Multimodal Protein Language Models for Flexibility Prediction and Loop Design

ICML 2025 Workshop FM4LS Submission54 Authors

Keywords: protein flexibility, molecular dynamics, root-mean-square fluctuation, loop modeling, ESM-3

TL;DR: We use multimodal embeddings from ESM-3 to improve predictions of protein flexibility and associated loop design.

Abstract: In this paper we use multimodal embeddings from the ESM-3 protein language model to predict protein flexibility. ESM-3's multi-modality allows for simultaneous integration of structural, sequence-based, and functional information into a residue-level embedding, yielding informative protein representations that can be used to predict protein flexibility. We demonstrate improved RMSF prediction compared to established methods such as CABS-flex. Additionally, we present a practical application of ESM-3 embeddings for loop design by sampling novel loop sequences conditioned on fixed structural and functional contexts. Such an approach may efficiently identify potential loop designs with increased rigidity.

Submission Number: 54