EGNN-CMutPred: Predicting Protein Mutational Effects by Integrating Primary and Tertiary Protein Structures

The study proposed the EGNN-CMutPred, Equivariant Graph Neural Network based Comprehensive Protein Mutational Effect Predictor, a novel approach for predicting the effects of protein mutations by integrating both primary and tertiary protein structures. By combining Evolutionary Scale Modeling 2 (ESM-2) for semantic embedding with Equivariant Graph Neural Network (EGNN) for structural encoding, the model improves its accuracy in predicting how mutations impact protein function and stability. The study aims to address the limitations of traditional sequence and structure-based prediction methods by incorporating both semantic and topological embeddings of proteins, allowing the model to capture a comprehensive understanding of each protein. EGNN-CMutPred was trained on non-redundant protein sequences from the CATH v4.3.0 database and evaluated against benchmarks, including ProteinGym (DMS) and the ProThermDB database, which measure changes in melting temperature (△Tm) and change in the variation of Gibbs free energy (△△G). The model effectively simulates mutations and predicts changes in protein stability, as demonstrated by strong performance in metrics such as Spearman’s Correlation and True Positive Rate. These results suggest that EGNN-CMutPred is a valuable tool for precision medicine and protein engineering, offering enhanced prediction capabilities over existing methods. Future research will refine the model’s computational techniques and expand its applicability to larger, more diverse datasets, furthering its potential in understanding protein mutations and their implications for disease and therapeutic development.