Simulations Plus, Inc. · NASDAQ:SLP
Simulations Plus — AIDD and EDDS webinar May 2023
45 slides from Simulations Plus's conference presentation, May 2023
Slide 1: Using AI-driven Drug Design to Shorten Your Drug Development Process
Using AI-driven Drug Design to Shorten Your Drug Development Process Jeremy Jones, Michael Lawless, David Miller, Marvin Waldman, Rafal Bachorz 1 | NASDAQ: SLP
Slide 2: Overview
Overview • How our AI-driven Drug Design module works • Example internal project: PPARy inhibitors • Early Drug Discovery Partnerships 2 | NASDAQ: SLP
Slide 3: De novo drug design at SLP
De novo drug design at SLP • Simulations Plus (SLP) is well known for our user-friendly software and providing expert consulting to support drug discovery, clinical development research, and regulatory submissions. – GastroPlus®: Our mechanistically based simulation software package that simulates I
Slide 4: Automating the de novo drug design process: AIDD
Automating the de novo drug design process: AIDD Initialize with K randomly generated analogs using chemical transforms starting from initial seed molecules Evaluate properties: ADMET_Risk SynthDiff Activity(s) HT-PK 3D shape matching Repeat N times Prune molecules using Pareto optimal layers Genera
Slide 5: 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 ~200 transforms h
Slide 6: Multi-parameter optimization
Multi-parameter optimization • Based on Pareto front calculations – 1000 normally distributed points • 2-dim: 7 Pareto optimal points • 5-dim: ~100 Pareto optimal points • Typically optimize using 4-5 parameters – Typical parameters: • Activity (QSAR model) • Docking Score • ADMET Risk • Synthetic A
Slide 7: ADMET Risk™
ADMET Risk™ Absorption Distribution Toxicity Metabolism 7 | NASDAQ: SLP
Slide 8: High-Throughput Physiologically Based Pharmacokinetic (HT-PBPK) Predictions
High-Throughput Physiologically Based Pharmacokinetic (HT-PBPK) Predictions • The algorithm uses our top-rated Advanced Compartmental Absorption and Transit (ACAT™) model from GastroPlus to simulate dissolution, transit, and absorption in the GI tract. • In a real-world trial, the HT-PBPK approach p
Slide 9: 3D Shape Matching
3D Shape Matching • Generates 100 conformations of each molecule 2 produced in AIDD and compares to reference ligand −𝑘𝑖𝑗 𝑅𝑖𝑗 𝑉 = 𝑝𝑖𝑗 𝑒 structure (X-ray or model) 𝑖,𝑗 – Can also screen a virtual 3D database instead of AIDD • Similarity scores consist of a shape term based on overlap…
Slide 10: Demonstration Time!
Demonstration Time! Seed Query Scaffold
Slide 11: Using AIDD to design PPARy inhibitors
Using AIDD to design PPARy inhibitors • Background: – Peroxisome proliferator activated receptor gamma (PPARy) is a member of the Nuclear Receptor class of transcription factors. – Best known for regulating fatty acid metabolism, adipocyte differentiation, and inflammatory responses. • PPARy agonist
Slide 12: Project Goals and Strategy
Project Goals and Strategy • Goals: – Design high potency inhibitors with good ADMET/PK properties – Pareto optimize: Activity (pIC50), ADMET risk, %Fb, SynDiff, 3D shape matching • Strategy: – First build QSAR models to predict activity using ADMET Modeler • Identify useful data: ensure all data co
Slide 13: PPARy inhibitor dataset
PPARy inhibitor dataset • Classification model (Yes/No) – 210 (127 mine) data points with categorical data (inhibit PPARy activity <> 10uM) – 90 (43%) neg, 120 (57%) pos • Regression model (predict IC50) – 113 (62 mine) data points with reliable IC50 data • Dominated by 2-3 classes – Indole sulfonam
Slide 14: Classification model
Classification model Using ADMET Modeler, I built many ANNE models using random stratified sampling to divide the data and tested reproducibility with random seeds. The best model was chosen for its high sensitivity and specificity in both the training and test sets. 14 | NASDAQ: SLP
Slide 15: Regression Model
Regression Model Using ADMET Modeler, I built many ANNE models using random stratified sampling to divide the data and tested reproducibility with random seeds. The best model was chosen for its high r2 and low RMSE scores in both the training and test sets. 15 | NASDAQ: SLP
Slide 16: Time to Design Some Molecules!
