LLNL-JRNL-737521

Chemical Attribution of Fentanyl Using

Multivariate Statistical Analysis of
Orthogonal Mass Spectral Data

B. P. Mayer, A. J. DeHope, D. A. Mew, P. E.

Spackman, A. W. Williams

August 25, 2017

Analytical Chemistry
Disclaimer

This document was prepared as an account of work sponsored by an agency of the United States
government. Neither the United States government nor Lawrence Livermore National Security, LLC,
nor any of their employees makes any warranty, expressed or implied, or assumes any legal liability or
responsibility for the accuracy, completeness, or usefulness of any information, apparatus, product, or
process disclosed, or represents that its use would not infringe privately owned rights. Reference herein
to any specific commercial product, process, or service by trade name, trademark, manufacturer, or
otherwise does not necessarily constitute or imply its endorsement, recommendation, or favoring by the
United States government or Lawrence Livermore National Security, LLC. The views and opinions of
authors expressed herein do not necessarily state or reflect those of the United States government or
Lawrence Livermore National Security, LLC, and shall not be used for advertising or product
endorsement purposes.
 Chemical Attribution of Fentanyl Using Multivariate Statistical Analysis of

Orthogonal Mass Spectral Data

LLNL-JRNL-737521

Brian P. Mayera,*, Alan J. DeHopeb, Daniel A. Mewa, Paul E. Spackmana, Audrey M.

Williamsa

a
Forensic Science Center, Lawrence Livermore National Laboratory, 7000 East Ave. L-

091, Livermore, CA 94550, United States

b
Materials Science Division, Lawrence Livermore National Laboratory, 7000 East Ave.

L-382, Livermore, CA 94550, United States

*
Corresponding Author
E-mail: mayer22@llnl.gov
Phone: 925-423-1128
Fax: 925-423-9014
 Abstract

Attribution of the origin of an illicit drug relies on identification of compounds

indicative of its clandestine production and is a key component of many modern forensic

investigations. The results of these studies can yield detailed information on method of

manufacture, starting material source, and final product - all critical forensic evidence. In

the present work, chemical attribution signatures (CAS) associated with the synthesis of

the analgesic fentanyl, N-(1-phenylethylpiperidin-4-yl)-N-phenylpropanamide, were

investigated. Six synthesis methods, all previously published fentanyl synthetic routes or

hybrid versions thereof, were studied in an effort to identify and classify route-specific

signatures. 160 distinct compounds and inorganic species were identified using gas and

liquid chromatographies combined with mass spectrometric methods (GC-MS and LC-

MS/MS-TOF) in conjunction with inductively coupled plasma mass spectrometry (ICP-

MS). The complexity of the resultant data matrix urged the use of multivariate statistical

analysis. Using partial least squares discriminant analysis (PLS-DA), 87 route-specific

CAS were classified and a statistical model capable of predicting the method of fentanyl

synthesis was validated and tested against CAS profiles from crude fentanyl products

deposited and later extracted from two operationally relevant surfaces: stainless steel and

vinyl tile. This work provides the most detailed fentanyl CAS investigation to date by

using orthogonal mass spectral data to identify CAS of forensic significance for illicit

drug detection, profiling, and attribution.

2
 Introduction

Critical to law enforcement and related intelligence efforts to combat illicit drug

abuse are methods that can assess the presence and the persistence of both the drug and

its associated compounds: its chemical attribution signatures (CAS). These signatures

often include synthesis precursors and byproducts, metabolites in various biological

matrices, and degradation products after exposure to laboratory and other operational

surfaces. Analytical tools for the characterization of such compounds generally include

various forms of chromatographic separation combined with spectrometric detection

schemes. Gas and high-pressure liquid chromatography (GC and HPLC, respectively)

combined with mass spectrometric (MS) detection have traditionally been the workhorses

of such studies.1 Though they each have specific merits and drawbacks, both are geared

exclusively towards organic speciation and are generally used independently of one

another. Another mass spectrometric technique that has found considerable success in

forensic applications, particularly when coupled to statistical chemometric analyses, is

isotope ratio mass spectrometry using, for example, time-of-flight secondary ion mass

spectrometry (TOF-SIMS), accelerator mass spectrometry (AMS), or elemental analyzers

coupled to isotope ratio mass spectrometers.2-4

Fentanyl, or N-(1-phenylethylpiperidin-4-yl)-N-phenylpropanamide, is a synthetic

opioid originally designed for anesthesia and analgesia. With a potency roughly 100

times that of morphine5, its strong euphoric effects have bred a significant potential for

misuse. Since its designation as a Schedule II narcotic, multiple domestic clandestine

fentanyl laboratories were seized by law enforcement. Between 2000 and 2010, there

3
were more than 1,000 confirmed or suspected fentanyl-related overdoses and deaths in

the United States.6

In almost all clandestine laboratory raids in the 2000s, manufacture of fentanyl

was found to rely on N-phenethyl-4-piperidone (NPP) as the starting reagent.6 The use of

NPP is driven by its commercial availability and the ease of its manufacture from

smaller, readily available compounds. Various sophisticated chromatographic and

spectrometric techniques7-11 have been previously exploited to obtain information on

fentanyl signatures. In a particularly detailed study by Lurie et al. two fentanyl synthesis

pathways were profiled using UHPLC-MS/MS.9 A substantial number of compounds was

detected and identified in this work, but unique CAS were readily classified due in part to

the small number of routes studied and the significant differences between them (one

route did not use NPP as an intermediate). Forensic chemical attribution of synthetic

schemes that rely on a common intermediate, however, are often difficult to discriminate

amongst due to a small number of unique signatures, particularly when present at trace

levels. The work of Lurie thus prompted us to investigate the degree to which highly

similar synthetic routes can be unambiguously discriminated.

