Generator Reward Function Optimization for Non-covalent Pan-KRAS Inhibitors Casey Galvin, Tasuku Ishida, Masakazu Atobe Elix, Inc. CBI Annual Meeting 2024 - October 31, 2024

Case Study: Generation of Non-covalent, Pan-KRAS Inhibitors 2 Copyright © Elix, Inc. All rights reserved. Problem Setting KRAS mutations are associated with many cancers, but has poor ligandability. After several decades of work, covalent inhibitors against KRAS finally reached market. However, covalent inhibitors target only one mutant (G12C). Therefore, we need non-covalent inhibitors that target multiple mutants. Image Sources: Fell, J. B. et al., J. Med. Chem., Vol. 63, pp. 6679-6693 (2020), Kim, D. et al., Nature, Vol. 619, pp. 160–166 (2023) Our input data is only the crystal structure of Adagrasib with G12C KRAS, as well as common molecular descriptors and filters. Input Data 1. Determine optimal reward function structure for generative models. 2. Identify generated structures that may be good candidates. Study Objectives

3 Evaluation Metric: Docking Score Correlation between activity and docking score of published inhibitors against G12C, G12D, G12V, WT. ● Published activity data from Eurofins KRAS assays ● Docking simulations performed on Elix Discovery™ using AutoDockVina ● Scores averaged over 3 docking simulations using RDKit and Gypsum as conformer generators Evaluation Metric: Maximize the number of generated molecules with docking score < -12.0 kJ/mol against Adagrasib Goal: Identify generator reward function that gives best performance based on evaluation metric Note: these data are not used in any of the generators, and only used to evaluate molecules after generation https://www.eurofinsdiscovery.com/solution/ras-drug-discovery https://www.eurofinsdiscovery.com/catalog/kras-wt-human-neucleotide-exchange-assay-leadhunter-assay-tw/125970 https://www.eurofinsdiscovery.com/catalog/kras-g12c-human-neucleotide-exchange-assay-leadhunter-assay-tw/125960 https://www.eurofinsdiscovery.com/catalog/kras-g12d-human-neucleotide-exchange-assay-leadhunter-assay-tw/125950 Evaluation Metric R2 = 0.66

Proposed Method • Integrate generative model with docking simulations and pharmacophore modeling based on Adagrasib. • Note: Activity data is not used as an input during generation. Generative Model and Reward Function • Using ReINVENT architecture, which has reinforcement learning mechanism, as implemented in Elix Discovery™. • Docking simulation using crystal structure for covalent inhibitor Adagrasib (MRTX-849; PDB ID = 6ut0). • Pharmacophore model based on same crystal structure, and edited to remove covalent interactions. Reward Function Construction 4 Copyright © Elix, Inc. All rights reserved. Molecular Descriptors Molecular Filters Reward Score + or x + or x = x 0 or 1 0 1 0 0 1 0 [0, 1] 1 Docking Pharmacophore

Reward Function Examples 5 Copyright © Elix, Inc. All rights reserved. Molecular Descriptors Molecular Filters Reward Score + or x + or x = x 0 or 1 0 1 0 0 1 0 [0, 1] 1 Docking Pharmacophore Reward Score = … Filters * mean(Descriptors) Filters * mean(Descriptors, Docking Score) Filters * mean(Descriptors) * Docking Score Filters * mean(Descriptors, Pharmacophore Score) * Docking Score Filters * mean(Descriptors) * Pharmacophore Score * Docking Score Which of these reward functions maximizes the number of generated molecules with a docking score < -12?

Reward Function Examples 6 Copyright © Elix, Inc. All rights reserved. Molecular Descriptors Molecular Filters Reward Score + or x + or x = x 0 or 1 0 1 0 0 1 0 [0, 1] 1 Docking Pharmacophore Reward Score = … Filters * mean(Descriptors) Filters * mean(Descriptors, Docking Score) Filters * mean(Descriptors) * Docking Score Filters * mean(Descriptors, Pharmacophore Score) * Docking Score Filters * mean(Descriptors) * Pharmacophore Score * Docking Score Which of these reward functions maximizes the number of generated molecules with a docking score < -12?

