Predicting Protein Thermostability with Equivariant Graph Neural Networks and Contrastive Learning

Accurately predicting the thermostability changes induced by single-point mutations is a core challenge in protein engineering and industrial enzyme modification. However, current computational methods often struggle to jointly integrate evolutionary information, structural context, and the interactions between mutation sites and their local microenvironments, presenting obvious limitations in guiding protein modification practice. To this end, this paper proposes the Equivariant Contrastive Learning Network (ECL-Net) for effectively predicting protein thermostability changes. Firstly, this method employs a joint sequence-structure representation strategy, deeply integrating sequence embeddings extracted by protein language models (PLMs) with spatial structural information captured by residue interaction networks, while an Equivariant Graph Neural Network (EGNN) is utilized for feature propagation and structural modeling. Furthermore, the self-attention mechanism of the language model is utilized to determine critical residues highly correlated with the mutation, thereby introducing structural microenvironment information. Building upon this, a contrastive learning method jointly guided by the mutation site and local microenvironment is designed, aiming to guide the model to learn more robust feature representations. Experimental results demonstrate that ECL-Net outperforms existing baseline models. Among the four predicted transglutaminase mutants, the model's predictions for three mutants were consistent with the experimental outcomes, demonstrating the potential applicability of ECL-Net in guiding protein engineering.