Time to Design Some Molecules! • Goals: – In general, derive novel, high potency inhibitors with good ADMET/PK properties – Pareto optimize: Activity (pIC50), ADMET risk, %Fb, SynDiff, 3D shape matching • Strategy: Try several AIDD approaches – Modify side chains from best class (indole sulfonamides
Slide 17: Strategy 1: Directed side chain optimization
Strategy 1: Directed side chain optimization • Seed molecule: core scaffold of sulfonamide + aromatic substituted indole • Query: Scaffold that forces aromatic substitutions on the indole N (with 1 C linker) and C2, allows any C substitution on sulfur • Optimization Parameters: – ADMET risk (0.9), %
Slide 18: Strategy 2: Scaffold hopping
Strategy 2: Scaffold hopping • Seed molecule: most potent indole sulfonamide • Query: Scaffold that forces aromatic ring to maintain orientation of side chains (SMARTS string or use MedChem Designer) • Optimization Parameters: – ADMET risk (0.9), %Fb (95%), syn diff (2.5), S+Sw (10), PPARGi_pIC50 (9
Slide 19: Strategy 3: 3D shape matching
Strategy 3: 3D shape matching • Reference ligand: Indole sulfonamide from PDB:2HFP • Various indole sulfonamide seed, no query molecule • Optimization Parameters: – 3D Tanimoto similarity (0.9), ADMET risk (0.9), %Fb (95%), syn diff (2.5), PPARGi_pIC50 (9.5) • Results: – Very good results in terms o
Slide 20: Selecting compounds for study
Selecting compounds for study • Overall goal: – Derive novel, high potency inhibitors with good ADMET/PK properties • Process: – Combine compounds from all AIDD runs that meet the following criteria into one file: • pIC50>6 (IC50< 1uM), ADMET risk < 6, bioavailability > 40% , solubility > 0.05mg/mL,
Slide 21: Success stories are on the way!
Success stories are on the way! 21 | NASDAQ: SLP
Slide 22: Large pharma company collaboration
Large pharma company collaboration • The partner company entrusted Simulations Plus with compound structure and activity data from one of their ongoing drug discovery programs. We worked with the partner’s team to define the multi-objective parameters against which the lead molecule(s) needed to be
Slide 23: Two new drug discovery partnerships
Two new drug discovery partnerships 23 | NASDAQ: SLP
Slide 24: What does a typical discovery project look like?
What does a typical discovery project look like? Project Schedule (weeks) 0 1 2 3 4 5 6 7 8 Review client data and target Curate relevant literature/data Build or integrate QSAR/QSPR models Initial medchem assessment Initial AIDD generation Novelty/availability search Additional AIDD as necessary Pr
Slide 25: Take home messages
Take home messages • AI approaches can accelerate the drug discovery and optimization process. • The results are only as good as the data and models on which they’re built. • The Simulations Plus AIDD module within the ADMET Predictor platform is a user-friendly, end-to-end drug discovery and optimi
Slide 26: TENA KOUTOU KATOA!
TENA KOUTOU KATOA! • Michael Lawless • David Miller • Rafal Bachorz • Cheminformatics Team 26 | NASDAQ: SLP
Slide 27: 27 | NASDAQ: SLP
SimulationsPlus 27 | NASDAQ: SLP Cognigen DiLisym Services | Lixoft
Slide 28: PCA of model compounds vs this 3DSM run vs best 126
PCA of model compounds (0) vs this 3DSM run (1) vs “best 126” (2) Compiled a list of 253 “best” compounds from all AIDD runs Only 2 duplicates! pIC50>6, ADMET risk <7, %Fb>80, syndiff<5, S+Sw>.005 28 | NASDAQ: SLP
Slide 29: PLS similarity model results
PLS similarity model results • As Michael suspected, current rules don’t appear to allow creation of sulfonamide – Edit rules: Jeremy_moleculeTransforms.crf • REACTIONNAME Add_SO2C • REACTIONCLASS ADD_FUNCTIONAL_GROUP • SMK [N;_H1:1]>>[N:1]S(=O)(=O)C • $$$$ – We do get phenyl added directly to S • A
Slide 30: PLS similarity model results
PLS similarity model results • New rule allows desired addition • AIDD now produces something very similar to best compounds – 2-8-23Similar2 30 | NASDAQ: SLP
