More Structures, Less Accuracy: ESM3's Binding Prediction Paradox

Abstract

This paper investigates the impact of incorporating structural information into the protein-protein interaction predictions made by ESM3, a multimodal protein language model (pLM). We utilized various structural variants as inputs and compared three widely used structure acquisition pipelines--EvoEF2, Gromacs, and Rosetta Relax--to assess their effects on ESM3s performance. Our findings reveal that the use of a consistent identical structure, regardless of whether it is relaxed or variant, consistently enhances model performance across various datasets. This improvement is striking in few-show learning. However, performance deteriorates when different relaxed mutant structures are used for each variant. Based on these results, we advise caution when integrating distinct mutant structures into ESM3 and similar models.This study highlights the critical need for careful consideration of structural inputs in protein binding affinity prediction.