A Genetic Algorithm for Flexible Molecular Overlay and Pharmacophore Elucidation

1
A Genetic Algorithm for Flexible Molecular
Overlay and Pharmacophore Elucidation

• Gareth Jones
• OpenEye Cup Feb 22, 2005, Santa Fe NM

2
What is a GA?

• A Genetic Algorithm is based on the process of
  Darwinian Evolution and is good for complex
  search problems.
• A Chromosome is an encoded representation of a
  problem, typically an integer or binary string.
• A chromosome encodes the conformation of each
  active and the superimposition of the set.
• Each chromosome can be assigned a fitness value,
  which is a measure of how good a solution the
  chromosome is.
• A GA starts with a random population of
  chromosomes.
• Genetic operators such as crossover or mutation
  are repeatedly applied to the population.
• Over time the average fitness of the population
  increases.

3
Overview of the GA
4
Roulette Wheel Parent Selection.

• Each member of the population gets a slice of the
  roulette wheel that is proportional to it's
  (ranked) fitness i.e. selection is biased
  towards fitter individuals

5
Genetic Operators

• Mutation. Randomly choose some portion of the
  chromosome string and randomly alter it.
• Crossover. Choose a random cross point after
  which the string is copied from the other parent.
  
• Mixing. Crossover for sparse strings

6
Genetic Algorithm for Pharmacophore Elucidation

• GAPE
• Based on GASP
• Distributed by Tripos Associates
• G. Jones, P. Willett and R. C. Glen, Journal of
  Computer-Aided Molecular Design 9 (1995) p532
• Incorporates ideas from GOLD
• Distributed by CCDC
• G. Jones, P. Willett, R. C. Glen, A. R. Leach and
  R. Taylor, Journal of Molecular Biology 267
  (1997) p727
• And new stuff
• Java code base (some JNI)

7
Molecular Overlay and Pharmacophore elucidation

• The problem is to overlay a series of actives
  (without any prior knowledge of the
  pharmacophore) such that common molecule receptor
  interactions can be identified.
• We use a GA to search simultaneously the
  conformational space of a series of actives as
  well as the possible superimpositions of each
  conformation.

8
Design of the Algorithm

• Identify important features in each molecule and
  label them.
• GA chromosome contains conformational information
  in binary bitstrings and encodes mappings as
  integer strings.
• When decoding select one molecule as the base
  molecule (the one with the most features) and use
  GA mappings to fit other molecules on top of the
  base molecule.
• Fitness function comprises a score for the
  overlay of similar features, internal steric
  energies and common molecular volume.
• Similarity scores based on hydrogen bond
  distributions (Mills and Dean).

9
Chromosome Encoding

• An overlay of N molecules requires a chromosome
  of 2N-1 strings.
• N binary strings encode conformational
  information one string for each molecule. Each
  byte encodes an angle of rotation about one
  rotatable bond and bits are used to flip free
  corners
• Integer strings map each molecule onto the
  fitting molecule.

10
Chromosome Decoding

• The N binary strings are decoded to produce
  molecular conformations for each molecule.
• Each of the N-1 integer strings is decoded
• Two passes of least-squares fitting is used to
  overlay molecules on top of the base molecule.
  Fitting is applied so that functional groups are
  positioned to interact with the same point in the
  receptor. The chromosome is then rebuilt.

11
GA Mapping
12
Scoring Function
13
Geometric Weight

• R (fitting radius) is annealed from 3.5 to 1.5
• Additional geometric constraints (linear drop-off
  between maximum and minimum angles)
• Acceptor geometries LP, Plane, Cone, None
• Donor, fitting point, donor angle.
• Desolvation correction

H1
r12
R
D2
H2
D1
14
Donor/Acceptor Weights

• J.E.J Mills and P.M. Dean, Three-Dimensional
  Hydrogen-bond Geometry and Probability
  Information from a Crystal Survey, Journal of
  Computer-Aided Molecular Design 10 (1996)
  pp.607-622.
• Ionization term

