Integrating Molecular Dynamics and Generative Modeling for Cryptic Pocket Drug Discovery

Identifying cryptic pockets and designing ligands that selectively target them could enable safer, more effective drugs. However, this remains a challenge for both traditional physics-based methods, which are computationally expensive, and machine learning approaches, which are constrained by limited training data. Recent flow-matching methods for pocket-conditioned ligand generation have shown promise when the binding pocket is known a priori, but it remains unclear how atomistic molecular dynamics (MD) can be incorporated to improve ligand design. We present a work-in-progress framework that couples MD sampling with a pre-trained generative model to identify relevant pocket conformations, guide ligand generation, and validate generated candidates. Applied to a known cryptic site in KRAS, the workflow produces chemically diverse candidates and identifies stable molecules for subsequent refinement. This provides a modular route toward iterative, simulation-guided molecular design in which generative models and physics-based tools can be composed in a closed loop.