Transcript Very Large Scale Virtual Screening of Chemical Databases
DockCrunch and Beyond...
The future of receptor-based virtual screening
Bohdan Waszkowycz, Tim Perkins & Jin Li Protherics Molecular Design Ltd Macclesfield, UK
Outline
Structure-based virtual screening – an achievable (and possibly useful) tool for drug discovery – the DockCrunch validation study Protherics’ experience since DockCrunch – methods: making VS a routine task – analysis: getting the most from your data – the future (and beyond)
Virtual Screening
compound collections virtual libraries computational screening targeted selection screen smaller focused libraries molecular docking receptor structure
Why Use Molecular Docking?
Most detailed representation of binding site – overcomes simplifications of pharmacophores – identify both conservative and novel solutions – impetus for de novo design/optimisation Broad range of analyses applicable – diverse scoring/selection criteria Quality/throughput of available methods – good enough, despite technical limitations
DockCrunch
Validation study for large-scale virtual screening – flexible ligand/rigid receptor docking – PRO_LEADS docking code using ChemScore scoring function – 1.1M druglike ACD-SC compounds – dock versus oestrogen receptor (agonist and antagonist structures) – collaboration with SGI
Oestradiol:Oestrogen Receptor Complex
DockedEnergy Profiles
Agonist Receptor Antagonist receptor 25 30 20
ACD-SC Agonists
25
ACD-SC Antagonists
20 15 15 10 10 5 5 0 -60 -50 -40 -30 -20
Docked Energy (kJ/mol)
-10 0 -60 -50 -10 -40 -30 -20
Docked Energy (kJ/mol)
• Achieve good separation in terms of predicted binding affinity
DockCrunch Results
Demonstrated technical feasibility – 1.1M cpds docked in 6 days/64 processor Origin – implemented automated pre- and post-processing Demonstrated potential for lead identification – successful discrimination of seeded known hits – activity for 21 out of 37 assayed compounds – ER binding affinities to 7nM Ki – novel non-steroidal chemistries
Since DockCrunch...
VS established as a routine CAMD task: – 2.2M structures docked in DockCrunch – 1.5M docked versus in-house target – 2.5M docked to date in external contracts – project 1: 0.25M Dec 2000 – project 2: – project 3: – project 4: – project 5: 0.25M Jan 2001 1M Feb 2001 1M March-April 2001 0.5M to do in May... – diverse targets/databases/project objectives
Virtual Screening within
Prometheus
Commercial databases Database preparation e.g. salt removal, protonation Virtual databases Database pre-filtering select drug-like profile Receptor structure Receptor-ligand docking predict binding mode/affinity Analysis graphical browsing, subset selection
PRO_LEADS Docking
Tabu search + extended ChemScore function – robust prediction of binding free energy – 85% success rate achieved across diverse test set Pre-calculated grids for energies/neighbour lists – defines extent of binding site – automatically/graphically defined Selection of PRO_LEADS docking protocol – use standard protocol across all receptors – specific constraints or modified energy terms available if desired
Example of Grid Definition
cAMP-dependent kinase (1YDS) contact surface coloured by lipophilicity
Docking Throughput
Standard protocols take 1–5 mins/ligand – e.g. typical VS run at ~4 min for 3M tabu steps – 250k cpds/week on 100 processor Linux cluster (VA Linux 750MHz PIII) PLUNDER script for parallelization – automatic processing of ligand batches – balances processor workload – works across heterogeneous architectures – supplies running time statistics – handles hardware failures
Data Analysis and Subset Selection
Intrinsic problems of scoring functions: – cannot parameterize all critical interactions – try to take account of induced fit effects – calibrated only versus good binders – ignore co-operativity in binding When applied to random datasets: – predicted affinity typically normal distributed – overestimates binding affinity of random set energy alone not ideal for subset selection
Achieving Better Selection
Need to supplement scoring function – consensus scoring schemes Explore more fundamental descriptors of receptor:ligand complementarity – capture characteristics of diverse receptor types – assess deficiencies of existing scoring functions – use as simple filters or as pseudo energy terms
Enrichment Rates
Effect of different selection criteria for ER set for recovery of seeded compounds
Selection criteria
none DockedEnergy < –30 kJ/mol DockedEnergy < –35 kJ/mol DockedEnergy < –40 kJ/mol DockedEnergy < -30 kJ/mol plus complementarity descriptors
% of total dataset
100.0
Reference ligands recovered
20 19.0
5.8
1.0
20 18 15
Enrichment Rate
1 5 15 76 0.9
20 108
Requirements for Analysis Package
VS generates huge data output – want to be able to browse through entire dataset Real-time navigation of large datasets – graphing property distributions – selections based on property filters – browsing of 3D models within selections – initiating additional property calculations – data transformations – writing subset/reports
PropertyViewer
Approach to Analysis
1. Preliminary exploration – browse property distributions – comparisons with known ligands 2. Initial elimination of poor structures – DockedEnergy, component energies – DE corrected for size/functionality – receptor:ligand steric complementarity – polar/lipophilic surface complementarity
Approach to Analysis
3. Further filtering define focused subsets – tighter 2D property filters – clustering by 2D chemistry – presence of key 3D binding interactions – specific H-bonds, specific lipo contacts, pocket occupancy, volume overlap with reference ligand/fragment, etc – similarity/diversity of 3D binding mode – 3D similarity descriptors – final ranking by DockedEnergy or hybrid energy/complementarity scoring function
DockedEnergy vs Size
Complementarity Space
ER and FXa datasets
Addressing More Difficult Cases - COX2
Knowns show clustering in property space despite modest DockedEnergy
Improvements in Docking Function
original docking function some misdocked knowns new docking function more consistent docking +ve shift in random energies
Comparison of filters in subset selection
87% pass 2D filters 37% pass energy filters Initial filtering to ~10% – energy filters – complementarity – 2D properties 43% 22%
12%
1% 2% 9% 22% pass complementarity filters 0% Selection of final ~1% subset – 3D structural features – preferred binding motifs – 2D/3D diversity
Conclusions
Established VS as a routine CAMD task – focused software development – achieved success in drug discovery projects VS is more than a black box – data mining is worthwhile – explore receptor-ligand complementarity to achieve good subset selection and point towards better scoring functions
Future Directions for VS
Exploit expanding computing resource – improved docking/scoring functions – improved receptor representations Broader application of VS – evaluation of drugability of early targets – screening of very large virtual libraries – routine screening across protein families – DMPK issues
Acknowledgements
Tim Perkins Richard Sykes Richard Hall David Frenkel David Sheppard Jin Li Martin Harrison Carol Baxter Chris Murray Thanks to: SGI, MSI, MDL, VA Linux http://www.protherics.com/crunch/