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

Area of Science:

• Computational biology
• Protein bioinformatics
• Structural biology

Background:

• The protein sequence-function gap hinders biological research due to the slow, costly experimental determination of protein functions.
• Advances in genome sequencing have generated vast amounts of protein sequence data, exacerbating the need for efficient function prediction methods.
• The recent availability of accurate protein structure predictions has opened new avenues for integrating structural information into function prediction models.

Approach:

• TransFun integrates protein sequence information, processed by a pre-trained protein language model (ESM), with 3D protein structures predicted by AlphaFold2.
• The method employs 3D-equivariant graph neural networks to effectively combine sequence-derived features and structural data for function prediction.
• Transfer learning from the ESM model enhances the extraction of relevant features from protein sequences.

Key Points:

• TransFun demonstrates superior performance on benchmark datasets (CAFA3 and a new dataset) compared to state-of-the-art methods.
• The combination of transformer-based language models and 3D-equivariant graph neural networks proves effective for leveraging both sequence and structure data.
• Integrating TransFun predictions with sequence similarity-based predictions further boosts protein function prediction accuracy.

Conclusions:

• TransFun offers a powerful computational approach to accurately predict protein function by integrating sequence and structural data.
• The method effectively addresses the protein sequence-function gap, accelerating biological discovery.
• The successful application of protein language models and 3D-equivariant graph neural networks highlights their potential in structural bioinformatics.