7 Results: Reward Function Optimization Filters * mean(Descriptors) Filters * mean(Descriptors) * Docking Score Filters * mean(Descriptors, Pharmacophore Score) * Docking Score Filters * mean(Descriptors) * Pharmacophore Score * Docking Score

8 Analysis of Generated Molecules Adagrasib Docking Simulations (RDKit and Gypsum) RDKit Gypsum without Tautomers Gypsum with Tautomers Mean -12 -12.2 -12.2 -12.1 +/- 0.1 A Promising Molecule Applied post-generation filtering ● MW > 450 ● Total number of rings = [3, 5] ● Fraction sp3 bonds = [0.2, 0.4]

9 A Promising Molecule The molecule also exhibits interactions known to be important for Adagrasib. RDKit conformer Ligand ILE100 VAL9 VAL103 MET72 TYR96 THR58 GLU62 Adagrasib (MRTX 849) o o o o o o o Docking pose comparison with Adagrasib in G12C shows that the torsion angle is replicated Cyan = Adagrasib, Grey = generated molecule

10 A Promising Molecule The molecule also exhibits interactions known to be important for Adagrasib. RDKit conformer Ligand ILE100 VAL9 VAL103 MET72 TYR96 THR58 GLU62 Adagrasib (MRTX 849) o o o o o o o Docking pose comparison with Adagrasib in G12C shows that the torsion angle is replicated This molecule seems like a promising candidate ● Docking score < -12 ● Significant coverage of important interactions ● Similar pose as Adagrasib So how does it perform against the other crystal structures?

11 A Promising Molecule Docking Simulation Docking Score (mean) G12C w/ BI-2865 -12.2 G12D w/ BI-2865 -12.5 G12V w/ BI-2865 -12.4 WT w/ BI-2865 -12.0 G12D w/ MRTX-1133 -12.5 Mean of all above -12.3 Comparison with BI-2865 Comparison with MRTX-1133 Ligand ILE100 VAL9 VAL103 MET72 TYR96 THR58 GLU62 MRTX1133 (non-covalent) o o o o o o o

12 Conclusion Reward Function Optimization ● We identified the combination of docking simulation and pharmacophore model that maximizes the yield of molecules with target docking scores generated by a reinforcement learning based archiecture ● Using a different generative architecture will likely lead to different results Identification of Promising Molecules ● Using our evaluation metric as a way to prioritize generated molecules, we identified a scaffold that exhibits a good docking score, comparable pose to the reference molecule (Adagrasib), and coverage of important interactions ● This molecule also performed well in docking simulations based on non-covalent inhibitors, suggesting an IC50 in the nM range Next Steps ● Can screen commercially available compounds for structurally similar compounds for in vitro testing ● Further in silico optimization using generators designed for side chain generation and ring replacement ● Understand the interplay between docking simulations and pharmacophore modeling

www.elix-inc.com

14 Results: Examples of Generated Molecules (6/6) G12C/G12D docking = -12.9/-13.2 kcal/mol G12C/G12D docking = -13.6/-13.6 kcal/mol G12C/G12D docking = -12.2/-12.6 kcal/mol MRTX-849 (covalent inhibitor) Used for crystal structure BI-2865 (pan-KRAS non-covalent) WT IC50 = 13 nM G12C/G12D docking = -10.97/-10.77 MRTX-1133 (non-covalent inhibitor) G12D IC50 = 0.14 nM G12C/G12D docking = -11.63/-12.33

15 Results: Chemical Diversity of Generated Molecules Diverse Chemical Space Exploration ● Molecules with docking scores < -12 from three sets of generated molecules were compared on Elix Discovery™. The same generator and reward function was used for each set. The generator used a multiplicative docking and pharmacophore reward. ● TMAP groups molecules based on their chemical structure. The different colors in the plots corresponds to the different sets of generated molecules. ● The separation of the colors indicates that for each run of the generator, a different area of the chemical space is explored.

16 Does Adagrasib Docking Score Predict Other Scores? # of Docking Scores Below Threshold -11.5 5 6 4 7 3 4 2 4 1 7 0 655 # of Docking Scores Below Threshold -11.5 5 6 4 7 3 4 2 4 1 7 0 12 All molecules Molecules with generator docking score < -11.5