Solution State Methyl NMR Spectroscopy of Large Non-Deuterated Proteins Enabled by Deep Neural Networks

Abstract

Methyl-TROSY nuclear magnetic resonance (NMR) spectroscopy is a powerful technique for characterising large biomolecules in solution. However, preparing samples for these experiments is arduous and entails deuteration, limiting its use. Here we demonstrate that NMR spectra recorded on protonated, uniformly 13C labelled, samples can be processed using deep neural networks to yield spectra that are of similar quality to typical deuterated methyl-TROSY spectra, potentially providing more information at a fraction of the cost. We validated the new methodology experimentally on three proteins with molecular weights in the range 42-360 kDa and further by analysing deep learning-processed NOESY spectra of MSG (81 kDa), where observed NOE cross-peaks were in good agreement with the available structure. The new method represents a substantial advance in the field of using deep learning to analyse complex magnetic resonance data and could have a major impact on the study of large biomolecules in the years to come.

Characterising Aromatic Side Chains in Proteins through the Synergistic Development of NMR Experiments and Deep Neural Networks

“…We have previously successfully trained the FID-Net ( 31 ) architecture to virtually decouple and enhance the resolution of 13 C- 1 H correlation spectra reporting on the methyl-groups of large proteins ( 35 ). An initial attempt to use the same strategy for the aromatic region of 13 C- 1 H correlation spectra of medium-to-large proteins was not satisfactory in our hands.…”

“…Previous DNNs devised to transform NMR spectra were not quantitative with respect to the intensities of cross-peaks ( 35 ) and were not useful to study chemical exchange, characterise binding or other studies where accurate peak intensities are necessary. FID-Net-2 was however trained to be quantitative in this regard.…”

Resolution Enhancement of Metabolomic J-Res NMR Spectra Using Deep Learning

“…Each HR J-Res spectrum had dimensions of 256 × 16,384 points corresponding to the F1 (J-coupling) axis from −39.1542 to 39.1542 Hz and the F2 (chemical shift) axis from −3.560 to 13.129 ppm, respectively. For each HR spectrum, a corresponding LR spectrum was generated by applying a Gaussian filter with kernel dimensions (5,7) to blur the HR spectrum, followed by down-sampling with areal interpolation, as illustrated in Figure 1a. As a result, the dimensions of each LR J-Res spectrum were reduced to 128 (F1) × 8192 (F2) pixels, effectively halving the resolution of the HR spectrum along both axes.…”

“…Deep learning and artificial intelligence are becoming of increasing importance in many areas of science including NMR spectroscopy. 7 In particular, deep learning offers a promising approach to addressing this deconvolution challenge. Li et al introduced a deep neural network (DNN)-based approach called DEEP Picker for peak picking and deconvolution in both 1D and 2D NMR spectra; 8 however, as it was designed for macromolecular spectra, it did not include 2D J-Res NMR spectra.…”

“…Deep learning and artificial intelligence are becoming of increasing importance in many areas of science including NMR spectroscopy . In particular, deep learning offers a promising approach to addressing this deconvolution challenge.…”