Steering Biomolecular Generative Models at Inference Time

Diffusion and flow-based models have become a standard approach for generating biomolecular sequences and structures. However, controlling these models to satisfy specific design constraints remains challenging, often requiring expensive retraining or fine-tuning. In this talk, I will present a general framework for inference-time control based on Feynman-Kac Correctors (FKCs), which enables sampling from different distributions at inference-time (such as annealed targets, product-of-experts compositions, and reward-tilted objectives) without modifying the underlying model. I will demonstrate how this framework enables controlled generation across a range of biomolecular design tasks using DISCO, a model we developed to jointly generate protein sequences and structures. These design tasks include specificity (e.g., predicted binding to a target while avoiding off-target interactions) and steering multimodal outputs, where sequence and structure must be jointly guided.