CrossPredGO: A Novel Light-Weight Cross-Modal Multi-Attention Framework for Protein Function Prediction.

Area of Science:

• Bioinformatics
• Computational Biology
• Protein Science

Background:

• Proteins can be represented in multiple ways, including sequence, 3D structure, and interaction data.
• Uni-modal and multi-modal approaches exist for protein function prediction, with multi-modal methods often increasing trainable parameters through feature concatenation.

Purpose of the Study:

• To propose a novel, light-weight cross-modal multi-attention (CrMoMulAtt) mechanism for efficient protein function prediction.
• To capture the relative contribution of each data modality with fewer trainable parameters.

Main Methods:

• Developed a novel cross-modal multi-attention (CrMoMulAtt) mechanism.
• Integrated protein sequence, 3D structure, and protein-protein interaction (PPI) data.
• Proposed the CrossPredGO mechanism for protein function prediction.

Main Results:

• The CrMoMulAtt mechanism demonstrated a higher contribution from PPI data and a lower contribution from structure data.
• The CrossPredGO mechanism achieved accuracy improvements of 3.29% to 7.20% compared to existing methods.
• The proposed method reduced trainable parameters by up to 31% compared to DeepGO and MultiPredGO.

Conclusions:

• The novel CrMoMulAtt mechanism offers an efficient approach to protein function prediction.
• This method effectively integrates multi-modal protein data while minimizing computational overhead.
• CrossPredGO represents a significant advancement in protein function prediction accuracy and parameter efficiency.