The Study Of Deep Learning Based Method To Predict Drug-target Binding Interaction

The accurate identification of the drug-target binding affinity and binding site plays a crucial role in accelerating drug discovery and development.The higher the interaction strength between drug and target,the higher the probability that the drug will bind the target.In the early stages of drug discovery,one of the most critical goals is to identify candidate small molecules that bind tightly to target proteins.In addition,the determination of drug binding sites on target can reveal the drug-target recognition mechanism,thereby accelerating the process of drug design and development.Although traditional experiments have more reliable results,they usually are costly and time-consuming.Therefore,developing effective computational methods to accurately predict the drug-target binding affinity and binding site is of great siginificance.This article utilizes public available datasets and integrates various physical and chemical features to construct different prediction models based on deep learning algorithms from a new perspective,which improves the accuracies of drug-target binding affinity prediction and drug binding sites prediction.The main research contents and innovations of the article are listed as follows:(1)A drug-target binding affinity prediction method(Deep DTAF),which integrates traditional convolution and dilated convolution,has been proposed.In order to overcome the limitation of extracting both global and local features,Deep DTAF constructs a fusion of the traditional convolution and dilated convolution to capture multiscale interactions.The binding pocket is first used as the local input feature in the proposed model to predict protein-ligand binding affinity.Deep DTAF has higher prediction accuracy compared with other classical methods.Moreover,the analysis shows that Deep DTAF is promising and valuable in the treatment of Alzheimer’s disease and human immunodeficiency disease.(2)A variant graph neural network based on a physical Vina correction term,named Graphscore DTA,is proposed to predict the drug-target binding affinity.To solve the problem of poor interpretability of existing deep learning models and difficulty in capturing mutual information between drug and target,bidirectional information transfer mechanism is constructed to capture the mutual information between drug and target;multi-head attention mechanism is introduced to mark important atoms of drug and residues of target for providing the interpretation of the model;gated recurrent unit is designed to allocate the proportion of information distribution between amino acid(atom)node and super node in protein(drug)graph;the physical Vina scoring function is used as a correction term to optimize the graph neural network variant.A series of analyses show that Graphscore DTA has both better target selectivity and better drug selectivity compared to other classical prediction models.(3)A fusion of convolutional neural network and graph neural network based on target protein sequence and structure,named CGraph DTA,is proposed to predict the drug-target binding affinity.To address the problems that poor interpretability of existing deep learning models and the target sequence and structure information cannot be well integrated in existing methods,CGraph DTA transforms the protein structure into graph and uses the evolutionary information of the target sequence as the node features of the graph in the target channel.Meanwhile,multi-head attention mechanism is introduced to provide the interpretation of the model.The pre-training model is used to extract high-dimensional features of small molecules for charactering different small molecules.The multiscale convolutional neural networks and graph neural networks are constructed to extract sequence features and structure features,respectively.The analysis shows that CGraph DTA can successfully capture noncovalent interactions in target-ligand pairs and can be used for drug screening of specific target.In conclusion,CGraph DTA is a reliable tool to predict the drug-target binding affinity.(4)A convolutional neural network model based on target sequence and structure,named RLBind,is proposed for predicting small molecule binding sites on target RNA.To overcome the difficulty of capturing global and local features in existing RNA binding site prediction methods,RLBind combines target sequence information and structure information to extract features from global and local sequence channels,respectively.RLBind is the first deep learning model to specifically predict RNAtargeted small molecule binding sites.The analysis shows that the global channel information can help capture long-range interactions,while the local channel information can help capture short-range interactions.RLBind remains a potential useful tool for predicting drug molecule binding sites on RNA even though RNA 3D structures are unavailable.