Erroneous predictions, QSARS

Only within these limits of the descriptor values can the relationship be assumed to hold. Beyond this domain, the model may reveal a different type of relationship of the activity on the structural descriptors, which evidently must result in erroneous predictions. A striking example of the pitfalls of inapt extrapolations can be demonstrated with the application of a fourth-order polynominal log BCF/log QSAR (Connell and Hawker, 1988) outside its parameter range (log 2.6-9.8). Because of the curvature of the model, higher BCF estimates result for compounds with log P = 0 than for... 

The main conclusion is that the troubles with QSARs are mainly in their application the erroneous use, in particular by non-experts, of models which are (i) not correctly validated, (ii) applied to chemicals out of the applicability domain or (iii) not pertinent (descriptive versus predictive models). 

A further assessment of the quahty of a QSAR can be made in terms of which compounds (if any) are poorly predicted by the model or do not fit it well. These compounds are commonly termed outliers, and a number of techniques exist to identify them (Eriksson et al., 2003). Thrae are a number of reasons for the occurrence of outliers (Cronin and Schultz, 2003). In aquatic ecotoxicology, outliCTs are frequently associated with chemicals acting by a different mechanism of toxic action, and much has been learned from their analysis (cf. Lipnick, 1991). Another reason for the occurrence of outliers is the creation of an insignificant model that is simply not predictive. Also, outlie may result from the use of poor (i.e., low-quality) or erroneous data in a model. 

---