Activity prediction of anti cancer drugs by using ensemble learning and drugs' signatures

Personalized cancer treatment is an ever-evolving approach due to complexity of cancer. As a part of personalized therapy, effectiveness of a drug on a cell line is measured at the laboratory environments. However, these experiments are backbreaking and money consuming. To surmount these difficulties, computational methods are used with the provided data sets. In the present study, we considered this as a regression problem and firstly designed an ensemble model by combining three different regression models to reduce prediction error for each drug-cell line pair. We defined our base models as gradient boosting regression, kernelized bayesian multi-task learning and trace-norm regularized multi-task learning. Two major data sets, genomics of drug sensitivity in cancer and cancer therapeutics response portal, were used to evaluate our method. Results of this evaluation show that predictions of ensemble method are significantly better than models per se. Furthermore, we report the cytotoxicty predictions of our model for the drug-cell line pairs that do not appear in the original data sets. The another method to predict anti cancer drug activity is similarity of signature regularized multi-task learning. To constitute this model, firstly, drug signature is generated by examining drug-cell line experiments found in LINCS data set. Then, by using these activity signatures of drugs, similarities between drugs are calculated and a powerful model which overperforms the models which ignore drug similarities is designed. Also this model is evaluated with genomics drug sensitivity in cancer and cancer therapeutics response portal data sets and results of the both data sets show that constituted model have significantly more predictive power than contrasted model.