Organosilicon chemistry, a subsection of organic chemistry with unique chemical properties, is typically underrepresented in the training sets of (Q)SAR models and often outside the applicability domain.
Using bacterial reverse mutation as an endpoint for a proof-of-concept study, we compiled a peer-reviewed reference dataset with publicly available reliable Ames tests of more than 100 organosilicon substances to assess the predictive performance of fourteen in silico methods, including mechanistic profilers, expert systems, hybrid and statistical models with varying underlying algorithms (Derek Nexus, Sarah Nexus, VEGA mutagenicity models, TEST mutagenicity models, OECD QSAR Toolbox profilers).
The assessment showed that most, but not all, tools predict mutagenic potential of organosilicon chemistry as accurately as for other organic datasets, with six tools providing balanced accuracies ≥ 80%. The nature of the algorithm was identified as the key driver for accuracy. The presence of the silicon atom in a molecule sometimes resulted in substances being outside model-specific applicability domains, but this did not necessarily correlate with predictive performance. Furthermore, the assessment showed that organosilicon chemistry does not possess intrinsic gene mutation potential. All positive Ames test results in the dataset could be linked to organofunctional structural alerts known to be related to mutagenicity.