Combining protein sequences and structures with transformers and equivariant graph neural networks to predict protein...

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Millions of protein sequences are available, but experimental function determination is slow and costly.
• A significant gap exists between the number of known protein sequences and their experimentally determined functions.
• Computational methods are needed to predict protein function accurately and efficiently.

Purpose of the Study:

• To develop a novel computational method for protein function prediction.
• To leverage both protein sequence and structure information for enhanced prediction accuracy.
• To address the limitations of sequence-only methods in protein function prediction.

Main Methods:

• Developed TransFun, a method combining transformer-based protein language models and 3D-equivariant graph neural networks.
• Utilized pre-trained protein language models (e.g., ESM) for sequence feature extraction.
• Integrated 3D protein structures predicted by AlphaFold2 with equivariant graph neural networks.

Main Results:

• TransFun demonstrated superior performance compared to state-of-the-art methods on benchmark datasets (CAFA3 and a new dataset).
• The integration of sequence and structure information significantly improved protein function prediction.
• Combining TransFun predictions with sequence similarity further enhanced prediction accuracy.

Conclusions:

• TransFun effectively utilizes protein sequences and structures for accurate function prediction.
• Transformer-based language models and 3D-equivariant graph neural networks are powerful tools for this task.
• The developed method helps bridge the protein sequence-function gap, facilitating biological research.