Novel Approaches To Molecular Similarity Molecular Similarity Searching Using COSMO Screening Charge

1
Novel Approaches To Molecular Similarity
(Molecular Similarity Searching Using COSMO
Screening Charges)

• Andreas Bender, ab454_at_cam.ac.uk
• Unilever Centre for Molecular Science
  Informatics,
• Chemistry Dept, University of Cambridge, UK

2
Outline

• Introduction to Molecular Similarity
• What is it and why is it hugely important?
• Our approach to molecular similarity Finding
  many active compounds vs. generalizability?
• How to compare molecules
• The algorithm
• Results
• Special Feature Discovery of Binding Patterns

3
Similarity Searching What is it?

• Describes the identification of molecules
  similar to a given molecule (in analogy to the
  psychological concept of similarity in
  perception)
• Complementary approach to substructural searching
  
• In substructure searching exact retrieval of a
  subgraph of a molecule is performed
• In similarity searching, an abstract molecular
  representation in descriptor space is calculated
  which is compared to abstract representations of
  other molecules
• For reviews see e.g.
• Bender, A. and Glen, R.C., Org. Biomol. Chem.,
  2004, (2) 3204 3218.
• (freely available from www.cheminformatics.org)

4
Similarity Searching Why is it relevant?

• The old complaint It is expensive to bring drugs
  to the market (800 Mio USD) and it takes long
  (10 years)
• Similarity searching can help find new drugs
  (rather their early stage companions, hits or
  leads) by picking the most promising compounds
  to synthesize, test
• Even de novo (computer-based) design of
  completely new structures is possible
• This decreases the need for animal testing, it
  saves time and money
• gt It is a good thing to do

Well well - but HOW?
5
Similarity Searching Illustration

• We have a red Porsche, a blue Ferrari and a red
  kettcar. Which ones are similar?
• 1. Abstract representation Top Speed The Porsche
  and the Ferrari are similar
• 2. Abstract representation Colour The Porsche
  and the kettcar are similar (!)
• Same problem with molecules but they dont have
  colours (means, the important properties are not
  obvious). How do you encode those molecules to be
  similar in some abstract representation?

6
Molecular Similarity The way it should be and
the way it is

• The God of Molecular Similarity (according to
  Google Picture Search)
• The Molecular Similarity Principle
• Small structural changes cause small property
  differences
• Basis of all current structure-property
  predictions
• A is active, B is not how about C?
• Solubility of A is known how does it change if
  we add group X here?

7
Does the Molecular Similarity Principle work?
8
The importance of shape
Slides courtesy of Hugo Kubinyi, Erlangen
Lectures see http//www.cheminformatics.org -gt
Links -gt Education
9
What is the relevant property?

• Usually no one knows / it depends on the
  particular system
• Current descriptors treat molecules as static
  entities but even by definition receptor
  binding involves dynamical motions of the protein
• No agreement exists which kind of interaction of
  the ligand with the receptor actually causes (for
  example agonistic or antagonistic) action Is it
  occupancy? Is it on-off rates? Or some completely
  different property?
• How do you encode shape / surface properties??
  (We are not dealing with 1-dimensional entities
  like proteins / DNA, we encounter rings,
  branching!)

10
Is it possible and sensible to define molecular
similarity?

• YES, but one needs to be careful
• Similarity depends on the Context (e.g. the
  particular receptor easy in case of
  non-directional properties, e.g. solubility)
• Similar changes may have different (even
  detrimental!) effects, depending on system
• Chiral molecules (same structure, but different
  stereochemistry, like your left and right hand)
  may have totally different activities sometimes
  problematic to capture

11
Our approaches to molecular similarity

• How to describe and compare molecules
• Description of the system in a suitable form
• Selection of important features
• Model generation and prediction
• Results
• Special Feature Discovery of Binding Patterns

12
Descriptor Choice
13
2D Environment around an atom (MOLPRINT 2D,
a.k.a. Atom Environments)

• E.g. 6-aminoquinoline

Assign Sybyl mol2 atom types find
connections find connections to
connections create a tree down to n levels bin
the atom types for each level create a
fingerprint for this atom
N2
Level 0 Level 1 Level 2
Car Car
Car, Car, Car
1
2
1
1
These features are created for every (heavy) atom
in the molecule (Bender, A. et al., J. Chem. Inf.
Comput. Sci. 2004, 44, 170-178)
14
Feature Selection

• E.g. comparing faces first requires the
  identification of key features.
• How do we identify these?
• The same applies to molecules.

15
B) Information-Gain Feature Selection

• How can we select the active compounds (red)?

