Toward ultimate NMR resolution with deep learning

TL;DR

This paper introduces a deep learning method to significantly improve NMR resolution by accurately identifying peak positions in complex spectra.

Contribution

The novel peak probability presentation (P3) and MR-Ai network enable near-theoretical precision in NMR peak localization.

Findings

P3 achieves peak-localization precision close to the Cramér-Rao lower bound using synthetic spectra.

MR-Ai enhances spectral quality by enabling coprocessing of multiple spectra with sparse sampling.

Validation on 60 proteins and challenging cases like Tau and MATL1 demonstrates P3's effectiveness.

Abstract

Resolution in NMR is defined as the ability to distinguish and accurately determine signal positions while mitigating overlap. In the pursuit of ultimate resolution, we introduce peak probability presentations (P3), a statistical spectral representation that assigns a probability to each spectral point, indicating the likelihood that a peak maximum occurs at that location. The mapping between the traditional spectrum and P3 is achieved using MR-Ai, a physics-inspired and computationally efficient deep-learning neural network. P3 is validated on 60 database proteins and showcased on the challenging Tau and MATL1 proteins. Using synthetic spectra, we show that the achieved peak-localization precision closely approaches the theoretical limits set by the Cramér-Rao lower bound and Bayesian Monte Carlo estimates. Furthermore, MR-Ai enables the coprocessing of multiple spectra, facilitating direct information exchange between datasets to enhance spectral quality, particularly in cases of highly sparse sampling.

Deep learning delivers super-resolution for multidimensional NMR spectra.

Figures