Establishing a Successful Virtual Screening Process Stephen Pickett Roche Discovery Welwyn Introduction • Challenges facing lead generation and lead optimisation • Overview of computational methods in lead generation • “Needle” screening • Model Validation • Conclusions Challenges Facing Lead Generation and Lead Optimisation • Reduce fall-out rate in development • Nature of compounds, not just number of compounds is important • Require leads not hits • Fail fast Challenges Facing Lead Generation and Lead Optimisation • Increase robustness of candidates in humans • Simultaneous optimisation of – – – – Biological activity Physicochemical properties Pharmaceutic properties Pharmacokinetic properties • In vitro screens - synthesised compounds • Computational screens - virtual compounds Role for Computational Techniques Overview Property Prediction Genern & Applicn of Predictive Models Compound Prioritisation Purchase SynthesisScreening Tasks Compound set comparisons Compound filtering Compound selection (virtual screening) Library Design Virtual screening • Application of computational models to prioritise a set of compounds for screening • Similarity to lead(s) – 2D › Substructural keys › BCUTS, topological pharmacophores (CATS) – 3D › Pharmacophores › Pharmacophore fingerprints › Shape, surface properties, MFA • Q/SAR models • Fit to protein binding site Process Targeted screening Reagents Reaction Ideas Property Filtering Library design Reagent Scoring Enumeration Property Filtering Compounds Docking / Pharmacophore Scoring Prioritised Syntheses Prioritised Screening Process Requirements • Robust and iterative – Flexibility – Reliability – Usability • Substructural filters – acid anhydrides, reactive alkyl halides ... – functional groups incompatible with chemistry • Price, supplier, availability • Reagent Scoring • Rapid calculation of product properties • Apply consistently across projects Computational Methods in Lead Generation at RDW • Biological Screening – Pharmacophore and/or docking for compound prioritisation. – Target families – Data analysis • Needle Screening – Selection of diverse compound set for NMR screening library. – Designing a focussed needle set. • Lead Generation libraries – Design of targeted libraries – Ligand-based design Needle Screening: An application • IMPDH – Inosine Monophosphate DeHydrogenase – Key enzyme in purine biosynthesis – Potential host target for halting viral replication. • Known inhibitors OH O O H N O O O H N O N H O O OH O O MPA 20nM N VX-497 7nM O O N O N F O H N O O N H BMS 17nM N “War-Head” 19mM MPA “warhead” bound to IMPDH Needle Screening: An application • Aim – Find novel replacements for phenyl oxazole “warhead”. › Low molecular weight, chemically tractable “needles”. • Methods – NMR screening – Structure-based virtual screening to select set of compounds for biological evaluation. Process • Optimise virtual screening protocol (FlexX) • Virtual screening of suitable small molecules – reagents available in-house • Biological evaluation • Develop chemistry around actives Overview of FlexX • Fragment based docking methodology – Break molecule into small fragments at rotatable single bonds – Dock multiple conformations of each fragment – Regenerate molecule from docked fragments • Scoring Function – Trade-off between speed and accuracy – Focussed on identifying good intermolecular interactions – Takes no account of absent or poor interactions • Post-processing of solutions required – Additional calculations – Visual inspection Optimisation of Virtual Screening Protocol • Dataset – 47 t-butyl oxamides (40nm to >>40mM). 21 with IC50. • Examine influence of O N N Y O R • Protein model – 2 X-ray structures › oxamide › MPA analogue • Crystal waters • Scoring functions – Flex-X, ScreenScore and PLP Binding site with four waters Binding site with oxamide Summary of Results • Prediction of pKi values of actives – ScreenScore best in this case – Less dependence on X-ray structure – Best results when incorporating crystal waters • Docked orientations good • Identified most appropriate model set up – Good correlation with actives but ... – Inactives cover range of scores • 2 sub-classes of inactives poorly predicted – Intramolecular terms. PCA analysis of docking scores Correlation of Docking Score with pKi (N=21) pKi vs FlexX score -1.5 -2.0 -2.5 pKi -3.0 -3.5 -4.0 -4.5 -5.0 -5.5 -60 -50 -40 D7W X -30 -20 Virtual Screening • Screening Sets – In-house available reagents: 3425 compounds after filtering • Dock into best model from each X-ray structure • Data analysis – Initial visual inspection of predicted binding mode – Clustering of structures – Further visual inspection and selection of 100 compounds • 74 compounds available for biological evaluation Frequency of Scores 30 100% 90% 80% 70% 20 60% 15 50% 40% 10 30% 20% 5 10% 0 0% -55 -50 -45 -40 -35 -30 -25 -20 -15 -10 Score -5 0 % cumulative % database 25 D5WX D7WX cum D5 cum D7 Screening results • 8 compounds with % inhibition > 65% @250mM. Cmpd Cmpd1 Cmpd2 Cmpd3 Cmpd4 IC50 mM 31 32 32 54 Cmpd Cmpd5 Cmpd6 Cmpd7 Cmpd8 IC50 mM 88 99 168 620 10% hit-rate with 50-fold reduction in compounds screened. Novel, patentable warheads Uncompetitive inhibition with respect to IMP Thoughts on Model Validation • Validate against known actives • Efficiency (enrichment) – Ratio No. Actives found/No. Hits : No. Actives/DB size • Effectiveness (coverage) – Ratio No. Actives found : No. Actives in DB • Beware of over-fitting – Coverage across structural classes Pharmacophore Hypothesis Validation Enrichment of hits and effectiveness of finding all possible hits. 100 effectiveness 90 enrichment factor 80 70 60 50 40 30 20 10 0 hypo1 hypo2 hypo3 hypo4 hypo5 hypo6 hypo7 hypo8 hypo9 hypo10 by-hand all Docking Model Selection Efficiency 100 100 90 90 80 80 70 70 M1 60 Hit rate Actives (%) Effectiveness M2 M3 50 M4 40 M1 M2 M3 M4 60 50 40 30 30 20 20 10 10 0 0 0 20 40 60 Screen (%) 80 100 0 20 40 60 Screen (%) 80 100 Conclusions • Effective virtual screening strategy established. • Successfully applied to lead generation. • Virtual needle screening powerful method for lead generation. Acknowledgements • Brad Sherborne, Ian Wall, John King-Underwood, Sami Raza • Phil Jones, Mike Broadhurst, Ian Kilford, Murray McKinnell • Neera Borkakoti