A Multimodal Transformer Model for comprehensive Structure Elucidation

Abstract

Structure elucidation is integral in the day-to-day operation of any organic chemistry laboratory, providing crucial insights into the composition and purity of diverse substances. This can be achieved through a number of spectroscopic measurements, such as Nuclear Magnetic Resonance (NMR), Mass Spectrometry (MS) or Infrared (IR) spectroscopy. While the acquisition of the spectra has been largely automated, the analysis of them is not straightforward making the analysis of spectra, particularly in large quantities, a time-consuming and tedious undertaking. We have recently shown that Transformer models are capable of predicting the exact chemical structure from NMR and IR spectra. However, these models use only one modality, i.e. either NMR or IR spectra as input. This is fundamentally different from how a chemist would determine the structure of an unknow compound, which involves combining the information present in multiple different orthogonal spectroscopic techniques. Here we present a model that is capable of emulating this approch. We first compile a dataset of simulated 1H-NMR, 13C-NMR, HSQC, IR, and MSMS spectra for ca. 1.1 million molecules sampled from USPTO. For each molecule six different spectra are generated. Utilising this dataset we train a multimodal Transformer model that can use exploit the information present in multiple different spectra. On the data generated our model achieves an accuracy of 79.6% and outperforms all baselines based on a single type of spectroscopic information.