR models, guidance documents as well as the QSAR
Application Toolbox. The OECD (Q)SAR Project is carried out with the financial
assistance of the European Union.”

The (Q)SAR project website includes links to a number of documents and

software relating to OECD (Q)SAR activities including the following:
• History (of the project)
• Introduction to (Q)SARs
• Grouping of chemicals
• Validation of (Q)SAR models
• OECD QSAR Toolbox

Rijksinstituut voor Volksgezondheid en Milieu (RIVM) [National

Institute for Public Health and the Environment — Netherlands]
http://www.rivm.nl/rvs/risbeoor/Modellen/QSAR.jsp

Page 124 of 186

The RIVM QSAR webpage provides a definition of QSAR and links to a number
of QSAR related publications produced by the Institute including:
• Report: (Q)SARs: gatekeepers against risk on chemicals?
• A literature review of (Q)SARs for human toxicological endpoints
• A literature review on (Q) SARs for ecotoxicological endpoints
• Report: The application of structure-activity relationships in human hazard
assessment: a first approach
• Report: Estimating the PBT-profile

While much of the information on the webpage is in Dutch, the majority of the
publications are available in English.

United States Environmental Protection Agency (US EPA) Office

of Pollution Prevention and Toxics (OPPT) Sustainable Futures
Initiative (SF)
http://www.epa.gov/oppt/sf/tools/methods.htm

A description of the SF initiative from the OPPT webpage is listed below.

“The goal of the Sustainable Futures Initiative (SF) is to make new chemicals
safer, available faster, and at lower cost. It works by giving chemical developers
the same risk-screening models that EPA uses to evaluate new chemicals before
they enter the market.”

“Using these computer-based models, companies can identify potentially risky

chemicals early in the development process and reduce risk by finding safer
substitutes and/or processes before submitting them to the EPA. Also, the
companies that take training and graduate from Sustainable Futures become
eligible for an expedited EPA review of their pre-screened chemicals.”

The computer-based models and tools freely available for download from the
OPPT webpage include the following:
• EPI Suite
• ECOSAR
• PBT Profiler
• Oncologic
• Analog Identification Methodology (AIM)
• NonCancer Screening Protocol
• E-FAST
• ChemSteer

Page 125 of 186

United States Food and Drug Administration (US FDA) Center for
Drug Evaluation and Research (CDER) Informatics and
Computational Safety Analysis Staff (ICSAS)
http://www.fda.gov/AboutFDA/CentersOffices/OfficeofMedicalProductsandTobac
co/CDER/ucm092125.htm

The following description of ICSAS is listed on the US FDA CDER webpage:

“The Informatics and Computational Safety Analysis Staff (ICSAS) is part of

CDER’s Office of Pharmaceutical Science. ICSAS is an applied regulatory
research unit that:
• Develops databases of toxicological and clinical endpoints
• Transforms data, developing rules for quantifying toxicological and clinical
effects
• Evaluates structure activity relationship (SAR) and data mining software
using ICSAS databases
• Works with software developers to develop toxicology and clinical effects
prediction programs through research leveraging partnerships
• Reduces the use of animals in testing by eliminating non-critical laboratory
studies
• Facilitates the review process by making better use of accumulated
scientific knowledge
• Supplies tools to the pharmaceutical industry to develop better means to
identify and eliminate compounds with potentially significant adverse
properties early in the drug discovery and development process”

The webpage also includes links to databases and further information on ICSAS
activities including:
• Database Projects
• Maximum Recommended Therapeutic Dose (MRTD) Database
• Human Liver Adverse Effects Database
• Genetic Toxicity, Reproductive and Developmental Toxicity, and
Carcinogenicity Database
• Salmonella Mutagenicity E-state Descriptors
• Chemical Structure Similarity Searching
• The Computational Toxicology Program and ComTox Consulting Service
• ComTox Regulatory Application of ICSAS MCASE/MC4PC-ES by the
Center for Food Safety and Applied Nutrition
• Application of Computational Toxicology to Assess Clinical Adverse Drug
Reactions
• Publications

Page 126 of 186

Additional Useful References
The following are some additional useful references on the development,
evaluation, and application of (Q)SAR tools. These references are also included
in the reference section of this document.

Bassan, A., E. Fioravanzo, M. Pavan, and M. Stocchero. 2011. Applicability of

physicochemical data, QSARs and read-across in Threshold of Toxicological
Concern assessment. Final report of a study carried out by Soluzioni
Informatiche (S-In, Vicenza, Italy) for the European Food Safety Authority
(EFSA). http://www.efsa.europa.eu/en/supporting/pub/159e.htm

Cronin, M. 2010. Quantitative structure-activity relationships (QSARs) —

applications and methodology. Chapter 10 in Recent Advances in QSAR Studies:
Methods and Applications. Puzyn, T., J. Leszczynski, and M.T.D. Cronin (eds.).
Springer, Heidelberg, Germany, pp. 3–11.

ECHA. 2010. Practical guide 5. How to report (Q)SAR. European Chemicals

Agency, Helsinki, Finland.
http://echa.europa.eu/documents/10162/13655/pg_report_qsars_en.pdf

Lo Piparo, E., A. Worth, M. Manibusan, C. Yang, B. Schilter, P. Mazzatorta, M.N.

Jacobs, H. Steinkellner, and L. Mohimont. 2011. Use of computational tools in
the field of food safety. Regulatory Toxicology and Pharmacology 60:354–362.

Mostrag-Szlichtyng, A. and A.P. Worth. 2010. In silico modelling of microbial and

human metabolism: a case study with the fungicide carbendazim. JRC Scientific
and Technical Report EUR 24523 EN. Publications Office of the European Union,
Luxembourg. Available at:
http://ihcp.jrc.ec.europa.eu/our_labs/computational_toxicology/publications/

Mostrag-Szlichtyng, A. and A. Worth. 2010. Review of QSAR Models and

Software Tools for predicting Biokinetic Properties. JRC Technical Report EUR
24377 EN. Publications Office of the European Union, Luxembourg. Available at:
http://ihcp.jrc.ec.europa.eu/our_labs/computational_toxicology/publications/

van Leeuwen, C.J., G.Y. Patlewicz, and A.P. Worth. 2007. Intelligent testing
strategies. In: Risk Assessment of Chemicals: An Introduction. 2nd Edition. Van
Leeuwen, C.J. and T.G. Vermeire, (eds.).

Vonk, A.J., R. Benigni, M. Hewitt, M. Nendza, H. Segner, D. van de Meent, and

M.T.D. Cronin. 2009. The Use of Mechanisms and Modes of Toxic Action in
Integrated Testing Strategies: The Report and Recommendations of a Workshop
held as part of the European Union OSIRIS Integrated Project. ATLA 37:557–
571.

Page 127 of 186

Worth, A., M. Fuart-Gatnik, S. Lapenna, and R. Serafimova. 2011. Applicability of
QSAR analysis in the evaluation of developmental and neurotoxicity effects for
the assessment of the toxicological relevance of metabolites and degradates of
pesticide active substances for dietary risk assessment. Report produced for
EFSA. Available at: http://www.efsa.europa.eu/en/supporting/pub/169e.htm

Worth, A., S. Lapenna, E. Lo Piparo, A. Mostrag-Szlichtyng, and R. Serafimova.

2010. The Applicability of Software Tools for Genotoxicity and Carcinogenicity
Prediction: Case Studies relevant to the Assessment of Pesticides. JRC
Technical Report EUR 24640 EN. Publications Office of the European Union,
Luxembourg. Available at:
http://ihcp.jrc.ec.europa.eu/our_labs/computational_toxicology/publications/

Worth, A.P. 2010. The role of QSAR methodology in the regulatory assessment
of chemicals, Chapter 13 in Recent Advances in QSAR Studies: Methods and
Applications. Puzyn, T., J. Leszczynski, and M.T.D. Cronin, (eds.). Springer,
Heidelberg, Germany, pp. 367–382.

Worth, A.P. and A.Mostrag-Szlichtyng. 2011. Towards a Common Regulatory

Framework for Computational Toxicology: Current Status and Future
Perspectives, in New Horizons in Predictive Toxicology: Methods and
Applications. Wilson, A.G.E. (ed.). The Royal Society of Chemistry, Cambridge,
UK, pp. 38–69.

Page 128 of 186

Appendix II

European Commission Joint Research Centre (JRC) QSAR

Model Reporting Format (QMRF) and QSAR Prediction Reporting
Format (QPRF)
The European Commission Joint Research Centre (JRC) has developed the
QSAR Model Reporting Format (QMRF), a template for summarizing and
reporting critical information on (Q)SAR models. The JRC also maintains a freely
accessible database of QMRFs. The QSAR Prediction Reporting Format
(QPRF), also developed by the JRC, is a template for summarizing and reporting
individual substance-specific predictions generated by (Q)SAR models. Both
templates have been designed to solicit information about (Q)SAR models and
predictions that corresponds to the OECD “Principles for the Validation, for
Regulatory Purposes, of (Q)SAR Models”. Also, the JRC has noted that the
QPRF and QMRF are complementary and that a QPRF should always be
associated with a QMRF.

While not specifically designed for documenting (Q)SAR models and predictions
for pesticides, the QMRF and QPRF can be viewed as examples of detailed
information templates of the type that may need to be considered when (Q)SAR
predictions are to be used as critical sources of data in pesticide assessments.

Tables 12.1 and 12.2 below list the information fields included in the QMRF and
the QPRF. The most current versions of the QMRF and the QPRF, guidelines for
reviewing the QMRF, a QMRF editor for filling in the QMRF, guidance on creating
SDF files associated with QMRFs, and examples of completed QMRFs can be
downloaded from the following website:

http://ihcp.jrc.ec.europa.eu/our_labs/computational_toxicology/qsar_tools/QRF

The website above also provides access to the JRC database of QMRFs.

In addition, the European Chemicals Agency has developed guidance for the use
of the QMRF and QPRF to report on (Q)SARs and to input information from
these templates into IUCLID 5 (ECHA, 2009).

Page 129 of 186

Table 1. JRC QMRF (version 1.2) Information

1. QSAR identifier
1.1 QSAR identifier (title)
1.2 Other related models
1.3 Software coding the model

2. General Information
2.1 Date of QMRF
2.2 QMRF author(s) and contact details
2.3 Date of QMRF update(s)
2.4 QMRF update(s)
2.5 Model developer(s) and contact details
2.6 Date of model development and/or publication
2.7 Reference(s) to main scientific papers and/or software package
2.8 Availability of information about the model
2.9 Availability of another QMRF for exactly the same model

3. Defining the endpoint — OECD Principle 1

3.1 Species
3.2 Endpoint
3.3 Comment on endpoint
3.4 Endpoint units
3.5 Dependent variable
3.6 Experimental protocol
3.7 Endpoint data quality and variability

4. Defining the algorithm — OECD Principle 2

4.1 Type of model
4.2 Explicit algorithm
4.3 Descriptors in the model
4.4 Descriptor selection
4.5 Algorithm and descriptor generation
4.6 Software name and version for descriptor generation
4.7 Descriptors/Chemicals ratio

5. Defining the applicability domain — OECD Principle 3

5.1 Description of the applicability domain of the model
5.2 Method used to assess the applicability domain
5.3 Software name and version for applicability domain assessment
5.4 Limits of applicability

Page 130 of 186

6. Defining goodness-of-fit and robustness — OECD Principle 4
6.1 Availability of the training set
6.2 Available information for the training sets
6.3 Data for each descriptor variable for the training set
6.4 Data for the dependent variable (response) for the training set
6.5 Other information about the training set
6.6 Pre-processing of data before modeling
6.7 Statistics for goodness-of-fit
6.8 Robustness- Statistics obtained by leave-one-out cross-validation
6.9 Robustness- Statistics obtained by leave-main-out cross-validation
6.10 Robustness- Statistics obtained by Y-scrambling
6.11 Robustness- Statistics obtained by bootstrap
6.12 Robustness-Statistics obtained by other methods

7. Defining predictivity — OECD Principle 4

7.1 Availability of the external validation set
7.2 Available information for the external validation set
7.3 Data for each descriptor variable for the external validation set
7.4 Data for the dependent variable for the external validation set
7.5 Other information about the external validation set
7.6 Experimental design of test set
7.7 Predictivity- Statistics obtained by external validation
7.8 Predictivity- Assessment of the external validation set
7.9 Comments on the external validation of the model

8. Providing a mechanistic interpretation — OECD Principle 5

8.1 Mechanistic basis of the model
8.2 A priori or a posteriori mechanistic interpretation
8.3 Other information about the mechanistic interpretation

9. Miscellaneous information
9.1 Comments
9.2 Bibliography
9.3 Supporting information

10. Summary for the JRC Inventory

10.1 QMRF number
10.2 Publication date
10.3 Keywords
10.4 Comments

Page 131 of 186

Table 2. EC QPRF (version 1.1) Information Fields

1. Substance
1.1 CAS number
1.2 EC number
1.3 Chemical name
1.4 Structural formula
1.5 Structure code
a. SMILES
b. InChi
c. Other structural representation
d. Stereochemical features

2. General information
2.1 Date of QPRF
2.2 QPRF author and contact details

3. Prediction
3.1 Endpoint (OECD Principle 1)
a. Endpoint
b. Dependent variable
3.2 Algorithm (OECD Principle 2)
a. Model or submodel name
b. Model version
c. Reference to QMRF
d. Predicted value (model result)
e. Predicted value (comments)
f. Input for prediction
g. Descriptor values
3.3 Applicability domain (OECD Principle 3)
a. Domains
i. Descriptor domain
ii. Structural fragment domain
iii. Mechanistic domain
iv. Metabolic domain
b. Structural analogues
c. Considerations on structural analogues
3.4 The uncertainty of the prediction (OECD Principle 4)
3.5 The chemical and biological mechanisms according to the model
underpinning the predicted result (OECD Principle 5)

4. Adequacy (Optional)
4.1 Regulatory purpose
4.2 Approach for regulatory interpretation of the model result
4.3 Outcome
4.4 Conclusion

Page 132 of 186

Appendix III
Listed below is a selection of example case studies of the application of (Q)SAR
tools and approaches to the prediction of the ecotoxicity and toxicity of pesticides
and other chemicals. These examples were prepared by various groups in the
US EPA and Health Canada for applications within their respective programs.
These examples were not specifically designed for use in this document, but they
were generously contributed in order to illustrate various issues discussed in
sections 3, 4, 5, 6, and 7 of the document.

