MS2Mol: A transformer model for illuminating dark chemical space from mass spectra

The ability to identify small molecules in complex samples from their mass spectra is among the grand challenges of analytical chemistry. Improvements to this ability could significantly advance fields as diverse as drug discovery, diagnostics, environmental science, and synthetic biology. A primary bottleneck is that standard structure elucidation technologies are limited to identifying only those molecules that are contained in databases of known spectra or molecular structures and are therefore not well suited to identifying the vast majority of potentially billions of natural metabolites, whose structures are not yet catalogued. To improve the identification of molecular structures within this vast dark chemical space, we present MS2Mol, a de novo structure prediction model based on a generative sequence to sequence transformer. We also release EnvedaDark, a first-of-its-kind data set for benchmarking identification performance on unknown metabolites. EnvedaDark contains experimental mass spectra from 226 natural products not currently found in major databases. We demonstrate on this challenging dataset that MS2Mol is able to predict 21% of molecular structures to within a close-match accuracy threshold and 62% to within meaningful similarity, both of which are significant improvements over the closest match retrieved using standard database methods. We further present a confidence scorer that enables practical usage for novel molecule discovery and enriches the accuracy on meaningfully-similar and close-match thresholds to 98% and 63%, respectively, for the top 10% most confident predictions.