Abstract :
During the early phase of drug discovery, machine learning methods are often utilized to select compounds to send for experimental screening. In order to accomplish this goal, any method that can provide estimates of error rate for a given set of predictions is an extremely valuable tool. In this paper we compare Platt Calibration Algorithm and recently introduced Conformal Algorithm to control the error rate in the sense of precision while preserving the ability to identify as many compounds as possible (recall) that are highly likely to be bio-active in a certain context. We empirically evaluate and compare the performance of Platt´s Calibration and offline Mondrian ICM in the context of SVM-based classification on 75 distinct classification problems. We perform this evaluation in the real world setting where the true class labels of compounds are unknown at the time of prediction and are only revealed after the biological experiment is completed. Our empirical results show that under this setting, offline Mondrian ICM and Platt Calibration are not able to bound precision rates very well on an absolute basis. Comparatively, Mondrian ICM, even though not theoretically designed to control precision directly, compares favorably with Platt Calibration for this task.
