Deep Reinforcement Learning for Molecular Inverse Problem of Nuclear Magnetic Resonance Spectra to Molecular Structure

Abstract

Spectroscopy is the study of how matter interacts with electromagnetic radiation. The spectra of any molecule are highly information-rich, yet the inverse relation of spectra to the corresponding molecular structure is still an unsolved problem. Nuclear magnetic resonance (NMR) spectroscopy is one such critical technique in the scientists’ toolkit to characterize molecules. In this work, a novel machine learning framework is proposed that attempts to solve this inverse problem by navigating the chemical space to find the correct structure given an NMR spectra. The proposed framework uses a combination of online Monte Carlo tree search (MCTS) and a set of graph convolution networks to build a molecule iteratively. Our method can predict the structure of the molecule ∼80% of the time in its top 3 guesses for molecules with <10 heavy atoms. We believe that the proposed framework is a significant step in solving the inverse design problem of NMR spectra.

Impact of noise on inverse design: the case of NMR spectra matching

“…56 As shift predictions need to become more accurate, limiting N through prior knowledge of the chemical space could be beneficial. Similar findings have been reported by Sridharan et al , 41 noting that brute force enumerations of chemical space lead to worse rankings than constrained graph generation. Note that while the trends in 13 C and 1 H elucidation are similar, less error is permissible when using 1 H shifts.…”

“…Recent machine learning approaches tackle the inverse problem using a combination of graph generation and subsequent chemical shi predictions for candidate ranking. [39][40][41] First explored by Jonas, 39 a Top-1 ranking with 57% reconstruction success-rate was achieved using deep imitation learning to predict bonds of molecular graphs. Sridharan et al 41 used online Monte Carlo tree search to build molecular graphs resulting in a similar Top-1 ranking of 57.2%.…”

“…[39][40][41] First explored by Jonas, 39 a Top-1 ranking with 57% reconstruction success-rate was achieved using deep imitation learning to predict bonds of molecular graphs. Sridharan et al 41 used online Monte Carlo tree search to build molecular graphs resulting in a similar Top-1 ranking of 57.2%. Huang et al 40 relied on substructure predictions from which complete graphs can be constructed, reaching 67.4% top-1 accuracy by ranking substructure proles instead of shis.…”

Conditional Molecular Generation Net Enables Automated Structure Elucidation Based on 13C NMR Spectra and Prior Knowledge

“…Sridharan et al used a Monte Carlo tree search to find the correct target molecule with 10 heavy atoms. 8 Matthew W. Kanan and his colleagues introduced a machine learning framework based on a convolutional neural network (CNN) to predict the structures of small (<11 nonhydrogen atoms) organic molecules using 1H and/or 13 C NMR spectra. However, their models were only able to elucidate structures with fewer than 11 heavy atoms and neglected the prior knowledge, which greatly limited their real-world applicability.…”

“…Machine learning methods have also shown great promise to solve the inverse problems of structure elucidation. Sridharan et al used a Monte Carlo tree search to find the correct target molecule with 10 heavy atoms . Matthew W. Kanan and his colleagues introduced a machine learning framework based on a convolutional neural network (CNN) to predict the structures of small (<11 nonhydrogen atoms) organic molecules using 1H and/or 13 C NMR spectra.…”