EIHR-PROT: explicit modelling of homology reliability for protein function prediction

Background: Protein function prediction remains a fundamental challenge in computational biology, with deep learning methods increasingly complementing traditional homology-based approaches. While protein language models (PLMs) such as ESM-2 have demonstrated sufficient understanding of evolutionary patterns to be competitive for downstream functional annotation tasks, homology-based inference continues to provide valuable information when reliable templates exist. However, existing hybrid approaches typically employ fixed fusion strategies that fail to account for the varying reliability of homology evidence across proteins. We introduce EIHR-PROT (Explicitly Incorporated Homology Reliability Signals for PROTein Function Prediction), a novel framework that adaptively integrates sequence-derived neural predictions with homology-based priors through a learned reliability-gating mechanism. Our model extracts frozen embeddings from ESM-2, processes them through an attention-pooling module and neural prediction branch, while simultaneously constructing homology priors from DIAMOND-based sequence similarity searches. Results: Compared with similar methods which employ aspect-specific weights, EIHR-PROT uses explicit alignment-derived features, including bitscore, query coverage, and hit count, to dynamically weight each modality's contribution on a per-protein basis. Evaluation on the PDB dataset demonstrates that EIHR-PROT outperforms existing sequence and homology-based methods, DeepGOPlus and TALE+, across all Gene Ontology (GO) aspects with Fmax scores of 0.604, 0.676, 0.731, for BP, MF, and CC, respectively. Ablation studies confirm that explicit reliability signals improve performance, particularly for Biological Process (BP) prediction, while gate weight analysis reveals appropriate down-weighting of homology contributions in low-similarity regimes. Conclusion: Beyond improving predictive performance, our findings demonstrate that explicitly modelling homology reliability provides a principled and interpretable framework for adaptive evidence integration in hybrid protein function prediction. Implementation of our model can be found at https://github.com/uwidia/EIHR-PROT.