Computational De Novo Design of KRAS Inhibitors through Generative Adversarial Networks

K-RAS is a key oncogene frequently mutated in cancer, historically considered “undruggable”. Recent breakthroughs (e.g. sotorasib) have demonstrated the feasibility of targeting KRAS (especially G12C). Parallel advances in deep generative modeling have transformed de novo drug design . Generative Adversarial Networks (GANs) in particular can efficiently sample novel molecular graphs with tailored properties. In this study, we present a GAN-based framework for KRAS inhibitor discovery. Our approach trains a graph-based GAN on known bioactive ligands, then iteratively refines the generator via active learning and molecular docking to enrich high-affinity candidates. We identify structural gaps in existing methods (few focus on KRAS-specific GAN design and end-to-end validation) and aim to address them. Computational screening of GAN-generated libraries yields top candidates with favorable binding scores and drug-like profiles; Figure 1 illustrates our GAN architecture. We further propose in silico ADMET filtering and outline preliminary in vitro assays for the most promising compounds. This integrated pipeline demonstrates that GAN-driven design can explore novel chemical space for KRAS, generating candidates dissimilar to known scaffolds. Our results provide a workflow and proof-of-concept for AI-driven KRAS inhibitor design, bridging computational innovation with preclinical validation needs.