Anti-Cancer Drug Combination Prediction Based On Multi-Source Data And Deep Forest Framework

Combination therapy has shown an obvious efficacy on complex diseases such as cancer.However,even with high-throughput screens,experimental methods are insufficient to explore novel combination therapies.The cost of screening technology is too high and difficult to cover all drug combinations.There is an urgent need for drug combination prediction based on multiple data and artificial intelligence algorithms to help researchers find a number of effective new drug combinations quickly and at low cost.In this study,two methods based on drug multi-source data and deep forest framework are proposed for the prediction of cell line specific anti-cancer drug combinations.Then,potential molecular mechanism of drug combination synergy is explored and analyzed.Firstly,a new dataset containing the physical,chemical and biological properties of drugs is proposed.Then,a deep learning method based on deep forest framework is proposed to achieve the two-classification task and predict whether the drug combination is synergistic.The unit of processing imbalanced data and dimensionality reduction based on complexity measure is introduced.This can well alleviate the adverse effects on the classification process caused by class imbalanced,high feature dimensions and small number of samples.This method also improves the prediction accuracy for minority samples(synergistic drug combinations).Through several comparison schemes,this method shows excellent classification performance on the dataset.Next,analysis of the performance of different hyperparameter settings performed.The applicability of deep forest framework in drug combination prediction is systematically evaluated.Secondly,an enhanced cascade-based deep forest regressor,EC-DFR,is designed to predict the quantitative score of drug combination synergy.The enhanced cascade module can adjust the training set from sample space and feature space.This effectively improves the prediction performance of deep forest algorithm.EC-DFR shows excellent prediction performance in multiple benchmark datasets and is adaptable to datasets of different sizes.The results of cellular experiment further confirmed the predictive ability of EC-DFR.Through the analysis of contribution for each feature,it is found that the transcriptome data of drugs have a prominent contribution to the prediction.Further analysis find that drugs can promote the synergistic or antagonistic effect by regulating the transcriptional expression of key genes,which provides a theoretical clue for the experimental study of the mechanism of synergy.Based on the analysis of the key genes in drug combination prediction,the synthetic lethality between genes is further explored,and it is predicted that the synthetic lethal gene pair can assist the design of combination therapy.A computational framework based on similarity network fusion is proposed to predict potential synthetic lethal pairs,which achieves excellent prediction performance.Case analysis shows that the predicted results of this algorithm are consistent with the literatures in reducing the drug resistance of cells and designing the combination therapy.This study can promote the design of combination therapy and accelerate the discovery of anti-cancer drug combinations.