Simulations Plus, Inc. · NASDAQ:SLP
Simulations Plus — Technical Presentation (September 2020)
39 slides from Simulations Plus's technical presentation, Sep 30, 2020
Slide 1: Delivering on the promise of AI-driven drug discovery with ADMET Predictor 10
Delivering on the promise of AI-driven drug discovery with ADMET Predictor® 10 Background and application examples Marvin Waldman Sr. Research Fellow 30-Sep-2020
Slide 2: Motivation
Motivation • In-house project to develop novel anti-malarial compounds as proof of concept to use in-silico models to design compounds
Slide 3: Compound Design Protocol
Compound Design Protocol Activity Models Analog Generation
Slide 4: Compound Design Protocol
Compound Design Protocol Predict/Measure Properties Conclusion: “It seems likely that further iterations and in vivo characterization would be productive.” Can this be automated?
Slide 5: Background
Background • History of Computer-Based (“De Novo”) Drug Design ‒ Early 1990’s • Structure (Receptor) Based ‒ Ludi, MCSS/Hook, Sprout ‒ Late 1990’s – Rule of 5, ADMET, Drug Design is multi-objective • “It ain’t just activity anymore” ‒ Early to late 2000’s • Multi-objective ligand and structure-based
Slide 6: Generating Virtual Molecules
Generating Virtual Molecules • Elementary “Transforms” ‒ Modify molecules with small changes or fragment additions/deletions • Change/Add single atom or bond • Add common fragments ‒ Carboxylic acid, sulfonamide, phenyl, etc. • Synthetic Feasibility? ‒ Filter molecules based on structural alerts, e.
Slide 7: Molecule Selection/Scoring Approaches
Molecule Selection/Scoring Approaches • Multi-objective Criteria ‒ Weighted Sum or “Combining” Function ‒ Pareto optimal ‒ Criteria • Activity/Docking Scoring • ADMET liabilities • Similarity to known active/lead • Synthetic Accessibility • Drug-likeness • Chemical Filters
Slide 8: AIDD Workflow
AIDD Workflow Initialize with K randomly generated analogs using chemical transforms starting from initial seed molecules Evaluate properties: ADMET_Risk SynthDiff Activity(s) HTPK Prune molecules using Pareto optimal layers Repeat N times Generate M more analogs using chemical transforms and random
Slide 9: Generating Analogs
Generating Analogs • Uses a library of chemically “intelligent” SMIRKS transforms ‒ Example: Non-fluorine_to_fluorine • Simple version: [!#9:1]>>[#9:1] • Problem (Need to avoid) X Highly reactive acid halide • Improved SMIRKS: [!#9;D1_S$(*~[#6])!$(*C=[O,N,S]):1]>>[#9:1] • Currently ~150 transforms
Slide 10: Transforms
ij SimulationsPlus SCIENCE + SOFTWARE = SUCCESS
Slide 11: Properties
~50 Built-in models %Fa, %Fb Synthetic Difficulty+ User Models Properties
Slide 12: Compromise Solution
Compromise Solution • Multi-objective, but combine related objectives (e.g. ADMET) into one : ADMET_Risk™ • Typical/Recommended use: 4-5 objectives ‒ ADMET_Risk ‒ Synthetic Difficulty ‒ 1-2 activity models (e.g., activity and selectivity) ‒ Good PK (e.g, bioavailability)
Slide 13: Pareto Selection
Pareto Selection • Pareto Results ‒ 1000 normally distributed points • 2-dim. : 7 Pareto optimal points • 5-dim. : ~100 Pareto optimal points Too Many objectives leads to too many molecules! Most of them are good in just one or a few properties.
Slide 14: ADMET Risk
ADMET Risk™ Absorption Distribution Toxicity Metabolism
Slide 15: PBPK Simulation: Methodology
PBPK Simulation: Methodology ACAT™ Model* + Compartmental Model Pure in silico simulation No exptl properties required Extremely rapid, multi-threaded * Advanced Compartmental Absorption and Transit Model
Slide 16: Synthetic Accessibility/Difficulty
Synthetic Accessibility/Difficulty SA Ertl Synth Diff Score = fragmentScore – complexityPenalty Training ~1 million ~47 million Heavy Atoms Outer Layer Any aromatic vs. aliphatic Fragment Complexity Same Same Macrocycles frequencies Stereocenters Range 1-10 0-10 Spiro centers Bridges Ertl and Shuffe
Slide 17: Comparison: SAScore vs. SythDiff
Comparison: SAScore vs. SythDiff Ertl and Shuffenhauer, J Cheminf, 2009, doi:10.1186/1758-2946-1-8 Li et al., J Cheminf, 2018, doi:10.1186/s13321-018-0287-6
Slide 18: Distribution of SynthDiff Scores
Distribution of SynthDiff Scores WDI Focused Set of 2260 compounds
Slide 19: SynthDiff+
SynthDiff+ • SynthDiff was designed to “score” real molecules • Performance on virtual molecules sometimes too “optimistic” ‒ Add additional penalties based on drug-likeness filters • Brenk et al., ChemMedChem, 3, 435 (2008) • Rishton, Drug Disc. Today, 2, 382 (1997) ‒ Maximum penalty of 4 ‒ Augment
Slide 20: Early Results
Structure OBJ_HIVI-ST ADMET_Risk SynthDiff 12.799 3.000 4.256 “o. 6.525 3.956 2.313 Boa 12.248 3.500 4.181 “ke A , 12.163 6.000 4.132 en i 11.952 4.000 4.082 gare ° 11.051 1694 3.674, SimulationsPlus SCIENCE + SOFTWARE = SUCCESS
Slide 21: Out-of-Scope Predictions
Structure OBJ_HIVI-ST ADMET_Risk_|SynthDiff_|HIVI-S My S28-40. bese So 12.799 6.525 12.248 12.163 11.952 11.051 3.000 3.956 3.500 6.000 4,000 1.694 4.256 2.313 4.181 4.132 4.082 3.674 12.799 E : = = = SimulationsPlus SCIENCE + SOFTWARE = SUCCESS
Slide 22: Applicability Domain
Applicability Domain 10% Buffer Training Set Range
Slide 23: After Applying Penalties
Structure OBJ_HIVI-ST__|ADMET_Risk |SynthDiff | HIVI-ST | HIVI-ST+ |ADMET_Risk+ |SynthDiff+ 12.799 3.000 4.256 12.799) 2.799 14.000 8.253 6.525 3.956 2.313 6.525) -3.475 11.956 6.312 12.248 3.500 4.181 12.248) 2.248 13.500 7.897 12.163 6.000 4.132 12.163) 2.163 15.000 7.848 11.952 4.000 4.082 11.952
Slide 24: Flip Side of the Coin: Capping Values
Flip Side of the Coin: Capping Values Trivially Simple Molecules: Very easy to make Very good in one objective Assigning a capping value tends to filter out such molecules. The capping value is assigned as the result when the actual result is “better”, because this value is “good enough”.
