ProGraphTrans: multimodal dynamic collaborative framework for protein representation learning.

Area of Science:

• Computational Biology
• Bioinformatics
• Protein Science

Background:

• Protein representation quality is crucial for accurate functional prediction.
• Multimodal deep learning improves protein representation by integrating sequence, structure, and chemical data.
• Existing methods struggle with guiding structural information and static fusion strategies, limiting accuracy in identifying key functional residues.

Purpose of the Study:

• To address challenges in multimodal protein representation learning.
• To explore guiding mechanisms for structural information in multimodal feature interaction.
• To develop a dynamic fusion strategy for sequence-structural features.

Main Methods:

• Proposed ProGraphTrans, a multimodal dynamic collaborative framework.
• Implemented a dynamic attention multimodal fusion mechanism.
• Utilized a multi-scale convolutional neural network to capture local sequential patterns.

Main Results:

• ProGraphTrans outperforms existing methods on four protein downstream tasks.
• The framework demonstrates superior performance across various indicators.
• Achieved excellent interpretability in identifying key functional residues.

Conclusions:

• ProGraphTrans is an effective protein representation method.
• The dynamic collaborative framework offers advantages over static approaches.
• The method shows significant potential for advancing protein functional prediction.