The Study Of Deep Learning Based Multi-drug Synergy Prediction Model

Cancer is one of critical illnesses that seriously threaten human life and health,and its induced mortality rate is second only to cardiovascular disease.The main treatment methods for cancer include surgical treatment,radiotherapy and drug treatment.Currently,drug treatment is still an important and effective method for cancer treatment.Due to the diverse carcinogenic pathogenesis and complex development process,treated with a single targeted agent,the tumors of most patients tend to develop drug resistance and eventually lead to disease relapse in clinical cancer treatment.Compared with singleagent therapy,combinational therapy has been explored as a better alternative with better clinical efficacy,more durable drug response and reduced host toxicity and adverse side effects.According to US.FDA,the number of approved anti-cancer drugs is more than 200 and the number of two-drug combination could reach up to 19900.If three or more drugs are combined in a synergy therapy,the undirected experimental screening for drug combinations tends to become cost expensive,time-consuming and labor-intensive,rendering it infeasible to cover all possibilities.To address these challenges,we developed a novel,effective and accurate computational method based on deep learning framework for the prediction of synergistic multi-drug combinations(Deep MDS)through using a large-scale dataset that integrated by targets information,drug response data and large-scale genomic profile of cancer cell lines from diverse tissues.Without the limitation of drug numbers,Deep MDS could be applied to screen specific synergistic multi-drug(chemical compounds and TCM)combinations for specific cancer cells systemically and effectively.?.ReviewThis chapter reviews the research progress in the main mechanisms of drug resistance in tumor,combinational therapy strategies for tumor,and the application of artificial intelligence algorithms such as machine learning in the biomedical research,especially for prediction of synergistic anti-tumor drug combinations.Firstly,the main mechanisms of drug resistance in tumor were summarized from tumor heterogeneity,tumor microenvironment,tumor stem cells,anti-tumor drug inactivation,increased release of extracellular drugs,reduced drug absorption,changed drug metabolism,gene mutations of drug targets,the role of mi RNA in tumor resistance,tumor-immune system.Then the combinational therapy strategies for tumor were reviewed from p-gp inhibitors combined with anti-tumor drugs,apoptosis sensitizers combined with anti-tumor drugs,si RNA /mi RNA targeting MDR combined with anti-tumor drugs,the combination of TCM monomer and chemotherapeutic drugs,the combination of TCM extracts,TCM compound and chemotherapeutic drugs,and the combination of TCM and chemotherapeutic for drug resistance reversal.In furthrer,the application of artificial intelligence algorithms such as machine learning in the biomedical research,especially for prediction of synergistic anti-tumor drug combinations was systematically reviewed.In sum,all the reviewed research progress provided the basic data support and theoretical foundation for this work.?.Modeling data collection and processingGene expression data of 1000 cancer cell lines from 29 different tissues were collected from Array Express database,then the downloaded data was preprocessed based on R and Bioconductor package.Based on gene sets defined by c Bio Portal consisted of the cancer-related pathways,a total number of 215 feature genes were selected and their corresponding gene expression data were used as the final genomics information of each cancer cell line.A total of 1839 anti-cancer compounds and their 1093 target information were collected,including 265 chemical compounds from GDSC and 1574 natural compounds from NPACT.Then drug response data(IC50)of above drugs to 1000 cancer cell lines were collected.Finally,all the collected data was effectively integrated to construct a total of 201405 large-scale sample datasets with 1308 dimensions(Expression data of feature genes of canacer cells-Compound target information-Drug response data).Then the constructed datasets were randomly divided into a training set(80%)and a test set(20%),which were used to construct multi-drug synergy prediction models.?.Construction of synergistic anticancer multi-drug combinations predictionDeep learning algorithm and f,which were applied to predict and evaluate synergy effect of multi-drug combinations.Based on Keras framework,DNN classification and regression models were developed using Anaconda5.1and Python 3.6.Genomic information(gene expression data)of cell lines and target information of drugs as input were loaded in the nodes(also called neurons)of