Deep Learning-based Classification Of Synergistic Anticancer Drug Combinations

Combination medication is a classic treatment proposed in the field of medicine for cancer and other major diseases.It is favored for its remarkable therapeutic effect and strong drug resistance.However,the combination of drugs in some cases will also produce negative effects such as enhanced toxic effects and weakened efficacy.At the same time,the traditional clinical analysis methods are far from enough to verify the massive drug combination schemes one by one.The different pathogenic factors and action paths among different cancers also bring about great differences in the effect of drug combinations.The analysis of drug synergy based on therapeutic targets provides a basis for the formulation of the optimal drug regimen,which has great practical significance.In the field of pharmacy,according to the therapeutic effect,the combination of drugs can be divided into three types: antagonism,additive action,and synergy,which correspond to the three states of mutual cancellation,superposition and promotion of the effects of each component.In recent years,drug interactions,especially synergy,have attracted great attention.With the development of high-throughput analysis technology,a considerable number of drug combination data has been obtained,which has spawned many methods for drug synergy analysis.However,the analysis of drug combination still faces many problems,such as difficulty in feature extraction and poor recognition effect.(1)This thesis proposes a new method MTP-DCNN based on molecular fingerprints and multi-target protein features to identify the combination of anti-cancer drugs.MTP-DCNN uses the molecular fingerprint characteristics of drug compounds and the multi target protein(MTP)of cancer cell lines to recognize the synergistic drug combinations.Firstly,multi-target genes and their expressed protein sequences specific to cancer cell lines are extracted.Then,deep convolution neural network(DCNN)is used to reduce the dimension of multi-target protein features.Finally,the features are concatenated with molecular fingerprint features and input into deep neural network for Collaborative drug combination classification.A benchmark data set involving 21 cancer cell lines and 35 drug compounds were tested and compared with the best performance method in the field.An overall accuracy of 93% and kappa correlation of 0.55 are achieved in the collaborative drug combination recognition,which are 4% and 0.04 higher than the compared best method respectively.(2)Based on the same characteristics of drug molecular fingerprints and multi-target proteins,this thesis combines attention mechanism and attention based gated recurrent unit(AGRU)to construct a deep recurrent neural network MTP-AGRU,which can fully extract context information.MTP-AGRU has achieved an overall accuracy of 93% and a kappa correlation of 0.53.The performance of this model is close to that of MTP-DCNN.(3)In order to facilitate the use of relevant researchers,an online prediction service website was developed to provide efficient collaborative drug combination identification service.