Multi-Feature Learning Approaches For Drug-Target Interaction Prediction

Drug-target interaction(DTI)prediction plays a crucial role in drug discovery.The traditional experimental approaches need to rely on a lot of prior knowledge,biochemical experiments,and clinical experiments to clarify the potential drug-target interaction.However,it is a high-cost and time-consuming process.Thus,the computational approach for DTI prediction is urgent.With the significant success of deep neural networks in different fields,many works focus on deep learning-based models to predict new DTIs.However,the deep learning methods for DTI prediction exist many challenges:(1)how to realize the content association between drugs and targets by neural network model,so as to improve the multi-feature fusion of drug and target data;(2)how to establish a neural network model to realize biochemical semantics understanding in drug and target data,so as to improve the representation of drugs and targets;and(3)how to design the neural networks to model the interaction of drug-target pairs,so as to improve the ability to predict DTIs.This paper also focuses on DTI prediction tasks and carries out the research on its deep models.The computational method for DTI prediction is to predict the association between a drugtarget pair or the degree of the interaction by processing and analyzing data of drug-target pairs and mining the relationship between them.According to the challenges and research contents of DTI prediction,this paper carries out research from the aspects of representation learning,biological language understanding,and feature fusion and proposes multi-feature learning methods under different task backgrounds.The main achievements are summarized as follow:1.A deep drug-target interaction prediction method based on gated residual networks is proposed.Combining two mechanisms for information enhancing in deep models: the gate mechanism of long short-term memory neural networks and residual connection of residual network,a novel deep model for information processing and enhancement is designed to improve DTIs prediction.In this paper,the proposed gated residual block is taken as the basic unit,and the gated residual network is formed through its’ hierarchical connection.Then,a deep DTI prediction method based on the gated residual network is established.Through the proposed information enhancement mechanism,the association features are learned and extracted to optimize the DTI prediction.Experiments show that the proposed computational method in this work shows good performance in the prediction task of whether the drug-target pair is associated and has the ability for multi-task classification tasks.In addition,this paper also analyzes the effects of gate function in the gated residual block on predicting DTIs.2.A deep drug-target interaction prediction method based on multi-source and multifeature representation is proposed.It redefines the task as a multi-source task and extracts more relative information between drug-target pairs,so as to further improve the classification accuracy of DTI prediction.In order to address the problem of the loss of essential association information caused by the simple combination of drug data and target data in the existing work,a multi-source neural network based on a multi-feature representation is proposed to predict DTIs.The model consists of two parts: representation learning and feature fusion.The former is committed to generating the deep representations of drugs and targets by convolution neural networks and fully connected neural networks,respectively;The latter uses the fully connected neural networks to fuse the deep representations of drugs and targets to optimize DTI prediction.Experiments show that the proposed computational method effectively improves the accuracy of DTI prediction,and has been successfully applied to the task of predicting drugs associated with Alzheimer’s disease-related genes.3.A deep drug-target binding affinity score prediction method based on multiple attention blocks is proposed.It models the correlations between atoms by the relative position information between atoms,and optimizes the modeling of the interaction between deep drug representations and deep target representations,so as to improve the prediction of DTIs.For the existing methods(1)ignore the correlations between atoms,and(2)lose the fusion information by simply combining deep drug representations and deep target representations,a deep computational model with multiple attention blocks is proposed.Firstly,in the drug representation learning process,the complex compounds are fully modeled by enhancing the correlations between atoms;Secondly,the interactions of drug-target pairs are encoded by the multi-head attention mechanism,so as to improve the DTI prediction ability of the deep model.The experimental results show that the modeled correlations and interactions contribute to drug-target binding affinity prediction.In addition,this paper applies the trained model to predict the binding affinity score between the COVID-19 gene protein and approved drugs,and then provides the experimental results as a reference.4.A deep drug-target binding affinity score prediction method with multi-granularity encoding and multi-scale self-attention is proposed.It models the correlations of chemical groups from two aspects of the encoding method and deep neural network model,so as to improve the ability of DTI prediction.In terms of encoding methods,this paper first utilizes the word segmentation algorithm to generate two multi-granularity dictionaries for drugs and proteins,respectively.Then,the dictionaries are used to encode the chemical functional groups which include biochemical information in drug and protein sequences.As for the deep learning model,a multi-scale self-attention model is established by allocating different window sizes to a multi-head self-attention mechanism.It is designed to generate deep drug representations and deep target representations which include the rich multi-scale features of correlations.Experiments show that the proposed computational method has good performance on DTI prediction.In addition,the proposed multi-granularity encoding and various local patterns extracted by the multi-scale self-attention model contribute to predicting the binding affinity score of drug-target pairs.