Accurate and\nEfficient Structure Elucidation\nfrom Routine One-Dimensional NMR Spectra\nUsing Multitask Machine Learning

Rapid determination of molecular structures can greatly\naccelerate\nworkflows across many chemical disciplines. However, elucidating structure\nusing only one-dimensional (1D) NMR spectra, the most readily accessible\ndata, remains an extremely challenging problem because of the combinatorial\nexplosion of the number of possible molecules as the number of constituent\natoms is increased. Here, we introduce a multitask machine learning\nframework that predicts the molecular structure (formula and connectivity)\nof an unknown compound solely based on its 1D <sup>1</sup>H and/or <sup>13</sup>C NMR spectra. First, we show how a transformer architecture\ncan be constructed to efficiently solve the task, traditionally performed\nby chemists, of assembling large numbers of molecular fragments into\nmolecular structures. Integrating this capability with a convolutional\nneural network, we build an end-to-end model for predicting structure\nfrom spectra that is fast and accurate. We demonstrate the effectiveness\nof this framework on molecules with up to 19 heavy (non-hydrogen)\natoms, a size for which there are trillions of possible structures.\nWithout relying on any prior chemical knowledge such as the molecular\nformula, we show that our approach predicts the exact molecule 69.6%\nof the time within the first 15 predictions, reducing the search space\nby up to 11 orders of magnitude.