Mass2SMILES: deep learning based fast prediction of structures and functional groups directly from high-resolution MS/MS spectra.

Modern mass spectrometry-based metabolomics generates vast amounts of mass spectral data as part of the chemical inventory of biospecimens. Annotation of the resulting MS/MS spectra remains a challenging task that mostly relies on database interrogations, in silico prediction and interpretation of diagnostic fragmentation schemes and/or expert knowledge-based manual interpretations. A key limitation is additionally that these approaches typically leave a vast proportion of the (bio)chemical space unannotated. Here we report a deep neural network method to predict chemical structures solely from high-resolution MS/MS spectra. This novel approach initially relies on the encoding of SMILES strings from chemical structures using a continuous chemical descriptor space that had been previously implemented for molecule design. The deep neural network was trained on 83,358 natural product-derived MS/MS spectra of the GNPS library and of the NIST HRMS database with addition of the calculated neutral losses for those spectra. After this training and parameter optimization phase, the deep neural network approach was then used to predict structures from MS/MS spectra not included in the training data-set. Our current version, implemented in the Python programming language, accurately predicted 7 structures from 744 validation structures and the following 14 structures had a Tanimoto similarity score above 0.9 when compared to the true structure. It was also able to correctly identify two structures from the CASMI 2022 international contest. On average the Tanimoto similarity is of 0.40 for data of the CASMI 2022 international contest and of 0.39 for the validation data-set. Finally, our deep neural network is also able to predict the number of 60 functional groups as well as the molecular formula of chemical structures and adduct type for the analyzed MS/MS spectra. Importantly, this deep neural network approach is extremely fast, in comparison to currently available methods, making it suitable to predict on regular computers structures for all substances within large metabolomics datasets.