A Machine-Learned "Chemical Intuition" to Overcome Spectroscopic Data Scarcity.

Abstract

Machine learning models for predicting IR spectra of molecular ions (infrared ion spectroscopy, IRIS) have yet to be reported owing to the relatively sparse experimental data sets available. To overcome this limitation, we employ the Graphormer-IR model for neutral molecules as a knowledgeable starting point and then employ transfer learning to refine the model to predict the spectra of gaseous ions. A library of 10,336 computed spectra and a small data set of 312 experimental IRIS spectra is used for model fine-tuning. Nonspecific global graph encodings that describe the molecular charge state (i.e., (de)protonation, sodiation), combined with an additional transfer learning step that considers computed spectra for ions, improved model performance. The resulting Graphormer-IRIS model yields spectra that are 21% more accurate than those produced by commonly employed DFT quantum chemical models, while capturing subtle phenomena such as spectral red-shifts due to sodiation. Dimensionality reduction of model embeddings demonstrates derived "chemical intuition" of functional groups, trends in molecular electron density, and the location of charge sites. Our approach will enable fast IRIS predictions for determining the structures of unknown small molecule analytes (e.g., metabolites, lipids) present in biological samples.