In the present work, six synthetic methods were selected to prepare fentanyl in a

manner believed to most closely mimic conditions used by a clandestine manufacturer.

All methods were found either in the open literature12,13, on illicit drug synthesis

websites14, or were hybrids of such methods. Intermediate purification was limited to

simple solvent extractions; and silica gel chromatography, crystallization, etc. were

eliminated both to retain unique CAS and to simulate the often more novice skills of

illicit drug manufacturers. Taking a cue from samples previously seized from clandestine

4
laboratories, all of the routes chosen here relied on NPP as the critical intermediate

compound.

Studies similar to that reported presently have sought to use a variety of analytical

techniques for chemical signal attribution and forensics related to acutely toxic

chemicals. In particular these studies have sought to link relevant signatures to synthetic

origin, particularly in work by Fraga and coworkers.15,16 The current work, however,

employs several orthogonal, but complementary techniques for the identification of

fentanyl CAS. Electron impact (EI) and chemical ionization (CI) GC-MS and HPLC

combined with time-of-flight mass spectrometry (MS/MS-TOF) were used to profile the

organic content in the crude fentanyl reaction mixtures. In addition, inorganic content

was profiled by inductively coupled plasma mass spectrometry (ICP-MS), which has, to

the best of our knowledge, never been employed in previous fentanyl CAS studies.

The three analytical techniques provided 160 unique synthesis-related signatures.

The complexity of the resultant data, however, demanded the use of statistical techniques

to extract relevant CAS. Therefore, a multivariate statistical model is presented that

highlights the main sources of variance among the six synthetic routes. These statistical

results directly relate to route-specific CAS and provide a model that can also be

extended to the prediction of synthesis method for samples extracted from common

laboratory surfaces, namely stainless steel ducting and vinyl tile flooring. Forensic

samples are typically collected from a wide variety of surfaces, so the applicability of

such statistical models towards data from traditional sampling strategies is therefore

highly important, particularly when linking a compound to a particular laboratory is

critical. Ultimately, we demonstrate the power of statistical methods for CAS analysis

5
using multiple orthogonal analytical techniques, and we provide what we believe to be

the most detailed investigation to date for organic and inorganic signatures of fentanyl

manufacture.

Experimental

Synthetic Approach

The synthesis of fentanyl-like compounds dates back to the 1960s in work by

Janssen.17 The procedure patented in 1965 involved techniques believed to be too

complicated or expensive for clandestine laboratories (e.g. hydrogenations with precious

metal catalysts). As discussed above it is likely that the illicit manufacture of fentanyl

would rely on NPP, and so the six routes chosen presently use this intermediate. In-house

synthesis of NPP was preferred, however, as use of commercially available NNP is

unlikely due to its controlled status (Drug Enforcement Administration (DEA) Schedule I

compound).

The six chosen methods offer variations on the attachment of the different

fentanyl functional groups on the piperidine ring. 1-phenethyl-N-phenylpiperidine-4-

amine (ANPP), the direct precursor to fentanyl, was formed either directly from NPP via

reductive amination (Methods 1, 3, and 5) or by a two-step condensation-reduction

process (Methods 2, 4, and 6). ANPP was transformed into fentanyl using propionyl

chloride (an unscheduled compound) using either pyridine or N,N-diisopropylethylamine

(DIPEA). With few modifications of the published routes, all synthetic methods

successfully yielded fentanyl with a range of purities. Due to the nature of the routes,

complete synthetic details are not given in the present work, though they may be found in

6
the literature cited. We have, however, provided some information and observations of

the six routes studies in Section 1 of the Supporting Information. In all, three replicate

syntheses were performed for each of the six methods yielding a total of 18 crude

reaction samples. Scheme 1 gives the overall synthetic strategy for this work.

Scheme 1. Fentanyl synthesis summary. Black compounds are those common to a

majority of the synthetic routes. Specific compounds in Orange: One Pot12, Method 1;
Blue: Siegfried14, Methods 2, 4, 6; Green: Valdez13, Method 3, 5; Purple: N-methyl-4-
piperidinone route18,19, Method 6. Methods 4, 5, and 6 are hybrid routes exploiting certain
parts of the Valdez and Siegfried syntheses. ACN = acetonitrile, AcOH = acetic acid,
Et3N = triethylamine, MeOH = methanol.

Method 1 was a so-called “One Pot” method taken from the open literature12 and

was slightly modified for the current research. Method 2 was taken from a drug

enthusiast website and is generally referred to as the “Siegfried” method due to the

webpage’s authorship.14 Method 3 was taken from the open literature and was reported to

7
be an efficient, high yielding synthesis route.13 It is referred to as the “Valdez” method.

Method 4 uses the Valdez method of generating NPP and continues on to fentanyl using

Siegfried’s synthesis (named “Valdez→Siegfried”) . Method 5 is the complement to

Method 4, using Siegfried NPP to generate fentanyl via the Valdez route

(“Siegfried→Valdez”). Method 6 is the only route that uses N-methylpiperidone to

generate NPP (“Alt NPP→Siegfried”).18,19 The reaction then uses the Siegfried synthesis

to form fentanyl.

Materials

Unless otherwise stated, all reagents and solvents were obtained from commercial

suppliers (Sigma-Aldrich (St. Louis, MO), Alfa Aesar (Ward Hill, MA), J.T. Baker

(Avantor Performance Materials, Center Valley, PA), Fisher Chemical (Fairlawn, NJ))

and used as received.

