Predictive analysis of protein structures: parsing PDB data for structural and sequential insight and active site prediction in protein-protein interaction at residue level

Abstract

The prediction of active sites in protein-protein interactions (PPIs) is a crucial aspect of computational biology, aiding in the understanding of molecular mechanisms and facilitating drug discovery. This study presents an integrated workflow utilizing DenseNet201, a deep learning architecture, for predicting active sites in PPIs. The process begins with the retrieval and parsing of PDB files to extract chain sequences, followed by the computational generation of interatomic distance matrices of protein chains at the residue level. Sequence information and distance matrices, representing structural data for both positive and negative PPI datasets, are considered. Heatmaps generated from these distance matrices visually depict the proximity of active sites, with a color spectrum ranging from green (indicating shorter PPI distances) to red (indicating longer PPI distances), which can be used for future validation purposes. A color encoding strategy maps bigram interactions of amino acids to a lookup table, resulting in the generation of images from protein sequence information parsed from PDB data. Using a sliding window approach with specific dimensions of 32x32 and a stride of 2, sub-images are created and fed into the DenseNet201 model for training and testing. The classification strategy evaluates PPIs based on confidence scores compared against a threshold of 0.5. The workflow includes the generation of performance metrics, ensuring a robust evaluation of the model's predictive capabilities. This comprehensive approach enhances the visualization and prediction of active sites in PPIs, significantly contributing to advancements in bioinformatics. Currently, the workflow, which includes PDB data parsing and active site prediction by training image data using the DensePPI model, has been trained and tested, resulting in an overall accuracy of 95.84%. Additionally, this research will contribute to another project focusing on protein-protein interaction prediction, employing Graph Convolutional networks and graphlet features.