Prediction of Drug Efficiency by Transferring Gene Expression Data from Cell Lines to Cancer Patients

The paper represents a novel approach for individual medical treatment in oncology, based on machine learning with transferring gene expression data, obtained on cell lines, onto individual cancer patients for drug efficiency prediction. We give a detailed analysis how to build drug response classifiers, on the example of three experimental pairs of data "kind of cancer/chosen drug for treatment". The main hardness of the problem was the meager size of patient training data: it is many many hundred times smaller than a dimensionality of original feature space. The core feature of our transfer technique is to avoid extrapolation in the feature space when make any predictions of the clinical outcome of the treatment for a patient using gene expression data for cell lines. We can assure that there is no extrapolation by special selection of dimensions of the feature space, which provide sufficient number, say M, of cell line points both below and above any point that correspond to a patient. Additionally, in a manner that is a little similar to the k nearest neighbor (kNN) method, after the selection of feature subspace, we take into account only K cell line points that are closer to a patient's point in the selected subspace. Having varied different feasible values of K and M, we showed that the predictor's accuracy considered AUC, for all three cases of cancer-like diseases are equal or higher than 0.7.