All conventional data analysis methods extract mass spectra (just few up to dozens of mass-intensity pairs) corresponding to individual compounds; the spectra are used as queries in spectral databases. Despite the databases are large (hundreds of thousands records) and steadily growing, their coverage of billions of existing natural or anthropogenic compounds is still poor. Therefore, these methods cannot detect compounds not present in the databases intrinsically.

In 2020-23, several mutually independent methods, which overcome these restrictions, were published. All are based on large-scale machine learning models, usually inspired by the natural language processing field. These models "translate" from a mass spectrum to the corresponding molecular formula directly.

The theoretical part of the thesis will summarize and compare those methods briefly. Further, a similar model usable for data coming from the prevailing experiments in the collaborating Recetox centre -- mass spectrometry by electronic ionization, coupled with gas chromatography (GC-EI-MS), where the published methods are not directly applicable. Besides the model development itself, the student will also propose methodologies for data preparation and model evaluation. The chosen evaluation techniques should be similar to those used in the recommended literature so that we are able to compare the performance of our model with the published ones.