Surface Sample Preparation

Coupons for the surface study were made out of either 0.01” thick 304 stainless

steel or small pieces of 1/8” thick vinyl tile (Standard Excelon vinyl composition tiles,

Pattern 51858, Imperial Texture, sandrift white, 1/8 inch thick, Armstrong Commercial

Flooring, Lancaster, PA). The coupons were tared on an analytical balance, and a small

droplet of each crude fentanyl mixture was deposited via a clean metal wire. The mass of

the droplet was measured to within a tenth of a milligram. The spiked coupons were then

allowed to sit for two hours.

8
 After exposure was complete, the spiked coupon was then placed in a clean 20 ml

glass vial and 10 ml of solvent was added. The solvents chosen were acetonitrile,

dichloromethane, and 3 M nitric acid for LC-, GC-, and ICP-MS, respectively. The vials

for GC and LC analysis were then shaken using a mechanical mixer for 15 minutes at 600

rpm, while the samples for ICP-MS analysis were instead sonicated due to safety reasons

(i.e. to prevent gas build up). The extracts were then transferred to clean glass vials.

Vinyl tile extractions were filtered through 0.45 µm filter cartridges and directly

analyzed. Stainless steel extracts showed no evidence of solid debris and were used as is.

Instrumentation

GC-MS

A quantitative weight of each crude fentanyl product was transferred to a 2 mL

glass vial and diluted in 1 mL dichloromethane. Dilutions were then performed in

dichloromethane to yield a series of 20 μg/mL solutions. An Agilent Technologies (Santa

Clara, CA) 7890A GC equipped with an Agilent HP-5MS column (5%-Phenyl-

methylpolysiloxane, 30 m x 0.25 mm x 0.25 μm) was used for the chromatographic

separation. A carrier gas of helium (99.999%, Praxair, Inc., Danbury CT) was used, and

the GC was operated in constant flow mode (2.0 mL/min). One microliter of the liquid

sample was introduced via an autosampler (Agilent 7890 series) to the injection port held

at 230°C, splitless injection. The oven temperature was held at 40°C for 3 minutes, then

ramped at 10°C/min to 300°C and held for 5 minutes. Detection was performed with an

Agilent 5975C MS detector and operated in EI (70 eV) and CI (positive, NH3) modes.

For both ionization modes, the system was operated in scan mode (m/z 29-600, 2.57

9
scans/s) with the source and quadrupole mass analyzer held at 230°C and 150°C,

respectively. A solvent delay of 3 minutes was used. The detector was auto-tuned using

the standard tune capability of the ChemStation software, and the tune was confirmed

before each set of experiments. Compounds were identified based on spectral comparison

to the National Institute of Standards and Technology (NIST) Mass Spectral Library

(NIST 08 MS Search 2.0, NIST, Gaithersburg, MD) as well as manual comparison to

mass spectra published in the literature. Peak areas were calculated from extracted ion

chromatograms of the base peak identified for each compound.

LC-MS/MS

A quantitative weight of each crude fentanyl product was transferred to a 4 mL

glass vial and diluted in 1 mL 50:50 acetonitrile:water. From each solution, dilutions

were performed in 50:50 acetonitrile:water to yield a series of 20 μg/mL solutions. An

Agilent 1260 LC equipped with an Atlantis T3 reverse phase column (C18, 150 mm x 2.1

mm, 3 μm particle size, Waters, Milford, MA) was used. Time-of-flight mass

spectrometric detection was performed in positive ion mode with a Bruker micrOTOF-Q

III (Bruker Daltonics, Billerica, MA) equipped with an electrospray ionization (ESI)

source and operated in Auto MS/MS mode (m/z 50-1000). Three precursor ions were

monitored at a given time (m/z 50-450) with active exclusion after three spectra. MS was

performed with a capillary voltage of 4500 V, a dry gas flow rate of 8 L/min at 200°C,

quadrupolar ion and collision energies of 4.0 eV and 8.0 eV, respectively, and a spectral

acquisition rate of 2 Hz.

10
 The mobile phase consisted of water with 0.1% formic acid (A) and acetonitrile

with 0.1% formic acid (B). The gradient profile started with 95% A for 2 min, ramped to

5% A at 18 minutes, held for 13.5 minutes, ramped quickly back to 95% over 0.5 min,

and held for 10 min for column regeneration. This method was used for all samples

except those in the vinyl tile surface study, for which the final 5% A was held for 18.5

min to allow background contaminants to be completely removed from the column. Ten

microliters of the liquid sample were introduced via an autosampler (Agilent B1329B) to

the injection port. The detector was tuned and calibrated using the 20 μL injection loop of

a 6-port valve on the MS using Agilent’s ESI-L Low Concentration Tuning Mix (G1969-

85000). Compounds relevant to each synthetic route were identified based on computer-

aided identification of MS/MS peaks using Bruker’s Compass for otofSeries 1.5

software. Detailed analysis of each sample was done manually with peak areas calculated

by manually integrating the extracted ion chromatogram of the base peak. After route-

specific compounds were identified for each route, a target table was created and

searched against all 18 samples.

ICP-MS

Elemental analysis was performed using an Agilent Technologies (Santa Clara,

CA) 8800 Triple Quadrupolar ICP-MS (ICP-QQQ). An initial semi-quantitative scan

was performed to determine elements of interest within the sample sets. Down-selected

analytes were then measured quantitatively with the following parameters: carrier gas

(0.65 L/min), nebulizer pump (0.50 rps), spray chamber temperature (15°C), and dilution

gas (0.40 L/min). Argon was used as plasma, carrier, and dilution gas. In the collision

11
cell, a helium flow was used as follows: 0.0 mL/min in No Gas tuning mode and 4.0

mL/min in He mode. The measurements were performed as three replicates, with 50

sweeps per replicate. Integration time per mass was held at 0.10 sec. The rinse time was

set to 30 sec at 0.3 rps of the nebulizer pump, followed by 10 sec at 0.3 rps. Sample

introduction was performed using an ASX-500 autosampler (Cetac, Omaha, NE).

