Research On Drug Relocation Method Based On Deep Learning

The development of a drug is complex and expensive,and repositioning of existing drugs can further unlock the potential value of the drug.With the advancement of pharmacological research and accumulation of pharmaceutical experimental data,computational drug repositioning methods were born.Computational prediction models the association between drug characteristics and disease phenotypes based on drug molecular structure,transcriptomics,genomics,metabolomics,and proteomics,and can screen potential indication candidates for specific drugs quickly and accurately.However,due to the heterogeneity and complexity of biopharmaceutical data,graph structures are used to organize and characterize various entities and relationships in biopharmaceuticals,and efficient and accurate computational prediction on graph structures is still a challenge.Given the great success of deep learning in artificial intelligence,deep neural networks are beginning to be used for drug-disease association modeling on graph structures.In this paper,deep learning theories and methods are used to process graph-structured biomedical data,and three innovative approaches are proposed to address the heterogeneity of nodes,the multiscale nature of local topology,the specificity of drugs themselves,and the lack of association in drug-disease heterogeneous networks.（1）Multi-scale topological drug repositioning method（MTDR）MTDR integrates topological structure information at different scales from multiple drug-disease heterogeneous networks for drug repositioning.Based on the known biological similarity laws and the accessibility of nodes in the network,MTDR proposes a way to construct features at the drug-disease pair level.Based on the attention mechanism and bidirectional long-and short-term memory network,MTDR discriminates the importance of topologies at different scales and encodes the topological features captured by random walking into a low-dimensional feature space to obtain a comprehensive representation of drug-disease.In addition,multiple heterogeneous networks constructed from drug molecular structures,drug targets,etc.reflect the drug-disease g from different perspectives,and MTDR improves the accuracy of drug repositioning by integrating various properties of drugs and taking advantage of multi-source data through a convolutional neural network with an attention mechanism.（2）Meta-learning based method for drug-sensitive relocalization（DSDR）DSDR takes into account the specificity of drugs and the sparseness of drug-disease association samples,based on small-sample learning theory,and divides the available association samples into meta-tasks according to different drugs.A meta-learner based on recurrent neural network and attention mechanism is trained and tuned to distinguish different drugs and efficiently learn the drug-disease association patterns through these meta-tasks.This meta-learner is able to tailor each drug to achieve drug-sensitive repositioning.Moreover,instead of relying too much on pre-assessed drug similarity and disease similarity,DSDR uses an encoder based on an attention mechanism to distinguish the contributions of heterogeneous neighbors.A low-dimensional dense representation of each node is obtained based on the learning of the neighbor topology.（3）Topology reconstruction drug repositioning method（TRBDR）There is a certain amount of noise in the drug-disease heterogeneous network and there may be some missing edges.TRBDR uses a subgraph-based graph convolution to encode drug-disease topological features,and a perturbation is actively added to simulate the wrong edge information based on presence and absence when constructing the features of the training samples.To further explore the potential of drug-disease associations,an association-aware module based on random walking and Transformer structures was used to reconstruct the topology in the heterogeneous network.TRBDR also uses a recurrent neural network to generate subgraph representations with the addition of perturbations.A multilayer perceptron based association inference module evaluates the association likelihood by integrating node-level,subgraph-level features and the amount of perturbation changes.Experimental results show that TRBDR can greatly improve the robustness of the model in relocation.In order to evaluate the reliability and accuracy of the proposed method,six classical methods were chosen as the reference baseline for the repositioning performance.In addition,the prediction results of the model were validated in real clinical data,and the repositioning associations predicted by the model were proved to be reliable.