A Multimodal Protein Representation Framework for Quantifying Transferability Across Biochemical Downstream Tasks.

Area of Science:

• Computational Biology
• Bioinformatics
• Deep Learning

Background:

• Proteins are fundamental to life, and effective computational representations are crucial for biological analysis.
• Current protein representation methods often rely on text-based language models, overlooking proteins' complex 3D structures and functions.
• A need exists for advanced protein representations that capture multi-modal biological data.

Purpose of the Study:

• To develop a multimodal deep learning framework (MASSA) for comprehensive protein representation.
• To integrate protein sequence, structure, and functional annotation data.
• To evaluate the framework's performance on diverse downstream biological tasks.

Main Methods:

• Proposed a multimodal deep learning framework named MASSA.
• Utilized a multitask learning process with five pretraining objectives.
• Incorporated approximately 1 million protein sequence, structure, and functional annotations.
• Introduced an optimal-transport-based metric to assess representation transferability.

Main Results:

• Achieved state-of-the-art performance in predicting protein properties (stability, fluorescence), protein-protein interactions, and protein-ligand interactions.
• Demonstrated competitive results in secondary structure prediction and remote homology detection.
• Showcased strong correlation between feature space distributions and adaptability across tasks.

Conclusions:

• The MASSA framework provides a powerful, fine-grained protein-domain feature representation.
• Multimodal protein representation significantly enhances performance in various critical biological applications.
• The optimal-transport metric offers valuable insights into the learning process and model transferability.