Example No. 1
Case Study:

Use of EcoSAR QSAR Models to Estimate the Acute Toxicity of Organophosphate and
Carbamate Pesticide Classes to Fish Species.

The following is a summarized version of a case study of reliability and validation

testing of a set of (Q)SAR models for predicting acute toxicity to fish species.
This case study involved comparing available high quality empirical data on acute
toxicity for an external test set of organophosphate and carbamate pesticides
with model predictions for the same pesticides generated by the US EPA’s
EcoSAR models. The case study was prepared by the US EPA’s National Health
and Environmental Research Laboratory (NHERL) and it provides a useful
example of application of the OECD (Q)SAR validation principles and
considerations of the adequacy of (Q)SAR predictions as discussion in Section 5
of this document.

Issue: Currently the US EPA’s Office of Pesticide Programs receives acute toxicity data
for fish species via the FIFRA registration process. OPP typically does not obtain test
data for degradate chemicals of active ingredients and relies on QSAR approaches to
determine the potential hazard associated with these substances. In addition, the Office
of Water is interested in using QSAR approaches to fill data gaps to meet minimum data
requirements in the development of water quality criteria for pesticide active
ingredients. These are typically for other chordate and arthropod taxa, but at times
data gaps exist for fish and salmonid species. A potential issue with using QSAR models
is that most tools were developed to support the TSCA legislation which deals primarily
with industrial organic compounds. To this end, an analysis was conducted to
determine the reliability and validity of QSAR models for use in estimating the acute
toxicity to fish for a set of organophosphate and carbamate pesticides acting via an
acetylcholinesterase inhibition mode of action.

Page 133 of 186

Approach: For this exercise Version 1.1 of the US EPA’s EcoSAR modeling application
was used (http://www.epa.gov/oppt/newchems/tools/21ecosar.htm). Independent
assessments of EcoSAR attest to the reliability for predicting toxicity for non-specific
modes of action (MOAs), and, with limited success, predicting toxicity for more specific
MOAs such as reactive mechanisms (Moore et al., 2003; Reuschenbach et al., 2008).
With Version 1.1 of EcoSAR, the US EPA’s Office of Pollution Prevention and Toxics
(OPPT), which maintains the tool, began augmenting models with a limited number of
pesticide data which have been reviewed and categorized as acceptable for fulfillment
of pesticide registration and re-registration guideline requirements as explained under
FIFRA Subdivision E, Parts 158.145 and 158.150. This case study was conducted to
determine whether these improvements led to reliable QSAR models for use in
estimating hazard associated with carbamate and organophosphate pesticides with an
acute mode of action of acetylcholinesterase inhibition. To make this determination, a
high quality empirical data set was compiled and used to evaluate how well the QSAR
models estimates agreed with empirical toxicity test data. Chemicals that were used in
the EcoSAR model training set were excluded from the final evaluations of model
performance, thereby being a validation of the existing models as they relate to use in
estimating toxicity of acetylcholinesterase inhibitors to fish. The OECD QSAR validation
principles are used as a template for examining the applicability, reliability, robustness,
and predictivity of models.

QSAR Prediction Reporting Format (QPRF):

1. Substance: Carbamates: Initially a list of seventeen (17) carbamates and forty-six

(47) organophosphate pesticides were examined (See Attachment 1 for chemical
name, CAS Registry Number, and SMILES string and Attachment 2 for the logP
(log of the octanol water partition coefficient.)

2. General information:

a. QPRF author and contact details: This analysis was compiled by Chris
Russom, US EPA, ORD, NHEERL, MED, Duluth, MN
(Russom.chris@epa.gov and completed using EcoSAR V 1.1 with EcoSAR
outputs completed on June 2011.

3. Validation of QSAR model vs. High Quality Empirical Data Set:

a. Empirical Data Set: A data set was compiled for use in validating model
predictions by selecting test results from the ECOTOX database
(www.epa.gov/ecotox), and the OPP database of studies submitted for
registration of active ingredients (Brian Montague, OPP/EFED, personal

Page 134 of 186

communication). Studies were used which met the following criteria: (a)
fish species identified; (b) endpoint is LC50 ; (c) test conducted in
freshwater; (d) compound purity >90%; (e) test duration of 96 hrs; (f)
effect is mortality; (g) studies conducted in a laboratory setting; (h)
concentrations are not indefinite values (i.e., exclude >, < or ~ values); (i)
documented temperature and dissolved oxygen measurements; and (j)
adherence to standard test procedures (ASTM 2007; U.S. EPA 1996) e.g.,
as they relate to organism life stage, water temperature and dissolved
oxygen. Duplicate studies were removed by identifying tests where the
CAS Registry number, species, age, and LC50 toxicity concentrations were
the same. Once all requirements had been met an analysis of outliers
was performed. In instances where more than one toxicity value was
available for a chemical, species, endpoint, duration combination, the
ratio of the maximum and minimum concentration values was calculated.
When ratios approached or exceed 10, the original publications were
examined, and if errors in data were identified these data were not
included in the analysis. If an outlier could not be determined (e.g., only
two data points), then all data records were removed if the ratio was
greater than 10.

The focus of this evaluation was the use of QSAR models to estimate
acute toxicity to fulfill minimum data requirements for use by OW in
deriving Agency benchmarks. Rainbow trout or bluegill sunfish tended to
be the most sensitive fish species (Figure 1) upon examination of the
empirical validation data set. Since the ECOSAR model is generic for fish
(see 3b below), and a critical minimum data requirement under the
Water Quality Criteria Guidelines is a salmonid (USEPA 1985),
comparisons of the QSAR model estimates were made using the average
test concentration of empirical test results for rainbow trout when
possible, and bluegill test data only when rainbow trout data were not
available (See Attachment 2).

Page 135 of 186

 Figure 1: Average toxicity value by for OP and carbamate acetylcholinesterase
inhibitors from ICE model data set
1000000
Organophosphates
Carbamates

100000

Acute Toxicity (mg/L)

10000

1000

100

10

L. macrochirus C. carpio P. pimephales P. reticulata O. mykiss

b. QSAR Model Endpoint (OECD Principle 1 — A defined biological

endpoint):

i. Species/Endpoint: This analysis compares QSAR estimates to a

high quality empirical data set using the EcoSAR QSAR models for
acute (96 hr) LC50 to freshwater fish species. The EcoSAR QSAR
models provide an estimate for a generic fish, not for a particular
species of fish, although frequently used species to develop the
QSAR models included bluegill sunfish (Lepomis macrochirus),
common carp (Cyprinus carpio), fathead minnow (Pimephales
promelas), guppy (Poecilia reticulata), rainbow trout
(Oncorhynchus mykiss), medaka (Oryzias latipes), or zebrafish
(Brachydanio rerio). Model output to the user is presented in
concentration units of mg/L.

ii. Test Protocols: Acute toxicity test data used in the training set
followed either ASTM or OPP standard testing procedures (ASTM,
2007; U.S. EPA 1996). Therefore the data were from several
laboratories.

iii. Dependent variable: The model calculates log millimole/liter LC50

and EcoSAR software converts the value to mg/L for the report
page provided to the user.

Page 136 of 186

3.1 Algorithm (OECD Principle 2)

a. Model, Version, and QSAR sub-model name:

i. QSAR and model name: Table 1 provides details on the linear

regression models from EcoSAR Version 1.1 used in this model
validation exercise. The supporting files within EcoSAR provide a
list of all test data included in the model training set. Attachment
3 identifies the models used for each chemical in the validation
set. None of the chemicals in the validation set exceeded model
limits. These are chemical class-based models and guidance in the
model documentation states:

1. The Carbamate Esters, phenyl chemical class model may

be used to estimate toxicity of O-phenyl substituted
carbamate esters (i.e., RNC(=O)OPh; where R is anything
except two hydrogens; and R is not a primary amine. The
phenyl (Ph) can have substitutions on the ring.)

2. The Carbamate Esters, Oxime chemical class model may be

used to estimate toxicity of O-oxime substituted
carbamate esters (i.e., RNC(=O)ON=C; where R is anything
except two hydrogens; and R is not a primary amine.)

3. The Esters, Phosphates chemical class model may be used

to estimate toxicity for phosphate esters (i.e.,
R1OP(=O)(R2)R3 OR R4SP(=O)(R2)R3 where R1 can be alkyl
carbon, olefinic carbon, acetylenic carbon, aromatic
carbon, a carbonyl, phosphorus, sulfur, oxygen or
nitrogen; R2 and R3 can be anything EXCEPT a hydroxy
group (OH), and R4 can be an alkyl carbon, olefinic carbon,
acetylenic carbon, aromatic carbon, or phosphorus.)
EcoSAR flagged halogenated tri-alkylphosphate esters as
being significantly more toxic than would be estimated by
this model. EcoSAR documentation also commented that
this model may over or under estimate toxicity for
acetylcholinesterase inhibitors, and that proper
classification of these substances is ongoing.

4. The Esters, Dithiophosphates chemical class model may be

used to estimate toxicity for dithiophosphate esters (i.e.,
R1P(=S)(SR2)OR3; where R1 can be anything except sulfur;
R2 must be a hydrogen, carbon (alkyl, olefinic, acetylenic,
aromatic, or carbonyl), or phosphorus, and R3 must be a

Page 137 of 186

 hydrogen, carbon (alkyl, olefinic, acetylenic, aromatic, or
carbonyl), nitrogen, oxygen, phosphorus, or sulfur.)

5. The Esters, Monothiophosphates chemical class-based

model may be used to estimate toxicity for
monothiophosphate esters (i.e., R1P(=S)(R2)OR3; where R1
and R2 can be anything except sulfur, and R3 must be a
hydrogen, carbon (alkyl, olefinic, acetylenic, aromatic, or
carbonyl), nitrogen, oxygen, phosphorus, or sulfur.)

Table 1: EcoSAR V 1.1 Fish Acute LC50 QSAR model information:

Model name QSAR Statistics

2
Carbamate Esters, phenyl Log 96-h LC50 (mmol/L) = -0.3478 (logP) - 0.9147 R = 0.1697; N=23
2
Carbamate Esters, oxime Log 96-h LC50 (mmol/L) = -0.4048 (logP) – 1.6878 R = 0.4475; N=18
2
Esters, Phosphate Log 96-h LC50 (mmol/L) = -0.3504 (logP) - 0.9625 R = 0.2551; N=27
2
Esters, Dithiophosphate Log 96-h LC50 (mmol/L) = -0.4981 (logP) – 1.2363 R = 0.2355; N=28
2
Esters, Monothiophosphate Log 96-h LC50 (mmol/L) = -0.5902 (logP) – 0.7618 R = 0.1508; N=44

Figure 2: Log molar LC50 for fish (empirical test data and EcoSAR V1.1
QSAR estimates) vs. the log of the octanol / water partition coefficient (logP)

4.00

2.00

Non-polar Narcosis
0.00
Log Molar LC50

-2.00

-4.00

-6.00

-8.00

-2 -1 0 1 2 3 4 5 6 7
-10.00
Log P
Empirical Test Data Phenyl Carbamate Ester Oxime Carbamate Ester

2 Monothiophosphate Ester Phosphate Ester Dithiophosphate Ester

Page 138 of 186

 ii. Predicted value: Attachment 3 provides model estimates as
provided in EcoSAR user outputs. As stated earlier, there were no
flags on any of the estimates, therefore these chemicals were all
estimated within the domain of the model as parameterized. Also,
the model estimates the millimolar/liter as the unit, but EcoSAR
converts the value to mg/L prior to presenting it in the EcoSAR
output. Figure 2 is a plot of the log molar LC50 (QSAR estimated
and empirical) vs. log P. The non-polar narcosis toxicity line is
provided as a baseline.

iii. Input for prediction: CAS Registry numbers were used as input for
EcoSAR model prediction, and SMILES strings were available
within EcoSAR for all chemicals except cis-Thiocarboxime. The
EcoSAR structures were verified against a second source
(Alanwood or ChemID). The SMILES string was written and used
as input for EcoSAR model prediction for cis-Thiocarboxime and
this structure was verified in a second source.

iv. Descriptor values: The logP value from KowWin Version 1.68 was
retrieved by EcoSAR as the descriptor variable for the resident
QSAR models.