Chemometric Data Analysis

Data analysis was performed using Solo (V8.0, Eigenvector Research Inc.,

Wenatchee, WA). Partial Least Squares Discriminant Analysis (PLS-DA), a supervised

technique that facilitates classification of unknown samples against a known calibration

data set was then performed on the entire data set. All processed data sets (LC, GC, and

ICP-MS) were mean-centered by sample and range scaled by compound for each

analytical method separately (technique by technique). Range scaling was performed by

normalizing the data for each compound by the difference between its maximum and

minimum mean-centered responses. This process ensures that 1) all compounds are given

equal weight and 2) responses from each analytical technique are given equal importance.

This preprocessing procedure has been employed previously for such fused datasets to

remove response factors and to express response as “concentrations” independent of

analytical technique and experimental conditions.20

Results and Discussion

PLS-DA Discrimination of Fentanyl CAS from Crude Reaction Mixtures

12
 From the analytical data, 126 unique compounds were identified by LC-MS/MS,

and GC-MS identified 29 compounds. Sample LC- and GC-MS chromatograms from

each route are given in Section 2 of the Supporting Information to demonstrate

chromatographic ability of the methods chosen. Unabridged lists of LC- and GC-MS

compounds are given in Section 4 of the Supporting Information. LC and GC

compounds were given a numerical or alphabetical designation, respectively, in order of

retention time. LC- and GC-MS detected the same compound if it is labeled by both a

number and a letter. Doubly charged compounds detected by LC-MS are designated with

a “D” followed by its retention time. Initial semi-quantitative ICP-MS runs were

performed and five elements were subsequently quantified for statistical analysis based
23
on relevance to the synthetic methods and signal over background: Na, 39K, 127I, 133Cs,

and 137Ba. In total, 160 species were identified from analytical data.

Statistical techniques are available that can make analysis of large, complex data

sets objective and procedurally much simpler. These methods seek to describe the

observed experimental data with a reduced set of new “latent” variables whose goal is a

more efficient description of the underlying sources of sample variance. For this study

partial least-squares discriminant analysis (PLS-DA) was chosen as the multivariate

statistical methodology. In PLS-DA a compromise is struck between describing the set

of explanatory variables and predicting the response variables. General descriptive

discriminant analysis (DA) techniques allow one to identify variables that best

discriminate among various classes. Predictive DA techniques, however, extend the

concept of class assignment to classification of “unknown” data.21 The current work

exploits such a statistical model to ultimately predict synthesis routes of “unknown”

13
samples taken from operational surfaces. First, we address the use of PLS-DA to identify

important compounds that are highly route-specific and to highlight routes that produce

similar CAS profiles.

The selectivity ratio (SR) is a ratio of explained to residual variance of a given

variable (i.e. chromatogram peak area) and is useful in the objective identification of

important CAS.22 High SR values indicate the spectral variable (i.e. compound)

contributes much towards discrimination of samples (i.e. reaction mixtures or routes).

Selecting an appropriate SR threshold results in a reduced data set containing the most

important compounds, i.e., a purely statistically determined list of CAS.

Choosing a SR threshold for CAS determination reflects a balance between

identifying real markers and excluding those that do not bear significance on describing

sample variance. Though various strategies such as the F-test or so-called nonparametric

DIVA plots23 exist to more objectively determine SR thresholds, our initial goal was

simply to generate a minimum of 2 signatures for each route, and a global value SRmin = 1

was shown to be sufficient. Table 1 gives compounds that were identified as CAS for

which their SR values were greater than unity. Table S-1 of Section 3 of the Supporting

Information connects the identified compounds to their retention times, formulae, and

tentative names when possible.

Table 1. Route specific CAS identified through PLS-DA using a selectivity ratio, SRmin =
1. Letters and numbers refer to GC- and LC-MS detected compounds, respectively.
Compounds represented as both a letter and a number were detected by both techniques.
CAS with a “D” followed by a number are doubly charged compounds detected by LC-
MS. See the Tables S-1 through S-3 in the Supporting Information for more details,
particularly assignments based on LC-MS/MS data.

14
 Method (Class) LC-MS GC-MS ICP-MS
1, One Pot 2, 3, 4, 6, 7-S, 8, 9, 15-S, A, G, 24-I, --
20, 23, 24-I, 26-AB, 31, 7-S, U, V, Z,
32-AA, 33-AC, 36, 48, 32-AA, 26-
52, 54, 59, 64, 81, 87, 89, AB
95, 98, 99, 102, 103, 104,
105, D2.7, D2.8, D2.8-2
2, Siegfried 19, 27, 28, 43, 50, 53, 58, -- --
77, 82, 91, 94, 97, D13.7,
D13.8, D14.1, D14.4,
D14.8, D14.8-3, D15,
D15.4, D15.7, D17.3,
D18
3, Valdez 73 67-E --
4, Valdez→Siegfried 11, 35, 60, 70, 84, 100 B, H, Y --
5, Siegfried→Valdez 39-K, 71, D15.3, D15.3-2 J, 39-K --
6, Alt NPP→Siegfried 45, 51, 83, 107 -- --

The vast majority of the 87 PLS-DA derived CAS (SRmin = 1) belongs to Methods

1 and 2. Many of the SR values for these compounds are high, indicating large

discriminatory ability. Considering the top 25% CAS in terms of SR value, almost all

belong to Method 1. Only four of these variables were classified to other routes, namely

Methods 2 and 6. The large SR values of these variables imply the model should

discriminate among these three synthesis methods. Conversely, the CAS identified for

Methods 3, 4, and 5 are low in number and/or have low SR values. These facts indicate

the chosen threshold may be too high to confidently classify samples from these methods.