Red Active Green Inactive
?
?
?
16
C) Naïve Bayesian Classifier (classification by
presumptive evidence)

• The next step is to identify which molecules
  belong to which class.
• Example from e-mail classification (spam
  detection)
• Training set from nerdy chemists inbox

• Assign weighting factors to individual features,
  depending on relative frequencies in training set

17
Classification

• To do this we use a Naïve Bayesian Classifier
  using the features (atom environments) we have
  identified as being important.
• Ratio gt 1 Class membership 1
• Ratio lt 1 Class membership 2
• F feature vector
• fifeature elements

18
Application lead discovery

• If I have one active compound will I find
  another one?
• Database MDL Drug Data Report (MDDR)
• 957 ligands selected from MDDR
• 49 5HT3 Receptor antagonists,
• 40 Angiotensin Converting Enzyme inhib. (ACE),
• 111 HMG-Co-Reductase inhibitors (HMG),
• 134 PAF antagonists and
• 49 Thromboxane A2 antagonists (TXA2)
• 574 inactives
• Briem and Lessel, Perspect Drug Discov Des
  2000, 20, 245-264.
• Calculated Hit rate among ten nearest neighbours
  for each molecule

19
Comparison Single Queries
Using Tanimoto Coefficient
Using Bayesian

• Grey bars Briem and Lessel, Perspect. Drug Disc.
  Des., 2000, 20, 245-264.
• Black Bender, A., et al., J. Chem. Inf. Comput.
  Sci., 2004, 44, 1708 1718.

20
Combining Information of 5 Actives
Bender, A., et al., J. Chem. Inf. Comput. Sci.,
2004, 44, 170 178.
21
TXA2, Graph-based Descriptors
Query
1
2
3
4
5
6
7
Very little diversity in heterocyclic systems
no patents, no money, no good!
22
3D Environment around a surface point solvent
accessible surface using local surface properties
Central Point (Layer 0)
Points in Layer 1

• Points in Layer 2

Etc.
Bender, A., et al., J. Med. Chem., 2004, (47)
6569-6583 IEEE SMC 2004 Proceedings
23
Overall Performance Comparable to 2D methods


Bender, A. et al., J. Chem. Inf. Comput. Sci.
2004 (44) 170 178.
24
TXA2, 7 Hits among Top 10 by MOLPRINT 3D
Query
1
2
3
4
5
6
7
25
Which features are selected for classification?

• Even if your classifier works, do the selected
  features make sense?
• Set of active vs. inactive molecules
• Information Gain calculated for each feature,
  those which are much more frequent among actives
  are suspicious and might constitute the
  pharmacophore
• Look at features from HMG-CoA Reductase Inhibitors

26
Selected Features - HMG

• Binding Site HMG rigid lipophilic ring

27
HMG-15
28
Disadvantages

• Multiple probes had to be employed to cover
  putative interactions sufficiently
• Force fields neglect polarization /
  back-polarization effects (that is, it calculated
  charges which are not seen in solution)
• Force fields (usually) employ point charges, thus
  they dont capture directionality of some
  interactions such as hydrogen bonds
• -gt Use more sophisticated QM method!

29
COSMO Calculation of screening charges in ideal
conductor

30
Why COSMO-RS Properties?

• Interactions derived from first principles on
  single scale
• Gives directionality of H-acceptor lobes (unlike
  most force fields exceptions are e.g. the XED
  force field by Andy Vinter / Cresset)
• Employs solvent model, polarization /
  back-polarization
• Classification in agreement with chemical
  intuition (e.g. O of ester, but not O- is
  H-bond acceptor)
• Inaccessible atoms not used (no accessible
  surface)
• Secondary effects captured which are not
  accounted for by atom-typing

31
COSMO ?-Profile
32
A HMG-CoA Reductase Inhibitor

• Statin binding to HMG-CoA reductase involves
  charge interactions of a carboxylic acid group
  and hydrogen bond donor/acceptor functions to the
  pyruvate binding site
• In addition large lipophilic groups of the ligand
  is required which binds to a floppy lipophilic
  pocket of the target protein.
• Features can be well distinguished from ?
  screening charges
• Carboxylate is shown to the right (purple),
  hydrogen bond acceptor functions beneath side
  chain (red)
• Hydrogen bond donor functions point towards
  viewer (blue) while the lipophilic bulk of the
  structure is given in green

33
Encoding as 3-Point Pharmacophores

• Average ?-values calculated for each heavy atom
  atom typing (pos, neg, lipo, acceptor, donor)
  according to heavy atom average

Courtesy of Martin Stahl (Roche)
34
Comparison to other methods
35
Scaffolds Found
36
Summary

• 2D Method Finds lots of active molecules but
  they are similar to what is known already
• (Bender, A., et al., J. Chem. Inf. Comput. Sci.
  (2004) 44, 170-178 Bender, A., et al., J. Chem.
  Inf. Comput. Sci., 2004, 44, 1708 1718.)
• 3D Method Find less active compounds but
  enables discovery of new chemotypes
• (Bender, A. et al., J. Med. Chem., 2004, (47)
  6569-6583.)
• Features shown to correlate with binding patterns
• Similarity searching using screening charges
  derived from first principles shows good
  performance and possesses a sound theoretical
  basis

37
Acknowledgements

• Robert C Glen (Unilever Centre, Cambridge, UK)
• Hamse Y. Mussa (Unilever Centre, Cambridge, UK)
• Stephan Reiling (Aventis, Bridgewater, USA)
• Software
• GRID, CACTVS, gOpenMol many, many others
• Funding
• Bill Gates, Unilever, Tripos