hen CHER adjusts the priors according to the distribution of frequency and the similarity between phenotypes. This iterative sharing between drugs is central to the learning power of CHER. During each iteration we utilize L0-norm regularized regression to select predictive features for sensitivity to each drug. In the L0-norm regularized regression, a penalty is applied proportional to the number of features added to the model, as in classical stepwise regression methods, but the features added to the model are not shrunk as in lasso or elastic-net. L0-norm regularization has several advantages. First, the regularization term in the regression is nonparametric, since the sparse selection of predictors in L0-norm regularization is guided by the minimum description length, where selection of each feature is encoded as a cost or penalty that ensures sparsity of the model. Second, the correspondence between MDL and Bayesian statistics allows us to iteratively adjust our belief by setting the cost of each feature according to the probability of that feature being selected. At each iteration, we use L0-norm regularized regression with bootstrapping to build a probability distribution for each feature based on the number of times it was selected. PubMed ID:http://www.ncbi.nlm.nih.gov/pubmed/19752732 This prior distribution is further adjusted by sharing information between drugs, constructing a penalty for feature selection in the next iteration. Third, we use a greedy algorithm to efficiently construct an L0-norm normalized regression; the models resulting from this search have been demonstrated to have excellent performance. The consideration of contextual predictors requires that the search space include the interaction between genomic features and contexts. While such large feature space may pose challenges for many algorithms, the greedysearch allows CHER to efficiently seek the relevant predictors in this large feature space. To evaluate CHER’s performance, we test it on a synthetic dataset that is simulated from the real data. We compare CHER to the elastic net algorithm previously used for this data and evaluate three Aphrodine supplier metrics: precision, recall, and F-measure. F-measure 4 / 22 Context Sensitive Modeling of Cancer Drug Sensitivity Fig 2. Comparison of performance of CHER and elastic net on synthetic data. Bootstrapped elastic net is compared to bootstrapped CHER. A threshold of 0.3 and 0.5 are applied to the relevant frequency to determine robust features in CHER and elastic, respectively. The precision, recall, Fmeasure of each phenotype from EN is plotted against that from CHER. The first row shows the results of CHER from the first iteration and the second row the results of CHER from the 10th iteration. Each dot represents a PubMed ID:http://www.ncbi.nlm.nih.gov/pubmed/19754356 phenotype, 
colored by the noise level added. doi:10.1371/journal.pone.0133850.g002 scores the harmonic mean of precision and recall and represents overall performance of the two algorithms. CHER trades off some recall to produce higher precision compared to the elastic net. In biological applications precision is often preferred to recall, since minimizing false positives saves future costly experimental validations. Thus, precision and F-measure scores in the final iterations suggest the overall superiority of CHER in identifying correct predictors. Application of CHER to CCLE datasets CHER takes advantage of pooling samples from similar cancers to increase power. We constructed test datasets based on prior knowledge of cancer similarity and the number of availabl