3.2 Applicability domain (OECD principle 3)

a. Domains:

i. Descriptor variable: The logP values for the validation data set
ranged from 0.123-2.552 for the carbamates and -0.096-5.863 for
the organophosphates. Examining training sets used for each
QSAR model against the chemicals with estimated values not
included in the training set found that all carbamate chemicals
except formetanate (logP=0.89 vs. logP for Phenyl carbamate
ester training set of 1.52 to 3.06) had logP values within the
model training set logP range. For the organophosphates
estimated using the Monothiophosphate ester QSAR models, all
chemical in the validation set were within the training set logP
ranges (i.e., logP range of 2.4 to 4.7) except Dichlofenthion
(logP=5.202), Fenchlorfos (logP=4.865), Iodophos (logP=5.387),
and Trichloronate (logP=5.863). All chemicals estimated using the
Dithiophosphate ester QSAR models had logP values within the
training set logP range. Dicrotophos (logP=-0.096) was the only
substance estimated using the Phosphate ester QSAR model that
was outside the logP range for the training set (-0.74 to 4.85).
EcoSAR documentation acknowledges that in general, above a

Page 139 of 186

 logP of 5, research has shown that the hydrophobicity of the
molecule leads to “no effects at saturation” and due to this,
ECOSAR documentation recommends this as an upper limit of
acute toxicity (Mayo-Bean et al., 2011.) This limitation is more
biological than model domain dependant and does not appear to
be an issue with these substances, so for this exercise they were
considered to be within the domain of the model.

ii. Structural fragment domain: The EcoSAR models are chemical

class-based, therefore the chemical structure domain as described
under Section 3.1 defines the chemical domain of the QSAR. The
model selected for each validation chemical agrees with the rules
presented in the documentation and a scan of the structures,
except the description provided for oxime carbamates appears to
not include terminal -ON=C fragment. Both oxime carbamates in
the validation set were identified by EcoSAR as oxime carbamates,
so it could be the documentation is not capturing the fragments
properly.

iii. Mechanism domain: All of the validation and training set

chemicals listed in Attachments 1-4 are known to inhibit
acetylcholinesterase, and documentation included in EcoSAR for
Phenyl carbamate esters, oxime carbamate esters confirms this as
well. The documentation for the other EcoSAR QSAR models does
not mention specific modes of action associated with the chemical
class.

b. Structural analogues: Training set structural analogues are presented in

Attachments 1.

c. Considerations on structural analogues:

i. Structural requirements for carbamates (Fukuto 1990):

X: Leaving group; typically aryloxy or oxime (i.e., -ONR); when

X is a phenyl group, activity increases with 3 substitution on
the ring from hydrogen, methyl, ethyl, isopropyl. Having a
quarternary ammonium ion in the 3 position on the ring has a

Page 140 of 186

 maximum activity. Tert-butyl is less active than the isopropyl
form.

ii. Structural requirements for organophosphates (Fukuto 1990):

R: Typically a methyl or ethyl group. Many times RO = R1.

When R1=RO, and R is either a propyl or isopropyl group
acetylcholinesterase activity is very low

R1: Methoxy, ethoxy, ethyl, phenyl, amino, substituted amino,

alkylthio

X: Leaving group; typically phenoxy or aromatic group

containing hetero atoms, substituted thioalkyl, or substituted
alkoxy groups. When X contains a thioether group, these are
susceptible to metabolic activation to sulfoxides which are
metabolized to sulfones, making these more active AChE
inhibitors.

O(S): Direct acting organophosphates have the oxon group;

thiophosphates require metabolic activation via mixed
function oxidases to the oxon prior to inhibition of AChE
activity. Thiophosphates are less reactive and are more stable
to hydrolytic degradation than the oxon form.

3.3 The uncertainty of the prediction (OECD principle 4)

a. Data sources: The model training set was not generated from one
laboratory, therefore the data sets may have variability amongst the data
related to differences in genetic stock for fish, consistency in application of
test protocols such as static vs flow through exposures; use of analytical
procedures to measure test concentration, or reporting of only nominal
values, etc.

b. Test Species: As mentioned earlier these models are generic for fish, as
can be seen in Figure 1, fish species can have a range of sensitivities to these
substances.

Page 141 of 186

 c. Model Predictivity: The coefficient of correlation between predicted and
empirical test data is R=0.485. As presented in Figure 3, most chemicals are
estimated within an order of magnitude. Predictivity could be improved by
the addition of new chemical categories and expanded sub-structure rules
for existing chemical categories, as further described under Section 4.0.

Figure 3: Comparison of EcoSAR QSAR Estimates of the Fish Acute LC50

To Empirical Test Data for Organophosphates and Carbamates not in the Model
Training Sets

-1

-2
EcoSAR Estimated Log Molar LC50 for Fish

-3

Chloropyrifos-methyl oxon Naled

-4
Methidathion

-5

-6 Azinophos-methyl

-7

-8

-9
-9.00 -8.00 -7.00 -6.00 -5.00 -4.00 -3.00 -2.00 -1.00
Empirically Derived Log Molar LC50 for Fish
Phenyl Carbamate Ester Oxime Carbamate Ester Phosphate Ester
1 Dithiophosphate Ester Monothiophosphate Ester

3.4 The chemical and biological mechanisms according to the model underpinning
the predicted result (OECD principle 5).

Organophosphate and carbamate pesticides are indirect inhibitors of

cholinesterase caused by blocking the site (serine hydroxyl group) on
acetylcholinesterase where the neurotransmitter, acetylcholine, would normally
attach, thereby blocking the breakdown of acetylcholine (Fukuto 1990; Mileson
1998). Acetylcholine is a major neurotransmitter in the autonomic and
cholinergic (CNS) nervous system. Inhibition of acetylcholine binding at the
serine site results in a build-up of acetylcholine in synapses, resulting in
overstimulation of muscles, glands, and CNS. Toxic effects associated with
exposure to acetylcholine inhibiting chemicals include muscle contraction and
secretion, cholinergic hyperactivity, lack of muscle coordination, and respiratory
depression. A rate limiting step for the thion (i.e., (RO-)3P=S rather than (RO-

Page 142 of 186

 )3P=O) forms of organophosphates is the oxidative desulfuration to create the
oxon form which is required for the OP to bind to the AChE enzyme (Fukuto
1990). Documentation within EcoSAR identified the phenyl carbamate ester, the
phosphate ester, and the oxime carbamate ester models as associated with an
acetylcholinesterase mode of action. Although the monothiophosphate esters
and dithiophosphate esters are not explicitly identified as associated with
acetylcholinesterase inhibition, the structural requirements as outlined in the
EcoSAR documentation align with structure requirements for
acetylcholinesterase inhibitors (Fukuto 1990.)

4. Adequacy

Scientific validity of the EcoSAR model used to estimate the toxicity of chemicals listed in
Attachment 2, based on the adherence to the OECD principles as outlined above, finds
that for carbamates the method appears to be valid, but issues exist for some of the
organophosphates. Both carbamate models provided acceptable estimates of toxicity
for the five substances in the validation set, with all chemicals estimated within an order
of magnitude of the empirical test value. Comparison of the EcoSAR QSAR estimates for
the freshwater fish acute LC50 to empirical data sets for the organophosphate chemicals
found that most values agreed within an order of magnitude (see Figure 3). One
significant outlier was chlorpyrifos-methyl oxon, the metabolically active form of
chlorpyrifos-methyl. EcoSAR categorized the degradate as a phosphate ester, and using
this model, estimated the toxicity to be nearly 4000 times less toxic than found in
empirical studies. Methidathion and azinphos-methyl were more than 2 orders of
magnitude more toxic than estimated by the EcoSAR dithiophosphate ester model. The
QSAR estimate for naled was more than an order of magnitude different than empirical
test data. Once again, this substance was estimated using the phosphate ester model.
The reliability of these models for use with organophosphates depends on the
development of new or refined chemical class to estimate toxicity more effectively.

Page 143 of 186

 Azinphos-methyl Methidathion

Log P is the only molecular descriptor used in the EcoSAR models evaluated in this case
study. The coefficient of determination for the EcoSAR models evaluated herein (see
Table 1 for R2 values) reflects that factors other than partitioning into the organism are
required to completely describe the toxic response. To this end, these QSAR models
would improve by the identification and inclusion of toxicologically relevant molecular
descriptors in the EcoSAR QSAR models, with linkages to key events within the
acetylcholinesterase inhibition adverse outcome pathway.

Risk assessors will need to determine how relevant these model outputs are to the
question being asked. For instance, are estimations within an order of magnitude
acceptable? What statistical criteria must be met for the internal performance of a
model to be considered acceptable (i.e., R2 > 0.60)? Similarly, what statistical
benchmarks should be used to determine the predictivity of a model? The answer to
these questions may depend on whether the risk assessment is for a screening and
prioritization, or for deriving a final benchmark value. The purpose of this analysis was
to determine whether QSAR estimates could fill data gaps used in deriving Agency
benchmark values, and it is recommended that the models undergo further refinement
prior to this use.

References:

American Society for Testing and Materials (ASTM). 2007. Standard Guide for
Conducting Acute Toxicity Tests on Test Materials with Fishes, Macroinvertebrates, and
Amphibians. ASTM Subcommittee E46.01; E729-96(2007). 22 pp. American Society for
Testing and Materials, West Conshohocken, PA.

Fukuto, T.R. 1990. Mechanism of action of organophosphorus and carbamate

insecticides. Environ. Health Perspect. 87:245-254.

Mayo-Bean, K., J.V. Nabholz, R. Clements, M. Zeeman,, T. Henry, D. Rodier, K. Moran, B.

Meylan, and P. Ranslow. 2011. Methodology Document for the ECOlogical Structure-

Page 144 of 186

Activity Relationship Model (ECOSAR) Class Program: Estimating Toxicity of Industrial
Chemicals to Aquatic Organisms using ECOSAR Class Program (Version 1.1). U.S.
Environmental Protection Agency, Office of Chemical Safety and Pollution Prevention,
Office of Pollution Prevention and Toxics, Washington DC.

Mileson, B.E., J.E. Chambers, W.L. Chen, W. Dettbarn, M. Ehrich, A.T. Eldefrawi, D.W.
Gaylor, K. Hamernick, E. Hodgson, A.G. Karczmar, S. Padilla, C.N. Pope, R.J. Richardson,
D.R. Saunders, L.P. Sheets, L.G. Sultatos, and K.B. Wallace. 1998. Common mechanism
of toxicity: A case study of organophosphorus pesticides. Tox. Sci. 41:8-20.

Moore, D.R.J., R.L. Breton, and D.B. MacDonald., 2003. A comparison of model
performance for six quantitative structure–activity relationship packages that predict
acute toxicity to fish. Environ. Toxicol. Chem. 22, 1799–1809.

Reuschenbach P, Silvani M, Dammann M, Warnecke D, Knacker T. 2008. ECOSAR model

performance with a large test set of industrial chemicals. Chemosphere 71:1986–1995.

U.S. EPA. 1996. Ecological Effects Test Guidelines: OPPTS 850.1075. Fish acute toxicity
test, freshwater and marine. U.S. EPA, Office of Prevention, Pesticides, and Toxic
Substances, EPA 712-C-96-118. Public Draft; 13 pp.

U.S. EPA. 1985. Guidelines for deriving numerical national water quality criteria for the
protection of aquatic organisms and their uses. U.S. Environmental Protection Agency,
Office of Research and Development, Environmental Research Laboratories. PB85-
227049.

Page 145 of 186

 Attachment 1: Carbamate and organophosphate pesticides structures and CAS Registry Numbers.

Greyed cells are substances that were not included in the final evaluation of model performance because the chemical was
part of the EcoSAR model training set.

