Learning a generalized graph transformer for protein function prediction in dissimilar sequences.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• High-throughput sequencing generates vast data, outpacing experimental validation.
• Deep learning offers promising solutions for rapid protein function prediction.
• Current deep learning models may struggle with novel proteins distant from training data.

Purpose of the Study:

• To introduce Graph Adversarial Learning with Alignment (GALA), a generalized deep learning approach for protein function prediction.
• To enhance the generalizability of protein function prediction models to novel, non-homologous proteins.
• To improve the interpretability of protein function prediction models.

Main Methods:

• GALA integrates a graph transformer architecture and attention pooling for unified protein sequence and structure representation learning.
• Adversarial learning with a domain discriminator ensures domain-invariant protein representations.
• Label embeddings are generated and aligned in hidden space to optimize with label information.

Main Results:

• GALA achieves performance comparable to state-of-the-art methods on PDB and Swiss-Prot datasets.
• The model demonstrates biological interpretability by identifying key functional residues using class activation mapping.
• GALA shows excellent generalizability for predicting functions of proteins in unseen sequence spaces.

Conclusions:

• GALA's adversarial learning and label embedding alignment yield domain-invariant representations, improving generalizability.
• Integrating AlphaFold2 structures with GALA shows potential for annotating newly discovered protein sequences.
• The GALA implementation is publicly available for research use.