Integrating Heterogeneous Biological Networks and Ontologies for Improved Protein Function Prediction with Graph Neural Networks

Elucidating protein functions is critical to advance our understanding of biological systems. However, the majority of proteins lack functional annotations due to the pace of manual knowledge-driven curation of the Gene Ontology (GO) terms. Automatic function prediction (AFP) aims to computationally infer these missing annotations by integrating sequence, interaction, and ontology data. Current AFP methods have yet to fully utilize heterogeneous relationships across multi-omics networks between non-coding RNAs, proteins, and the GO hierarchies. To address this, we introduce LATTE2GO, a novel graph neural network that integrates protein, RNA, and GO function entities and their interactions into a unified representation. By extracting higher-order associations with attention mechanism, LATTE2GO achieved significant gains over previous graph-based AFP techniques on CAFA4 benchmarks. Our analyses revealed modeling specific protein-protein interactions (PPI) and GO relationships increases accuracy in predicting molecular functions and biological processes. Overall, LATTE2GO demonstrates that heterogeneous graph neural networks can integrate diverse omics knowledge to advance systems-level understanding of protein roles within the complex milieu of functional and physical interaction networks.