Global versus local QSPR models for persistent organic pollutants: balancing between predictivity and economy

Experimentally determined data on the key physicochemical parameters for halogenated congeners of persistent organic pollutants (POPs) are available only for a limited number of compounds. In the absence of experimental data, a range of computational methods can be applied to characterize those species for which experimental data is not available. One of the techniques widely used in this context is quantitative structure–property relationships (QSPR) approach. There are two ways to develop the QSPR models: using a more complex global model or fitting a simple local model that covers a specific class of chemically related compounds. The essence of the study was to investigate, if local models have significantly better explanatory and predictive ability than global models with wider applicability domains. Based on the obtained results, we concluded that whenever global models fulfill all quality recommendations by OECD, they would be applied in practice as more efficient ones in state of more time consuming procedure of modeling the particular groups of POPs one-by-one. On the contrary, local models are applicable to solve specific problems (i.e., related to only one group of POPs), when high-quality experimental data are available for a sufficient number of training and validation compounds.