Formally, the mass spectrum of a single compound can be represented as a smaller (cardinality ranging from tens to low hundreds) set of pairs (mass, intensity). Due to various factors (such as imperfect deconvolution, noise, etc.), it may happen that some of those pairs are missing. As a result, the precision of database retrieval declines.

The goal of this work is to apply machine learning methods (with a particular focus on deep learning methods coming from the natural language processing area) in such a way that they can suggest what are the missing parts. With this kind of advice, it will be possible to confirm or disprove the presence of the signal in original data and to improve the precision of database retrieval significantly.

Twofold focus of the work is expected. Firstly, the prediction of peaks with a lower intensity (that are often obscured by noise in measured data) will be addressed. Secondly, the correction of (deconvolution, ...) errors caused by an overlap of different compounds and expressed as randomly missing peaks even in higher intensity sections in measured data should be targeted.

The output of the work should be methods implemented in a suitable form (e.g Jupyter notebooks, R-Studio), evaluation of those on the dataset available and discussion of the results achieved.