Structure-aware protein self-supervised learning

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Protein representation learning is crucial for biological applications.
• Self-supervised learning addresses limited protein labels.
• Current methods often overlook vital protein structural information.

Purpose of the Study:

• To develop a novel structure-aware protein self-supervised learning method.
• To effectively capture and integrate protein structural information into representation learning.
• To enhance protein representation by combining sequential and structural data.

Main Methods:

• A graph neural network model is pretrained using self-supervised tasks focusing on pairwise residue distance and dihedral angles.
• Leveraged a pre-trained protein language model for sequential information.
• Integrated sequential and structural information via a pseudo bi-level optimization scheme.

Main Results:

• The proposed method effectively captures protein structural information.
• Integration with protein language models improved self-supervised learning.
• Demonstrated effectiveness on membrane protein classification, cellular compartment localization, and enzyme reaction classification tasks.

Conclusions:

• The novel structure-aware self-supervised learning method enhances protein representation.
• Integrating structural and sequential information is beneficial for downstream biological tasks.
• The approach provides a powerful tool for analyzing protein functions and properties.