Analyzing drug sensitivity prediction based on dose response curve characteristics

Precision medicine for cancer involves design of drug sensitivity prediction models that can predict patient response to various drugs. The drug response is usually represented by a single feature such as Area Under the Curve or IC50 derived from the experimental dose response curve. In this article, we consider the idea that predicting the dose response curve and generating the curve features instead of directly predicting the curve characteristics can increase prediction accuracy. Using the cancer cell line encyclopedia database, we illustrate that predicting dose response curve points to calculate AUC instead of directly predicting AUC can reduce prediction mean square error and increase correlation between experimental and predicted values.