An Ensemble Based Top Performing Approach for NCI-DREAM Drug Sensitivity Prediction Challenge

Abstract

We consider the problem of predicting sensitivity of cancer cell lines to new drugs based on supervised learning on genomic profiles. The genetic and epigenetic characterization of a cell line provides observations on various aspects of regulation including DNA copy number variations, gene expression, DNA methylation and protein abundance. To extract relevant information from the various data types, we applied a random forest based approach to generate sensitivity predictions from each type of data and combined the predictions in a linear regression model to generate the final drug sensitivity prediction. Our approach when applied to the NCI-DREAM drug sensitivity prediction challenge was a top performer among 47 teams and produced high accuracy predictions. Our results show that the incorporation of multiple genomic characterizations lowered the mean and variance of the estimated bootstrap prediction error. We also applied our approach to the Cancer Cell Line Encyclopedia database for sensitivity prediction and the ability to extract the top targets of an anti-cancer drug. The results illustrate the effectiveness of our approach in predicting drug sensitivity from heterogeneous genomic datasets.

Algorithms for Drug Sensitivity Prediction

“…Random forest regression [58] has become a commonly-used tool in multiple prediction scenarios [25,48,[85][86][87][88][89][90][91][92] due to its high accuracy and ability to handle large features with small samples. Random forest [58] combines the two concepts of bagging and random selection of features [93][94][95] by generating a set of T regression trees where the training set for each tree is selected using bootstrap sampling from the original sample set, and the features considered for partitioning at each node are a random subset of the original set of features.…”

“…Random forests are sufficiently robust to noise, but the biological interpretability of random forests is limited. Random forest was one of the top performing algorithms in the NCI-DREAM drug sensitivity prediction challenge [25,48], and it has been used in multiple other drug sensitivity studies [34,66,85,97]. Improvements to random forests in terms of incorporating multi-task learning that utilizes the relationships between output drug responses has been considered in [34].…”

“…Relative measurements, such as the Akaike Information Criterion (AIC) [118], Bayesian Information Criterion (BIC) [70], minimum description length [119] and Vapnik-Chervonenkis (VC) dimension [74], are also occasionally considered [67,120,121]. The sample selection approaches commonly considered in drug sensitivity predictive model evaluation involve techniques, such as k-fold cross-validation [32], bootstrap [48] or hold-out [25].…”

“…For instance, if the drugs considered are kinase inhibitors, a potential approach to feature selection might involve modeling from a set of around 500 available kinases [47,48]. Other approaches can include incorporating pathway knowledge in data-driven feature selection, such as pathway-based elastic net regularization [49], biological pathway-based feature selection integrating signalizing and regulatory pathways with gene expression data to select significant features that are minimally redundant [50] or the use of the activation status of signaling pathways as features [51].…”

Combination therapy design for maximizing sensitivity and minimizing toxicity

“…Design of drug therapies for cancer have primarily been considered from the perspective of sensitivity prediction using genetic characterizations as the predictor variables [1][2][3]. The genetic characterization based methodologies have severe limitations when the cancer type shows numerous aberrations among the samples and consequently predicting sensitivity based on similar steady state genetic characterizations provide limited accuracy.…”