COMPUTER AIDED DRUG

DEVELOPMENT SYSTEM
Unit 1- Computers in
Pharmaceutical Research and
Development
General Overview- HISTORY OF COMPUTERS IN
PHARMACEUTICAL RESEARCH AND
DEVELOPMENT
• INTRODUCTION
In today’s scenario computers are so important in pharmaceutical
R&D. generating, managing and transmitting information. Earlier days
the information it was a time consuming process for the chemist or
biologist to share the R&D data’s.

The usage of computers was introduced in the Pharmaceutical

Companies in the 1940’s predominantly for purpose of accounting and
not for the scientific purpose. Since the development of software's at
the infantry stage. So, it may be a hard time for a medical chemist or
biologist to handle data’s with out computers (all the data’s were stored
in papers).
 HISTORY- 1960’s
• Computers were first employed by IBM- 7094 for the science
purpose using FORTRAN (Formula Translation) II
• In 1962, QCPE [ Quantum Chemistry Program Exchange ]
was started to use.
• The chemists could understand the fundamental concept with
respect to the chemical structure and molecular properties
associated with biological activities.
 HISTORY- 1970’s
• QCPE was initiated for used for exchanging subroutines and programs for
ab initio and approximate electronic structure calculations.
• ORTEP [ oak ridge thermal ellipsoid program ] – Molecular graphics
• IBM-360 and 370 series were also commenced.
• IBM – 3278 computers used by the Computational Chemistry Group.
• These groups were attained small success such as finding correlations
between the calculated and experimental property.
• The DE NOVO design of pharmaceutical was attempted.
• Two computer based resources were launched:
• 1. Cambridge structural database
• 2. Protein data bank
 HISTORY- 1980’s
• In the decade of 1980’s, various approaches of quantum
chemistry, molecular mechanics, molecular simulations,
QSAR, molecular graphics are incorporated into modern
computational chemistry.

VAX 11/780 computer were the first to gain approval for

small molecule x-ray crystallography. [super mini computers]
 HISTORY- 1980’s
• 1982 - QSAR band was used.
• 1985 - quantum mechanics were studied by chemists
• Some of pharmaceutical companies had implemented
computer aided molecular design application and
software development.
• Computer Aided Drug Design and Structural Based
Drug Design were introduced.
 1986
• There are three importance for using of computers
1. Sending communications through mailing.
2. Implementation of software - Chemical draw – to
create a two dimensional chemical diagrams.
3. Aesthetic element – by using computer graphics -i.e
3D structures and molecular graphics.
DECADE OF FRUITION- [1990’s]
• R&D Managers believed that a super computer would enable
more ideas than would be practical with older computers

• Promising faster identification of new drug candidates.

• Molecular dynamics simulations and quantum mechanical

calculations on large molecules.

• Computational chemists were becoming nearly equal partners in

the quest to discover drugs.

• Programs such as SYBYL, BIOSYM, CHARM were popular

around the world for molecular modeling and simulations .
DECADE OF FRUITION- [1990’s]
• In 1991, the bio design was implemented by
molecular simulations Inc., [MSI].
• ISIS [ integrated scientific information system
is used to manage soft wares.
• SAR are used widely.
• In 1997, LIPINSKI Rule 5 was published.
Also, Computer aided ligand design was
proposed.
• In 1998, QCPE operation is changed and
web based operation started and continues.
 The Lipinski rule of five
• Its molecular weight is less than 500.
• The compound's lipophilicity, expressed as a quantity known as log P (the
logarithm of the partition coefficient between water and 1-octanol), is less
than 5.
• The number of groups in the molecule that can donate hydrogen atoms to
hydrogen bonds (usually the sum of hydroxyl and amine groups in a drug
molecule) is less than 5.
• The number of groups that can accept hydrogen atoms to form hydrogen
bonds (estimated by the sum of oxygen and nitrogen atoms) is less than 10.
• The rule is based on the 90-percentile values of the drugs' property
distributions, apply only to absorption by passive diffusion of compounds
through cell membranes.
• Exception to this rule: compounds that are actively transported through cell
membranes by transporter proteins are excempted.
 MARKETED PHARMACEUTICALS WHOSE DISCOVERY
WAS AIDED BY COMPUTERS

GENERIC NAME BRAND NAME YEAR OF DISCOVERY

APPROVAL ASSISTED BY
(U.S)
Norfloxacin Noroxin 1983 QSAR
Losarton Lozaar 1994 CADD

Nelfinavir Viraceot 1997 SBDD

Ximelagalran Exanta 2004 SBDD

QSAR - (QSAR models) are regression or classification models used in the
chemical and biological sciences and engineering.
CADD – Using computational methods to simulate drug-receptor interaction
SBDD - crystallographic information is utilized to identify and optimize novel lead
compounds, for accelerating the overall drug discovery process.
Statistical modeling
in pharmaceutical
research and
development.
 DESCRIPTIVE VERSUS
MECHANISTIC MODELING.
Need of performing statistical investigations:
1. To predict what the responses are going to be to
future input variables.
2. To extract some information about how these
responses are associating the response variables
3. To understand the foundations of mechanisms from
which the data are getting generated.
• Breiman identified 2 cultures in the statistical
modeling:
1. Data modeling culture [ assumes that the data are
generated by the given stochastic data model (estimating
probability distributions of potential outcomes) ].

