Abstract
The development of KRAS G12D inhibitors represents an effective therapeutic strategy for treating oncological pathologies. Existing quantitative structure-activity relationship (QSAR) models for KRAS G12D inhibitors have several limitations, primarily the lack of applicability domain determination and virtual screening implementation. In this study, we propose a set of regression QSAR models for KRAS G12D inhibitors by employing various molecular descriptors and machine learning methods. Our consensus model achieved a Q(2) test value of 0.70 on an external test set, covering 78% of the data within the applicability domain. We integrated this consensus model into our Python-based framework KRASAVA. The platform predicts inhibitory activity while considering the applicability domain, assesses compounds for compliance with Muegge's bioavailability rules, and identifies PAINS, toxicophores, and Brenk filters. Furthermore, we structurally interpreted the QSAR models to propose several promising inhibitors and performed molecular docking on these candidates using GNINA. For the reference inhibitor MRTX1133, we reproduced the crystal structure pose with an RMSD of 0.76 Å (PDB ID: 7T47). The key interactions with amino acid residues Asp12, Asp69, His95, Arg68, and Gly60, identified for both MRTX1133 and our proposed compounds, demonstrate a strong consistency between the molecular docking and QSAR results.