Targeting Drug-like properties in Chemical Libraries

1
Targeting Drug-like properties in Chemical
Libraries

• David Winkler, Frank Burden, Mitchell Polley
• Centre for Complexity in Drug Design
• CSIRO Molecular Science and
• Chemistry Department, Monash University

VICS
2
Complexity in Drug Design Group

• Prof. Frank Burden - Scimetrics Ltd -consultant
  to CSIRO
• Dr. Mitchell Polley - CSS postdoctoral fellow
• Darryl Jones - CSS PhD top-up student - Flinders
  University (Physics)
• Prof. Dave Winkler - CSIRO Molecular Science and
  Monash University/VICS

3
Overview of Project

• Aims to develop a method for evolving a chemical
  library of heterogeneous agents (molecules) using
  'drug-like' fitness functions
• Chemical space is vast (gt1080 possibilities)
• Method must explore drug-like chemical space and
  identify islands of activity and novelty
• Application in the discovery of novel bioactive
  agents such as drugs, crop care products
• Methodology applicable to design of new materials
  and nanomachines using different fitness functions

4
Overview of Project

• Steps
• Devise sparse, informative mathematical
  representations of molecules
• Devise sparse methods of selecting these for
  models
• Use agent-based methods (Bayesian neural nets) to
  map representations to properties and use models
  as fitness functions
• Develop methods for evolving chemicals using
  mutation operators so that maximum chemical space
  can be traversed
• Evolve chemical libraries using drug-like fitness
  functions

5
Highlights

• Representations
• Novel charge fingerprint descriptor devised and
  tested
• Theory of eigenvalue descriptors cracked
• momentum space descriptor work started
• Novel selectivity index developed

6
Sparse Descriptors

• Many thousands of descriptors have been devised
  (e.g. CoMFA fields, DRAGON)
• Many are highly correlated with other descriptors
  - contain the same information
• Some (e.g. molecular weight) are information-poor
• Models using sparse descriptors can be more
  predictive
• We work to the premise that it is possible to
  devise sparse, information-rich descriptors from
  which suitable subsets could be drawn for a wide
  variety of modelling problems

7
Charge fingerprints

• These are widely applicable, easily computed
  descriptors calculated by binning charges on
  atoms in different environments

8
EEM-based property descriptors

• Density Functional Theory (DFT) proposes that
  knowledge of electron density allows computation
  of many other properties
• Electronegativity equalization methods (Mortier,
  Bultinck and others) is a rapid, approximate DFT
  method
• All work to date has concentrated on charges or a
  few other observables.
• Main strength will probably lie with calculation
  of other molecular properties, when method is
  generalized and parameterized for more atom types

9
Generalized eigenvalue matrices
10
Why do eigenvalue descriptors work?
Eigenvalue matrix
EEM matrix
AT TL \ A TLT' A-1 TL-1 T' since T'T-1
for an orthogonal transformation i.e. inverse of
A is related to the eigenvalues
11
Momentum space descriptors

• the more interesting part of the electron density
  distribution in terms of biological activity is
  located near to the k-space origin. The
  corresponding r-space density distribution is
  associated with the outermost valence regions of
  the molecule
• k-space descriptions of electron density are more
  compact and simpler

12
Optimum Selectivity Index So
13
Highlights

• Sparse feature selection
• Automatic Relevance Determination (ARD) method
  refined
• Sparse Bayesian feature detection theory mastered
• Linear sparse feature detection using an EM
  algorithm and Jeffrey's prior
• Nonlinear Bayesian feature detection achieved but
  needs more work
• Novel variable selection when number of
  descriptors is much larger than the number of
  molecules in the data set

14
Sparse Bayesian variable selection
Descriptor
15
Highlights

• Optimum nonlinear modelling
• Bayesian regularized neural networks working well
• Linear sparse feature detection and modelling
• Nonlinear Bayesian feature detection and
  modelling using radial basis function regression
• Use of sparse Bayesian methods in neural networks
  under study

16
Highlights

• Models built
• Blood-brain barrier partitioning
• Drug intestinal absorption
• Acute toxicity
• Phase II metabolism - substrates and inhibitors
  (Flinders medical school collaboration) - SVM
• Several drug target models - e.g. farnesyl
  transferase

17
Blood-brain barrier model
Topological descriptors- 3 hidden nodes Training
set 85 compounds, test set 21 compounds
18
Intestinal absorption QSAR model
Property-based descriptors- 5 hidden nodes-
optimum model
19
Acute toxicity model
Burden index/binned charge descriptors 8 hidden
nodes Training set 450 compounds, external test
set 53 compounds
20
Using SVM and EEM descriptors to model phase II
metabolism
21
COX 1 and 2 QSAR and selectivity

• Built QSAR model for cyclooxygenase 1 and 2, and
  S0 using a large data set from Tom Stockfisch at
  Accelrys (454 compounds obtained from
  http//www.accelrys.com/references/datasets/)
• Used atomistic (A), Burden eigenvalue (B) and
  charge fingerprint (C) descriptors together with
  a Bayesian regularized neural net to build model
• Compared MLR with a Bayesian neural net with 3
  nodes in the hidden layer

22
COX 1 and 2 QSAR and selectivity
Selectivity of cyclooxygenase 1 and 2 inhibitors
23
Selectivity Index So QSAR Model
MLR R20.77 Q20.69
BRANN (3 nodes) R20.92 Q20.74