CAS Empirical EcoSAR

Pesticide Registry Fish 96 hr Training Chemical Class: Specific Structure: SMILES
Number LC50 data Set

Carbamates (N=17 total, with N=12 included in the EcoSAR model training set, N=5 used in model validation.)

Bendiocarb 22781-23-3 Yes No Carbamate CNC(=O)Oc1cccc2OC(C)(C)Oc12

cis-Thiocarboxime 29118-87-4 Yes No Oxime carbamate CNC(=O)ON=C(C)SCCC(#N)

endo-3-Chloro-exo-6-
cyano-2-
norbornanone, o- 15271-41-7 Yes No Oxime carbamate CNC(=O)ON=C1C(C2)C(C#N)CC2C1Cl
(Methylcarbamoyl)
oxime

Formetanate
23422-53-9 Yes No Formamidine CNC(=O)Oc1cccc(N=CN(C)C)c1
hydrochloride

Pirimicarb 23103-98-2 Yes No Dimethylcarbamate CN(C)C(=O)Oc1nc(nc(C)c1C)N(C)C

3,4,5-
Trimethylphenyl 2686-99-9 Yes Yes Phenyl methylcarbamate CNC(=O)Oc1cc(C)c(C)c(C)c1
methylcarbamate

Aldicarb 116-06-3 Yes Yes Oxime carbamate CNC(=O)ON=CC(C)(C)SC

Aldoxycarb 1646-88-4 Yes Yes Oxime carbamate CNC(=O)ON=CC(C)(C)S(C)(=O)=O

Aminocarb 2032-59-9 Yes Yes Phenyl methylcarbamate CNC(=O)Oc1ccc(N(C)C)c(C)c1

Carbaryl 63-25-2 Yes Yes Carbamate CNC(=O)Oc1cccc2ccccc12

Benzofuranyl
Carbofuran 1563-66-2 Yes Yes CNC(=O)Oc1cccc2CC(C)(C)Oc12
methylcarbamate

Methiocarb 2032-65-7 Yes Yes Phenyl methylcarbamate CNC(=O)Oc1cc(C)c(SC)c(C)c1

Methomyl 16752-77-5 Yes Yes Oxime carbamate CNC(=O)ON=C(C)SC

Page 146 of 186

 Attachment 1: Carbamate and organophosphate pesticides structures and CAS Registry Numbers.

Greyed cells are substances that were not included in the final evaluation of model performance because the chemical was
part of the EcoSAR model training set.

CAS Empirical EcoSAR

Pesticide Registry Fish 96 hr Training Chemical Class: Specific Structure: SMILES
Number LC50 data Set

Mexacarbate 315-18-4 Yes Yes Phenyl methylcarbamate CNC(=O)Oc1cc(C)c(N(C)C)c(C)c1

Oxamyl 23135-22-0 Yes Yes Oxime carbamate CNC(=O)ON=C(SC)C(=O)N(C)C

Propoxur 114-26-1 Yes Yes Phenyl methylcarbamate CNC(=O)Oc1ccccc1OC(C)C

Thiodicarb 59669-26-0 Yes Yes Oxime carbamate CSC(C)=NOC(=O)N(C)SN(C)C(=O)ON=C(C)SC

Organophosphates (N=47 total, with N=28 included in the EcoSAR model training set, N=19 used in model
validation.)

Acephate 30560-19-1 Yes No Phosphoramidothioate COP(=O)(NC(C)=O)SC

Benzotriazine
Azinphos-methyl 86-50-0 Yes No S=P(OC)(OC)SCN1N=Nc2ccccc2C1(=O)
organothiophosphate

Carbophenothion 786-19-6 Yes Yes Phenyl organothiophosphate CCOP(=S)(OCC)SCSc1ccc(Cl)cc1

Chlorfenvinphos 470-90-6 Yes No Organophosphate CCOP(=O)(OCC)OC(=CCl)c1ccc(Cl)cc1Cl

Chlorpyrifos 2921-88-2 Yes Yes Pyridine organothiophosphate CCOP(=S)(OCC)Oc1nc(Cl)c(Cl)cc1Cl

Chlorpyrifos-methyl
5598-52-7 Yes No Degradate — Oxon O=P(OC)(OC)Oc1nc(Cl)cc(Cl)c1Cl
oxon

Crufomate 299-86-5 Yes No Phosphoramidate CNP(=O)(OC)Oc1ccc(cc1Cl)C(C)(C)C

Dichlofenthion 97-17-6 Yes No Phenyl organothiophosphate CCOP(=S)(OCC)Oc1ccc(Cl)cc1Cl

Dicrotophos 141-66-2 Yes No Organophosphate COP(=O)(OC)OC(C)=CC(=O)N(C)C

Heterocyclic
Dioxathion 78-34-2 Yes No CCOP(=S)(OCC)SC1OCCOC1SP(=S)(OCC)OCC
organothiophosphate

Fenchlorphos 299-84-3 Yes No Phenyl organothiophosphate COP(=S)(OC)Oc1cc(Cl)c(Cl)cc1Cl

Fenitrothion 122-14-5 Yes No Phenyl organothiophosphate COP(=S)(OC)Oc1ccc(N(=O)=O)c(C)c1

Heterocyclic
Fosthiazate 98886-44-3 Yes No CCOP(=O)(SC(C)CC)N1CCSC1=O
organothiophosphate

Page 147 of 186

 Attachment 1: Carbamate and organophosphate pesticides structures and CAS Registry Numbers.

Greyed cells are substances that were not included in the final evaluation of model performance because the chemical was
part of the EcoSAR model training set.

CAS Empirical EcoSAR

Pesticide Registry Fish 96 hr Training Chemical Class: Specific Structure: SMILES
Number LC50 data Set

Iodofenphos 18181-70-9 Yes No Phenyl organothiophosphate COP(=S)(OC)Oc1cc(Cl)c(I)cc1Cl

Thiadiazole
Methidathion 950-37-8 Yes No COc1nn(CSP(=S)(OC)OC)c(=O)s1
organothiophosphate

Methyl
953-17-3 Yes No Phenyl organothiophosphate COP(=S)(OC)SCSc1ccc(Cl)cc1
carbophenothion

Monocrotophos 6923-22-4 Yes No Organophosphate CNC(=O)C=C(C)OP(=O)(OC)OC

Naled 300-76-5 Yes No Organophosphate COP(=O)(OC)OC(Br)C(Cl)(Cl)Br

SD-7438 2782-70-9 Yes No Organothiophosphate COP(=S)(OC)SCc1ccc(SP(=S)(OC)OC)cc1

Sulfotep 3689-24-5 Yes No Aliphatic organothiophosphate CCOP(=S)(OCC)OP(=S)(OCC)OCC

Trichloronate 327-98-0 Yes No Phenyl ethylphosphonothioate CCOP(=S)(CC)Oc1cc(Cl)c(Cl)cc1Cl

Chlorpyrifos-methyl 5598-13-0 Yes Yes Pyridine organothiophosphate COP(=S)(OC)Oc1nc(Cl)c(Cl)cc1Cl

Heterocyclic
Coumaphos 56-72-4 Yes Yes S=P(OCC)(OCC)Oc1ccc2C(C)=C(Cl)C(=O)Oc2c1
organothiophosphate

S=P(OCC)(OCC)OCCSCC +
Demeton 8065-48-3 Yes Yes Aliphatic organothiophosphate
O=P(OCC)(OCC)OCCSCC

Pyrimidine
Diazinon 333-41-5 Yes Yes CCOP(=S)(OCC)Oc1cc(C)nc(n1)C(C)C
organothiophosphate

Dichlorvos/DDVP 62-73-7 Yes Yes Organophosphate COP(=O)(OC)OC=C(Cl)Cl

Aliphatic amide
Dimethoate 60-51-5 Yes Yes CNC(=O)CSP(=S)(OC)OC
organothiophosphate

Disulfoton 298-04-4 Yes Yes Aliphatic organothiophosphate CCOP(=S)(OCC)SCCSCC

Phenyl
EPN 2104-64-5 Yes Yes CCOP(=S)(Oc1ccc(cc1)N(=O)=O)c2ccccc2
phenylphosphonothioate

Ethion 563-12-2 Yes Yes Aliphatic organothiophosphate CCOP(=S)(OCC)SCSP(=S)(OCC)OCC

Page 148 of 186

 Attachment 1: Carbamate and organophosphate pesticides structures and CAS Registry Numbers.

Greyed cells are substances that were not included in the final evaluation of model performance because the chemical was
part of the EcoSAR model training set.

CAS Empirical EcoSAR

Pesticide Registry Fish 96 hr Training Chemical Class: Specific Structure: SMILES
Number LC50 data Set

Ethoprophos 13194-48-4 Yes Yes Aliphatic organothiophosphate CCCSP(=O)(OCC)SCCC

Fensulfothion 115-90-2 Yes Yes Phenyl organothiophosphate CCOP(=S)(OCC)Oc1ccc(cc1)S(C)=O

Fenthion 55-38-9 Yes Yes Phenyl organothiophosphate COP(=S)(OC)Oc1ccc(SC)c(C)c1

Fonofos 944-22-9 Yes Yes Phenyl ethylphosphonothioate CCOP(=S)(CC)Sc1ccccc1

Isazofos 42509-80-8 Yes Yes Triazole organothiophosphate CCOP(=S)(OCC)Oc1nc(Cl)n(n1)C(C)C

Isofenphos 25311-71-1 Yes Yes Phosphoramidothioate CCOP(=S)(NC(C)C)Oc1ccccc1C(=O)OC(C)C

Malathion 121-75-5 Yes Yes Aliphatic organothiophosphate CCOC(=O)CC(SP(=S)(OC)OC)C(=O)OCC

Oxydemeton-methyl 301-12-2 Yes Yes Aliphatic organothiophosphate CCS(=O)CCSP(=O)(OC)OC

Parathion 56-38-2 Yes Yes Phenyl organothiophosphate CCOP(=S)(OCC)Oc1ccc(cc1)N(=O)=O

Parathion-methyl 298-00-0 Yes Yes Phenyl organothiophosphate S=P(OC)(OC)O-c(ccc1N(=O)=O)cc1

Phorate 298-02-2 Yes Yes Aliphatic organothiophosphate CCOP(=S)(OCC)SCSCC

Isoindole
Phosmet 732-11-6 Yes Yes COP(=S)(OC)SCN2C(=O)c1ccccc1C2=O
organothiophosphate

Profenofos 41198-08-7 Yes Yes Phenyl organothiophosphate CCCSP(=O)(OCC)Oc1ccc(Br)cc1Cl

Propetamphos 31218-83-4 Yes Yes Phosphoramidothioate CCNP(=S)(OC)OC(C)=CC(=O)OC(C)C

Pyrimidine
Tebupirimfos 96182-53-5 Yes Yes CCOP(=S)(OC(C)C)Oc1cnc(nc1)C(C)(C)C
organothiophosphate

COP(=S)(OC)Oc2ccc(Sc1ccc(OP(=S)(OC)OC)cc1)
Temephos 3383-96-8 Yes Yes Phenyl organothiophosphate cc2

Trichlorfon 52-68-6 Yes Yes Phosphonate COP(=O)(OC)C(O)C(Cl)(Cl)Cl

Page 149 of 186

Attachment 2: Empirical data set of acceptable quality data for acetylcholinesterase inhibitors for use in
comparison to QSAR model estimates. The ECOSAR Version 1.1 tool included generic freshwater fish
QSAR models. For this exercise, fish model estimates were compared to either rainbow trout
(Oncorhynchus mykiss) or bluegill (Lepomis macrochirus) empirical data.

Minimum Maximum Average Number of

CAS Registry
Pesticide Species Toxicity Toxicity Concentration Toxicity
Number
(ug/L) (ug/L) (ug/L) Values

3,4,5-Trimethylphenyl
2686999 Oncorhynchus mykiss 4700 4700 4700 1
methylcarbamate

30560191 Acephate Oncorhynchus mykiss 110000 110000 110000 1

116063 Aldicarb Oncorhynchus mykiss 560 560 560 1

1646884 Aldoxycarb Oncorhynchus mykiss 42000 42000 42000 1

2032599 Aminocarb Oncorhynchus mykiss 12000 25000 18314.29 7

86500 Azinphos-methyl Oncorhynchus mykiss 4.3 6.3 5.3 2

22781233 Bendiocarb Oncorhynchus mykiss 1200 1200 1200 1

741582 bensulide Oncorhynchus mykiss 720 1400 1073.33 3

63252 Carbaryl Oncorhynchus mykiss 780 3500 1796.63 16

1563662 Carbofuran Oncorhynchus mykiss 380 600 466.67 3

786196 Carbophenothion Lepomis macrochirus 13 13 13 1

470906 Chlorfenvinphos Oncorhynchus mykiss 510 510 510 1

2921882 Chlorpyrifos Oncorhynchus mykiss 7.1 25 13.37 3

5598130 Chlorpyrifos-methyl Oncorhynchus mykiss 120 301 210.5 2

5598527 Chlorpyrifos-methyl oxon Oncorhynchus mykiss 1.7 2 1.85 2

29118874 cis-Thiocarboxime Oncorhynchus mykiss 1500 1500 1500 1

56724 Coumaphos Oncorhynchus mykiss 890 890 890 1

299865 Crufomate Lepomis macrochirus 1800 1800 1800 1

Page 150 of 186

Attachment 2: Empirical data set of acceptable quality data for acetylcholinesterase inhibitors for use in
comparison to QSAR model estimates. The ECOSAR Version 1.1 tool included generic freshwater fish
QSAR models. For this exercise, fish model estimates were compared to either rainbow trout
(Oncorhynchus mykiss) or bluegill (Lepomis macrochirus) empirical data.

