Prediction of Protein-DNA Binding Sites Based on GraphSAGE

Accurately pinpointing the locations where proteins bind to DNA is crucial for the discovery of novel drug targets and the development of targeted therapeutic strategies. In this study, a graph neural network architecture called GraphSAGE has notably improved the precision of predicting protein-DNA binding sites. This model represents protein sequences as a graph, with amino acid residues serving as nodes, and employs a pre-trained, large-scale protein language model as ESM to derive residues’ features. The effectiveness of the model was rigorously evaluated using ten-fold cross-validation on benchmark datasets, including PDNA-316, PDNA-335, and PDNA-543, yielding Matthews Correlation Coefficients (MCC) of 0.741, 0.750, and 0.735, respectively. These results demonstrate considerable improvements over the existing state-of-the-art methods, with respective increments of 0.207, 0.185, and 0.174. Furthermore, the methodology is not only confined to binding site prediction but is also extensible to applications in protein functional annotation, drug discovery, and pharmaceutical design. The entire codebase for this project can be accessed without restriction through the subsequent URL: https://github.com/primrosehry/iProtDNA-SAGE.