Exploring the dark metabolome – how MetGenX and DeepMet complement each other: A structure classresolved comparison of in-silico MS/MS annotation tools

Untargeted metabolomics routinely detects thousands of features. However, currently only 15-20 % can be assigned with high confidence. To detect new metabolites and to fill up empty feature lists, more and more in-silico tools are introduced and compete with classical workflows for structure elucidation being based on de-novo synthesis and re-measurement. While in-silico tools are promising, the choice of the tools matters. This work shows the benchmark of two novel and conceptually different approaches (DeepMet, MetGenX) and compares their performance against longer existing tools (SIRIUS and BioTransformer 3.0). As testing data, we used a large in-house library of known metabolites that were measured on a timsTOF LC-MS platform. Acquired spectra were cross-checked with DeepMet [1], a chemical language model trained on structures of known mammalian metabolites, MetGenX [2] a structure-informed encoder-decoder that generates candidates conditioned on spectral templates and SIRIUS [4] with MSNovelist [5] for de novo structure generation. Biotransformer 3.0 was added as a rule-based reference that predicts metabolites of known anchor products and possible adducts that occur in biological systems.