Exploring the black box: structural and functional interpretation of QSAR models.

Size: px

Start display at page:

Download "Exploring the black box: structural and functional interpretation of QSAR models."

Oswald Marshall
5 years ago
Views:

EMBL-EBI Industry workshop: In Silico ADMET prediction 4-5 December 2014, Hinxton, UK Exploring the black box: structural and functional interpretation of QSAR

1 EMBL-EBI Industry workshop: In Silico ADMET prediction 4-5 December 2014, Hinxton, UK Exploring the black box: structural and functional interpretation of QSAR models. (Automatic exploration of datasets using QSAR) Pavel Polishchuk A.V. Bogatsky Physico-Chemical Institute of NAS of Ukraine, Odessa, Ukraine

2 Outline Introduction and existed approaches Structural QSAR interpretation theory and practical examples Functional QSAR interpretation theory, practical examples and comparison with docking studies Automatic exploration of chemical dataset 2

3 QSAR interpretation: interpretability vs. complexity Model interpretability Popular misbelief MLR PLS DT knn RF SVM ANN models ensembles Model complexity 3

4 Importance of QSAR structural interpretation Extract SAR information in a chemically meaningful way detection of structural alerts, creation of structural filters or set of rules fragment-based drug design Model validation interpretation results should not contradict with experimental observations 4

5 QSAR interpretation approaches Model-specific approaches: Rule-based (Decision tree) Regression coefficients (MLR, PLS) Latent variables (PLS) Weights and biases (ANN) Model-independent approaches: Local gradients or partial derivatives I.I. Baskin et al., SAR QSAR Environ Sci, 2002, G. Marcou et al., Molecular informatics, 2012, C i f(x i ) f(x x i i Δx i ) 5

relationship f(x) Var_1 Var_2 Mol_1-0.23 1.

6 QSAR interpretation: common workflow Model Variables contributions Structureproperty relationship f(x) Var_1 Var_2 Mol_ Mol_ Mol_ Mol_

7 Matched molecular pairs & molecular transformations logs = logs = ΔlogS = 2.58 H OH ΔlogS = 1.59 logs = logs = Leech A.G. et al, J. Med. Chem. 2006, 49, Sheridan R.P. et al., J. Chem. Inf. Model. 2006, 46,

8 Exemplified dataset 8

9 Structural QSAR interpretation - = logs pred = logs pred = ΔlogS pred = = logs pred = logs pred = ΔlogS pred = Polishchuk P.G. et al. Molecular Informatics 2013,32,

10 Structural QSAR interpretation - = logs pred = logs pred = ΔlogS pred = 2.39 Polishchuk P.G. et al. Molecular Informatics 2013,32,

11 Limitations of existed descriptors (Dragon, etc) Dragon fails - = Dragon is OK - = Computational MMP H COOH Polishchuk P.G. et al. Molecular Informatics 2013,32,

12 Simplex representation of molecular structure (SiRMS) Simplex generation example Atom-property labeling Kuz min, V. E. et al, Journal of Molecular Modeling 2005, 11, Kuz min, V. E. et al, Journal of Computer-Aided Molecular Design 2008, 22,

13 Local and global interpretation Local interpretation analysis of single compounds Global interpretation reveal trends 13

14 Interpretation: fragmentation Case number Do specific interactions of a ligand with its target exist or important? Is an orientation of a ligand relatively its target known? Fragments selection and grouping 1 NO (e.g. passive diffusion through membranes, solubility, lipophilicity, etc) not relevant can be done by the researcher based on his own knowledge 2 YES (ligand-receptor interactions, host-guest complexes, etc) YES 3 NO consider fragments positions relatively to the target and observed or predicted interactions MMP can be applied, silently assumed that all compounds have the same interaction mode 14

15 Examples of structural interpretation 15

16 Solubility (1033 compounds) Endpoint Solubility, logs 5-fold external cross validation results SiRMS Dragon Model R 2 CV RMSE R 2 CV RMSE PLS RF SVM Polishchuk P.G. et al., Molecular Informatics, 2013,

17 Mutagenicity (Ames, 4361 compounds) 5-fold external cross validation results Descriptors Algorithm Balanced Accuracy SiRMS RF SVM Dragon RF SVM Polishchuk P.G. et al. Molecular Informatics, 2013,

18 Combined contribution (effect) of fragments RF+SiRMS Contribution = 0 (non-mutagen) Contribution = 0 (non-mutagen) Contribution = 1 (mutagen) 18