Minimum Maximum Average Number of

CAS Registry
Pesticide Species Toxicity Toxicity Concentration Toxicity
Number
(ug/L) (ug/L) (ug/L) Values

8065483 Demeton Oncorhynchus mykiss 520 600 560 2

333415 Diazinon Lepomis macrochirus 136 460 254.67 3

97176 Dichlofenthion Oncorhynchus mykiss 1250 1250 1250 1

62737 Dichlorvos/DDVP Oncorhynchus mykiss 100 100 100 1

141662 Dicrotophos Oncorhynchus mykiss 6300 6300 6300 1

60515 Dimethoate Oncorhynchus mykiss 6200 8600 7433.33 3

78342 Dioxathion Oncorhynchus mykiss 69 69 69 1

298044 Disulfoton Oncorhynchus mykiss 1850 1850 1850 1

endo-3-Chloro-exo-6-cyano-2-
15271417 norbornanone, o- Oncorhynchus mykiss 13000 13000 13000 1
(Methylcarbamoyl) oxime

2104645 EPN Oncorhynchus mykiss 210 210 210 1

563122 Ethion Oncorhynchus mykiss 500 500 500 1

13194484 Ethoprophos Oncorhynchus mykiss 1150 1150 1150 1

299843 Fenchlorphos Oncorhynchus mykiss 550 645 597.5 2

122145 Fenitrothion Oncorhynchus mykiss 1000 2700 2050 7

115902 Fensulfothion Lepomis macrochirus 72 72 72 1

55389 Fenthion Oncorhynchus mykiss 550 840 740 3

944229 Fonofos Oncorhynchus mykiss 20 50 35 2

23422539 Formetanate hydrochloride Oncorhynchus mykiss 4400 4400 4400 1

98886443 Fosthiazate Oncorhynchus mykiss 111000 111000 111000 1

Page 151 of 186

Attachment 2: Empirical data set of acceptable quality data for acetylcholinesterase inhibitors for use in
comparison to QSAR model estimates. The ECOSAR Version 1.1 tool included generic freshwater fish
QSAR models. For this exercise, fish model estimates were compared to either rainbow trout
(Oncorhynchus mykiss) or bluegill (Lepomis macrochirus) empirical data.

Minimum Maximum Average Number of

CAS Registry
Pesticide Species Toxicity Toxicity Concentration Toxicity
Number
(ug/L) (ug/L) (ug/L) Values

18181709 Iodofenphos Oncorhynchus mykiss 16.2 16.2 16.2 1

42509808 Isazofos Oncorhynchus mykiss 18.7 18.7 18.7 1

25311711 Isofenphos Lepomis macrochirus 1400 1400 1400 1

121755 Malathion Oncorhynchus mykiss 30 200 105.6 5

950378 Methidathion Oncorhynchus mykiss 10 14 12 2

2032657 Methiocarb Oncorhynchus mykiss 436 750 593 2

16752775 Methomyl Oncorhynchus mykiss 1050 1600 1308.33 6

953173 Methyl carbophenothion Oncorhynchus mykiss 760 760 760 1

315184 Mexacarbate Oncorhynchus mykiss 4450 15000 9887.5 4

6923224 Monocrotophos Lepomis macrochirus 12100 12100 12100 1

300765 Naled Oncorhynchus mykiss 132 195 167.33 3

23135220 Oxamyl Oncorhynchus mykiss 4200 4700 4450 2

301122 Oxydemeton-methyl Oncorhynchus mykiss 730 730 730 1

56382 Parathion Oncorhynchus mykiss 864 1430 1214.67 3

298000 Parathion-methyl Oncorhynchus mykiss 2750 3700 3225 2

298022 Phorate Oncorhynchus mykiss 13 21 17 2

732116 Phosmet Oncorhynchus mykiss 105 4700 1142.08 12

23103982 Pirimicarb Oncorhynchus mykiss 79000 79000 79000 1

41198087 Profenofos Oncorhynchus mykiss 21 23.5 22.25 2

31218834 Propetamphos Oncorhynchus mykiss 940 2600 1770 2

Page 152 of 186

Attachment 2: Empirical data set of acceptable quality data for acetylcholinesterase inhibitors for use in
comparison to QSAR model estimates. The ECOSAR Version 1.1 tool included generic freshwater fish
QSAR models. For this exercise, fish model estimates were compared to either rainbow trout
(Oncorhynchus mykiss) or bluegill (Lepomis macrochirus) empirical data.

Minimum Maximum Average Number of

CAS Registry
Pesticide Species Toxicity Toxicity Concentration Toxicity
Number
(ug/L) (ug/L) (ug/L) Values

114261 Propoxur Oncorhynchus mykiss 3700 3700 3700 1

2782709 SD-7438 Oncorhynchus mykiss 34 34 34 1

3689245 Sulfotep Oncorhynchus mykiss 1000 1000 1000 1

96182535 Tebupirimfos Oncorhynchus mykiss 2220 2220 2220 1

3383968 Temephos Oncorhynchus mykiss 3490 6800 5145 2

59669260 Thiodicarb Oncorhynchus mykiss 2650 2650 2650 1

52686 Trichlorfon Oncorhynchus mykiss 370 8800 1882.57 28

327980 Trichloronate Oncorhynchus mykiss 140 140 140 1

Page 153 of 186

 Attachment 3: Comparison of ECOSAR model (Version 1.1) estimates for fish to the average LC50 value (mg/L) from
acceptable empirical data collection (See Supplemental Information Table 2 for details related to empirical test
data.) ECOSAR Class Model is the QSAR chemical class-based equation used to estimate toxicity for this exercise.

ECOSAR Version 1.1 Model Estimates and Average Empirical Test Data — Fish

Average of Ratio of
Version 1.1
Version Acceptable ECOSAR
CAS General Version 1.1 — ECOSAR Fish-
1.1 — Empirical Estimated
Registry Pesticide 96 hr LC50
KowWin Fish 96 hr value by
Number Chemical Class ECOSAR Class Model Estimate
LogP LC50 values Empirical
(mg/L)
(mg/L) Test data

Carbamate Validation Data Set: Chemical was NOT part of EcoSAR Model Training Set

22781233 Bendiocarb Carbamate Carbamate esters, phenyl 2.552 3.521 1.20 2.93

29118874 cis-Thiocarboxime Carbamate Oxime Carbamate Ester 0.123 3.684 1.50 2.46

endo-3-Chloro-exo-6-
cyano-2-norbornanone,
15271417 Carbamate Oxime Carbamate Ester 1.089 1.797 13.00 0.14
o-(Methylcarbamoyl)
oxime

Formetanate
23422539 Carbamate Carbamate esters, phenyl 0.879 13.318 4.40 3.03
hydrochloride

23103982 Pirimicarb Carbamate Carbamate esters, phenyl 1.399 9.456 79.00 0.12

Organophosphate Validation Data Set: Chemical was NOT part of EcoSAR Model Training Set

30560191 Acephate Organophosphate Esters (phosphate) -0.902 41.359 110.00 0.38

86500 Azinphos-methyl Organophosphate Esters, Dithiophosphates 2.532 1.01 0.01 190.57

470906 Chlorfenvinphos Organophosphate Esters (phosphate) 4.146 1.382 0.51 2.71

5598527 Chlorpyrifos-methyl oxon Organophosphate Esters (phosphate) 1.911 7.151 0.00 3865.41

299865 Crufomate Organophosphate Esters (phosphate) 3.299 2.22 1.80 1.23

97176 Dichlofenthion Organophosphate Esters, Monothiophosphates 5.202 0.046 1.25 0.04

141662 Dicrotophos Organophosphate Esters (phosphate) -0.096 27.934 6.30 4.43

78342 Dioxathion Organophosphate Esters, Dithiophosphates 3.446 0.509 0.07 7.38

299843 Fenchlorphos Organophosphate Esters, Monothiophosphates 4.865 0.075 0.60 0.13

122145 Fenitrothion Organophosphate Esters, Monothiophosphates 3.296 0.544 2.05 0.27

98886443 Fosthiazate Organophosphate Esters (phosphate) 2.471 4.207 111.00 0.04

Page 154 of 186

 Attachment 3: Comparison of ECOSAR model (Version 1.1) estimates for fish to the average LC50 value (mg/L) from
acceptable empirical data collection (See Supplemental Information Table 2 for details related to empirical test
data.) ECOSAR Class Model is the QSAR chemical class-based equation used to estimate toxicity for this exercise.

ECOSAR Version 1.1 Model Estimates and Average Empirical Test Data — Fish

Average of Ratio of
Version 1.1
Version Acceptable ECOSAR
CAS General Version 1.1 — ECOSAR Fish-
1.1 — Empirical Estimated
Registry Pesticide 96 hr LC50
KowWin Fish 96 hr value by
Number Chemical Class ECOSAR Class Model Estimate
LogP LC50 values Empirical
(mg/L)
(mg/L) Test data

18181709 Iodofenphos Organophosphate Esters, Monothiophosphates 5.387 0.047 0.02 2.90

950378 Methidathion Organophosphate Esters, Dithiophosphates 1.584 2.851 0.01 237.58

953173 Methyl carbophenothion Organophosphate Esters, Dithiophosphates 4.463 0.109 0.76 0.14

6923224 Monocrotophos Organophosphate Esters (phosphate) -0.307 31.165 12.10 2.58

300765 Naled Organophosphate Esters (phosphate) 1.605 11.374 # 0.17 67.97

2782709 SD-7438 Organophosphate Esters, Dithiophosphates 4.545 0.128 0.03 3.76

3689245 Sulfotep Organophosphate Esters, Monothiophosphates 3.98 0.25 1.00 0.25

327980 Trichloronate Organophosphate Esters, Monothiophosphates 5.863 0.02 0.14 0.14

Page 155 of 186

Attachment 4: Structures for chemicals in the Validation set (not in the EcoSAR training set for the models and
high quality empirical test data were available to use in model validation). Structures are from Alan Wood
website (www.alanwood.net) with the exception of endo-3-Chloro-exo-6-cyano-2-norbornanone, o-
(Methylcarbamoyl) oxime , Chlorpyrifos-methyl oxon, and SD-7438 which are from ChemIDPlus
(http://chem.sis.nlm.nih.gov/chemidplus/).

Bendiocarb (CAS: 22781-23-3) cis-Thiocarboxime (CAS: 29118-87-4)

Carbamates that were not

part of the EcoSAR Training
Set

endo-3-Chloro-exo-6-cyano-2- Pirimicarb (CAS: 23103-98-2) Formetanate hydrochloride (CAS:

norbornanone, o- 23422-53-9)
(Methylcarbamoyl) oxime
(CAS: 15271-41-7)

Page 156 of 186

Attachment 4: Structures for chemicals in the Validation set (not in the EcoSAR training set for the models and
high quality empirical test data were available to use in model validation). Structures are from Alan Wood
website (www.alanwood.net) with the exception of endo-3-Chloro-exo-6-cyano-2-norbornanone, o-
(Methylcarbamoyl) oxime , Chlorpyrifos-methyl oxon, and SD-7438 which are from ChemIDPlus
(http://chem.sis.nlm.nih.gov/chemidplus/).

Acephate (CAS: 30560-19-1) Azinphos-methyl (CAS: 86-50-0)

Organophosphates that were

not part of the EcoSAR
Training Set

Chlorfenvinphos Chlorpyrifos-methyl oxon Crufomate (CAS: 299-86-5)

(CAS: 470-90-6) (CAS: 5598-52-7)

Page 157 of 186

Attachment 4: Structures for chemicals in the Validation set (not in the EcoSAR training set for the models and
high quality empirical test data were available to use in model validation). Structures are from Alan Wood
website (www.alanwood.net) with the exception of endo-3-Chloro-exo-6-cyano-2-norbornanone, o-
(Methylcarbamoyl) oxime , Chlorpyrifos-methyl oxon, and SD-7438 which are from ChemIDPlus
(http://chem.sis.nlm.nih.gov/chemidplus/).