2. Algorithmic modeling culture [ uses algorithmic

models and treat the data mechanism as unknown ].
• Statistic data as being generated by a black box into which a
vector of input variables (x) i.e (independent variables) & out
of which a vector of response variables (y) i.e (dependent
variables).
• To understand the mechanism, the use of modeling concept is
essential.
• The purpose of model is essentially that of translating the
known properties about the black box as well as some new
hypothesis into a mathematical representation.
• The way in which the family of models of 9 set of [parameter
values] are usually selected depend upon the main purpose of
exercise.
1. If the purpose is data generating mechanism, the family
models are selected based on its adequacy to represent the
data structure. (Descriptive model) - It is useful for
discriminative between alternative hypothesis but not useful
for capturing the fundamental characteristics of mechanism.
2. If the purpose is to get some insight on or to increase our
understanding of the underlying mechanisms, the family of
models must be selected based on the reasonable assumption
with respect to the nature of mechanism.
 Understanding the mechanism of action

• DESCRIPTIVE METHOD – it must provide

updated, rich and reliable information about the
problem.
• MECHANISTIC METHOD – translating scientific
information in the mathematical models and for
approximately describing probabilistic or stochastic
components indispensable to handling the variability
inherently contained in the data generated process.
 EXAMPLE – ANIMAL TUMOR GROWTH
DATA.
• Used for the representation of the different concepts encountered
during the development of the model and its after identification use.

• An example of rat as a subject in 80 days – the tumor grows from

nearly 0 – 3000 cm³.

• The first evaluation of the data can be done by running NON –

PARAMETRIC ESTIMATION TECHNIQUES.

• Eg. The Nadaraya – Watson kernal regression equation [ does not

provide possibility of interpreting the outcome and are not all reliable
when extrapolating ].

• These techniques are essentially descriptive in nature and useful in

summarizing the data by providing interpolating values .
 PARAMETRIC MECHANISTIC
MODELING APPROACH
• To imagine using a parametric family of models
capturing the potential exponential growth of the
tumor volumes.
• The exponential parametric family is mathematically
fully characterized by the family of equations
V(t) = α exp (λt)
• A particular model is fully specified by fixing the
values for its two parameters α and λ.
 STATISTICAL PARAMETER ESTIMATION

Once the model functional form has been decided upon

and the experimental data have been collected, a value
for model parameters ( point of estimation ) and a
confidence region for the value ( interval of estimation )
must be estimated from the available data.
This is being done to have a non linear model, with a
known functional relationships.
 where, yi denotes the 1st observation corresponding to
the vector xi of independent variable,
Where θ∗ is the true but unknown parameter value
belonging to some acceptable domain
Where u is predicted value as a function of independent
variables and parameters
Where εi are true errors
 The ordinary least square estimate ( OLSE) θ^ of θ∗
minimizes globally over θ ϵ domain

we may indicate with θ^ , the weighted least square

estimator ( WLSE ), which is the value of θ
minimizing
 where e is the vector of ‘residuals’ [ y – u (θ)]
In this case in which the errors are independent of eah other
their co variances, then D = σ2 I, with the constant σ2 being
the common variance and I being the identity matrix
In this case, the same θ minimizing would also minimize
eqn 2 and the DLSE (distributed least-squares estimation (DLSE) method to
solve the chemical source localization (CSL) problem) can therefore be
seen as a particular case of WLSE (weighted least square
distribution- robust regression method-maximises the efficiency of parameter
distribution vs amount of influence)
 CONFIDENCE REGION

Find a set of values Iθ such that the probability that we

are correct in asserting that the true value of θ∗ of the
parameter lies in Iθ is 95%
In the present section we will assume that the mode; is
not too far from linear around the optimum found
if D = cov (ε) ; known
Indicate c vector θ^ the weighted least squares estimator
( WLSE ), that is let θ^ minimize
Expanding u in Taylor series around the true θ∗ and
neglecting terms of 2nd and higher order we may write

