In an article published on March 20, 2018 in the peer-reviewed journal Toxicological Sciences, Melissa Van Bossuyt and colleagues from the Department of Food, Medicines and Consumer Safety, Scientific Institute of Public Health, Brussels, Belgium evaluated the performance of several in silico models for prediction of mutagenicity of chemicals found in food contact materials (FCMs).
The authors built on the results of their previous studies, where chemicals that could be constituents of printed paper and board FCM (FPF reported) were screened using four in silico models for mutagenicity (FPF reported). These models were either rule-based (also called structure-activity relationship (SAR) models, represented by Toxtree and Derek Nexus) or statistics-based (also called quantitative SAR (QSAR) models, represented by VEGA Consensus and Sarah Nexus). In the present study, the scientists collected experimental data on the substances’ mutagenicity from several different sources. These data were then compared with in silico predictions to evaluate the models’ accuracy, sensitivity, specificity, and predictivity.
The four individual models showed varying prediction capacity for substances within their application domain. In particular, all models had high accuracy, ranging from 83% (Toxtree) up to 96% (Sarah Nexus), but varied widely in sensitivity (from 62% to 90%) and positive predictivity (from 49% to 87%). In general, QSAR models showed higher sensitivity, particularly for the substances inside their application domain. Among these, both the free VEGA Consensus and the commercial Sara Nexus performed similarly well. The performance could be further improved when the two QSAR models’ results were combined. However, higher sensitivity often coincided with lower specificity.
Much poorer model performance was observed for substances outside the models’ applicability domain, and in general for new compounds outside the training set. Commenting on the latter, the authors noted that, “[a]lthough combining the two models improves the prediction performance, it is important to note that still only one out of two mutagens is picked up.”
Irina Julia Adaktylou (October 10, 2018). “Data and toxicology: the challenges ahead.” Chemical Watch
Van Bossuyt, M., et al. (2018). “Performance of in silico models for mutagenicity prediction of food contact materials.” Toxicological Sciences (published March 20, 2018).