Dichlofenthion (CAS: 97-17-6 Dicrotophos (CAS: 141-66-2) Dioxathion (CAS: 78-34-2)

Fenchlorphos (CAS: 299-84-3) Fenitrothion (CAS: 122-14-5) Fosthiazate (CAS: 98886-44-3)

Iodofenphos (CAS: 18181-70-9) Methidathion (CAS: 950-37-8) Methyl carbophenothion

(CAS: 953-17-3)

Page 158 of 186

Attachment 4: Structures for chemicals in the Validation set (not in the EcoSAR training set for the models and
high quality empirical test data were available to use in model validation). Structures are from Alan Wood
website (www.alanwood.net) with the exception of endo-3-Chloro-exo-6-cyano-2-norbornanone, o-
(Methylcarbamoyl) oxime , Chlorpyrifos-methyl oxon, and SD-7438 which are from ChemIDPlus
(http://chem.sis.nlm.nih.gov/chemidplus/).

Monocrotophos (CAS: 6923-22-4) Naled (CAS: 300-76-5) SD-7438 (CAS: 2782-70-9)

Sulfotep (CAS: 3689-24-5) Trichloronate (CAS: 327-98-0)

Page 159 of 186

 Example No. 2

The following case study was prepared by the Office of Pesticide Programs
(OPP), US EPA.

Case Study:

Use of Analog Data to Determine Whether Additional Data Should be

Required for a Pesticide Degradate

In pesticide risk assessment there is often an abundance of toxicity data on the parent active
ingredient and very little, if any, data on pesticide metabolites or environmental degradation
products. This can be a problem in trying to assess the risks of metabolites or environmental
degradates. In the case of environmental degradates, a screening level risk assessment may be
performed to determine if additional toxicity data on the degradate should be called in. The
hazard component of the screening level assessment is often based on structural analogy of a
degradate to the parent active ingredient. If parent and degradate are closely related
structurally then toxicity data on the parent ai can be used to estimate the toxicity of the
degradate. If the margin of exposure between estimated toxicity and estimated exposure is not
considered large enough, additional toxicity data may be called in to enable a more
comprehensive risk assessment of a degradate. On occasion, the metabolite or degradate bears
little resemblance to the parent and an alternative analog with associated data must be found to
support the screening level risk assessment.

The herbicide dichlobenil is relatively stable in the environment except for aqueous photolysis.
A major photodegradate of dichlobenil in water (up to 19% of applied dichlobenil) has been
identified as 4-chloro-2(3H)benzoxazolone (BZZ). This photodegradate bears little resemblance
to the parent dichlobenil and therefore toxicity data on the parent are not considered useful for
assessing the toxicity of BZZ. In the absence of appropriate toxicity data the degradate, termed
BZZ, was determined to be of potential concern.

A number of online sources of information were consulted in search of appropriate analogs to

BZZ including EPA/OPPT's Analog Identification Methodology (AIM) (http://aim.epa.gov/),
ChemSpider (http://www.chemspider.com/) and Chemicalize (http://www.chemicalize.org/). A
close analog (and isomer) of BZZ, 5-chloro-2(3H)benzoxazolone, was identified and this
particular analog has a wealth of health effects data since it is an often prescribed muscle
relaxant (common name chlorzoxazone).

Page 160 of 186

 Cl Cl
H H
CN N Cl N
O O
Cl O O
Dichlobenil BZZ Chlorzoxazone

Chlorzoxazone has estimated physical properties very close to BZZ (see table below) and so
bioavailability is likely to be similar. Chlorzoxazone is pharmacologically active as a muscle
relaxant at doses of 10-30 mg/kg/day (http://www.drugs.com/monograph/chlorzoxazone.html),
so the potential exists for BZZ to be biologically active at the same dosage, although no
assumption is made about the kind of effects that might be observed at this dose of BZZ.

Estimation of selected physical-chemical properties using Episuite v4.1

Property Dichloben Chlorzoxazo Benzoxazolone

il BZZ ne * Model
log P 1.68 1.59 1.59 0.95 KOWWIN v1.68
KOC (L/kg) 257 15.06 14.76 10 KOCWIN/MCI v2.00
water sol.
(mg/L) 36 361 361 1414 Water NT v1.01
pKa n/a 8.38 8.43 8.91 SPARC v4.6
Henry's Law Henry v3.20, bond
(atm-m3/mole) 2.86E-05 2.72E-08 2.72E-08 3.66E-08 estimation

*Benzoxazolone is the unsubstituted fused ring structure common to BZZ and chlorzoxazone

Screening Risk Assessment. The theoretical upper limit for BZZ based on estimated surface
water concentration of parent dichlobenil is 0.005 mg/L and the corresponding dosage in a
young child is approximately 0.0005 mg/kg/day for a 10 kg child ingesting 1 liter of BZZ
contaminated water per day. The MOE between the lowest effective pharmacological dose of
chlorzoxazone (10 mg/kg/day and the theoretical intake is 10 mg/kg/day divided by 0.0005
mg/kg/day or 20,000. Although there are many uncertainties in a screening level risk
assessment such as this, the MOE is sufficiently large to conclude that additional toxicity data
are unlikely to result in risks of concern from ingestion of BZZ formed in drinking water as a
result of registered uses of dichlobenil.

Page 161 of 186

 Example No. 3
The following example has been provided by the Office of Pollution Prevention
and Toxics (OPPT) and the Office of Pesticide Programs (OPP) at the US EPA. It
serves to illustrate the potential use of SAR analysis as one line of evidence to
support a mode of action analysis for a pesticide active ingredient.

Case Study: Fomesafen cancer assessment and mode of action: use of mechanism-
based SAR

Description of the case:

Fomesafen, a diphenyl ether herbicide, was submitted to OPP’s Cancer Assessment

Review Committee (CARC) for re-evaluation of its carcinogenic potential to humans.
The herbicide was previously shown to be a mouse hepatocarcinogen by the submitter
and classified as a Category C possible human carcinogen by OPP. The new data
provided by the submitter included: (a) consistent negative genotoxicity data, (b) some
evidence of involvement of peroxisome proliferator-activated receptor alpha receptor
(PPARα) as a possible nongenotoxic mode of action for carcinogenicity, and (c)
metabolism data. No SAR study was attempted. The Committee concurred that the
pesticide should be nongenotoxic but considered the PPARα evidence inadequate.

SAR approaches conducted:

Several structurally related diphenyl ether pesticides with carcinogenicity data were
identified. Among these, Nitrofen, Lactofen, Acifluorfen and Oxyfluorfen were
considered the closest. Like Fomesafen, all four were hepatocarcinogenic in mice with
Oxyfluorfen being weakly/marginally active. The chemical structures are shown in the
figure below.

Page 162 of 186

 Cl

Cl O NO2

Nitrofen

Cl Cl Cl

CF3 O NO2 CF3 O NO2 CF3 O NO2

CO-NH-SO2-CH3 COOH COOCH(CH3)COOC2H5

Fomesafen Acifluorfen Lactofen

Cl

CF3 O NO2

OC2H5
Oxyfluorfen

The presence of a nitro group in aromatic ring is generally considered a genotoxic

structural alert. Indeed, there was some evidence that Nitrofen was positive in the Ames
test but the evidence was complicated by the presence of impurities. In addition to mouse
liver tumors, there was some evidence that Nitrofen may induce pancreatic tumors in the
rat. The mode of action of Nitrofen has not been thoroughly studied.

The mode of action of rodent hepatocarcinogenesis for both Lactofen and Acifluorfen
(HED MTARC; TXR #s 0051907 and 0052006, respectively) has been extensively
studied and shown to involve PPARα-medicated peroxisome proliferation. Lactofen can
be readily hydrolyzed by esterases to yield Acifluorfen as its primary metabolite.
Structure-activity relationships studies have shown that one of the major structural
requirements/alerts of most peroxisome proliferators is the presence of an acidic
functional group (e.g., carboxylic, sulfonic) either in the parent compound or a metabolite
(Woo and Lai 2003). The key question is whether Fomesafen can be hydrolyzed to a
carboxylic acid metabolite. In general, the amide (-CO-NH-) bond is quite resistant to
enzymatic hydrolysis. However, in Fomesafen, the presence of a sulfonyl group adjacent
to the amide linkage can significantly facilitate hydrolysis. Indeed, a metabolism study
by the submitter showed that up to 10% of Fomesafen may be hydrolyzed to yield a
carboxylic acid metabolite as the most significant metabolite. Thus, Fomesafen,
Acifluorfen, and Lactofen may actually have common carboxylic acid metabolite(s). It is
interesting to note that, despite structural similarity, Oxyfluorfen, which cannot be
metabolized to a carboxylic acid metabolite, is only weakly/marginally active as a
hepatocarcinogen . Attempts to demonstrate possible PPARα-mediated activity were
unsuccessful for Oxyfluoren. Overall, these findings strengthen the biological

Page 163 of 186

plausibility of PPARα mode of action for Fomesafen-induced liver tumor formation in
mice.

Outcome of the SAR study:

The SAR study provided significant support to the weight of evidence of a PPARα mode
of action of Fomesafen-induced mouse liver tumors. Based on the current scientific
understanding of peroxisome proliferation (e.g., Klaunig et al., 2003) and previous EPA
decisions on structurally related herbicides (e.g., Lactofen and Acifluorfen), the level of
confidence in this assessment is high. While the proposed mode of action for liver
tumors in mice is theoretically plausible in humans, it is quantitatively implausible and
unlikely to take place in humans based on quantitative species differences in PPAR
activation and toxicokinetics. In accordance with the EPA Final Guidelines for
Carcinogen Risk Assessment (March 29, 2005), the CARC classified Fomesafen as “Not
Likely to be Carcinogenic to Humans”.

References:

Klaunig JE, Babich MA, Baetcke KP, Cook JC, Corton JC, David RM, DeLuca JG, Lai
DY, McKee RH, Peters JM, Roberts RA, Fenner-Crisp PA. (2003). PPARalpha agonist-
induced rodent tumors: modes of action and human relevance. Crit Rev Toxicol
33(6):655–780.

Woo, Y.T., and Lai, D.Y. (2003). Mechanism of action of chemical carcinogens and their
role in structure-activity relationships (SAR) analysis and risk assessment. In:
Quantitative Structure-Activity Relationship (QSAR) Models of Mutagens and
Carcinogens, R. Benigni, ed., CRC Press, p. 41.

Page 164 of 186

 Example No. 4
The following example has been abstracted from the Screening Assessment for
the Challenge for Methylium, {4-(dimethylamino)phenyl]bis{4-(ethylamino)-3-
methyphenyl]-, acetate prepared by Environment Canada and Health Canada in
July, 2010 pursuant to section 74 of the Canadian Environmental Protection Act
1999 (CEPA, 1999).

This example does not include the entire text or conclusions of the screening
assessment document. The example is only an abstract of the following sections:
Substance Identity, Physical and Chemical Properties, Health Effects
Assessment, Appendix 6, and Appendix 7.

These sections have been abstracted to illustrate the application of (Q)SAR and
information on analog substances to assess the toxicity of a substance in a
weight-of-evidence type approach.

For a copy of the complete screening assessment document, please consult the
following website:

http://www.ec.gc.ca/ese-ees/default.asp?lang=En&n=403207BF-1

Page 165 of 186

 Substance Identity

For the purposes of this document, this substance will be referred to as MAPBAP acetate,
derived from the DSL name. MAPBAP acetate belongs to a class of dyes known as
cationic triarylmethanes. The class can be further sub-divided into those where the
charge on the cation (triarylmethane moiety) is localized or delocalized. MAPBAP
acetate belongs to the latter sub-category (Hunger 2003) implying that the bond holding
the cationic and anionic components of the structure together is at least partly covalent.

Table 2. Substance identity for MAPBAP acetate.