Slide 25: Applying Capping Values: Example
Applying Capping Values: Example Molecules on right “dominate” molecules on left after applying capping to SynthDiff
Slide 26: Restricting Chemical Space
Restricting Chemical Space User Definable Scaffold • Drawn using MedChem Designer • Specified via SMARTS Filtering Criteria • Default file supplied • Can be modified/replaced or omitted • SMARTS patterns or more general queries SLQ [#6]C(=O)N U >= 4 SCORE 1 NPQ AtomCount >= 66 SCORE 1
Slide 27: Scaffold Query Examples
Scaffold Query Examples Structure-Based scaffold SMARTS-Based scaffold
Slide 28: Run Parameters
Run Parameters Minimum Size?
Slide 29: Pareto Selection
Pareto Selection • Pareto Results ‒ 1000 normally distributed points • 2-dim. : 7 Pareto optimal points
Slide 30: Pareto Selection – Multiple Layers
Pareto Selection – Multiple Layers If initial Pareto-based selection results in too few molecules, repeat the selection using the remaining candidate molecules => Pareto Layers. Enforced after running 50% of generations. Avoids “bad” molecules entering the population too early.
Slide 31: Summary/Key Points
Summary/Key Points • Highly automated customizable protocol commercially available • ~150 chemically intelligent transforms ‒ Customizable, user-controllable • Property objectives ‒ ~50 built-in property models available including ADMET Risk to combine them ‒ HTPK simulations ‒ User models from ADME
Slide 32: Evaluation
Evaluation @S Simulations: Simulations Plus Partners with Large Pharmaceutical Company to Validate Al-Driven Drug Design Capabilities in ADMET Predictor® New AIDD Module applied to active therapeutic program to provide lead optimization support LANCASTER, Calif.-(BUSIN! Simulations Plus, Inc. (Nasda
Slide 33: Working Example/Demo
Working Example/Demo • Design BACE1 inhibitors starting from a known active “seed” ‒ BACE1 inhibition has been investigated as a potential target for treating Alzheimer’s disease • Use activity model based on 370 known BACE1 inhibitors • Augment ADMET Risk with an additional Risk model blood-brain b
Slide 34: BACE1 inhibitors paper, Journal of Medicinal Chemistry
Journal of Medicinal Chemistry 1,4-Oxazine #-Secretase 1 (BACE1) Inhibitors: From Hit Generation to Orally Bioavailable Brain Penetrant Leads Frederik J. R. Rombouts,*" ‘Gary Tresadern,” Oscar Delgado," Carolina Martinez: Lamenca," . Michiel Van Gool,' Aranzazu Garcia- Molina," Sergio A. Alonso de D
Slide 35: BACE1 Pharmacophore/Scaffold
BACE1 Pharmacophore/Scaffold
Slide 36: Customized Risk Rule for BBB
Customized Risk Rule for BBB Identifier Weight Rule logic BBB 1.0 BBB_Filter = Low OR LogBB < -2{-1} Pgp 1.0 Pgp_Substr = Yes Feat 1.0 ‘Cation > 0.5{0.975} SCIENCE + SOFTWARE = SUCCESS
Slide 37: BACE1 Activity Model
Predicted Performance for 36 inputs and 7 neurons (© Training Paints (N= 207) Test Pomnts (N= 63) 7 3 3 10 observed ALL: Slope=0.816 Intercept=1.259 QSqd=0.851 RMSE=0.418 MAE=0.326 RMSU=0.424 TRAIN; Slope=0.817 Intercey SimulationsPlus SCIENCE + SOFTWARE = SUCCESS
Slide 38: Acknowledgments
Acknowledgments David Miller Robert D. Clark Michael Lawless
Slide 39: For more information, visit our website
For more information, visit our website at: www.simulations-plus.com
Simulations Plus — Technical Presentation (September 2020) on 6ix