Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Molecular Generation of Non-covalent KRAS Inhibitor Candidates Using Machine Learning on Elix Discovery™, Elix, 8th Autumn School of Chemoinformatics, Nara

November 29, 2023

Molecular Generation of Non-covalent KRAS Inhibitor Candidates Using Machine Learning on Elix Discovery™, Elix, 8th Autumn School of Chemoinformatics, Nara


November 29, 2023

More Decks by Elix

Other Decks in Research


  1. Molecular Generation of Non-covalent KRAS Inhibitor Candidates Using Machine Learning

    on Elix Discovery™ 8th Autumn School of Chemoinformatics, Nara Casey Galvin, Ph.D. Elix, Inc. November 29th, 2023 © Elix, Inc.
  2. 2 Elix Discovery™ Capabilities Predictive Models • Classical • Graph

    neural networks Generative Models • FBDD, RL, CLM, Scaffold hopping Docking Simulations • 3rd party engine support Pharmacophore Modeling • Automatic building from PDB Data These features work together on a no-code GUI (JP/EN versions). Stop by our booth to learn more and see demos. Elix Discovery can be customized to specific requirements, including new feature development. Cloud or on-premise deployment with GPU acceleration © Elix, Inc.
  3. 3 Generating Molecules in a Low Data Environment - KRAS

    Case Study Adagrasib (Mirati) Covalent “warhead” https://doi.org/10.1021/acs.jmedchem.9b02052 Can we generate candidates of non-covalent inhibitors when such inhibitors do not yet exist? KRAS mutations are associated with many cancers • inhibitors are only available for G12C mutations • inhibitors use covalent binding • cannot work against other mutants (G12D, etc.) Low data strategy on Elix Discovery™: • Use crystal structure of G12C mutant in complex with covalent inhibitor (Adagrasib, shown at right) • Construct pharmacophore model to guide generators • Assess generated molecules using docking simulations • Compare against baseline of starting scaffold © Elix, Inc.
  4. 5 Pharmacophore Model Interactions © Elix, Inc. Just specify the

    PDB id, and Elix Discovery™ automatically downloads the necessary files and detects interactions in the model
  5. How Accurate is the Model? - Identifying Interactions 6 Compared

    to Mirati’s own analysis Hydrophilic • H-bond: Acrylamide carbonyl x Lys16 amine • H-bond: Pyrimidine N-1 x His95 NE2 • Salt bridge: Pyrrolidine N x Glu62 carboxylate (model specifies H-bond) • H-bond: Cyanomethyl x Gly10 backbone NH Hydrophobic • Naphthyl ring x hydrophobic pocket (Val9, Met72, Phe78, Tyr96, Ile100, Val103) Extra Interactions • 3 sets of hydrophobic interactions Screenshot from Elix Discovery™ © Elix, Inc.
  6. 7 Pharmacophore Model Used for Molecular Generation Removed interactions associated

    with the covalent binding mechanism The pharmacophore model can be used directly in generative models on Elix Discovery™, and is expressed as a normalized pharmacophore score Screenshot from Elix Discovery™ © Elix, Inc.
  7. Comparison of Docking Pose with Crystal Structure 9 Docking simulations

    performed directly on Elix Discovery™ with AutoDock Vina • Napthyl ring is same location and orientation • Pyrrolidine group in same location, with slight differences in rotation • Piperidine group shows the most deviation ◦ The ring is similarly positioned ◦ Cyano group is rotated in opposite direction ◦ Fluoro and carbonyl in same location, but with some spacing Reference pose (pink) Docked pose (white) Docking simulations (scores) can be incorporated directly into generative model reward functions. However, we use docking in this study for analysis after generation. Screenshot from Elix Discovery™ © Elix, Inc.
  8. 11 Baselines and Starting Scaffolds Parameter Value Pharmacophore Score 0.92

    Docking @ G12C -11.1 Docking @ G12D (ref. only) -9.7 Parameter Value Pharmacophore Score 0.92 Docking @ G12C -7.8 Docking @ G12D (ref. only) -8 Parameter Value Pharmacophore Score 0.86 Docking @ G12C -12.1 Docking @ G12D (ref. only) -11 Adagrasib Generator Starting Point MRTX1133 (Mirati’s answer) “Test Data” © Elix, Inc.
  9. 12 Generator Reward Function 1. Select the generator architecture 2.

