A method to predict molecular identity from routine H1 and C13 NMR

Stanford scientists have developed a machine learning framework that predicts the complete molecular structure of an unknown compound directly from routine 1D NMR spectra with no additional information required, automating one of the most time-consuming and expertise-dependent steps in chemical analysis.

Nuclear magnetic resonance (NMR) spectroscopy is the workhorse technique for characterizing small molecules across the pharmaceutical, chemical, and materials industries, yet interpreting the resulting spectra remains a laborious manual exercise that depends heavily on trained chemists. Structure elucidation from one-dimensional ¹H and ¹³C spectra, the most routinely acquired NMR data, is especially difficult because the number of possible structures grows combinatorially with the number of heavy atoms, quickly reaching into the range of 10³⁵ possible structures for commonly encountered molecular sizes. Researchers at Stanford addressed this challenge by building an end-to-end multitask machine learning framework that predicts both molecular formula and atomic connectivity directly from 1D NMR spectra, with no prior chemical knowledge required. The system pairs a convolutional neural network for spectral feature extraction with a transformer-based generative model that assembles molecular fragments into complete candidate structures. On molecules containing up to 40 heavy atoms (C, N, O, P, S, Si, B, F, Cl, Br, I), in which > 10³⁵ of structures are theoretically possible, the current prototype of the model identifies the exact correct structure within its top 15 predictions 60.4% of the time, reducing the chemist's search space by up over 35 orders of magnitude. Further improvements in accuracy are expected with ongoing work to access larger training datasets. By automating an analysis step that has historically gated everything from drug discovery to quality control, this technology offers chemical and pharmaceutical organizations a powerful tool to accelerate compound identification and unlock new applications in automated chemical synthesis within a global NMR market currently estimated at $760 million and growing.

Stage of Development
Prototype