Abstract
Accurate protein function prediction is fundamental to advancing drug discovery, precision medicine, and understanding complex biological systems. While gene ontology (GO) provides a standardized framework for protein annotation, a critical challenge persists: the imbalance between low-specificity GO terms and high-specificity GO terms. This imbalance creates blind spots in our understanding of protein function landscapes, particularly in clinically relevant pathways. We present ProGO-PSL, a novel large graph architecture designed to resolve this imbalance. ProGO-PSL simultaneously leverages explicit domain identifier from InterPro and implicit evolutionary context from Multiple Sequence Alignments, fusing these complementary data sources within a powerful imbalance learning framework. Our model consistently outperforms state-of-the-art methods by 5-15% across all specificity levels and on both benchmark dataset and independent test set, demonstrating robust generalization. Furthermore, ProGO-PSL generates interpretable representations that clarify relationships between low- and high-specificity GO terms, enabling a more complete functional characterization of the proteome. This work accelerates the identification of therapeutic targets in previously uncharacterized biological pathways.