A Chemical Perspective of Fentanyl CAS

Though extremely powerful, statistical methods of data analysis are often of a

“black box” nature and require a contextualization from a chemical perspective if a true

understanding of the underlying data is desired. To that end we now turn to a brief

discussion of some of the statistically determined CAS given in Table 1 and their

15
relevance to specific synthetic routes. We emphasize here that if structures/chemical

names are given, they have been tentatively assigned through MS/MS data in the case of

LC data or through GC library comparisons. They have not, however, been verified by

authentic reference standards. The synthetic endeavor that would require is outside of the

scope of the current work, especially considering the statistical focus. Work is on-going

to unambiguously assign structures to compounds considered highly important to

statistical discrimination. For compounds that were part of the synthesis (reaction

components, isolatable intermediates, etc.) their identity was easily confirmed through

comparing MS/MS spectra and retention times.

Fentanyl synthesized via Method 1 contained many early eluting compounds in

the LC-MS analyses. Many of these were alcohols that did not apparently undergo further

reaction. Under the LC analytical conditions used, most of these compounds are not well

retained on the column, and therefore elute at or near the system dead volume. This

observation was shown to result from the mismatch in initial mobile phase and sample

solvent compositions. This poor separation was deemed acceptable, however, as the high

organic component of the solvent system is required to preserve sample stability and to

ensure complete sample dissolution.

Due to the high resolution and exact mass capabilities of the LC-MS/MS-TOF, all

compounds could be deconvolved from the data. One alcohol in particular, 1-

phenethylpiperidin-4-ol, was detectable in relatively large amounts by both LC- (89%

relative to fentanyl) and GC-MS (~200-300% relative to fentanyl). Method 1 also

produced a large amount of acetylfentanyl (roughly 3.6x that of the desired fentanyl

product by LC-MS). In fact, greater than 25% of the total LC-MS base peak

16
chromatogram area was due to this byproduct. This finding agrees reasonably well with

the GC data, which showed roughly a 4.5-fold increase over fentanyl. Also present in

large quantities were unreacted ANPP and aniline – two compounds found in extremely

low amounts, if at all, in the other routes.

Additional unique attribution signatures of the one pot method are large amounts

of other acetate- or acetamide-based compounds. For example, 1-phenethylpiperidin-4-yl

acetate, N-phenylacetamide and N-phenethyl-N-phenylacetamide were deemed CAS by

PLS-DA. Also unique to this reaction method was a series of bipiperidine compounds

presumably formed through reductive amination of the phenethylpiperidinone and

unreacted piperidine hydrochloride. Again, it is important to keep in mind, as with most

assignments, these tentative identities are based merely on MS/MS data, GC-MS library

matches, and most probable chemical reaction pathways. They have not been confirmed

by authentic reference standards or in-house syntheses.

Method 2 only had two unique CAS based on LC-MS/MS and GC-MS data: N-

phenethyl-N-phenylpropionamide and N-phenethyl-N-propionylpropionamide, both

derivatives of 2-bromoethylbenzene. Also, the Siegfried method is the only one to use a

potassium-containing base (K2CO3) in the formation of NPP. Only one potassium adduct,

C10H23KN3, was deemed a CAS, though a structure was not proposed. ICP-MS data was

expected to reflect potassium use, as well, but no statistical importance was observed

from the PLS-DA analysis. Lastly, fentanyl synthesized via Method 2 yielded a

significant number of LC CAS that are doubly charged. Due to inconclusive MS/MS

fragmentation patterns we were not able to posit structures. Their presence, however, was

found to be indicative of the Siegfried method.

17
 Only two compounds were classified as unique CAS for the Valdez synthesis

method, and only one structure was determined. A library match of GC-MS data

identified this compound as N,N-diisopropylpropionamide. This compound most likely

forms as a result of reaction between propionyl chloride and DIPEA impurities.

Exclusive to the Siegfried route is a biphasic reaction mixture where water-soluble

compounds are removed into an aqueous phase via polyethylene glycol. The Valdez

method lacks such a phase transfer catalyst, so one may expect additional signatures to

appear, particularly charged quaternary amines. In fact, two such CAS were revealed by

LC-MS, namely 1,1-diphenethyl-4-(N-phenylpropionamido)piperidin-1-ium and 1-

phenethylpyridin-1-ium. These compounds were classified as CAS for Method 4,

Valdez→Siegfried.

PLS-DA identified two propionamides as specific to Method 5. Based on MS/MS

fragmentation patterns, these compounds were tentatively identified as N-

phenylpropionamide and N-ethyl-N-phenylpropionamide. A GC mass spectral library

match confirmed the assignment of N-phenylpropionamide, but the assignment of the

latter compound was not confirmed. Finally, several CAS were associated with Method 6

but none was assigned a specific structure. This method’s alternative route to making

NPP should result in a variety of byproducts, but few expected compounds were detected

by any means. Iodine was expected to be a relevant CAS for this route as well,

considering its exclusive presence in this route. The PLS-DA analysis, however, did not

consider it important. This may result from iodine being dominated by compounds with

higher selectivity ratios. In all, quantitative ICP-MS data affected data analysis very little.
139
I seemed to have some influence on discrimination of Method 6, but its SR value was

18
relatively low (SR = 0.47). Overall, it was observed that ICP-MS data was not useful in

the current study. That fact does not discount it as useful for other synthetic methods or

forensic studies, however.

