A new metric for the evaluation of model performance in the field of virtual screening and quantitative structure-activity relationship applications is described. This metric has been termed the power metric and is defined as the fraction of the true positive rate divided by the sum of the true positive and false positive rates, for a given cutoff threshold. The performance of this metric is compared with alternative metrics such as the enrichment factor, the relative enrichment factor, the receiver operating curve enrichment factor, the correct classification rate, Matthews correlation coefficient and Cohen's kappa coefficient. The performance of this new metric is found to be quite robust with respect to variations in the applied cutoff threshold and ratio of the number of active compounds to the total number of compounds, and at the same time being sensitive to variations in model quality. It possesses the correct characteristics for its application in early-recognition virtual screening problems.
Keywords: Area under the curve (AUC); Cohen’s kappa coefficient (CKC); Correct classification rate (CCR); Enrichment factor; Matthews correlation coefficient (MCC); Metric; Model performance; Power metric (PM); Receiver operating curve enrichment factor (ROCE); Relative enrichment factor (REF); Virtual screening.