Machine Learning-Driven Drug Repurposing for KRAS G12C and KRAS G12D Inhibition

KRAS is a predominant oncogenic driver across multiple cancers and was long considered undruggable due to its high nucleotide affinity and lack of classical binding pockets. Although recent advances have led to covalent inhibitors such as Sotorasib and Adagrasib for the KRAS G12C mutation, effective therapies for other common variants, particularly KRAS G12D, which is highly prevalent in aggressive pancreatic cancers, remain limited. In this study, we employed machine learning approaches to identify potential inhibitors of KRAS G12D and G12C by screening FDA-approved compounds curated from the ChEMBL database. Random Forest and Neural Network models were trained on bioactivity data from three BindingDB data sets: wild-type KRAS GTPase, KRAS G12C, and KRAS G12D. The trained models demonstrated strong predictive performance, achieving high correlation coefficients on independent test sets. To further validate the predictive capability of the models, two compounds identified as high-confidence candidates, Cobimetinib and Etrasimod, were selected for experimental evaluation. In vitro testing revealed measurable IC₅₀ and E _max values, with both compounds exhibiting preferential activity in KRAS G12D cellular backgrounds. While additional biochemical and pathway-level studies are required to confirm direct target engagement, these results support the model's utility in prioritizing candidate compounds with allele-specific activity profiles. Overall, this study provides a data-driven framework for identifying potential KRAS-targeted therapies and highlights the value of integrating machine learning predictions with experimental validation.