More Related Content
Viewers also liked (6)
Similar to 9th ICCS Noordwijkerhout (20)
Recently uploaded (20)
9th ICCS Noordwijkerhout
- 1. Real World Applications of Proteochemometric Modeling The Design of Enzyme Inhibitors and Ligands of G-Protein Coupled Receptors
- 2. Contents Our current approach to Proteochemometric Modeling Part I: PCM applied to non-nucleoside reverse transcriptase inhibitors and HIV mutants Part II: PCM applied to small molecules and the Adenosine receptors Conclusions
- 3. What is PCM ? Proteochemometric modeling needs both a ligand descriptor and a target descriptor Descriptors need to be compatible with each other and need to be compatible with machine learning technique... Bio-Informatics GJP van Westen, JK Wegner et al. MedChemComm (2011),16-30, 10.1039/C0MD00165A
- 4. What is PCM ? Proteochemometric modeling needs both a ligand descriptor and a target descriptor Descriptors need to be compatible with each other and need to be compatible with machine learning technique... Bio-Informatics GJP van Westen, JK Wegner et al. MedChemComm (2011),16-30, 10.1039/C0MD00165A
- 5. What is PCM ? Proteochemometric modeling needs both a ligand descriptor and a target descriptor Descriptors need to be compatible with each other and need to be compatible with machine learning technique... Bio-Informatics GJP van Westen, JK Wegner et al. MedChemComm (2011),16-30, 10.1039/C0MD00165A
- 6. Ligand Descriptors Scitegic Circular Fingerprints Circular, substructure based fingerprints Maximal diameter of 3 bonds from central atom Each substructure is converted to a molecular feature
- 7. Target Descriptors Select binding site residues from full protein sequence Each unique hashed feature represents one amino acid type (comparable with circular fingerprints)
- 8. Machine Learning Using R-statistics as integrated with Pipeline Pilot Version 2.11.1 (64-bits) Sampled several machine learning techniques SVM Final method of choice PLS Random Forest
- 9. Real World Applications of PCM Part I: PCM of NNRTIs (analog series) on 14 mutants Output variable: pEC50 Data set provided by Tibotec Prospectively validated Part II: PCM of small molecules on the Adenosine receptors Output variable pKi ChEMBL_04 / StarLite Both human and rat data combined Prospectively validated
- 10. Part I: PCM applied to NNRTIs Which inhibitor(s) show(s) the best activity spectrum and can proceed in drug development? 451 HIV Reverse Transcriptase Sequence Mean StdDev n pEC50 pEC50 (RT) inhibitors 1 (wt) 8.3 0.6 451 14 HIV RT sequences 2 6.9 0.7 259 3 7.6 0.6 444 Between zero and 13 point 4 7.5 0.7 443 mutations (at NNRTI 5 7.4 0.8 429 6 6.0 0.6 316 binding site) 7 6.5 0.6 99 Large differences in 8 6.9 0.7 147 compound activity on 9 8.3 0.6 222 10 7.9 0.7 252 different sequences 11 7.5 0.7 257 12 8.0 0.6 242 13 7.4 0.8 244 14 8.2 0.8 220
- 11. Binding Site Selected binding site based on point mutations present in the different strains 24 residues were selected
- 12. Used model to predict missing values C o m p o u n d s Mutants Original Dataset Completed with model
- 13. Prospective Validation Compounds have been experimentally validated Predictions where pEC50 differs two sd from compound average (69 compound outliers) Predictions where pEC50 differs two sd from sequence average (61 sequence outliers) Assay validation Completed with model
- 14. Prospective Validation Model: R02 = 0.69 RMSE = 0.62 log units Assay Validation R02 = 0.88 RMSE = 0.50 log units
- 15. The Applicability Domain Concept Still Holds in Target Space Prediction error similarity shows a direct correlation with average sequence similarity to training set R022 R0 RMSE 1 1 0.8 0.8 R0 2 0.6 0.6 RMSE 0.4 0.4 0.2 0.2 0 0 -0.2 -0.2 0.5 0.6 0.7 0.8 0.9 1 Average Sequence Similarity with Training Set
- 16. The Applicability Domain Concept Still Holds in Target Space Prediction error similarity shows a direct correlation with average sequence similarity to training set R022 R0 RMSE 1 1 0.8 0.8 R0 2 0.6 0.6 RMSE 0.4 0.4 0.2 0.2 0 0 -0.2 -0.2 0.5 0.6 0.7 0.8 0.9 1 Average Sequence Similarity with Training Set
- 17. The Applicability Domain Concept Still Holds in Target Space Prediction error similarity shows a direct correlation with average sequence similarity to training set R022 R0 RMSE 1 1 0.8 0.8 R0 2 0.6 0.6 RMSE 0.4 0.4 0.2 0.2 0 0 -0.2 -0.2 0.5 0.6 0.7 0.8 0.9 1 Average Sequence Similarity with Training Set
