Research On Key Algorithms And Applications Of Artificial Intelligence For Multidimensional Biomedical Data Integration Mining

In recent years,the rapid development of high-throughput omics technologies,highcontent screening technology,pharmaceutical informatics,and cheminformatics has led to the accumulation of large amounts of multi-source biomedical big data.On the basis of multidimensional biomedical data integration mining,Artificial Intelligence(AI)has gradually become a key technology for drug research and development.Specifically,AI technology has facilitated the progress of two critical areas in drug development,namely drug repositioning and drug combinations.These developments have opened up new opportunities for innovative drug development and disease treatment.However,despite the potential benefits of AI,the field of biomedicine still faces several challenges in applying AI algorithms.These challenges include the complexity of biomedical entity relationships and the interpretability of algorithms,which currently limit the practical application of AI.Currently,new graph network methods and interpretable artificial intelligence methods are emerging in many fields.Their application in the field of drug repositioning and drug combination has great potential to aid researchers in comprehending the mechanism of action and pharmacodynamic characteristics of drug,which may lead to revolutionary progress in drug research and development.Based on the integration of multidimensional biomedical data,this study aims to explore key algorithms of graph network and interpretable AI for drug repositioning and drug combination.Firstly,to effectively integrate multi-dimensional biomedical data,this study proposes a networkbased algorithm for multi-dimensional data integration.Then,leveraging the integrated multi-dimensional biomedical data,graph network AI methods are employed for data mining,and a set of data association mining frameworks are developed for drug repositioning and drug combination.Finally,to elucidate the prediction of biomedical associations,interpretable computing frameworks for biomedical entity association prediction are constructed using knowledge and data dual-driven AI algorithms.Firstly,aiming at multidimensional biomedical data integration,this study established a multidimensional data integration algorithm based on the random walk process on the network.As an example,the algorithm is applied to cancer multi-omics data integration,exhibiting strong anti-noise ability and high sensitivity to network structure perception,thus performing well.Based on the integration of multidimensional biomedical data,in terms of drug repositioning and drug combinations based on graph network,this study designed and implemented a series of graph neural network algorithms for drug-target association prediction,drug response prediction and drug combination prediction.First,for drugtarget association prediction,a heterogeneous network-based graph neural network model,DTI-HETA,is proposed.DTI-HETA employs the graph convolution strategy and graph attention mechanism to highlight the contributions of different neighborhood nodes.The model has excellent drug-target interaction prediction capabilities that can provide clues for drug repositioning.Second,for drug response prediction,Deep DRP,a graph neural network framework based on the multi-view attention mechanism,is proposed.By using the graph attention model and the multi-head self-attention mechanism,Deep DRP can effectively capture internal and external associations between biochemical entities.Deep DRP outperforms existing methods in multiple prediction scenarios,and in vitro experiments have successfully verified highly sensitive drugs predicted by the model.Finally,aiming at the prediction of drug combinations,a graph recurrent neural network model is developed to predict synergistic effect of drug combinations.The model uses gating mechanisms and gate recurrent unit to ensure that the representation of each node can absorb local and global information and potential association information more comprehensively.The model significantly outperforms state-of-the-art drug combination prediction algorithms.In addition,the novel drug combination predicted by the model was verified by in vitro experiments.These drug combinations can be further evaluated as potential combination therapy strategies.Based on the integration of multidimensional biomedical data,in terms of drug repositioning and drug combinations based on interpretable artificial intelligence,this study established prediction frameworks for drug treatment attributes prediction,anticancer drug combination prediction and biological entity association prediction based on AI algorithms which combines data-driven and knowledge-driven approaches.First,to predict drug treatment attributes,a contrast-activated knowledge embedding interpretable neural network is designed.The model achieves good performance using only 0.73% of the network parameters of a fully connected neural network.By embedding prior knowledge of biological pathways into neural networks and extracting importance based on “contrast activation”,the model captures key pathways in drug treatment attribute prediction and provides interpretable analysis.Secondly,to develop anti-cancer drug combination therapies,the synthetic lethality prediction model KNNSL and the drug combination prediction model SEEnergy are proposed.KNNSL achieves a balance between predictive and interpretive ability and identifies promising drug combinations such as combinations of MDM2 and CDK9 inhibitors,which exhibited significant anticancer effects in in vitro experiments.The explanation provided by the model further clarifies the related biological processes.SEEnergy provides explanatory analysis of drug synergies mechanism through biological knowledge-embedded neural networks and attention-based graph networks.The verification based on calculations and in vitro experiments shows that SEEnergy provides mechanism elucidation with application potential while having high prediction accuracy.Finally,for the association prediction of biological entities,a knowledge-embedded interpretable deep forest framework is proposed,which further promotes the development of knowledge-embedded interpretable AI models.The model is suitable for various biological entity association prediction tasks,such as drug combination prediction,drug-target association prediction and synthetic lethal prediction.The model uses the special properties of the tree structure to explain the model at multiple levels from the aspects of important biological process localization and decision rule extraction.The model exhibits excellent predictive power in various indicators and tasks and is helpful to reveal the underlying biological mechanisms.The innovation of this study mainly includes the following three aspects.Firstly,a network-based data integration algorithm is proposed for multidimensional biomedical data integration.Secondly,based on the integration of multidimensional biomedical data,a series of graph-based models are designed and implemented for drug repositioning and drug combinations,which provide efficient computational tools for drug-target association prediction,drug response prediction,and drug combination prediction.By leveraging the association information between biological entities,these models improve drug discovery efficiency and reduce costs.Thirdly,this study proposes new explainable AI models that integrate data-driven and knowledge-driven approaches.By incorporating prior biological knowledge,these models achieve an optimal balance between predictability and interpretability in drug treatment attribute prediction,anti-cancer drug combination development,and biological entity association prediction.Moreover,the models’ interpretability enables the discovery of hypotheses about drug mechanisms of action,which can be further validated.This feature supports more accurate drug discovery,facilitates the understanding of drug mechanisms,and guides clinical applications of drugs.