S2DA-GO: enhancing protein function prediction via gradient-decoupled cross-attention and semantic priors

Abstract

Accurate automated protein function prediction is essential for bridging the widening gap between the exponential accumulation of uncharacterized protein sequences and the limited repository of experimentally verified functional annotations. However, despite recent advances in incorporating Gene Ontology (GO) priors and sequence-label interactions, existing computational methods still face substantial challenges in long-tailed multi-label settings, particularly in handling optimization instability caused by rare-label noise and in learning reliable representations for sparsely annotated GO terms. To address these challenges, we propose S2DA-GO, a sequence-based model using pre-trained protein language model embeddings and GO textual semantic priors for protein function prediction. S2DA-GO integrates a global contextual stream with a local target-aware stream to capture multi-scale functional patterns. To alleviate optimization instability caused by long-tailed label noise, we introduce a Gradient-Decoupled Cross-Attention Module (GDCAM), which reduces the interference of label-specific gradients on the shared backbone. In addition, we incorporate learnable residual semantic priors derived from BioBERT-encoded GO definitions, enhancing the model’s adaptability to rare functional terms. On the benchmark dataset, S2DA-GO outperformed the strong baseline GDTGO across all three GO branches, achieving notable relative AUPR improvements of 4.0% Molecular Function (MF), 6.0% Biological Process (BP), and 8.6% Cellular Component (CC), with AUPR scores reaching 65.1%, 33.4%, and 41.7%, respectively. Notably, S2DA-GO remained robust in low-homology settings and provided interpretable residue-level signals that accurately aligned with experimentally verified binding sites. Overall, S2DA-GO alleviates feature interference and improves prediction for sparsely annotated GO terms, providing a promising framework for large-scale annotation of uncharacterized proteins.