A scalable equivariant graph network framework for precise protein function prediction

BACKGROUND: Protein function research helps in understanding the complex biological processes that occur within cells. However, the intricate nature of protein structures and functions, along with the rapid growth of protein sequence data, presents a pressing challenge to develop efficient computational methods for accurate protein annotation. RESULTS: In this study, we propose ENGINE, a multi-channel deep learning framework designed for robust protein function prediction. ENGINE integrates an equivariant graph convolutional network model to capture geometric features from protein 3D structures, leverages the large language model ESM-C to encode evolutionary and sequence-derived information, and combines an innovative 3D sequence representation that unifies spatial and sequential signals. We demonstrate that ENGINE consistently surpasses current state-of-the-art methods across diverse protein function prediction benchmarks, demonstrating robust generalisation and high predictive accuracy. Beyond performance, ENGINE provides interpretable insights into key sequence features and structural motifs, enabling the identification of functionally critical residues and substructures within proteins. This facilitates a deeper mechanistic understanding of protein function annotation outcomes and supports hypothesis generation for downstream biological studies. CONCLUSION: By offering reliable predictions with biological interpretability, ENGINE contributes to advancing research into cellular processes and disease mechanisms. The model is available at GitHub ( https://github.com/ABILiLab/ENGINE ) and Zenodo ( https://doi.org/10.5281/zenodo.17221153 ), serving as a valuable tool for the broader scientific community.