Improving protein generation by taking functional constraints (binding free energy) into account using diffusion models

Abstract

This project undertakes the task of improving an existing protein generation diffusion model (RFdiffusion) by taking into account the binding free energy of the protein-ligand complex that the generated protein will form with a specified ligand. In pursuit of scalability and speed, computed docking scores that are empirical approximations of binding free energy are used in their place. Beyond the suggested improvements, the report also explores certain constraints and possibilities in enhancing the current workflow.