Building an efficient classifier for drug-induced cell death
University of California, San Diego
Human cancers are notoriously heterogeneous systems. In most cases, uncontrolled cell division results in genomic instability and massive variance in the expression of functional gene products. A consequence of this heterogeneity is that a tumor’s response to therapy is variable and generally unpredictable. For unresponsive tumors, there also may be different mechanisms of resistance. There is tremendous interest in being able to predict whether a given tumor will respond to a particular therapy. The question that remains is: Which features of a tumor yield the most predictive model? We address this problem computationally. Specifically, we propose that Monte Carlo sampling of a model of a drug mechanism can generate sufficient data to build an accurate predictor of efficacy and identify different mechanisms of resistance. To illustrate this, we first developed a method by which to derive an analytical expression for the steady state of an arbitrary drug mechanism. Letting the independent parameters of the model take on values from a distribution and then sampling from those distributions results in different numerical realizations of the model, reflecting the heterogeneity observed in human cancer. Using a small array of nVidia GTX 460’s, we rapidly sampled and simulated the response of the heterogeneous system to the therapeutic agent TRAIL. Quadratic Programming Feature Selection (QPFS) was then used to identify a minimal, non-redundant set of features from which to build a logistic model of TRAIL efficacy. Using only the top four or seven features identified by QPFS, we were able to achieve classification accuracies above 77 percent and 81 percent, respectively, compared to the full mechanistic model.