Collaborative Filtering For Drug-Target Interaction Prediction Based On Heterogeneous Biological Network

Most drugs interact with their target proteins in-vivo to exert efficacy.Drug-target interaction(DTI)prediction research presents important significance for promoting the development of modern medicine and pharmacology.Traditional biochemical experiments for DTI prediction confront the challenges including long time period,high cost and high failure rate,finally leading to a low drug productivity.Chemogenomic-based computational methods can realize high-throughput prediction,thus narrow the scope of research for drug candidates in downstream biochemical experiments.In addition,integrating heterogeneous auxiliary data can provide diverse information and multi-dimensional perspectives for prediction task.In this research,the DTI prediction problem is regarded as a link prediction task.The specific research contents are as follows:(1)We construct a semantically rich heterogeneous network from eight kinds of biomedical data.There are several problems in raw biomedical data,such as high dimensionality,high noise,and incompleteness.To address these issues,we employ two different representation learning algorithms to extract heterogeneous network features.The generated node low-dimensional dense vectors are then used as inputs for interaction prediction models.(2)We construct a deep collaborative filtering prediction model with multi-embeddings,named DCFME,which can jointly utilize the feature information provided by aforementioned two embeddings.The model learns the global/local and deep drug-target coupling interactions through two independent sub-modules.In addition,the model uses an improved focal loss function that concentrates the loss on sparse and hard examples in the training process.Extensive experiments have shown that DCFME achieves more significant performance improvement on sparse data sets,along with much better robustness and generalization capacity against changes in the data set.(3)We construct a meta-path-based prediction model with a co-attention mechanism,namely MPCNet.The model can effectively learn explicit representations for drug,target and their meta-path-based context.Then these representations are enhanced via a co-attention mechanism.The experimental results prove the superiority of the model in predicting performance and interpretability.In addition,we establish a drug-target interaction prediction platform based on model to provide public prediction services.In summary,we learn the complex structure and semantic information of the networks through feature extraction methods.Moreover,we construct prediction models to screen the drug-target interactions.We have experimentally demonstrated that the two representation learning methods can capture the rich features of the heterogeneous network.In addition,the two prediction models display not only the effectiveness in prediction,but also advantages in robustness,generalization ability and interpretability.