NMR spectrum reconstruction as a pattern recognition problem.

Area of Science:

• Biophysics
• Computational Chemistry
• Spectroscopy

Background:

• Non-uniform sampling (NUS) in Nuclear Magnetic Resonance (NMR) spectroscopy accelerates data acquisition but poses reconstruction challenges.
• Traditional methods often struggle to fully exploit the information embedded in the sampling schedule and spectral characteristics.

Purpose of the Study:

• To introduce a deep neural network (DNN) based on the WaveNet architecture (WNN) for enhanced NMR spectra reconstruction.
• To demonstrate the WNN's capability in recognizing patterns within NUS NMR spectra for improved data processing.

Main Methods:

• Development of a WaveNet-based neural network (WNN) for NMR spectral pattern recognition.
• Training the WNN on fixed non-uniform sampling (NUS) schedules to learn point spread function (PSF) patterns.
• Application of WNN to reconstruct 2D ¹H-¹⁵N correlation spectra of Ubiquitin, Azurin, and Malt1 proteins.

Main Results:

• The WNN achieved high-quality and robust reconstruction of NUS spectra for globular proteins of varying sizes.
• Demonstrated successful virtual homo-decoupling in a 2D methyl ¹H-¹³C - HMQC spectrum of MALT1 using WNN's pattern recognition.
• Showcased that WNN effectively utilizes prior knowledge of the NUS schedule for advanced NMR processing.

Conclusions:

• The WNN offers a powerful new technique for processing NMR data acquired with non-uniform sampling.
• Exploiting NUS schedule information via WNN surpasses existing algorithmic methods in spectral reconstruction quality.
• This deep learning approach opens avenues for developing novel and more powerful NMR data processing techniques.