Protein-protein Interaction Prediction Method Based On Label Co-occurrence Graph Guidance

The investigation into protein-protein interaction(PPI)holds substantial worth and meaning.Firstly,PPI is the core of many biological processes in cells,including signal transduction,metabolic regulation,and cell cycle.Gaining insights into PPI can provide valuable comprehension of the intricate mechanisms governing these fundamental biological processes.Secondly,PPI is a potential drug target.In the process of drug discovery,one or more members of protein-protein interaction are usually selected as targets to explore the interaction between drugs and proteins,and thus discover drugs for treating diseases.Therefore,studying PPI can provide valuable information and ideas for drug discovery.In addition,studying PPI can also provide important help for the fields of biotechnology and bioengineering.By understanding PPI,we can better design gene editing tools,develop new protein materials,improve enzyme catalysis and other aspects of performance,and play an important role in applications such as agriculture,medicine,and environment.The thesis mainly aims to explore the interaction relationships between proteins,by constructing a label relationship graph and introducing effective label information.Secondly,by constructing a protein interaction graph,a graph neural network is used to discover the interaction relationships between proteins.The main contents of the research work are as follows.Previous works predict PPI by constructing protein interaction graphs while ignoring the noise existing in the original dataset.In addition,in the task of multi-label classification,there has been no analysis of the relationship between multiple types,such as the existence of certain connections between labels.Therefore,the thesis proposes a label-guided multiscale graph neural network PPI prediction method(LGMG-PPI),which obtains multiple scales of protein interaction graphs through the graph data augmentation method,improves the anti-noise ability of protein interaction graphs,and then uses graph neural networks to learn protein feature representation,while introducing contrastive learning to ensure the consistency of the learning direction of multi-scale protein interaction graphs.In addition,a label relationship graph is constructed to obtain the hidden connections between learning labels,and to guide the prediction of PPI.Experimental analysis on three publicly available datasets showed that the model LGMG-PPI proposed in this article is better than other models in previous works.Existing PPI prediction methods only construct a protein interaction graph,but different forms of protein interaction graphs have different potential information.In addition,in the previous work,the learned label information is directly guided,which led to the lack of generalization ability of the model.Based on this,this thesis proposes a label-aware dual-view graph neural network PPI prediction method(LADV-PPI),which first constructs a protein feature graph based on protein features,and then combines the original protein topology graph and protein feature graph to construct a higher confidence consensus graph,and then introduces a graph neural network to learn from two perspectives(topology graph and consensus graph).At the same time,just like in the previous work,a label relationship graph is constructed to learn label information.A dualchannel multi-layer perceptron is then designed to introduce label information,mapping PPI features to the label space from explicit and implicit channels respectively to enhance the model’s generalization ability.The results show that the method LADV-PPI proposed in this thesis is always better than all baseline methods.