Accurate and efficient structure elucidation from routine one-dimensional NMR spectra using multitask machine learning

TL;DR

The paper tackles structure elucidation from 1D NMR data, a combinatorially hard inverse problem. It develops a multitask transformer framework that first pretrains on substructure-to-structure translation and then applies end-to-end spectrum-to-structure learning using minimal spectral preprocessing. Key results show 93.2% exact recovery within 15 predictions for up to 19 heavy atoms in the substructure-to-structure task, and 69.6% structure accuracy with both 1H and 13C NMR inputs in the multitask model, with 1H NMR being more informative. The approach dramatically narrows the search space (up to 11 orders of magnitude), runs in under 3 seconds on CPU, and is extensible to larger molecules and stereochemistry, representing a scalable, unsupervised tool for rapid structure elucidation.

Abstract

Rapid determination of molecular structures can greatly accelerate workflows across many chemical disciplines. However, elucidating structure using only one-dimensional (1D) NMR spectra, the most readily accessible data, remains an extremely challenging problem because of the combinatorial explosion of the number of possible molecules as the number of constituent atoms is increased. Here, we introduce a multitask machine learning framework that predicts the molecular structure (formula and connectivity) of an unknown compound solely based on its 1D 1H and/or 13C NMR spectra. First, we show how a transformer architecture can be constructed to efficiently solve the task, traditionally performed by chemists, of assembling large numbers of molecular fragments into molecular structures. Integrating this capability with a convolutional neural network (CNN), we build an end-to-end model for predicting structure from spectra that is fast and accurate. We demonstrate the effectiveness of this framework on molecules with up to 19 heavy (non-hydrogen) atoms, a size for which there are trillions of possible structures. Without relying on any prior chemical knowledge such as the molecular formula, we show that our approach predicts the exact molecule 69.6% of the time within the first 15 predictions, reducing the search space by up to 11 orders of magnitude.

Figures (5)

• Figure 1: Overview of the full multitask structure elucidation workflow (top) and the substructure-to-structure workflow (bottom). Weights from a transformer pretrained on the substructure-to-structure task are used to initialize the multitask model. Specific details regarding the transformer model architecture and multitask model architecture can be found in SI Section 1.
• Figure 2: (Left) Transformer and the best multitask model test accuracy as a function of the problem size. The problem size is determined by extrapolating an exponential fit to the number of molecules in GDB-9ramakrishnan_quantum_2014, GDB-11fink_virtual_2007, GDB-13blum_970_2009, and GDB-17, and the plot begins with the number of possible structures for 10 heavy (non-hydrogen) atoms. (Right) Distribution of Tanimoto similarities of the best incorrect predictions relative to the target molecule after removing correct predictions and invalid SMILES.
• Figure 3: (Top) Examples of molecules that were correctly predicted by the transformer model. The molecules shown have between 23 to 65 substructures, and an example of a correctly predicted molecule and its constituent substructures is shown in SI Figure 4. (Bottom) Examples of molecules with incorrect predictions from the transformer model. The number beneath each predicted molecule is the Tanimoto similarity between the prediction and target.
• Figure 4: (Top) Examples of molecules that were correctly predicted using the multitask spectrum-to-structure model. (Bottom) Examples of molecules with incorrect predictions from the multitask model. The number beneath each predicted molecule is the Tanimoto similarity between the prediction and target.
• Figure 5: Distribution of true/false positives and true/false negatives as a function of the probability predicted by the multitask model using both 1H and 13C NMR as inputs and a pretrained transformer. A decision boundary of 0.5 is used to distinguish between positives and negatives.