A statistical analysis of an effective method to conduct in silico screening for active compounds

Abstract

Statistical analysis was conducted to study the efficiency of methods for active compound selection by comparing the method of screening a test compound set by an in silico method DOCK using a target protein (receptor) of a known structure versus the method of screening by the standard in vitro assays and also to determine how best to utilize the DOCK output variables. In this study we used DOCK output data on 327 compounds and the in vitro assay data on synthetic product resulting from an enzymatic reaction of a given substrate, and those compounds giving greater than 50% inhibition activity in an in vitro assay were considered to be active compounds. The representative variables were selected from a group of variables with mutually high correlation in the 108 DOCK output variables and subjected to liberal variable selection or conservative variable selection by the stepwise selection-elimination method of the logistic regression model, yielding 16 and 3 variables, respectively. These variables were then used for screening by the logistic regression method, and the performance was evaluated by the jackknife method (a performance evaluation method in which a measured value predicted from the n-1 observations removing the own predicted observation). The results indicated that elimination of about 80% of test compounds by DOCK in silico screening gave 80% sensitivity and 15% false positive rate. We demonstrate the usefulness of in silico screening using a prediction model by logistic regression. Copyright 2004 Chem-Bio Informatics Society.