Hit to Lead Discovery of Benzylpiperidine Acetylcholinesterase Inhibitors Using Generative Models: a Retrospective Case Study Nazim Medzhidov, Ph.D & Joshua Owoyemi, Ph.D Elix, Inc. Chem-Bio Informatics Society (CBI) Annual Meeting 2022, Tokyo, Japan | October 26th, 2022

2 Background ● Challenges associated with the traditional drug discovery process have facilitated the application of machine learning approaches in this domain. ● Generative AI approaches for molecular design are actively investigated. ● Evaluating generative models in silico is challenging, confirmation requires experimental validation (expensive) ● Majority evaluate model performance based on optimizing computable properties (logP, QED, SA score, etc.) ● How to select generated candidates efficiently? Objectives ★ Design a scenario and a pipeline to evaluate generative models in silico: ○ Hit-to-lead campaign ○ Novel chemotype discovery ★ Test our Elix Discovery™ Platform ★ Candidate prioritization pipeline

3 Study Workflow AI Model Development Dataset Preparation Post-processing & Prioritization Result Analysis ● Pre-training dataset: ○ ChEMBL ○ AChE inhibitors removed ○ Target chemotype scaffold removed ● Training Set: ○ AChE inhibitors from ChEMBL ● Elix Discovery™ Platform ● Elix Predict: ○ AChE inhibitory activity prediction model ○ Blood Brain Barrier (BBB) Permeability prediction model ● Elix Create: ○ SmilesFormer Generative Model ○ 10 sampling runs ● 30K molecules generated in each of 10 sampling runs ● Post-processing: ○ Phys-Chem Filters (RO5) ○ MCF filters ○ Novelty ○ BBB Permeability ● Prioritization: ○ QED score ○ Predicted activity ○ Binding affinity (docking) ● Quality assessment: ○ Target scaffold discovery ○ Documented potent compound discovery ● Short list of best 200 molecules from each run ● Final short-list of 20 most frequently selected best compounds

4 Dataset Preparation

5 Datasets: Acetylcholinesterase inhibitors ChEMBL (~2.2M) Training dataset (1076): A + B + C AChE Inhibitors with IC50 values (4,238) AChE inhibitors before 1992 (120) More recent molecules with same chemotypes present in A (847) A Pre-training dataset (~2.2M): Physostigmine Tacrine Rivastigmine B C Hit & hit expansion compounds (109) D Molecules containing piperazine, piperidine or indan (357) (Hidden) Established chemotypes before 1992 AChE inhibitors removed (15.5K mols) 48 mols with the scaffold removed B 1992 First appearance of donepezil chemotype in ChEMBL database A B C D (Hit compound) (Target chemotype) A

6 Datasets: Chemical Space Visualization Physostigmine Tacrine Rivastigmine Hit Compound Target Chemotype

AI Model Development 7

8 Elix Discovery™ Platform Generative Model ● SmilesFormer ○ Pre-trained on ChEMBL dataset without AChE inhibitors and target scaffold ○ Trained on: datasets A + B + C (1076 samples) ● Multiobjective Optimization Problem: ● SA score ● QED score ● Favorable physical-chemical properties ● Novelty (distance from the training set) ● Activity Predictive Models ● AChE inhibitory activity prediction model: ○ GCN ○ Trained on: datasets A + B + C (1076 samples) ● Blood Brain Barrier (BBB) Permeability prediction model: ○ GCN ○ Trained on an in-house dataset (9059 samples)

Molecule Generation, Post Processing & Prioritization 9

10 Generation strategy and post-processing pipeline 30K mols / run Run 1 Run 2 Run 5 Run 3 Run 4 Run 6 Run 7 Run 10 Run 8 Run 9 Random Sampling One seed Group seed No seed 1 2 3 Filtering 5 6 ~4000 mols / run ● RO5 ● MCF ● Novelty ● BBB Permeability Run 1 Run 2 Run 5 Run 3 Run 4 Run 6 Run 7 Run 10 Run 8 Run 9 Prioritizing 5 6 ● QED ● Predicted activity ● Binding affinity (docking) 200 mols / run Run 1 Run 2 Run 5 Run 3 Run 4 Run 6 Run 7 Run 10 Run 8 Run 9 Aggregation 20 most frequently selected candidates ● Recommendation score: ○ Consistency of selection ○ Min = 1, Max = 10

Results 11

12 Discovering reported potent scaffold and molecules Reported scaffold discovery success (number of runs) Reported potent compound discovery success (number of runs) No Seed 0 / 10 0 / 10 One Seed 4 / 10 5 / 10 Group Seed 9 / 10 9 / 10 D Molecules from hidden dataset D (target chemotype) containing represented substructures Reported potent scaffold A B Molecules rediscovered with One Seed setting Molecules rediscovered with Group Seed setting IC50 = 81 nM Rank = 8 IC50 = 6.7 nM Rank = 31 IC50 = 58 nM Rank = 107 IC50 = 94 nM Rank = 166 IC50 = 81 nM Rank = 56 IC50 = 30 nM Rank = 71 IC50 = 6.7 nM Rank = 124 IC50 = 94 nM Rank = 393 Random Sampling

13 Final 20 Candidates by Recommendation Score No seed One seed Legend: top 1% recommendation score (max = 10)

14 Conclusion ● Designed a retrospective case study of novel chemotype discovery for generative models (quality assessment) ● Tested our Elix Discovery™ platform in a hit-to-lead discovery campaign ● Given an early hit compound, optimized the scaffold to a more complex diverse scaffolds including a reported potent indanone-piperidine scaffold ● Multiple sampling runs and recommendations score analysis helped to focus on consistently top ranked candidates ● Among the prioritized candidates, reported indanone-piperidine containing potent molecules were discovered ● These molecules were included in the top 1% of the generated molecules ● Final 20 top ranked candidates included at least one known potent AChE inhibitor ● Potential presence of yet unknown potent compounds among final recommendations

www.elix-inc.com

APPENDIX 16

17 Final 20 candidates: One seed vs Group Seed Group seed One seed

18 Chemical Space Visualization Physostigmine Tacrine Rivastigmine Hit Compound Target Chemotype