FKSFold: Improving AlphaFold3-Type Predictions of Molecular Glue–Induced Ternary Complexes with Feynman–Kac–Steered Diffusion

This paper discloses technical details of an early approach we explored for molecular glue induced ternary complex structure prediction using Feynman-Kac (FK) steering in the diffusion modules of AlphaFold3-type models. While we recently reported the successful results of our YDS-GlueFold model in predicting 8 novel molecular glue induced ternary structures validated by experimental data, here we share insights from our earlier FK steering experiments which were conducted prior to developing YDS-GlueFold. Notably, the unmodified AlphaFold3-type baseline failed to predict any of the eight complexes, underscoring the difficulty of sampling ternary conformations. By contrast, our FK-steered diffusion approach—FKSFold—successfully predicted 3 of the 8 cases in these early trials. We detail how we leveraged FK formalism to modify the diffusion process, using interface predicted TM-score (ipTM) as a reward function to guide sampling toward high-quality protein-protein interfaces in ternary complexes. This strategy of using ipTM as a reward function aligns well with the mechanism of action (MOA) of molecular glues, which function by inducing or stabilizing protein-protein interfaces. By implementing stochastic control theory within the diffusion module, our approach attempted to steer the reverse diffusion process toward conformations with superior interface quality in ternary protein-protein-small molecule complexes. In these early trials, FK steering successfully predicted 3 out of the 8 cases that were later all correctly predicted by YDS-GlueFold. This work provides insights into the evolution of our technical approaches toward molecular glue-induced complex prediction, documenting an important step in the development pathway that eventually led to YDS-GlueFold. Code is available at https://github.com/YDS-Pharmatech/FKSFold .