Kirsten Rat Sarcoma (KRAS) is a frequently mutated oncogene in human cancers, yet the structural effects of its common mutations (G12C, G12D, G12V) and their druggability remain poorly understood. Here, we integrate molecular dynamics (MD) simulations with deep learning to map the conformational landscape of wild-type (WT) KRAS and its G12C, G12D, and G12V mutants. Extensive all-atom MD simulations were performed to sample the conformational ensembles of KRAS-WT and each mutant form. Atomic coordinates from these trajectories served as input features to a message-passing neural network (MPNN), which learned a low-dimensional embedding of KRAS conformations. Projecting these embeddings onto a two-dimensional manifold revealed distinct clusters that highlights mutation-specific conformational states. By extracting representative structures from these clusters and applying binding pocket detection, we identified unique mutation-specific pockets and potentially druggable conformations. These findings underscore the profound influence of oncogenic mutations on KRAS dynamics and illustrate the power of an integrated MD–deep learning approach to reveal cryptic, druggable states. Our integrative strategy provides insight into mutation-induced structural changes in KRAS and offers a novel avenue for precision oncology, guiding the design of mutation-selective therapeutics for KRAS-driven cancers.