PLS-DA Model Validation and Application to Surface Samples

From the complete data set of 160 variables, five components were chosen for the

PLS-DA calibration model, covering an explained variance of 95.0% of the analytical

data. The calibration data from the 18 samples derived from crude reaction mixtures are

displayed as a scores plot of the two dominant components given in Figure 1. The graph

reveals Method 1 can be separated well with the first component. The second component,

in turn, can separate Method 2 well; but there exists poor separation between Methods 3

and 5 and Methods 4 and 6. Plotting these data in conjunction with the third component’s

scores (Figure 2), though, allows for reasonable separation for all but Methods 3 and 5.

12
10
8
Scores on LV 2 (17.84%)

6
4
2
0 Method 1
-2 Method 2
-4 Method 3
Method 4
-6 Method 5
-8 Method 6
-20 -16 -12 -8 -4 0 4 8
Scores on LV 1 (29.49%)

Figure 1. Scores plot for the first and second components for the crude fentanyl mixture
calibration data.

19
Figure 2. Three-dimensional scores plot for components 1, 2, and 4 showing good
separation of all classes except for data from synthesis Methods 3 and 5. Color-coding
was done according to class membership. The X-Y projection is equivalent to those data
plotted in Figure 2.

The observed statistical results can be easily rationalized in terms of the reaction

chemistries. Method 1 generated quite a number of CAS with large SR values and was

therefore easily separated by the first component. The same discriminatory ability was

observed for Method 2 by a second component. Methods 4 and 6, which share the

Siegfried method of ANPP→fentanyl synthesis, could not be distinguished by two

components alone, and a third component was required for acceptable discrimination.

Particularly problematic was the separation by scores values of Methods 3 and 5, routes

that both share the Valdez method of ANPP→fentanyl synthesis. Since these methods

share various precursors and byproducts in relatively low amounts, common signatures

result in few CAS, which is reflected in similar component scores. Note the close

clustering of replicates for a given route. This is due to the consistency in reaction

20
profiles between replicates for both the GC- and LC-MS data sets. Section 5 of the

Supporting Information gives sample raw peak areas for the One Pot route to demonstrate

the excellent reproducibility between syntheses.

Though scores plots are convenient ways to conceptualize the statistical results,

PLS-DA also provides a quantitative analysis of the data and can classify unknown

samples based on their similarity to calibration data sets. PLS-DA model generation

begins with cross-validation, which provides objective quantitation of the model’s ability

to generalize to independent data sets (i.e. those resulting from the follow-up surface

study).

The current model was shown to perform very well in the discrimination of

samples into their corresponding classes during cross-validation. In other words, the

likelihood that a given validation set corresponds to its known specific synthesis route is

extremely high. In fact, for almost all cases there was no significant probability of

misclassification. Only a single sample from Method 3 had any significant non-zero

probability of belonging to another method (Sample M3-3 as Method 5, p = 0.094). This

observation is understandable, though, as Method 5 (Siegfried NPP/Valdez

ANPP→fentanyl) is a hybrid method of Methods 2 and 3 (Siegfried and Valdez,

respectively) and therefore shares a variety of common signatures with those routes.

Nevertheless, correct classification for all 18 samples was observed since the probability

of correct classification outweighs any underlying similarities to other methods. A heat

map of classification data is shown in Figure S-3 in Section 6 of the Supporting

Information.

21
 It is important to recognize that the ability to discriminate amongst different

syntheses is in part reflective of the different nature of the routes themselves. Minor

changes in synthesis, whether in the form of material source, reaction conditions, or

synthetic chemist (novice or professional), may have a significant impact on the end

discriminatory ability. Unfortunately, the matrix of such conditions that can be tweaked

and/or investigated is infinite and can therefore not be fully addressed, particularly with

regards to the scope of the current work. We reemphasize the fact that the routes chosen

were intended to reflect clandestine synthesis routes, but that sample-to-sample

variability (though shown to be small) was inherently incorporated via completely

independent syntheses from start to finish. It is part of on-going work to assess the

influence of the abovementioned factors towards ultimate route discrimination. For not,

we feel it is sufficient to use the current reaction matrix to demonstrate the proof-of-

concept nature of the statistical CAS methodology.

With a validated PLS-DA calibration model, we can now attempt classification of

samples that have been exposed to stainless steel and vinyl tile. These materials aim to

mimic protocols commonly found in forensic investigations, namely surface-based

sampling from laboratories, warehouses, containers, etc. It is important to demonstrate

that any additional signals associated with surface swipes or extractions especially for

porous and/or organic matrices do not interfere with those from compounds of interest.

Determination of the presence of these signatures on such surfaces is critical for linking a

batch of seized fentanyl to its synthesis location, even if elucidating the synthetic

pathway from surface data is not required or possible.

22
 Figure 3 shows scores results of the PLS-DA analysis for the “unknown” stainless

steel surface data and compares those to results from the calibration set. It is evident that

despite having been exposed to the metal surface, the relevant signatures are persistent

enough that correct classification is likely possible. In other words, surface sample and

calibration scores values cluster qualitatively well. As additional evidence of CAS

persistence over a given period of exposure, Section 7 of the Supporting Information

gives LC-MS chromatographic traces of the time dependence of the CAS profile over a

24 hour period. CAS intensity invariance over this period suggests that this model should

be useful even if sampling is delayed.

