AlphaFold2-aware protein-DNA binding site prediction using graph transformer.

Area of Science:

• Computational Biology and Bioinformatics
• Structural Biology
• Genomics and Molecular Biology

Background:

• Protein-DNA interactions are fundamental to biological processes like transcription and DNA repair.
• Accurate identification of protein-DNA binding sites is critical for mechanistic studies and drug discovery.
• Existing sequence-based methods struggle to capture spatial information crucial for binding site prediction.

Purpose of the Study:

• To develop an accurate predictor, GraphSite, for identifying DNA-binding residues using protein sequence and structure.
• To leverage AlphaFold2-predicted protein structures for enhanced binding site prediction.
• To improve upon existing sequence-based and structure-based methods for DNA-binding residue identification.

Main Methods:

• GraphSite frames DNA-binding site prediction as a graph node classification task.
• A transformer-based variant model is employed, incorporating protein structural information.
• The method utilizes predicted protein structures from AlphaFold2 and a graph transformer architecture.

Main Results:

• GraphSite demonstrates substantial improvement over the latest sequence-based and structure-based prediction methods.
• On an independent test set of 181 proteins, GraphSite exceeded state-of-the-art structure-based methods.
• Performance gains include a 16.4% increase in area under the precision-recall curve and an 11.2% increase in Matthews correlation coefficient.

Conclusions:

• GraphSite offers a highly accurate method for identifying DNA-binding residues by integrating predicted protein structures.
• The approach provides valuable insights for understanding protein-DNA interactions and facilitates drug design.
• The study provides open-access datasets, predicted structures, source codes, and a web server for GraphSite.