Slide 31: PLS similarity model results
PLS similarity model results • Pare back to C to see if N + sulfone +phenyl added – Did not find entire addition, and not where I wanted it – Did get carbonyl + N + sulfone, in a different place 31 | NASDAQ: SLP
Slide 32: – Corey began by tinkering with substitutions on the S of the sulfonamide (series
– Corey began by tinkering with substitutions on the S of the sulfonamide (series 2) then dropped the indole phenyl ether and tinkered further with S-substitution (early series 4) and N-indole substitution (late series 4) for novelty’s sake. – With me, brought back phenyl ether and focused on indole
Slide 33: Protein-ligand interaction
Protein-ligand interaction • Compound 2a was co-crystallized with the ligand-binding domain of the protein (PPAR-c LBD) and the co- activator peptide fragment (SRC-1) and subjected to X-ray structure determination (PDB-id, 2HFP). • Interestingly, two molecules of compound 2a were seen to span the bi
Slide 34: Activity cliffs
Activity cliffs p I C 50 6.54 8.52 7.04 8.70 Phenyl ether is important I suspect differences are due to lab assay variance (show same trend with other substitutions) 34 | NASDAQ: SLP
Slide 35: Corey’s thought was that only meta substituted phenyl really affected activity. Looking back, not
7.222 8.523 8.699 9.0 Corey’s thought was that only meta substituted phenyl really affected activity. Looking back, not sure why we didn’t use the ester substituted phenyl for KVA class optimization. 35 | NASDAQ: SLP
Slide 36: Activity cliffs
Activity cliffs p I C 50 5.03 7.10 36 | NASDAQ: SLP
Slide 37: KVA-E series SAR
KVA-E series SAR 7.036 5.416 6.614 5.973 5.482 6.529 Meta >> para CF3 > OCH3~CH3 > Halogen(s) 37 | NASDAQ: SLP
Slide 38: History of targeting PPARs
History of targeting PPARs • PPAR gamma agonists have long been used to treat Type II diabetes – Exactly how they work is not completely understood, but it involves increased lipid storage in fat depots, which helps regulate blood sugar and insulin sensitization – Several were approved by the FDA an
Slide 39: Screening strategy
Screening strategy PPRE FF luc • Started by screening targeted collections from collaborators • Quantified activity/selectivity against PPAR A/G/D using luciferase reporter assays • Determined effects against relevant cancer Antag Agonist models in culture • Confirmed effects against endogenous gene
Slide 40: Synthetic Accessibility/Difficulty
Synthetic Accessibility/Difficulty SA Ertl Synth Diff Score = fragment_Score – complexity_Penalty Training ~1 million ~47 million Fragment Heavy Atoms Outer Layer Any aromatic vs. aliphatic Macrocycles Complexity Same Same frequencies Stereocenters Range 1-10 0-10 Spiro centers Bridges Ertl and Shuf
Slide 41: PPAR agonists
41 | NASDAQ: SLP ZC CH; PPARo/PPARy dual agonists ° fenofibrate [ECsy 30 1M] ! bezafibrate [Em 50 uM] a H CH; } gemfibrozil [ECso 190 uM] y- rere som sitet olay (a) 4 uM; (y) 150 nM] PPAR, agonists H ° \- ° 5 N20 $ ~ Ss Ha we o. fis CH wok HO a Hy OH CH, \, y, 15d-PGJ2 troglitazone ciglitazone [ECs
Slide 42: History of targeting PPARs
History of targeting PPARs 42 | NASDAQ: SLP Cheng, et al 2019
Slide 43: PPARG pathway activation in bladder cancer
PPARG pathway activation in bladder cancer 20-30% of advanced bladder cancers have PPARγ pathway alterations. • Biton, Cell Reports (2017) • In silico analysis of BlCa transcriptomes reveals PPARG as a major driver of luminal cancers. • Halstead eLIFE (2017) • TCGA analysis suggests that hyperactive
Slide 44: Strategy 4: 3D shape matching
Strategy 4: 3D shape matching • Reference ligand: Indole sulfonamide from PDB:2HFP • No Query molecule: • Optimization Parameters: – 3D Tanimoto similarity (0.9), ADMET risk (0.9), %Fb (95%), syn diff (2.5), PPARGi_pIC50 (9.5) • Results: – Very good results in terms of similarity, activity, and ADME
Slide 45: Background
Background • History of Computer-Aided 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 des