Discovering organic reactions with a machine-learning-powered deciphering of tera-scale mass spectrometry data

The accumulation of large datasets by the scientific community has surpassed the capacity of traditional processing methods, underscoring the critical need for innovative and efficient algorithms capable of navigating through extensive existing experimental data. Addressing this challenge, our study introduces a machine learning (ML)-powered search engine specifically tailored for analyzing tera-scale high-resolution mass spectrometry (HRMS) data. This engine harnesses a novel isotope-distribution-centric search algorithm augmented by two synergistic ML models, assisting with the discovery of hitherto unknown chemical reactions. This methodology enables the rigorous investigation of existing data, thus providing efficient support for chemical hypotheses while reducing the need for conducting additional experiments. Moreover, we extend this approach with baseline methods for automated reaction hypothesis generation. In its practical validation, our approach successfully identified several reactions, unveiling previously undescribed transformations. Among these, the heterocycle-vinyl coupling process within the Mizoroki-Heck reaction stands out, highlighting the capability of the engine to elucidate complex chemical phenomena. Mass spectrometry generates vast amounts of data in chemistry labs. Here, authors developed a machine learning-driven search engine that analyzes archived data to discover chemical reactions without performing additional experiments.