15
Other Terms

• Volume Integral
• Uses atomic Gaussians
• Grant Pickup J. Phys. Chem. 1995, v99, p3503.
• VDW
• Parameters from TAFF
• Clark et al. J. Comp. Chem. 1989 v10 p892.
• User defined Feature Sets

16
Torsional Distributions

• Restrict conformational search to observed
  torsions
• Examples from GOLD
• Can also use NUMIMBA Klebe and Meitzer, J.
  CAMD, 1994, V8, p583

17
Incorporation of Activities

• Acitivities can guide creation of the overlay
• Use pKi to specify activity.
• Weight pair-wise terms in volume integral and
  pharmacophore feature score.
• pKi values are used as initial weights then the
  weights are normalized so that the average weight
  is 1.

18
Examples

• Examples use standard protocol
• Steady state operator-based GA
• Selection pressure of 1.001
• Score Weights
• Donor Hydrogen Weight 1500
• Acceptor Atom Weight 1750
• Aromatic Ring Weight 2500
• Volume Weight 10
• Energy Weight 10
• 60000 operations (crossovers 47.5, mutations
  47.5 and migrations 5)
• Island model- 5 populations of 100 chromosomes
• Input structures minimized and randomized
• Best guess at atom protonation
• 10 runs per test system (avg run time 256s)
• Display best scoring system.

19
5-HT3 Antagonists
20
Angiotensin II Antagonists
21
5-HT2A Antagonists
22
Incorporation of Activities
8.76 8-OH_DPAT7.94 Buspirone7.56
Spiperone7.16 MCPP5.75 Quipazine5.38
Remoxipride
http//www.gpcr.org 5HT1A pKi human activity
23
Overlay with Activities
Overlay without Activities
24
Examples from PDB

• A comparison of the pharmacophore identification
  programs Catalyst, DISCO and GASP Patel,
  Gillet, Bravi Leach JCAMD V16 (2002) 653-681.
• DHFR 1boz, 1drf, 1ohk, 1dlr, 1hfp, 2dhf
• CDK2 1aq1, 1di8, 1e1v, 1e1x, 1fin, fvv
• Thrombin 1c4v, 1d4p, 1d6w, 1dwd, 1fpc
• HIV RT 1bqm, 1tvr, 1ddt, 1fk9, 1rt5, 1rt1, 1ep4,
  1klm, 1rt3, 1vru
• Thermolysin 1hyt, 4tmn, 5tmn, 1qf1, 5tln, 7tln

25
DHFR
26
DHFR
27
Thrombin
28
Thrombin
29
CDK 2
30
CDK 2
31
HIV RT

• Shared pharmacophore 1bqm, 1tvr, 1ddt, 1fk9,
  1rt5, 1rt1
• Others 1ep4, 1klm, 1rt3, 1vru.

32
HIV RT
33
HIV RT
34
Thermolysin
35
Thermolysin
36
Pharmacophore Search

• Query molecules are held rigid.
• Query can be the output of a GAPE overlay.
• Database of target compounds sequentially matched
  using GAPE against queries.
• Sort alignments using GAPE score
• Test on 20 5HT2A ligands, overlaid on a
  conformation of Sarpogrelate.
• Used in-house on datasets of 800 compounds.

37
5HT2A Example
38
Rigid Pharmacophore Search

• Query pharmacophore and structure(s) from GAPE.
• Use clique detection to screen conformations.
• Rank structures that pass using GAPE feature and
  volume score.
• Example 5HT3 query. Database of 18M conformers
  (ChemNavigator collection).
• 650 conformers/second.
• Proximity filter.
• Work in progress.

39
5-HT3 Example
40
Conclusions

• The GA is a good method for molecular overlay and
  pharmacophore generation.
• Development of chromosome representation and
  scoring function required considerable effort.
• Believe that GAPE is a significant improvement
  over GASP.
• Pharmacophore features do not have to be present
  in all molecules.
• Incorporation of activities.
• More improvements.

41
Acknowledgements

• Paul Watson, Carleton Sage, Runtong Wang Arena
  Pharmaceuticals.
• Val Gillet, Simon Cottrell, University of
  Sheffield, UK PDB test systems.
• Robin Taylor, CCDC GOLD torsional library.
• PyMol
• AstexViewer