Inference Time Search for Protein Structure Prediction

Abstract

Scaling inference-time compute for deep learning models has led to superhuman performance in games and enhanced reasoning capabilities for language models. However, similar gains have not yet been made in the field of biomolecular structure prediction. We introduce a new paradigm for inference-time search by adding architectural components and a finetuning procedure to state-of-the-art structure prediction models that give rise to a discrete latent space. We implement algorithms for searching and sampling in this discrete latent space and conduct experiments on a small model, demonstrating an increase in oracle and top-1-selected accuracy for predicted protein-protein complex structures.