Is it possible to improve the quality of predictions from an “intelligent” use of multiple QSAR/QSPR/QSTR models?

Abstract

AbstractQuantitative structure‐activity/property/toxicity relationship (QSAR/QSPR/QSTR) models are effectively employed to fill data gaps by predicting a given response from known structural features or physicochemical properties of new query compounds. The performance of a model should be assessed based on the quality of predictions checked through diverse validation metrics, which confirm the reliability of the developed QSAR models along with the acceptability of their prediction quality for untested compounds. There is an ongoing effort by QSAR modelers to improve the quality of predictions by lowering the predicted residuals for query compounds. In this endeavor, consensus models integrating all validated individual models were found to be more externally predictive than individual models in many previous studies. The objective of this work has been to explore whether the quality of predictions of external compounds can be enhanced through an “intelligent” selection of multiple models. The consensus predictions used in this study are not simple average of predictions from multiple models. It has been considered in the present study that a particular QSAR model may not be equally effective for prediction of all query compounds in the list. Our approach is different from the previous ones in that none of the previously reported methods considered selection of predictive models in a query compound specific way while at the same time using all or most of the valid models for the total set of query chemicals. We have implemented our approach in a software tool that is freely available via the web http://teqip.jdvu.ac.in/QSAR_Tools/ and http://dtclab.webs.com/software‐tools.