19 Questions How does the fragment influence the property? WHY?! 19

20 Functional interpretation of QSAR models Structural interpretation A pic 50 = f(a 1, A 2, A 3 ) = x - = B pic 50 = f(b 1, B 2, B 3 ) = y C Contribution(C) = x - y Functional interpretation A - = B C pic 50 = f(a 1, A 2, A 3 ) = x pic 50 = f(a 1, A 2, B 3 ) = y Contribution 3 (C) = x - y 1, 2, 3 groups of descriptors represented different physico-chemical factors (charge, H-bonding, etc) of compound A and B. 20

21 Antagonists of fibrinogen receptor (functional interpretation example) 21

22 Fragment examples Antagonists of fibrinogen receptor: dataset Arg-Gly-Asp Arg-mimetic Linker Asp-mimetic 338 compounds 22

23 Antagonists of fibrinogen receptor: models 5-fold external cross validation results Algorithm R 2 RMSE RF SVM (RBF kernel) SVM (linear) PLS

24 Structural interpretation (global) Linker Asp-mimetic Arg-mimetic 24

25 Functional interpretation (global) Arg-mimetic Linker Asp-mimetic 25

26 Functional interpretation of RF model (local) RF model Asp224 Phe160 Tyr190 Arg Ser225 Mg electrostatic H-bonding hydrophobicity polarizability 26

27 Functional interpretation of SVM model (local) SVM-RBF model Asp224 Phe160 Tyr190 Arg Ser225 Mg electrostatic H-bonding hydrophobicity polarizability 27

28 Automatic exploration of datasets of chemical compounds (dataset mining) 28

29 SiRMS-QSAR software QSAR model building 1 2 utilizes ncpu

30 SiRMS-QSAR software Calculation of fragments contributions not implemented yet 30

31 SiRMS-QSAR software Plot fragments contributions structural interpretation functional interpretation 31

32 SiRMS-QSAR software External visualization tool 32

33 Interpretation workflow scheme Create sdf file with property values Build models (regression or classification) Look at models stat (if all models are bad reconsider dataset) Calculate fragment contributions Plot contributions of desired models selected from statistically significant ones 33

34 ADME/Tox examples (SAR trends, global interpretation) Datasets taken from: 1) Cheng W. et al., J. Chem. Inf. Model., 2012, ) Kovdienko N.A. et al., Molecular informatics, 2010, ) Polishchuk P.G. et al., J. Chem. Inf. Model., 2009, ) in-house data 34

35 Permeability (structural interpretation) consensus of RF, GBM, SVM models 35

36 Permeability (functional interpretation) consensus of RF, GBM, SVM models 36

37 Toxicity (structural interpretation) consensus of RF, GBM, SVM models 37

38 Toxicity (functional interpretation) consensus of RF, GBM, SVM models 38

39 Summary SiRMS Descriptors Others (Dragon, CDK, etc) Models Fragments Interpretation Regression + + Classification + + Terminal (substituent) + + Scaffold/linker + - Structural + + Functional +? 39

40 Conclusions Almost any QSAR model can be interpreted using the proposed schemes. Results of structural and functional interpretation obtained from different models are well correlated between models and correspond to observed trends. Structural interpretation allows to reveal trends in SAR, rank fragments, find potential structural alerts, etc. Functional interpretation may provide a guess about factors which are dominated and influence on the investigated property. 40

41 Perspectives Smart automatic fragmentation approaches Detection of potential activity cliffs in local interpretation Testing on other types of descriptors Usage of datasets which include mixtures of compounds Application of this approach for wider range of structurally diverse datasets with different end-points and comparison to MMP 41

42 Useful web links A.V. Bogatsky Physico-Chemical Institute, Chemoinformatic group: SiRMS project on GitHub: SiRMS-QSAR (dataset analysis): External web-based visualization: 42

43 Acknowledgement A.V. Bogatsky Physico-Chemical Institute (Odessa, Ukraine) Strasbourg University (France) Prof. V. Kuz min Dr. T. Khristova Dr. L. Ognichenko A. Kosinskaya E. Mokshina M. Kulinskiy Prof. A. Varnek Dr. D. Horvath 43

Structural interpretation of QSAR models a universal approach

Methods and Applications of Computational Chemistry - 5 Kharkiv, Ukraine, 1 5 July 2013 Structural interpretation of QSAR models a universal approach Victor Kuz min, Pavel Polishchuk, Anatoly Artemenko,