DA-HGL: a domain-augmented heterogeneous graph learning framework for protein function prediction

Accurate protein function prediction is critical for deciphering disease mechanisms and advancing precision medicine, yet remains challenging for proteins with sparse annotations. Traditional methods struggle with annotation sparsity and fail to integrate multimodal data holistically. We propose DA-HGL, a heterogeneous graph learning framework that integrates protein sequences, domain architectures, and Gene Ontology (GO) hierarchies through a multilayered graph and non-negative matrix factorization with dual biological constraints. DA-HGL uniquely models domain-function coherence, GO semantic consistency, and topological congruence. Evaluated on yeast and human proteomes, DA-HGL achieves Fmax gains of 9.0% (yeast CC) and 17.2% (human BP) over state-of-the-art methods. By dynamically learning domain-context associations and resolving annotation sparsity, DA-HGL excels in cold-start scenarios and disease-specific predictions (e.g. Parkinson's "ubiquitin-dependent catabolism"). This framework offers a robust tool for accelerating functional genomics and precision medicine. Code/data: https://github.com/husaiccsu/DA-HGL.