Chemical Abstracts
Service Registry 72102-55-7
Number (CAS RN)
Methylium, [4-(dimethylamino)phenyl]bis[4-(ethylamino)-3-
DSL name methylphenyl]-, acetate

Methylium, [4-(dimethylamino)phenyl]bis[4-(ethylamino)-3-
National Chemical methylphenyl]-, acetate (1:1) (TSCA)
Inventories (NCI)
names1 Methylium, [4-(dimethylamino)phenyl]bis[4-(ethylamino)-3-
methylphenyl]-, acetate (AICS, PICCS, ASIA-PAC, NZIoC)

[4-(Dimethylamino)phenyl]bis[4-(ethylamino)-3-
Other names
methylphenyl]methylium acetate

Chemical group
Discrete organics
(DSL Stream)

Major chemical class or

Cationic triphenylmethanes; anilines;
use

Major chemical sub- Secondary Aromatic Amines, Secondary Amines, Tertiary

class Amines, Tertiary Aromatic Amines

Chemical formula C27H34N3.C2H3O2

Page 166 of 186

 Chemical structure 2

CN(c2ccc(cc2)C[(OC(=O)C)](c3cc(c(cc3)NCC)C)c1cc(c(cc1)N
SMILES 3
CC)C)C

Molecular mass (g/mol) 459.64

1
National Chemical Inventories (NCI). 2006: AICS (Australian Inventory of Chemical Substances);
ASIA-PAC (Asia-Pacific Substances Lists); PICCS (Philippine Inventory of Chemicals and Chemical
Substances); NZIoC ( New Zealand Inventory of Chemicals) and TSCA (Toxic Substances Control Act
Chemical Substance Inventory).
2
This substance is an organic salt, comprising a carbocation and an acetate anion.
3
SMILES stands for : Simplified Molecular Line Input Entry System. This SMILES notation was used to
generate predictions. It is for the neutral form of the molecule and indicates a covalent bond between the
carbocation and acetate anion. This is typically how they are shown in EPIWIN. It is not fully established
what effect using this SMILES will have on the predictions. The acetate part of the SMILES is placed in
square brackets here to highlight the fact that the molecule is at least partly ionic.

Page 167 of 186

I.

II. Physical and Chemical Properties

No experimental data are available for MAPBAP acetate. At the Environment Canada-
sponsored Quantitative Structure-Activity Relationship (QSAR) Workshop in 1999
(Environment Canada 2000) modelling experts identified many structural classes of
pigment and dyes as being “difficult to model” using QSARs. Some physical and
chemical properties of many of the structural classes of dyes and pigments are not
amenable to prediction by models. Under such circumstances, a "read-across" approach
is considered which employs close analogues, to determine the approximate physical and
chemical properties of MAPBAP acetate. A search of the ChemIDPlus (2009) database
yielded a number of suitable analogues which are described in Table 2. Experimental
data for these analogues, when available, were used as extrapolated (read-across) values
for MAPBAP acetate or as supporting values for the weight of evidence.

A limited number of read-across data were found for the selected analogues and,
therefore, predicted values are also used for MAPBAP acetate and the uncertainties of the
predictions are noted.

Table 3 below contains predicted physical-chemical properties of the neutral form of

MAPBAP acetate that are relevant to its environmental fate. Analogue data are available
for water solubility and log Kow. The water solubility of Ethyl Violet (CAS RN 2390-59-
2) is 9000 mg/L (Green 1990). There is an indication that triphenylmethane acetates are
more soluble than the chlorides (Pfenninger and Bruttel 1985) indicating the water
solubility of MAPBAP acetate is high.

Page 168 of 186

 Table 2. MAPBAP acetate and its structural analogues

Methylium, [4- Analogue 1 Analogue 2 Analogue 3

(dimethylamino)phenyl]bis[4-(ethylamino)- Ethanaminium, N-[4- N-(4-(Bis(4-(dimethyl Methanaminium, N-(4-
3-methylphenyl]-, acetate [bis[4- amino) phenyl) ((4-
(diethylamino)phenyl]met methylene) -2,5-cyclo (dimethylamino)phenyl)p
hylene]-2,5- hexadien-1-ylidene)-N- henyl-methylene)-2,5-
cyclohexadien-1-ylidene]- methyl cyclohexadien-1-ylidene)-
N-ethyl-, chloride methanaminium, N-methyl-, chloride
Chloride
Ethyl Violet
(CAS RN 2390-59-2)
MAPBAP acetate Gentian violet Malachite Green
(CAS RN 72102-55-7) (CAS RN548-62-9) (CAS RN 569-64-2)

Page 169 of 186

Comparative analysis:
The differences between the chemical structures of MAPBAP acetate (i) and analogues 1,2 and 3 are:
• the number and position of the methyl, or ethyl, groups;
• the counteranions: acetate, for MAPBAP acetate and chloride (Cl-) for the analogues.
For all substances, the charge on the cation is de-localized. Resonance hybrids can occur and these affect the position of the
counteranion (acetate for (i) and chloride for the analogues).

Page 170 of 186

Table 3. Physical and chemical properties for the neutral form of MAPBAP acetate and analogues

Property Substance Type Value1 Temperature (°C) Reference

Melting point (ºC) MAPBAP acetate Modelled 236.73 - MPBPWIN 2008

Boiling point MAPBAP acetate

Modelled 551.67 - MPBPWIN 2008
(ºC)

MAPBAP acetate 9.13 x 10-10

Vapour pressure EPIWIN 2004
Modelled (6.85 x 10 -12 25
(Pa)
mmHg)

MAPBAP acetate 1.92 x 10-10

Henry’s Law constant Modelled
25 HENRYWIN 2008
(1.895 x 10-15
(Pa·m3/mol)
atm·m3/mole)

Log Kow Analogue

0.51
(Octanol-water partition
Experimental - Tsai et al. 1991
coefficient)
(C.I. Basic Violet 3 CAS
(dimensionless) RN 548-62-9)

Page 171 of 186

 Property Substance Type Value1 Temperature (°C) Reference

Koc

(Organic carbon-water MAPBAP acetate

Modelled 10.26 2 - PCKOCWIN 2008
partition coefficient)

(dimensionless)

Analogue 9000

Experimental - Green 1990

(CAS RN 2390-59-2)
Water solubility 3 (mg/L)

MAPBAP acetate WSKOWWIN 2008

Modelled 475 2 25

1
Values and units in brackets represent those originally reported by the authors or estimated by the models.
2
This value was modelled using the experimental analogue logKow of 0.51 as input,.
3
Importer of MAPBAP acetate has indicated that it is completely soluble at environmental pHs (eg. pH 7).

Page 172 of 186

 Health Effects Assessment

No empirical toxicity data were identified for MAPBAP acetate. Sources of health
hazard information considered included examination of available international reviews,
assessments or classifications, reviewing the available empirical data where available and
the use of predictive models as appropriate. The outputs of predictive models were also
considered using five different QSAR models: TOPKAT (2004), CASETOX (2008),
Toxtree (2009), DEREK 2008, and Model Applier (2009).

Using the representative molecular structure of MAPBAP (with the acetic acid fragment
(acetate) attached to the carbon atom (attached to three aromatic rings)), the following
results were obtained. Positive predictions were obtained on five different genotoxicity
endpoints and only one of these (i.e. rodent micronucleus assay) is corroborated by more
than one model (CASETOX and Toxtree). The Benigni-Bossa model within the Toxtree
also predicts it to be a Salmonella typhimurium TA100 mutagen with metabolic
activation. On the other hand, the female rat cancer models of both CASETOX and
Model Applier gave positive predictions. The male rat cancer model of Model Applier as
well as both mice models (male and female) of CASETOX gave positive predictions. The
presence of a structural alert indicative of genotoxic carcinogenicity is another piece of
supporting information that has been obtained from the Benigni-Bossa model within
Toxtree. Applying the OncoLogic model to a nearly similar structure containing hydroxyl
group in place of the acetate group results in a positive carcinogenicity prediction. This
prediction is based on presence of Nitrogen substituted groups on the aromatic rings.

It is important to note that the Toxtree micronucleus model is a coarse grain filter for
preliminary screening of potential in vivo mutagens and the OncoLogic does not use the
identical structure for prediction purposes. Also, the Ames point mutation models of
CASETOX and Model Applier predict negative results whereas TOPKAT and DEREK
fail to provide any information. However, in the case of cancer models, there are at least
three models (CASETOX, Model Applier and Toxtree) that classify this chemical as a
potential carcinogen. The CASETOX, Model Applier and the Toxtree models are based
on unique methodologies for making predictions and since they point towards a similar
outcome, it carries more weight.

Thus the model predictions were mixed for carcinogenicity (6 positive and 4 negative),
genotoxicity (6 positive and 7 negative), developmental (2 positive; 18 negative and 10
no result) and reproductive toxicity (1 positive and 12 no result).

Potential structural analogues of MAPBAP acetate for the purposes of read-across for
human health toxicity information were identified using Leadscope (Leadscope 2008)

Page 173 of 186

and ChemID (ChemIDPlus 2009) along with professional judgement. As the main
structure would distribute the positive charge across the molecule through resonance
structures, the acetate counter ion would likely interchange with other ions or substrates
when the dye is used. Therefore, the moiety of interest from a human health
toxicological perspective would be the parent molecule itself. Other similar triarylamine
substances, that have empirical data, include gentian violet (CAS 548-62-9), malachite
green (CAS 569-64-2), C.I. Basic Violet 4 (CAS 2390-59-2) and leucomalachite green
(CAS 129-73-7) as shown in Appendix 7.

Gentian violet has been classified by the European Union as Carcinogenicity Category 2
(ECB 2002) based on carcinogenicity in experimental animals. One study did report
negative in vitro genotoxicity for mutations in a reverse mutation assay in several S.
typhimurium strains after exposure to gentian violet at concentrations ranging from 5 –
1000 µg/plate (NICNAS 1999). Malachite green has been classified by the European
Union as Reproductive Toxicity Category 3(ECB 2003) based on developmental toxicity
in experimental animals. Also, the U.S. NTP (2005) reported equivocal evidence of
carcinogenicity in female rats and negative results for genotoxicity from an in vivo
micronucleus assay and an in vitro assay in S. typhimurium (NTP 1997, 1994). C.I Basic
Violet 4 had negative in vitro genotoxicity data for chromosomal aberrations in Chinese
Hamster Ovary cells (NICNAS 1999) and was also found to be predominately negative in
vitro in assays conducted in S. typhimurium and mouse lymphoma cells (CCRIS 2009).
Leucomalachite green was found to have some evidence of carcinogenicity in female
mice and had positive in vivo genotoxicity data (NTP 1996, 2005).

The information obtained from the QSAR models as well as potential analogues, suggest
that there may be potential carcinogenic or developmental toxicity hazards associated
with the substance.

The confidence in the toxicity database is considered to be low due to the lack of
available data for MAPBAP acetate.

Page 174 of 186

 References
CASETOX [Prediction module]. 2008. Version 2.0. Beachwood (OH): MultiCASE.
[cited 2009 September 30]. Available from:
http://www.multicase.com/products/prod03.htm [restricted access].

[CCRIS] Chemical Carcinogen Research Information System. 2009. Mutagenicity studies

for C.I basic violet 4. Available from:
http://toxnet.nlm.nih.gov/cgi-bin/sis/search/r?dbs+ccris:@term+@rn+2390-59-2

ChemIDPlus. 2009. Tool for searching the NLM ChemIDplus database using Chemical
Name, CAS Registry Number, Molecular Formula, Classification Code, Locator Code,
and Structure or Substructure. http://chem.sis.nlm.nih.gov/chemidplus/

[DEREK] Deducting Estimation from Existing Knowledge [Prediction module on CD

ROM]. 2008. Version 10.0.2. Cambridge (MA): Harvard University, LHASA Group.
[cited 2010 05 31]. Available from: http://lhasa. harvard.edu/?page=toxicology.htm
[restricted access]

[ECB] European Chemicals Bureau. 2002. Summary record. Commission Working

Group on the Classification and Labelling of Dangerous Substances. Meeting at ECB
Ispra,16–18 January 2002. European Commission, Directorate-General Joint Research
Centre, Institute for Health and Consumer Protection, European Chemicals Bureau.
ECBI/15/02-Rev. 3. Available from:
http://apps.kemi.se/hclass/DocumentDownload.aspx?DocId=922706

[ECB] European Chemicals Bureau. 2003. Summary record. Commission Working

Group on the Classification and Labelling of Dangerous Substances. Meeting at ECB
Ispra,15–17 January 2003. European Commission, Directorate-General Joint Research
Centre, Institute for Health and Consumer Protection, European Chemicals Bureau.
ECBI/30/03-Rev. 3. Available from:
http://apps.kemi.se/hclass/DocumentDownload.aspx?DocId=923632

Environment Canada. 2000. Chemicals Evaluation Division. Environmental

Categorization for Persistence, Bioaccumulation and Inherent Toxicity of Substances on
the Domestic Substances List Using QSARs. Final Report. Environment Canada.

Green, FJ. 1990. The Sigma-Aldrich handbook of stains, dyes and indicators. Aldrich
Chemical Company, Inc. Milwaukee, Wisconsin.

Hunger, K. (ed.). 2003. Industrial Dyes: Chemistry, Properties, Applications. Wiley-

VCH, Weinheim.

Page 175 of 186

Model Applier [Prediction module] 2009. Columbus (OH): Leadscope Inc. [cited 2009
September 30]. Available from: http://www.leadscope.com/all_products.php [restricted
access].

[NCI] National Chemical Inventories [database on a CD-ROM]. 2006. Issue 1. Columbus

(OH). American Chemical Society, Chemical Abstracts Service Columbus, Ohio.
Available from: http://www.cas.org/products/cd/nci/require.html

[NICNAS] National Industrial Chemicals Notification and Assessment Scheme 1999.

