(Q)SAR Approaches to Predict the Extent of Nitrosation in Pharmaceutical Compounds

Since their discovery as impurities in numerous pharmaceuticals beginning in 2018, there has been a strong push to predict and prevent the formation of mutagenic nitrosamines. Several experimental methods, particularly the Nitrosation Assay Procedure, have been developed to predict a molecule's susceptibility to nitrosation. Here, we have compiled the results of hundreds of these experiments from the literature to construct two structure-activity relationship models: a statistical model and an expert rule-based model. The statistical model has been built with graph neural networks and was trained on a dataset of 207 nitrogen-containing molecules. This model makes a binary call for each nitrogen center, predicting if it is likely to be nitrosated or not. Conversely, the rule-based model labels each possible nitrosamine product as one of four categories, ranging from "unlikely" to "very likely". It makes this determination based on 15 rules, which cover 12 deactivating (inhibit nitrosation) and 3 activating (favor nitrosation) features that have been drawn from the literature. Both models perform remarkably well, with accuracies of ∼80%. The rule-based model is generally biased toward favoring nitrosation while the statistical model is more likely to classify an amine as un-nitrosatable due to the makeup of the dataset. Using the models together can balance these biases and further improve the reliability of both.