Research On Drug Discovery Methods Based On Multi-view Network Representation Learning

Drug development is of great significance to human health and disease.However,the development of a new pharmaceutical is very time consuming,expensive and high risk.Re-cently,with the explosive growth of biomedical data,deep learning as a typical artificial intelligence technology,is promoting the development of drug discovery.Especially,bio-logical network-based representation learning technologies have achieved great success in the field of drug discovery.However,there are many challenges including low accuracy,over-fitting phenomenon,and poor generalization ability in representation learning-based drug discovery,because of sparsely labeled data,and multi-level relationships,multi-scale semantic,and multi-view information among biological entities.Therefore,based on bi-ological heterogeneous network,this thesis proposes multi-view network representation learning technologies for precise drug discovery.A multi-view attention-based deep learning framework for anticancer drug response prediction:The internal correlations of feature items and external relationships of biomedi-cal entities indicate different biological mechanisms.However,previous methods can only capture a single level of relationships,resulting in poor performance.Therefore,this work proposes Multi DRP that is a multi-view attention-based deep learning framework for an-ticancer drug response prediction.On the one hand,cell line network-and drug network-driven graph attention models are developed to capture the external relationship among entities.On the other hand,we also integrate multi-head self-attention models to enhance the correlation among internal feature items in biomedical entities.Finally,the multi-view representations are fused via residual networks.In various prediction scenarios,Multi DRP had achieved great performance.Based on network proximity,GOBP enrichment,and drug pathway association analysis,we found that Multi DRP could accurately identify sensitive and resistant drugs for cancer cell lines.The in vitro experiments for lung cell line NCI-H23 showed that the predicted eight drugs revealed high sensitivity values,seven of which exhibited the IC₅₀values of less than 10n M.Biological network-based representation learning to discover anti-inflammatory agents for COVID-19:Excessive inflammatory response is a key factor leading to the death of pa-tients with COVID-19.However,in the early stage of COVID-19 outbreak,there is the sparse data of anti-inflammatory agents for COVID-19 and complex semantic relationships among biomedical entities,thus limiting the deep learning-based precision drug discov-ery.Therefore,this study proposes a deep representation on biological networks,termed Deep R2cov,to discover potential agents for treating the excessive inflammatory response in COVID-19 patients.By exploring the multi-hub characteristics of biological heteroge-neous networks,we construct the specific meta-path of biological networks,and design a path entity mask-based self-supervised learning algorithm to learn complex semantic re-lationships among entities.Next,Deep R2cov performs Connectivity Map and Pub Med publication analysis to further narrow the space for candidate drugs.Finally,we predicted22 potential drugs,and verified the anti-inflammatory mechanism of action and the binding modes via Pub Med literature data,clinical reports and molecular docking technologies.The self-supervised representation algorithm were applied to five biomedical applications,and achieved great prediction performance.A self-supervised learning approach integrating local-global representations for biomed-ical link prediction:Link prediction in biological networks is of great significance to life sciences,such as drug repositioning and adverse drug reaction prediction.Self-supervised representation learning techniques have been successfully applied to biomedical link pre-diction.However,existing representation learning methods are difficult to simultaneously capture local and global associations among entities.In addition,there is a poor correlation between most of self-supervised tasks and link prediction,thus leading to the serious neg-ative transfer phenomenon.Therefore,this work proposes Bio ERP that is a self-supervised learning approach integrating local-global representations for biomedical link prediction.Based on Transformer encoders,entity mask-and path detection-driven self-supervised rep-resentation learning techniques are developed to simultaneously capture the local and global semantic among biological entities.The path detection task can alleviate the negative trans-fer phenomenon,because it is an extension of link prediction.The experimental results on seven datasets suggested that Bio ERP achieved the better performance than 28 baselines.Multi-task adversarial learning-based self-supervised molecular representation approaches for drug discovery:Most molecular representation methods are single task-based self-supervised models,thus ignoring the multi-view information within biological heteroge-neous networks.How to effectively combine multiple self-supervised task to generate more generalization representations is the key to further promote drug discovery.Therefore,this paper proposes MSSL2drug that is a multi-task adversarial learning-based self-supervised molecular representation method for precise drug discovery.Inspired by various view fea-tures including structure,semantic,and attribute in heterogeneous biomedical networks,we design six basic self-supervised tasks and 15 multi-task combinations.Importantly,these combinations generate molecular representations by a graph attention-based multi-task ad-versarial learning framework and are systematically evaluated in drug discovery.Finally,we found that（1）combinations of multi-view tasks achieved the best performance compared to other multi-task joint models.（2）The local-global combination models yielded higher performance than random two-task combinations when there are the same view features.The multi-view self-supervised model achieved higher results in different test scenarios than nine baselines,and found that vandetanib（K_D=28.6μM）and pazopanib（K_D=20.7μM）could bind to IL-6 with high affinity by surface plasmon resonance（SPR）assay.Based on the internal-external view,local-global view and structure-semantic-attribute view of biological networks,this paper designs a supervised multi-view attention frame-work（i.e.,Multi DRP）and three self-supervised representations learning technologies（i.e.,Deep R2cov,Bio ERP and MSSL2drug）are applied to different drug discovery tasks,such as anticancer drug response prediction and drug repositioning prediction.Experiment results showed that the proposed representation learning technology can improve the prediction accuracy of drug discovery,and is a key foundation for precise drug development.