We have,

we observe that the WLSE estimation θ^ is approximately

unbiased ( in a small neighborhood of θ∗ ) and,
Where we have obviously denoted
• A first approach to the definition of the confidence
region in parameter space follows the linear
approximation to the parameter joint distribution.
• If the estimates are approximately normally
distributed around θ∗ with dispersion (U.T D-1 U)-1
then an approximate 100 (1-α) % confidence region
for θ∗ is
 A second approach considers that the regions of
equivalent parameter values must enclose parameters
for which the loss function is nearly the same or at any
rate less different than some threshold.
In this case Taylor’s expansion of S(θ^) around θ∗,

so that an approximate 100 (1-α) % confidence region

would be
or
 The above figure shows the estimates of parameters a
and v0 obtained for subject 4, conditional on the relative
estimates value on b.
The oblique elongation on both in the asymptotic and
in the exact region depends on the strong correlation
between parameter estimates for a and v0.
 NON LINEARITY AT THE OPTIMUM

Briefly there exist methods of assessing the maximum

degree of intrinsic non linearity that the model exhibits
around the optimum found.
If maximum non linearity is excessive, for one or more
parameters the confidence regions obtained applying the
results of the classic theory are not to be trusted.
In this ase, alternative simulation procedures may be
employed to provide empirical confidence regions.
 SENSITIVITY ANALYSIS

• The goal here is to determine the (relative)

effect of a variation in a given parameter value
on the model prediction.
Let y= u(X, q) + e be the considered model, with y
∈ Δm and q ∈ Θ ⊂ Δq. We study the sensitivity
of the modelling function u with respect to the
parameter q by means of the (absolute)
sensitivity coefficient, which is the partial
derivative
• or by means of the normalized sensitivity
coefficient

The normalization serves to make sensitivities

comparable across variables and parameters. In
this context, by sensitivity we would mean the
proportion of model value change due to a given
proportion of parameter change.
Absolute and normalized sensitivity coefficients
can be computed analytically or approximated
numerically.
• above figure shows the time course of the absolute
sensitivity coeffi cients of the Gompertz model with
respect to the parameters,
• which can be computed analytically
• Above figure shows the same sensitivity coefficients
expressed as a percentage of their maximum value.
• From these graphs it is apparent that at time zero only
the V0 parameter influences model output.
• On the other hand, as time progresses, V0 rapidly
loses importance and growth and decay parameters a
and b prevail in determining (much larger) variations
in predicted volume, each in the expected direction
(with a increasing and b decreasing expected
volume).
• Above figure shows the time course of the Pearson
MCCC between V and the three model parameters for
such a Monte Carlo simulation with n = 10,000,
having generated values for the three parameters out
of uniform distributions respectively on the intervals
[0.9, 1.1], [0.165, 0.195], [0.55, 0.65].
• The main advantage of this scheme is that the
required number of samples does not grow as
fast as a regular Monte Carlo sampling from the
joint distribution of the parameters as the
number of parameters increases.
 OPTIMAL DESIGN

• Several design optimization criteria exist, the obvious

approach is to choose the time points so as to
minimize the parameter estimate dispersion
(variance-covariance) matrix, which in our case, for
ordinary least squares estimation, is approximated by
the inverse of the Fisher information matrix (FIM) at
the optimum.
• Our criterion therefore becomes to “maximize” in
some sense, the FIM.
• Depending on the specific objective we pose for
ourselves, we might want to maximize one of the
eigenvalues of the FIM, thus obtaining maximum
precision on one parameter; maximize its trace (the
sum of the eigenvalues); or maximize its determinant
(the product of the eigenvalues). This last method,
called D-optimal design (D as in determinant) is
possibly the most widely utilized method of optimal
design.
 POPULATION MODELING

• The sample mean and dispersion of these estimates would then constitute
our estimate of the population mean and dispersion.
• we could find the mean and dispersion in the “Control” and “Treated”
subsamples.
• There are two problems with the above procedure,
• The first is that it is not efficient, because the inter subject parameter
variance it computes is actually the variance of the parameters between
subjects plus the variance of the estimate of a single-subject parameter.
• The second drawback is that often, in real-life applications, a complete
data set, with sufficiently many points to reliably estimate all model
parameters, is not available for each experimental subject
 There are two-stage model:
Stage 1 (intra individual variation)
In the first stage let the jth observation on the ith individual be modelled as
follows:
yij = fi (βi , xij ) + eij , i = 1, . . . , m, j = 1, . . . , ni
•where the function f is a non linear function of the subject- specific parameter
vector βi,
•xij is the observed variable, eij is the normally distributed noise term,
•m is the total number of subjects and ni is the number of observations for the
ith subject.
Stage 2 (inter individual variation)
In the second stage the subject-specific parameter vector is
modelled as:
βi = g(β,bi ) = β + bi , bi ~ N(0,D)
•where β is a p-dimensional vector of fixed population
parameter,
•bi is a k dimensional random effect vector associated with
the ith subject (not varying with j),
•D is its general variance-covariance matrix.
Thank you….