Introduction

KRAS is one of the most intensively pursued oncology drug targets, but it remains difficult to drug broadly across the most prevalent mutations. Covalent KRAS G12C inhibitors validated KRAS druggability, yet other common mutants (e.g., G12D, G12V) lack a cysteine suitable for covalent bonding and therefore require non-covalent approaches. In early discovery for non-covalent KRAS inhibitors, the challenge is not only to find high-affinity ligands, but to do so efficiently while exploring diverse chemotypes.

This study addresses that challenge by using a structure-based molecular generation workflow to propose novel non-covalent KRAS inhibitor candidates. A 3D binding hypothesis (pose and interaction pattern) was used as the steering signal: generated molecules may differ substantially in 20 structure but are optimized to reproduce critical 3D interactions in the binding site. The main objective is to determine how reward-function wiring (additive vs. multiplicative combinations) changes outcomes, and to identify a formulation that most efficiently enriches promising non-covalent KRAS candidates.

Methods

Docking was configured using the adagrasib-bound KRAS G12C co-crystal structure (PDB 6UT0). A pharmacophore was derived from the reference pose and edited to remove covalent-warhead-specific features to reflect a non-covalent binding hypothesis.

A generative model optimized multi-objective reward functions combining docking score, pharmacophore score (additive vs. multiplicative), and additional property constraints. The docking score is not universally a reliable surrogate for biochemical potency. To avoid overinterpreting docking outputs, the study first benchmarked docking scores against published in vitro IC50 values for known KRAS inhibitors. After confirming docking score correlates with IC50 in this setup, we used a docking-score cutoff as a practical proxy for nM-range activity when evaluating generated compounds.