Full Public Report. Component of Basazol Violet 46 L. File No: NA/538. Available at:
http://www.nicnas.gov.au/publications/car/new/na/nafullr/na0500fr/na538fr.pdf

[NTP] National Toxicology Program 1994. Study report number A90714. Available
from: http://ntp-
apps.niehs.nih.gov/ntp_tox/index.cfm?fuseaction=salmonella.salmonellaData&
endpointlist=SA&study%5Fno=A90714&cas%5Fno=569%2D64%2D2&activetab=sum
mary

[NTP] National Toxicology Program 1996. Study report number A31156. Study start
date 07/20/1996

[NTP] National Toxicology Program 1997. Study report number A82983. Study start
date 01/16/1997 Available from: http://ntp-
apps.niehs.nih.gov/ntp_tox/index.cfm?fuseaction=micronucleus.micronucleus
Data&cas_no=569%2D64%2D2&endpointlist=MN

[NTP] National Toxicology Program 2005. Toxicology and Carcinogenesis Studies of

Malachite Green Chloride and Leucomalachite Green (CAS Nos. 569-64-2 and 129-73-7)
in F344/N Rats and B6C3F1 Mice (Feed Studies). Study report number C93029. Study
Report date February, 2005.

Pfenninger Heinz and Bruttel Beat. 1985. Process for converting sparingly soluble
inorganic salts of cationic dyes and brighteners into more soluble salts of organic acids.
United States Patent 4559144. Available from:
http://www.freepatentsonline.com/4559144.html

[TOPKAT] Toxicity Prediction Program [Internet]. 2004. Version 6.2. San Diego (CA):
Accelrys Software Inc. Available from:
http://www.accelrys.com/products/topkat/index.html

Toxtree version 1.60. 2009. Tool to estimate toxic hazard by applying decision tree
approach. Developed by Ideaconsult Ltd Bulgaria.

Page 176 of 186

Appendix 6: Summary of (Q)SAR Results

(Q)SAR PREDICTIONS ON CARCINOGENICITY

Mice Rat
Model/
Rat Mice Rodent Mammal
Species
Male Female Male Female

Model
Applier N N P P P N N -

Multicase
Casetox P P ND* P - - - -

Topkat
NR NR NR NR - - - -

Derek - - - - - - - NR

* This one is weakly positive (30 case units & 81 % probability)

Page 177 of 186

 Model/endpoints

TT
CT

TK
MA
chrom. ab.

-
-
N
ND
chrom. ab. other rodent

-
-
-
ND
chrom. ab. rat

-
-
-
ND
micronucleus mice

-
-
P
ND
micronucleus rodent

-
-

P
ND
drosophila

-
-
ND
ND
drosophila HT

-
-
-
ND
drosophila SLRL

-
-
-
ND
(Q)SAR PREDICTIONS ON GENOTOXICITY

mam. mutation

-
-
-
ND

- mam. mutation DL
-
-
ND

UDS
-
-
P

NR

UDS human
-
-
-
N

lymphocytes

Page 178 of 186

UDS rat hepatocytes
-
-
-
ND

-
-
-

mouse lymphoma mut

ND

s. cerevisiae
-
-
-
N

yeast
-
-
-
N

hgprt
-
-
-
ND

e. coli
-
-
-
N

e. coli w
-
-
-
ND

microbial
-
-
-
P

salmonella
P
N
N

NR
-
-
-

BB cancer alert
(Q)SAR PREDICTIONS ON REPRODUCTIVE TOXICITY

Model Applier

Model/ Female
Male
endpoint

Species mice rat Rodent mice rat rodent

repro ND ND ND ND ND ND

sperm - - - ND ND ND

Multicase Casetox

mice rat rabbit human

NR P NR NR

(Q)SAR PREDICTIONS ON DEVELOPMENTAL TOXICITY

Model Applier

Endpoint/ Species mice rabbit rat rodent

Retardation N ND N N

Weight decrease N ND N N

Fetal death N ND N N

Post impl. loss ND ND N N

Pre impl. loss P ND N N

Structural N ND ND N

Visceral N - N N

Page 179 of 186

Multicase Casetox

Endpoint/Species Hamster Mammal Miscellaneous

Teratogenicity - P NR

Developmental NR - -

MA – model applier;

CT – Multicase Casetox;

TK – Topkat;

TT – Toxtree;

BB – Benigni-Bossa rule;

ND – not in domain;

'-' no model available in QSAR suite

NR – no result

P – positive

N – negative

Page 180 of 186

Appendix 7: Analogues of MAPBAP acetate considered in Human Health portion of assessment

Structure Name / CAS RN Data/Classifications

European Union Carcinogenicity Category 2 (ECB 2002)

based on evidence in experimental animals

Genitian violet
Genotoxicity
548-62-9
In-vitro reverse mutation:

Negative in S.typhimurium TA98, TA100, TA1535,

TA1537 with and without activation (NICNAS 1999).

European Union Reproductive Category 3 (ECB 2003)

Carcinogenicity
Equivocal evidence of cancer in female rats (NTP 2005)

Genotoxicity
Malachite green
In-vitro gene mutation:
569-64-2
Negative in S.typhimurium TA97, TA98, TA100, TA102,
TA104, TA1535 with and without activation (NTP 1994).

Chromosome aberration:

In vivo: Negative in mouse micronucleus bone marrow

and peripheral blood (NTP 1997).

Page 181 of 186

 Genotoxicity

In vitro gene mutation:

Negative in S.typhimurium TA98, TA100, TA1537,

TA1538 with and without activation; TA1535 without
activation (CCRIS 2009):

C.I Basic Violet 4 Positive in TA1535 with S9 activation (CCRIS 2009)

`
2390-59-2
Negative in Mouse Lymphoma L5178Y with and without
activation (CCRIS 2009)

Chromosome aberration:

Negative in Chinese Hamster Ovary Cells V79 with and

without S9 activation (NICNAS 1999).

Carcinogenicity

Some evidence of carcinogenicity in female mice (NTP

2005)
Leucomalachite green
Genotoxicity
129-73-7
In vivo chromosome aberration:

Positive in female mouse micronucleus peripheral blood

study (NTP 1996).

Page 182 of 186

 Example No. 5
The following example has been provided by the Office of Pesticide Programs (OPP) at
the US EPA. It serves to illustrate the potential use of SAR analysis and weight of
evidence approaches in risk assessment decision making.

Case Study: Use of a weight of Evidence (WOE) approach, including SAR information, to
waive the chronic toxicity/carcinogenicity study requirement in a biocide reregistration decision.

1, 2-benzisothiazolin-3-one (BIT) is a member of the isothiazolone class of biocides. Some of the

registered uses of BIT involve chronic/long term exposures (e.g., use in metal working fluids). To
address this type of potential exposure scenario, chronic and/or cancer studies would usually be
required.

The chemical structures of the isothiazolone biocides can be divided into two sub-classes (Figure
1):

1. General Isothiazolone Class: Isothiazolone pesticide chemicals without a benzene ring

attached at the 4-5 position of the isothiazolone ring. For example: Kathon RH287,
Kathon RH886 and/or OIT, which have been registered by OPP.

2. 1,2-Benzoisothiazolone Class: Isothiazolone pesticide chemicals with a benzene ring

attached at the 4-5 position of the isothiazolone ring. In this case, BIT is the chemical
being discussed.

Page 183 of 186

 Figure 1. Chemical Structures of the Isothiazolone Biocides

All isothiazolone biocides contain an isothiazolone ring (Figure 1).

The issue to be discussed in the example case is whether chronic/cancer studies can be waived
based on existing conditions. The issue is discussed from following different aspects.

• Pesticidal Mode of Action: BIT, CMIT, MIT, OIT, and DCOIT all share a common
pathway for antimicrobial activity:
• All inhibit cell respiration
• All inhibit the same class of dehydrogenase enzymes

These biocides react with microbial cells through cleavage of the S-N bond to form an
S-S linkage with the thiol group on target enzymes. Biocidal activity is a function of the
inhibition of cell respiration.

• Structure Activity Consideration: According to the pesticide mode of action

consideration, the antimicrobial activity for all isothiazolone classes is due to the
isothiazolinone ring and the sulfur nitrogen bond in the isothiazolinone ring plays a key
role for efficacy as a biocide. The current issue is whether the benzene ring will be a
concern for potential toxic effects of BIT. Based on SAR information, the benzene ring
may prolong the biological half-life of the metabolic intermediate moieties in the body.

• Toxicity Profile for BIT: There are no carcinogenicity or chronic toxicity studies for
any of the benzene ring-isothiazolone chemicals (such as BIT). BIT is not mutagenic as
all acceptable guideline mutagenicity studies were negative. The toxicity profile of BIT
shows that it is an irritant following oral and dermal exposures, and this is the effect
observed following repeated dosing in subchronic toxicity studies.

o In two oral subchronic rat studies, gastrointestinal irritation was reported at

10 mg/kg/day (lowest dose tested), and there were no other treatment-related
systemic effects.

o In a 90-day rat dermal study, skin irritation and histopathology were noted at
all doses of 100, 300 and 1000 mg/kg/day, while systemic toxicity was only
reported at the limit dose (1000 mg/kg/day). Gastrointestinal

Page 184 of 186

 irritation/histopathology was also reported at 100 mg/kg/day which may be
attributed to grooming.

o The repeated-dose metabolism and disposition study indicates the

metabolites associated with BIT exposure may remain in the body much
longer than the parent compound. The evidence from this study indicates
benzene containing metabolites may accumulate in the liver, kidneys and
thyroid gland.

• Toxicity Profile for the non-benzene ring isothiazolone pesticides: The mutagenicity
data for the non-benzene ring containing isothiazolones were largely negative except for
a few positive observations in vitro with CMIT/MIT and DCOIT. Three
chronic/carcinogenicity studies are available for non-benzene ring isothiazolone
pesticides and all were negative for carcinogenicity, although two of these were found to
have major deficiencies for the chronic toxicity portion of the studies.

o One study was conducted using drinking water administration of a 14.2%

CMIT/MIT mixture at doses of 2.0/3.1, 6.6/9.8, and 17.2/25.7 mg/kg/day in
rats males/females. This study reported hyperplasia of the GI tract but no
other systemic effects.

o The second study used dermal administration of a single dose of 400 ppm
CMIT/MIT to the skin of mice for 30 months and the only significant finding
was dermal irritation.

o A carcinogenicity study was conducted using dietary administration of OIT.

There were no reported carcinogenic effects following oral exposure to up to
1000 ppm in the diet for 78 weeks. Although some tumors were reported,
the incidences were within the ranges for the control animals.

All of the isothiazolones produced toxicity at the site of contact, i.e. irritation of the
gastrointestinal tract, skin and respiratory tract, when administered at high doses. These
biocides produce minimal to no significant systemic toxicity; no histopathological change
distant from the site of dosing was observed, which appears correlated with rapid
metabolism and excretion for these chemicals. Based on read-across comparison, it is
concluded that:

o Skin irritation: Similar findings in all dermal studies (BIT, CMIT/MIT, OIT)
although at different dose levels.
o Skin histopathology: Similar for BIT and OIT, none found in rabbit study on
CMIT/MIT
o Similar clinical chemistry findings with BIT and OIT and similar to BIT oral dog
study
o Severe skin irritation in BIT dermal study

Page 185 of 186

 The relative potency is as follows:

Skin Irritation: CMIT, DCOIT > OIT, MIT, BIT

Skin Sensitization: CMIT > DCOIT >OIT > MIT > BIT

• Risk assessment considerations: As a class, the isothiazolone pesticides are irritants by

all routes of exposure, and are dermal sensitizers. For BIT, gastrointestinal irritation
provides the basis for points of departure for short, intermediate, and chronic/long-term
exposure scenarios.

Final Recommendation:

Based on a read-across comparison and weight of evidence (WOE) approach, that the chronic
toxicity/carcinogenicity study for BIT is not required at this time if the risk assessment is
protective of irritation. This recommendation was based on the following considerations: 1)
available cancer studies for the isothiazolone pesticides are negative; 2) there is a lack of
mutagenicity concern for BIT, and the other isothiazolone pesticides; 3) BIT and the other
isothiazolones are irritants following oral, dermal and inhalation exposures and produce similar
effects following subchronic exposures; 4) the isothiazolones as a group have a known mode of
action for antimicrobial activity; 5) irritation is the predominant effect and is the basis of the
PODs; 6) although the metabolism study for BIT showed an increased half life and accumulation
of radioactivity in thyroid compared to other isothiazolone chemicals, these observations were
determined to be not of toxicological significance, as the toxicological effects of BIT up to 90
days were not different than the effects observed with the other isothiazolone chemicals.

It is recommended that the available data be evaluated to inform the need for a UF to account for
subchronic to chronic exposure durations for BIT.

Page 186 of 186