Benchmarking docking and ML re-scoring screening performance for KRAS G12D in pancreatic cancer.

The KRAS G12D mutation is commonly found in pancreatic cancer and is integral to the cell signaling pathway, making it a critical target for drug development. Structure-based virtual screening (SBVS) is a conventional strategy for discovering new inhibitors and expanding the chemical space against KRAS G12D. However, the SBVS efforts needed to be evaluated and benchmarked. Herein, in the first stage of the study, we evaluated two popular docking tools (FRED and AutoDock Vina) utilizing DeepCoy decoys-using DEKOIS 2.0 parameters-against KRAS G12D. Furthermore, re-scoring performance of the docking outcome via two popular pretrained machine learning scoring functions (ML SFs), such as CNN-Score and RF-Score-VS v2 were explored. While FRED exhibited the best screening performance based on pROC-AUC value, both tools showed high early enrichment indicated by EF 1%. Interestingly, both FRED and AutoDock Vina displayed superior performance to re-scoring using the ML SFs. This highlights the target-specific nature of the screening performance. In the second stage, accordingly, a VS effort was performed using FRED on specs world diversity database. Selected hits identified as potential KRAS G12D binders, were subjected to cell viability assays against the pancreatic cancer cell line PANC-1. Molecule (CP3) exhibited a promising antiproliferative activity with IC