Protein Design

While reinforcement learning has been applied to align both continuous and discrete diffusion models at training time, more work needs to be done in investigating test-time alignment strategies. In this project, we study the tradeoffs between training-time and test-time alignment in the context of protein design and explore how to maximize test-time alignment performance. In particular, we look at how to efficiently utilize reward-guided sampling methods while leveraging the interactions between amino acids that form the resulting proteins.

Publications

1. M. Cemri, A. Jalal, K. Ramchandran, “Discrete Diffusion Posterior Sampling for Protein Design”, ICML Workshop 2024.