Drug-Target Binding Affinity Prediction Based On Deep Learning

Drug discovery is an important topic related to global human health.Pharmaceutical companies and drug scientists have focused on relying on existing drug knowledge to develop targeted therapies.Traditional screening of drug-target interactions is time-consuming and expensive when conducted based on biological experimental assays.Moreover,increasingly efficient processing is necessary to accumulate extensive histological data.In recent years,deep learning has been able to effectively explore drug-target interaction mechanisms by applying enriched datasets as well as its powerful feature learning capability.Most current computational approaches for predicting drug-target interactions are viewed as binary classification problems.However,this ignores the advantage of binding affinity strength.Furthermore,the use of binding affinity values can continue to precisely limit options when searching for candidate drug-target pairs.In a study of drug-target binding affinity prediction,the input information cannot effectively represent drug and target characteristics,and the model optimization problem still exists.Therefore,this thesis mainly emphasizes two aspects of network construction based on sequential representation as well as the network construction based on topology and multiple target representations.In terms of network construction based on sequential representation,as the 3D structure for most target proteins is unknown,input information based on sequencing is the most intuitive and easily obtained representation.Therefore,this thesis uses the amino acid sequence,secondary structure sequence,as well as drug molecule sequence and molecular fingerprint information to describe input features.The sequence and structure features are learned by designing a two-channel sequence analysis module as well as a structure information analysis module.The predicted binding affinity values are obtained based on an ensemble bagging LightGBM regression model.This method delivers 1.5% CI increase on KIBA dataset and 1% increase on Davis dataset.In terms of the network construction based on topology and multiple target representations,considering that drugs’ topology representation is closer to molecules’ nature,and learning contextual semantics from extensive unlabeled target protein data can more effectively represent homologous features,a baseline model is constructed for regression analysis.This model combines a topological representation of drugs by using a graph convolution neural network,a priori representation,a contact map with a variable-length recurrent neural network,as well as a residual neural network for feature extraction.This thesis also compares the performance of utilizing different target sequences,structure information representations,and neural network variants.From a drug discovery perspective,the methods provided in this thesis can accurately predict drug target binding affinity values.They can also provide computational methods for screening candidate pairs with high binding affinity values,which can promote new drug development and drug repositioning.According to the computational perspective,the two network construction methods proposed in this thesis,based on sequential representation and topology along with multiple target representation,validate the importance of drug and target sequence and structure representation,and are also helpful references for designing models to extract input information.