The exponential growth of protein sequence data has outpaced experimental functional characterization, resulting in a widening annotation gap. Addressing this requires both the discovery of novel roles and the refinement of existing Gene Ontology annotations. Current computational approaches to this problem operate either at the atomic residue scale or the systemic network scale, which prevents a holistic understanding of protein functional roles. To address this limitation, we propose MS-GNN, a graph neural network framework that bridges these scales. At the protein level, we construct spatial graphs of amino acids through language model embeddings and AlphaFold-derived structural contact maps. At the systemic level, these representations are integrated into a global protein–protein association network. This unified architecture achieves state-of-the-art performance across all three sub-ontologies. Beyond architectural innovation, we demonstrate that complementing strict experimental labels with broader, curated functional data drastically improves performance, showing that annotation sparsity, rather than algorithmic capacity, is the primary bottleneck in model improvement.
Finally, ablation studies confirm the necessity network-level information, particularly functional features, highlighting the necessity of multi-scale integration.

Co-authors: Pierre Larmande, Jérôme Azé