PF2PI: Protein Function Prediction Based on AlphaFold2 Information and Protein-Protein Interaction

Protein-protein interaction networks and predicted protein structures in the AlphaFold2 database provide valuable information for annotating protein functions. In this study, we construct a model, PF2PI, for predicting protein functions within a single species. The core idea of PF2PI is the utilization of the attention mechanism to fuse information from two data sources in order to facilitate subsequent predictions. The model first performs network processing on PPI data and uses it as an input of a data source. Then the protein residue position information obtained from the AlphaFold2 database is aggregated. After obtaining the protein contact map, random walk and pooling are used to obtain its feature vector as second source input. We used a Transformer model with attention mechanism to pre-train our model, and a multi-label classification task is added behind it to complete protein function prediction. We compared our model with selected baseline model, and the experiments showed that most of the indicators of our model have improved, confirming the effectiveness of the model.