SSAS-GO: structure-sequence adaptive synergy network for protein function prediction

Area of Science:

• Computational Biology
• Bioinformatics
• Deep Learning

Background:

• Protein function prediction is crucial for drug discovery and disease treatment.
• Current deep learning methods often struggle to integrate diverse protein features (sequence and structure) effectively.
• Existing models face limitations in capturing global protein topologies and integrating multimodal data without interference.

Purpose of the Study:

• To develop a novel deep learning framework that synergistically integrates protein sequence and structural information for enhanced function prediction.
• To overcome the limitations of existing methods in handling multimodal data and capturing global protein topologies.
• To improve the accuracy and efficacy of protein function prediction for applications in drug discovery and disease treatment.

Main Methods:

• Proposed the Structure-Sequence Adaptive Synergy network (SSAS-GO), a novel deep learning architecture.
• Employed a Multi-Scale Motif Block for extracting localized sequence features.
• Utilized a parallel Dual-Stream Graph Encoder to capture spatial protein topologies.
• Introduced a Task-Adaptive Cross-Modal gating mechanism for dynamic feature recalibration.

Main Results:

• Achieved state-of-the-art Area Under the Precision-Recall Curve (AUPR) scores on the PDBch test set for Biological Process (0.463) and Cellular Component (0.559).
• Demonstrated a significant 17.7% relative AUPR improvement for Molecular Function tasks on the AFch test set.
• The Task-Adaptive Cross-Modal gating mechanism effectively balanced and integrated sequence and structural features.

Conclusions:

• The SSAS-GO network effectively integrates multimodal protein data, outperforming existing methods in function prediction.
• The proposed architecture offers a promising approach for advancing protein function prediction, with significant implications for biological research and medicine.
• Dynamic feature recalibration is key to overcoming modality interference and maximizing model performance.