Uncertainty Quantification and Flagging of Unreliable Predictions in Predicting Mass Spectrometry-Related Properties of Small Molecules Using Machine Learning.

Mass spectral identification (in particular, in metabolomics) can be refined by comparing the observed and predicted properties of molecules, such as chromatographic retention. Significant advancements have been made in predicting these values using machine learning and deep learning. Usually, model predictions do not contain any indication of the possible error (uncertainty) or only one criterion is used for this purpose. The spread of predictions of several models included in the ensemble, and the molecular similarity of the considered molecule and the most "similar" molecule from the training set, are values that allow us to estimate the uncertainty. The Euclidean distance between vectors, calculated based on real-valued molecular descriptors, can be used for the assessment of molecular similarity. Another factor indicating uncertainty is the molecule's belonging to one of the clusters (data set clustering). Together, all three factors can be used as features for the uncertainty assessment model. Classification models that predict whether a prediction belongs to the worst 15% were obtained. The area under the receiver operating curve value is in the range of 0.73-0.82 for the considered tasks: the prediction of retention indices in gas chromatography, retention times in liquid chromatography, and collision cross-sections in ion mobility spectroscopy.