Study On Computational Modeling Of Protein Interaction Sites

Proteins play an important role in life activities,and protein-ligand interactions are an important way for proteins to perform their biological functions.Therefore,protein interaction site identification is important for understanding the regulatory mechanisms of organisms.In this thesis,a modeling study of protein interaction site computation is carried out based on the interaction data of proteins with three different ligands,respectively,as follows:Firstly,the protein sequence-based deep model ctP2ISP is proposed to predict proteinprotein interaction sites by Convolutional Neural Network and Transformer.The method employs a weighted loss function to suppress prediction preferences while utilizing an improved sampling strategy for data augmentation.Superior performance is demonstrated on all six benchmark test sets,while the prediction results on open test sets related to novel coronaviruses and influenza viruses are consistent with their biological views.Secondly,the deep models GraphPPepIS and SeqPPepIS based on protein 3D structure are proposed to predict both protein-peptide interaction sites by Graph Convolutional Network and Transformer.SeqPPepIS uses peptide sequences instead of GraphPPepIS peptide structures,avoiding the limitation of inaccurate or missing structural information caused by peptide structure flexibility.And it improves the protein-peptide docking task by accurately identifying the interaction sites.Finally,GraphPLIS,a deep model based on small molecule ligand integration data,is proposed to predict protein-small molecule ligand interaction sites by Graph Convolutional Network and Graph Attention Network.The method uses a unified model for learning different small molecule data to obtain protein features across different ligands.The model shows good competitiveness on benchmark datasets of all five small molecule ligands,providing an effective practical solution for multi-source information fusion.The proposed method makes full use of deep neural networks to model protein sequences and structures with biochemical semantics,respectively,so as to accurately identify protein interaction sites and provide research clues for umderstanding protein interaction networks,while expanding the application scenarios of deep learning in bioinformatics.