- 18. Does PCM outperform scaling and QSAR? PCM outperforms QSAR models trained with identical descriptors on the same set When considering outliers, PCM outperforms scaling PCM can be applied to previously unseen mutants Validation pEC50 10-NN 10-NN 10-NN Assay PCM QSAR Experiment scaling (both) (target) (cmpd) R02 (Full plot) 0.88 0.69 0.69 0.31 0.41 0.21 0.28 R02 (Outliers) 0.88 0.61 0.59 0.36 0.34 0.32 0.18 RMSE (Full plot) 0.50 0.62 0.57 0.96 0.90 1.29 1.16 RMSE (Outliers) 0.50 0.52 0.58 1.06 0.72 1.39 1.29
- 19. Model Interpretation (Sequences) Effect of mutation presence on compound pEC50 High impact mutations are K101P, V179I and V179F
- 20. Model Interpretation (Compounds) Effect of substructure presence on compound pEC50
- 21. Model Interpretation (Compounds) Example of positively correlated substructure and negatively correlated substructure
- 22. Conclusions PCM can guide inhibitor design by predicting bioactivity profiles, as applied here to NNRTIs We have shown prospectively that the performance of PCM approaches assay reproducibility (RMSE 0.62 vs 0.50) Interpretation allows selection between preferred chemical substructures and substructures to be avoided
- 23. Part II: PCM applied to the Adenosine Receptors Model based on public data (ChEMBL_04) Included: Human receptor data (Historic) Rat receptor data Defined a single binding site (including ELs) Based on crystal structure 3EML and translated selected residues through MSA to other receptors Looking for novel A2A receptor ligands taking SAR information from other adenosine receptor subtypes into account
- 25. Adenosine Receptor Data Set Little overlap between species Validation set consists of 4556 decoys and 43 known actives External Receptor Human Rat Overlap Range (pKi) Decoy Validation A1 1635 2216 147 4.5 - 9.7 130 1139 A2A 1526 2051 215 4.5 - 10.5 57 1139 A2B 780 803 79 4.5 - 9.7 11 1139 A3 1661 327 82 4.5 - 10.0 255 1139
- 26. In-silico validation External validation on in house compound collection Lower quality data set leads to less predictive model Inclusion of Rat data improves model (RMSE 0.82 vs 0.87) Our final model is able to separate actives from decoys 33 of the 43 known actives were in the top 50
- 27. Prospective Validation Scanned ChemDiv supplier database ( > 790,000 cmpds) Selected 55 compounds with focus on diverse chemistry Compounds were tested in-vitro
- 28. Conclusions We have found novel compounds active (in the nanomolar range) on the A2A receptor Hit rate ~11 % PCM models benefit from addition of similar targets from other species (RMSE improves from 0.87 to 0.82) PCM models can make robust predictions, even when trained on data from different labs
- 29. Further discussion Poster # 47 A. Hendriks, G.J.P. van Westen et al. Proteochemometric Modeling as a Tool to Predict Clinical Response to Antiretroviral Therapy Based on the Dominant Patient HIV Genotype Poster # 51 E.B. Lenselink, G.J.P. van Westen et al. A Global Class A GPCR Proteochemometric Model: A Prospective Validation Poster # 54 R.F. Swier, G.J.P. van Westen et al. 3D-neighbourhood Protein Descriptors for Proteochemometric Modeling
- 30. Acknowledgements Prof. Ad IJzerman Prof. Herman van Vlijmen Andreas Bender Joerg Wegner Olaf van den Hoven Anik Peeters Rianne van der Pijl Peggy Geluykens Thea Mulder Leen Kwanten Henk de Vries Inge Vereycken Alwin Hendriks Bart Lenselink Remco Swier
- 31. Real World Applications of Proteochemometric Modeling The Design of Enzyme Inhibitors and Ligands of G-Protein Coupled Receptors
- 33. Leave One Sequence Out By leaving out one sequence in training and validating a trained model on that sequence, model performance on novel mutants is emulated
- 34. Best performing compounds Sequence Compound with highest Activity Full Model Difference pEC50 (pEC50) (pEC50) (Activity and Model) All 326 8.39(賊 0.61) 8.53(賊 0.73) 0.14 1 365 9.16 9.55 0.39 2 221 8.19 8.38 0.19 3 79 8.71 8.81 0.10 4 321 8.83 8.79 0.04 5 321 9.12 8.73 0.39 6 221 8.01 7.93 0.08 7 364 untested 7.50 n/a 8 221 untested 8.42 n/a 9 365 untested 9.43 n/a 10 326 untested 9.23 n/a 11 151 9.05 8.86 0.19 12 321 untested 9.29 n/a 13 100 9.06 8.87 0.19 14 79 9.51 9.62 0.11 Average 0.18
- 35. Worst performing compounds Sequence Compound with Lowest Activity Full Model Difference pEC50 (pEC50) (pEC50) (Activity and Model) All 109 5.85(賊0.54) 5.82(賊0.66) 0.03 1 248 6.09 6.01 0.08 2 109 untested 4.87 n/a 3 422 untested 5.78 n/a 4 84 5.84 5.67 0.17 5 84 5.65 5.54 0.11 6 109 4.60 4.06 0.54 7 439 5.01 5.20 0.19 8 84 4.74 5.20 0.46 9 248 untested 5.96 n/a 10 181 5.82 6.01 0.19 11 181 5.42 5.61 0.19 12 109 5.90 6.09 0.19 13 181 5.11 5.29 0.18 14 181 5.62 5.81 0.19 Average 0.21