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

Both the root mean square error (RMSE) and the mean absolute error (MAE) are regularly employed in model evaluation studies. Willmott and Matsuura (2005) have suggested that the RMSE is not a good indicator of average model performance and might be a misleading indicator of average error, and thus the MAE would be a better metric for that purpose. While some concerns over using RMSE raised by Willmott and Matsuura (2005) and Willmott et al. (2009) are valid, the proposed avoidance of RMSE in favor of MAE is not the solution. Citing the aforementioned papers, many researchers chose MAE over RMSE to present their model evaluation statistics when presenting or adding the RMSE measures could be more beneficial. In this technical note, we demonstrate that the RMSE is not ambiguous in its meaning, contrary to what was claimed by Willmott et al. (2009). The RMSE is more appropriate to represent model performance than the MAE when the error distribution is expected to be Gaussian. In addition, we show that the RMSE satisfies the triangle inequality requirement for a distance metric, whereas Willmott et al. (2009) indicated that the sums-of-squares-based statistics do not satisfy this rule. In the end, we discussed some circumstances where using the RMSE will be more beneficial. However, we do not contend that the RMSE is superior over the MAE. Instead, a combination of metrics, including but certainly not limited to RMSEs and MAEs, are often required to assess model performance.

The main utility of QSAR models is their ability to predict activities/properties for new chemicals, and this external prediction ability is evaluated by means of various validation criteria. As a measure for such evaluation the OECD guidelines have proposed the predictive squared correlation coefficient Q(2)(F1) (Shi et al.). However, other validation criteria have been proposed by other authors: the Golbraikh-Tropsha method, r(2)(m) (Roy), Q(2)(F2) (Schüürmann et al.), Q(2)(F3) (Consonni et al.). In QSAR studies these measures are usually in accordance, though this is not always the case, thus doubts can arise when contradictory results are obtained. It is likely that none of the aforementioned criteria is the best in every situation, so a comparative study using simulated data sets is proposed here, using threshold values suggested by the proponents or those widely used in QSAR modeling. In addition, a different and simple external validation measure, the concordance correlation coefficient (CCC), is proposed and compared with other criteria. Huge data sets were used to study the general behavior of validation measures, and the concordance correlation coefficient was shown to be the most restrictive. On using simulated data sets of a more realistic size, it was found that CCC was broadly in agreement, about 96% of the time, with other validation measures in accepting models as predictive, and in almost all the examples it was the most precautionary. The proposed concordance correlation coefficient also works well on real data sets, where it seems to be more stable, and helps in making decisions when the validation measures are in conflict. Since it is conceptually simple, and given its stability and restrictiveness, we propose the concordance correlation coefficient as a complementary, or alternative, more prudent measure of a QSAR model to be externally predictive.