Robust Uncertainty Estimates for Unbalanced Data Sets
Mathematical models of quantitative structure-activity relationships (QSARs) play a key role in qualifying synthesis ideas, drug candidates and leads. In many cases (e.g.,solubility), these are regression models but classification models–e.g., for mutagenicity or CYP inhibition–are also important. Overall sensitivity and specificity are commonly used as performance metrics for such models. Some predictions are more clear cut than others, however, so it is often also important to obtain robust confidence estimates for individual predictions, especially in regulatory contexts.
13th German Conference on Chemoinformatics, November 5-7, 2017, Mainz, Germany
By Robert D. Clark, Marvin Waldman