the input layer.The output of the constructed classification model were binary classification results,whose positive results represent the synergistic anti-cancer effect and negative results represent no synergistic anti-cancer effect.Then for the regression models,the output were the IC50 values,which were used to measure the synergistic anti-cancer effect.To further evaluate and compare model performance of DNN,four other traditional machine learning algorithms were applied to construct classification and regression prediction models based on the same datasets.Based on the KNeighbors Classifier and KNeighbors Regressor functions,KNN classification and regression models were constructed.Based on the Random Forest Classifier and Random Forest Regression functions,RF classification and regression models were constructed respectively.SVM classification and regression prediction models were constructed using svm.SVC and svm.SVR functions respectively.Gradient Boosting Classifier and Gradient Boosting Regressor functions were used to construct GBM classification and regression models.All models were trained based on 161124 training datasets for the initial ones.?.Parameter optimization and model performance evaluationGrid search algorithm and 10-fold cross validation were applied to seek the optimal parameters combinations for all the constructed classification and regression models.Meanwhile,a series of performance metrics were used to evaluate performance of optimized models.According to the optimization results,compared with other traditional machine learning algorithms,Deep MDS based on deep learning algorithm exhibited significant improved prediction performance.In specific,for the regression task,Deep MDS(with two hidden layers having 200 nodes in the first layer and 100 nodes in the second layer,dropout rate of 0.5,learning rate of 10-5,batch size of 128,epoch number of 300)achieved MSE of 2.50,RMSE of 1.58 and R2_score of 0.86.For the classification task,Deep MDS(with two hidden layers having 200 nodes in the first layer and 100 nodes in the second layer,dropout rate of 0.5,learning rate of 10-3,batch size of 32,epoch number of 500)achieved SEN of 95%,SPE of 93%,ACC of 94%,MCC 0f 0.88 and AUC of 0.97.?.Literature validation of model performanceThe synergy effect of 17 drug pairs from literature were predicted using Deep MDS,and the predicted results were subsequently validated by the experiment results of the literature.The experiment results of the literature showed that among the predicted top five drug pairs,three pairs(gefitinib and tamoxifen,sorafenib and tamoxifen,erlotinib and sorafenib)exerted strong synergy effect and two pairs(sorafenib and dasatinib,gefitinib and toremifene)presented synergy effect.On the other hand,among the predicted bottom six drug pairs,all the pairs(tamoxifen and flavopiridol,gefitinib and erlotinib,gefitinib and sorafenib,everolimus and BIBW-2992,sorafenib and everolimus,erlotinib and flavopiridol)showed additive effects and even antagonism.?.Prediction of synergistic effect of Multi-drug(chemical compounds and TCM)combinationsTo evaluate the applicability and accuracy of Deep MDS,Deep MDS were applied to predict the optimal synergistic drug cominations for MCF-7,MDA-MB-468 and MDAMB-231.Among the 120 combinations of chemical compounds,the top three synergy combinations for the MCF-7 cell line were ? 12(doxorubicin,docetaxel and gemcitabine),?7(doxorubicin,5-Fluorouracil and docetaxel)and ?18(doxorubicin,gemcitabine and paclitaxel).Also,for the MDA-MB-468 cell line,the top three combinations were ?12(doxorubicin,docetaxel and gemcitabine),?3(doxorubicin and docetaxel)and ? 7(doxorubicin,5-Fluorouracil and docetaxel).Then for the MDA-MB-231 cell line,the top three were ? 33(doxorubicin,gemcitabine,methotrexate and paclitaxel),?32(doxorubicin,docetaxel,gemcitabine,methotrexate and paclitaxel)and ? 59(5-Fluorouracil,docetaxel,methotrexate and epirubicin).Among the 26 component combinations of Prunella vulgaris.L,the top three synergy combinations for the MCF-7 were D2(ursolic acid and oleanolic acid)?D4(ursolic acid and kaempferol)and F4(ursolic acid,luteolin,quercetin and kaempferol);For the MDA-MB-468,the top three were D11(oleanolic acid and kaempferol),D7(luteolin and quercetin)and T11(luteolin,oleanolic acid and quercetin);For the MDA-MB-231,the top three were D4(ursolic acid and kaempferol),D11(oleanolic acid and oleanolic acid)and D2(ursolic acid and oleanolic acid).?.Experimental evaluation of predicted resultsIn vitro cell viability assay for each cancer cell line was conducted to evaluate the predicted results.For the nine drug combinations of MCF-7,the predicted results were almost consistent with experimental results,where the predicted top three ?12(30.88 n M),?7(56.52 n M)and ?18(88.62 n M)shown strong synergistic effect and the predicted bottom three ?9(2121.31 n M),?44(1499.70 n M)and ?41(6022.00 n M)exhibited additive effect even antagonism.Meanwhile,the optimal combination ?12(doxorubicin,docetaxel and gemcitabine,30.88 n M)of MCF-7 shown better synergy effect than clinical used combinations ?3(doxorubicin and docetaxel,377.18 n M)and ?55(gemcitabine,epirubicin and paclitaxel,814.10 n M).For the MDA-MB-468,the predicted top three were experimentally validated with strong synergy effect.In specific,the IC50 of ?12(doxorubicin,docetaxel and gemcitabine)was 115.50 n M;the IC50 of?3(doxorubicin and docetaxel)was 207.60 n M;the IC50 of ? 