A set of nine biological endpoints, comprising data for bioconcentration, biotransformation, AChE inhibition and acute aquatic toxicity, for a set of twelve organophosphorothionates was modelled using a set of ten non-empirical, mainly quantum-chemical, descriptors, employing two different techniques, viz. Multiple Linear Regression and Partial Least Squares projection to latent variables regression.

For these compounds and biological endpoints, PLS models are not only quantitatively better than MLR models, as judged by model fit (r²) and ‘predictivity’ (PRESS/SS_Y), but it is also shown that there exist ways of mechanistically interpreting the results of the PLS models. These interpretations can be used to come to a better understanding of the processes involved in organophosphorothionate toxicity and to generate even better, more mechanistically based models. This finding clearly refutes the commonly held belief that PLS models are inherently more difficult to interpret than comparable MLR models.