Figure 3. PLS-DA scores from stainless steel data color-coded by the predicted “most
probable” class. Spheres are the surface data sets, whereas pyramids represent data from
the calibration sets taken from Figure 2. Vinyl tile data given in Figure S-6 of the
Supporting Information.

23
 Figure 4a presents quantitative results from predicted class membership through a

heat map for the stainless steel samples. Ideally there would only be on-diagonal intensity

indicating complete confidence in correct class assignment. Indeed, for samples within a

class (i.e. synthesis route), classification probability is essentially unity. The model does,

however, predict that several samples bare similarities to others outside of its class (i.e.

synthesis route). This fact is reflected by off-diagonal intensity. That these methods are

all interrelated (i.e. permutations of the Valdez and Siegfried routes) may account for the

multiple classifications. However, because the probability of making the correct

assignment is generally higher than that of an incorrect classification, the PLS-DA

analysis assigns the “most-probable” class correctly for every sample except for one.

Only one Method 3 sample, M3-SS2, was classified as belonging to Method 5 but the

probability was relatively low (p = 0.680). Despite surface matrix effects, incubation time

and conditions, potential volatilization of compounds, etc., the statistical analysis still is

able to assign the large majority of the stainless steel surface samples to their

corresponding known synthetic origins.

24
 a) p
0.0
M6
0.1
0.2
M5
0.3
M4 0.4
0.5
M3 0.6
0.7
M2
0.8
M1 0.9
1.0
M1-SS1
M1-SS2
M1-SS3
M2-SS1
M2-SS2
M2-SS3
M3-SS1
M3-SS2
M3-SS3
M4-SS1
M4-SS2
M4-SS3
M5-SS1
M5-SS2
M5-SS3
M6-SS1
M6-SS2
M6-SS3
b) p
0.0
M6
0.1
0.2
M5
0.3
M4 0.4
0.5
M3 0.6
0.7
M2
0.8
M1 0.9
1.0
M1-VT1
M1-VT2
M1-VT3
M2-VT1
M2-VT2
M2-VT3
M3-VT1
M3-VT2
M3-VT3
M4-VT1
M4-VT2
M4-VT3
M5-VT1
M5-VT2
M5-VT3
M6-VT1
M6-VT2
M6-VT3

Figure 4. PLS-DA probabilities of class membership for Method 1 (M1, bottom) to

Method 6 (M6, top) for stainless steel (a) and vinyl tile (b) samples. On-diagonal
intensity reflects probability of correct classification. Off-diagonal intensities indicate
that there is non-zero likelihood of that sample to be misclassified.

PLS-DA demonstrated that the CAS profile necessary for accurate classification

persisted relatively well on stainless steel, a relatively inert surface. In contrast, we also

investigated the ability of vinyl tile, a complex polymeric matrix, to retain the signatures

needed for the statistical analysis. Like the stainless steel data, striking visual similarities

between the test and calibration data sets can be seen in the scores plot given in Figure S-

3 in Section 8 of the Supporting Information. The remarkable similarity between

calibration and test scores may be surprising considering the large number of new

25
compounds introduced through exposure to the organic surface. New signatures were

automatically ignored, however, since the variable set was pre-determined by the

calibration data. Ultimately, signatures relevant to fentanyl synthesis appear persistent

against even complex matrices like vinyl tile.

Examination of the class prediction plot given in Figure 4b shows that, again, the

model has a tendency to confuse samples from Methods 3 and 5. Probabilities of

misclassification of samples from Method 3 as Method 5 exceed 0.85 for all three

samples. Again, the similar, trace CAS profiles of these routes explain the tendency for

misclassification. Correct classification generally dominates any chance of the model’s

incorrectly assigning samples, however. Again, there was only a single erroneous PLS-

DA assignment – Method 3 sample M3-VT1 was classified as a Method 5 sample.

PLS-DA of Individual Data Sets

As a final, though important, consideration, we investigated the degree to which

individual techniques were able to discern differences amongst the six synthesis methods.

PLS-DA analyses were performed GC, LC, and GC+LC data sets. The motivation behind

this was to show the level of classification that could be performed in a laboratory with

limited equipment and to highlight the benefits of using multiple sources of mass spectral

data.

For the GC data alone, nine components were necessary to surpass the 90%

cumulative variance threshold. This may indicate that differences among samples may

not be sufficient to provide a robust predictive model. Analysis of LC data, however,

needed only five components to exceed the variance threshold. The same is true for the

26
combined LC and GC data. Figure S-7 in Supporting Information shows scores plots

from these PLS-DA analyses. Scores plots show that for the GC data, relatively poor

separation was achieved by the first three, dominant components. LC data, however, was

able to separate well many of the samples into their respective methods, though, again,

Methods 3 and 5 and Methods 4 and 6 were difficult to resolve.

The overall ability of the data to statistically resolve the methods was gauged by

the number of both incorrect and/or multiple (i.e. ambiguous) class assignments made by

the model. The results from considering GC and LC scores data alone were mirrored

during application of the predictive model, which often failed to unambiguously or

correctly assign samples to the respective synthesis methods. In fact, of the 36 surface

samples, six stainless steel and six vinyl tile samples were misclassified or multiply

classified when using just GC data. LC data only provided a modest improvement - four

stainless steel and six vinyl tiles samples were not classified properly. Combining the GC

and LC data resulted in better discriminatory ability among the six routes. Three samples

from each surface set were incorrectly assigned a route. The full, unabridged data set

combing GC, LC, and ICP-MS results resulted in one stainless steel and three vinyl tile

classification errors, which, again, includes both wrong and multiple assignments. These

improvements in discriminatory ability highlight the power of using data from multiple

orthogonal techniques for unambiguous classification of unknown samples.