7(doxorubicin,5-fluorouracil and docetaxel)was 522.90 n M.The optimal combination ?12(115.50 n M)of MCF-7 shown better synergy effect than clinical used combinations ?3(207.60 n M)??55(774.20 n M).For the MDA-MB-231,the predicted top three were also experimentally validated with strong synergy effect.In specific,the IC50 of ? 33(doxorubicin,gemcitabine,methotrexate and paclitaxel)was 52 n M;the IC50 of ? 32(doxorubicin,docetaxel,gemcitabine,methotrexate and paclitaxel)was 219 n M;the IC50 of ?59(5-fluorouracil,docetaxel,methotrexate and epirubicin)was 244 n M.Also,the optimal combination ?33(52 n M)of MCF-7 shown better synergy effect than clinical used combinations ?3(4567 n M)and ?55(1506 n M).To evaluate that the developed model could screen specific drug combinations for specific cancer cells,the optimal combinations ?12(MCF-7 and MDA-MB-468)and ?33(MDA-MB-231)were used to conduct cross-validation on other cancer cells.According to the final results,the developed model was evaluated that it could predict the specific optimal drug combinations for specific cancer cells based on genomics information.In further,the predicted results of TCM was experimentally evaluated to identify the applicability and accuracy of Deep MDS.For the component combinations of Prunella vulgaris.L,IC50 of D2(ursolic acid and oleanolic acid)on MCF-7 was 50.95 ?M;IC50 of D11(oleanolic acid and kaempferol)on MDA-MB-468 was 21.51 ?M;IC50 of D4(ursolic acid and kaempferol)on MDA-MB-231 was 58.04 ?M.In sum,the applicability and accuracy of Deep MDS were well evaluated by above experimental results.?.The study of synergistic anti-cancer mechanisms of optimal drug combinationsIn order to identify the potential mechanisms of synergistic anti-cancer effect of optimal drug combinations,Gene Ontology(GO)enrichment analyses were carried out in terms of KEGG pathway and biology process.?12 of MCF-7 may include KEGG pathways of TGF-beta signaling pathway and Progesterone-mediated oocyte maturation and biology processes of positive regulation of fibroblast proliferation,transmembrane receptor protein tyrosine kinase signaling pathway and negative regulation of cell differentiation.? 12 of MDA-MB-468 may include KEGG pathways of Thyroid hormone signaling pathway,Transcriptional misregulation in cancer and Gn RH signaling pathway and biology processes of leukocyte differentiation and gliogenesis.?33 of MDA-MB-231 may include KEGG pathways of EGFR tyrosine kinase inhibtor resistance and RIG-I-like receptor signaling pathway and biology processes of phosphatidylinositol-mediated signaling and positive regulation of transcription of Notch receptor target.D2 of MCF-7 may include KEGG pathway of TGF-beta signaling pathway and biology process of regulation of protein serine/threonine kinase activity.D11 of MDA-MB-468 may include KEGG pathway of Progesterone-mediated oocyte maturation,HIF-1 signaling pathway and Endocrine resistance and biology processes of apoptotic signaling pathway,positive regulation of cell death and cytokine-mediated signaling pathway.D4 of MDA-MB-231 may include KEGG pathway of Endocrine resistance,Viral carcinogenesis and VEGF signaling pathway and biology process of negative regulation of cell proliferation;molecular function of protein heterodimerization activity.The results indicated that for the different cancer,the different drug combinations exerted synergistic anti-cancer effect via different mechanisms of action,and provide the basis for individualized cancer treatment.In sum,the deep learning based multi-drug synergy prediction model was successfully developed.Without the limitation of drug numbers,it could be applied to screen specific synergistic multi-drug(chemical compounds and TCM)combinations for specific cancer cells systemically and effectively.This work is expected to provide more effective combinational therapies for individualized cancer treatment as well as guide clinical rational drug use.Meanwhile,this work could also provide basic data support and theoretical foundation for the study and quality control of effective substances from traditional Chinese medicine.