    Specify the reward function 3. Specify scaffold and generate! Screenshots from Elix Discovery™ © Elix, Inc.
  10. 13 Generator Reward Function Reward function used for the generators

    Screenshot from Elix Discovery™ © Elix, Inc.
  11. 14 Parameter Value Pharmacophore Score 0.92 Docking @ G12C -11.1

    Docking @ G12D (ref. only) -9.7 Elix Discovery Mirati Optimization A Top-scoring Example for Constrained Scaffolds with LibINVENT Parameter Value Pharmacophore Score 0.89 +/- 0.04 Docking @ G12C -13.9 Docking @ G12D (for ref.) -11.9 © Elix, Inc.
  12. 15 Regenerating Known Scaffolds In some cases, our generators reproduce

    known molecules (in this case, when generating pyrrolidine moieties) • Elix Discovery’s Novelty detector identifies these molecules automatically (including the original scaffold) • Links to Pubchem and Chembl entries enable further investigation, and identification of original sources © Elix, Inc.
  13. Generation of Central Scaffold with IsoMol 16 Not yet disclosed

    Connect these fragments with a new scaffold Assess interactions against G12C, G12D, and other KRAS mutants using docking simulations Fragment libraries built on the Elix Discovery™ platform from ChEMBL and Enamine libraries In collaboration with CRO, synthesizing and assaying top scoring generated ligands against G12C, G12D © Elix, Inc.
  14. 18 Drugging KRAS is Difficult KRAS is difficult to drug

    • The typical KRAS mutation site has a shallow pocket that leads to poor ligandability • The first successful inhibitors targeted the G12C mutation using covalent inhibitors ◦ The first KRAS inhibitor, Sotorasib (AMG-510), was approved in 2021 for non-small cell lung cancer ◦ Adagrasib (MRTX-849) from Mirati Therapeutics was approved in December 2022 for the same indication Sotorasib Adagrasib Covalent “warhead” https://doi.org/10.1021/acs.jmedchem.9b02052 © Elix, Inc.
  15. 19 Drugging KRAS is Difficult Other KRAS mutants remain difficult

    to drug • Covalent inhibitors cannot target mutations outside of G12C (thiol functionality of cysteine is targeted) • The more common G12D mutation does not have a functional group that can be targeted for covalent bonds • Recent efforts have focused on non-covalent inhibitors that target additional mutants • A June 2023 Nature article reported BI-2865 as a non-covalent pan-KRAS inhibitor Remove covalent warhead Structure based optimization https://www.nature.com/articles/s41586-023-06123-3 © Elix, Inc.
  16. 20 Why “pan” Inhibitors? KRAS mutations can be acquired •

    Treatment with G12C inhibitors can lead to acquired resistance • In a study on patients under G12C inhibitor treatment for lung cancer (NCSCLC), colorectal cancer (CRC) and appendix cancer (AC) that acquired resistance, numerous acquired mutations were detected (figure on right, and citation in lower right) • Mutations at the G12 codon other than “C” would lead to a failure of G12C inhibitors. Concurrent mutations in switch II binding pocket (Q, R, H, Y) also affect sotorasib, adagrasib • Supports need to develop pan-KRAS inhibitors, and drugs with alternative modes of binding for combo therapy https://www.nejm.org/doi/10.1056/N EJMoa2105281 © Elix, Inc.
  17. 21 Pharmacophore Model Interactions The pharmacophore model comprises several possible

    interactions between ligand and receptor that are automatically detected based on the atoms and functional groups present in the ligand and receptor, as well as the relative positioning of these moieties Interaction Name Criteria Ionic Distance Cation - Pi Distance, Angle Pi Stacking - Face to Face Distance, Angle Pi Stacking - Edge to Edge Distance, Angle Hydrogen Bond Distance, Angle Halogen Bond Distance, Angle Metal Complexation Distance Hydrophobic Distance © Elix, Inc.