Dynamics-informed RFdiffusion for <i>de novo</i> design of binders against the KRAS conformational ensemble

Protein-protein interaction surfaces pose a significant challenge for therapeutic design, particularly for oncogenes like KRAS, which have historically been considered challenging to target due to their lack of well-defined binding pockets. Although <i>de novo</i> protein design algorithms such as RFdiffusion have recently produced promising results, most approaches rely on a single static structure, thereby neglecting the native conformational dynamics of the protein. In this study, a 'dynamics-informed RFdiffusion' strategy was implemented using conformational clusters obtained from a 100 ns molecular dynamics (MD) simulation of KRAS, and this approach was directly compared with traditional static structure-based design. Following sequence optimization with ProteinMPNN and structural validation with AlphaFold2, the dynamic protocol was shown to produce stable designs at a rate of 60%, whereas the static protocol remained at 30%. The dynamic designs also exhibited larger interface areas, an increased number of hydrogen bonds, and strong electrostatic complementarity; comparative MD simulations confirmed their high binding stability. Bootstrapping (95% CI), permutation tests (<i>p</i> < 0.05), and Bayesian posterior analyses (P(θ_dynamic > θ_static) = 0.94) demonstrated that the findings are statistically robust. The results show that accounting for conformational dynamics significantly increases the design success rate and offers a robust methodological framework for <i>de novo</i> binder discovery against difficult targets like KRAS.