Conclusions

27
 A statistical analysis of chemical attribution signatures associated with the

synthesis of fentanyl was presented. The results demonstrated that the combination of

orthogonal techniques provided better discriminatory ability than that from individual

analyses alone. Though we recognize that many more aspects of this method need to be

addressed to assess its ultimately utility (synthetic variability, reaction conditions, CAS

surface persistence, matrix effects, environmental conditions, etc.) this work provides a

solid proof-of-concept work that aims to identify fentanyl CAS and to subsequently link

them synthetic source. Through the application of the PLS-DA model, classification of

“unknown” surface-exposed samples was possible even for synthetic strategies that are

relatively similar or even share common synthetic steps. Also important to stress is that

many important CAS are present at trace levels, further highlighting the power of

statistical techniques for forensic attribution studies. Furthermore, purely objectively

determined CAS were discussed in terms of their potential to be rationalized in terms of

the specifics of the synthesis procedures. The authors feel it is critical to contextualize the

“black box” nature often associated with statistical techniques with chemical knowledge

to truly gain a deeper understanding of the origins of chemical attribution signatures.

Again, we are working towards complete CAS identification and, also, this method’s

application towards other potential toxic materials. Ultimately, however, we believe this

work further reinforces the beneficial synergy between multivariate statistical analysis

and contextualizing results within a synthetic chemical framework while providing the

most comprehensive attribution study of fentanyl to date.

Acknowledgements

28
The authors would like to thank the Department of Homeland Security, Science and

Technology Directorate, Chemical Forensic Program for their funding of this work.

Livermore National Laboratory is operated by Lawrence Livermore National Security,

LLC, for the U.S. Department of Energy, National Nuclear Security Administration

under Contract DE-AC52-07NA27344.

29
 References

(1) Smith, F. P.; Siegel, J. A. Handbook for Forensic Drug Analysis, 1st ed.; Academic

Press: San Diego, 2004.

(2) Meier-Augenstein, W. Stable Isotope Forensics: An Introduction to the Forensic

Application of Stable Isotope Analysis, 1st ed.; John Wiley & Sons Ltd.: West Sussex,

2010.

(3) Volpe, A. M.; Singleton, M. J. Forensic Sci. Int. 2011, 209, 96-101.

(4) Kreuzer, H. W.; Horita, J.; Moran, J. J.; Tomkins, B. A.; Janszen, D. B.; Carman, A.

J. Forensic Sci. 2011, 57, 75-79.

(5) Peng, P. W. H.; Sandler, A. N. Anesthesiology 1999, 90, 756-799.

(6) Leonhart, M. M. Federal Register 2008, 73, 43355-43357.

(7) Lurie, I. S.; Iio, R. J. Chromatogr. A 2009, 1216, 1515-1519.

(8) Lurie, I. S. J. Chromatogr. A 2005, 1100, 168-175.

(9) Lurie, I. S.; Barrier, A. L.; Casale, J. F.; Iio, R.; Bozenko, J. S. Forensic Sci. Int.

2012, 220, 191-196.

(10) Garg, A.; Solas, D. W.; Takahashi, L. H.; Cassella, J. V. J. Pharmaceut. Biomed.

2010, 53, 325-334.

(11) Lambropoulos, J.; Spanos, G. A.; Lazaridis, N. V.; Ingallinera, T. S.; Rodriguez, V.

K. J. Pharmaceut. Biomed. 1999, 20, 705-716.

(12) Gupta, P. K.; Ganesan, K.; Pande, A.; Malhotra, R. C. J. Chem. Res-S 2005, 2005,

452-453.

(13) Valdez, C. A.; Leif, R. N.; Mayer, B. P. PLoS ONE 2014, 9, e108250.

30
(14) Seigfried. Synthesis of Fentanyl.

https://www.erowid.org/archive/rhodium/chemistry/fentanyl.html (accessed Aug 31,

2015).

(15) Fraga, C. G.; Pérez Acosta, G. A.; Crenshaw, M. D.; Wallace, K.; Mong, G. M.;

Colburn, H. A. Anal. Chem. 2011, 83, 9564-9572.

(16) Fraga, C. G.; Wahl, J. H.; Núñez, S. P. Forensic Sci. Int. 2011, 210, 164-169.

(17) Janssen, P. A. J.; Gardocki, J. F., Method for Producing Analgesia. U.S. Patent

3,141,823, July 21, 1964.

(18) Mustazza, C.; Borioni, A.; Sestili, I.; Sbraccia, M.; Rodomonte, A.; Ferretti, R.; Del

Giudice, M. R. Chem. Pharm. Bull. 2006, 54, 611-622.

(19) Grishina, G. V.; Potapov, V. M.; Abdulganeeva, S. A.; Korchagina, E. Y. Chem.

Heterocyc. Compd. 1985, 21, 1355-1362.

(20) Smilde, A. K.; van der Werf, M. J.; Bijlsma, S.; van der Werff-van der Vat, B. J. C.;

Jellema, R. H. Anal. Chem. 2005, 77, 6729-6736.

(21) Varmuza, K.; Filzmoser, P. Introduction to Multivariate Statistical Analysis in

Chemometrics, 1st ed.; CRC Press: Boca Raton, 2009.

(22) Rajalahti, T.; Arneberg, R.; Berven, F. S.; Myhr, K.-M.; Ulvik, R. J.; Kvalheim, O.

M. Chemometr. Intell. Lab. 2009, 95, 35-48.

(23) Rajalahti, T.; Arneberg, R.; Kroksveen, A. C.; Berle, M.; Myhr, K.-M.; Kvalhelm,

O. M. Anal. Chem. 2009, 81, 2581-2590.

31