In this case study, based on the notorious PCBs, the performance of selected empirical (EM) versus non-empirical and quantum chemical (NEM-QC) descriptors in multivariate QSARs have been evaluated. The informational content of the EM descriptors has been evaluated with respect to the physical understandable NEM-QC descriptors. Models for estimating key parameters for risk assessment have been developed, based on two-dimensional EM and three-dimensional NEM-QC descriptors, respectively. In spite of the simplicity of the two-dimensional descriptors, no evidence of lower predictive ability of the EM descriptors compared to the NEM-QC descriptors was observed. Homogeneity of variance within and between experimental methods, species and PCB congeners has been analysed, disclosing the importance for the need for handling uncertainty aspects of the results obtained by different laboratories and methods. The latter appears crucial for the model developments when using data from more than one source.