RaFT-DM: A Residue-Aware Fusion Transformer With Domain-Wise Memory for Accurate Multi-Label Protein Function Prediction.

Abstract

Accurate protein function prediction via modeling protein sequences as long structured symbolic sequences is essential for understanding cellular mechanisms and guiding drug discovery. However, conventional deep learning models often use global pooling, which weakens domain-specific signals and fails to capture contextual dependencies across domains. Here, we propose Residue-aware Fusion Transformer with Domain-wise Memory (RaFT-DM), a domain-conditioned information routing framework for multi-label protein function prediction. Rather than treating domain annotations as auxiliary inputs, RaFT-DM uses functional domains as structured intermediate memory units between residue-level multimodal evidence and protein-level GO labels. It first aligns sequence and structural signals at the residue level, then routes the fused evidence through domain-wise bottlenecks that preserve local functional cues and model contextual dependencies among domains. Experiments on standard benchmarks show that RaFT-DM consistently outperforms state-of-the-art baselines. By replacing global pooling with domain-conditioned routing, RaFT-DM reduces false negatives and misclassifications, enabling more accurate and interpretable predictions. The implementation of RaFT-DM is available at https://github.com/AI4DS-wk/RaFT-DM.