A combined NMR and Deep Neural Network approach for enhancing the spectral resolution of aromatic side chains in proteins

<p><span><span>Nuclear magnetic resonance (NMR) spectroscopy has become an important technique in structural biology for characterising the </span><span>structure, </span><span>dynamics</span> <span>and interactions </span><span>of </span><span>macromolecules. Wh</span><span>ile</span><span> a plethora of </span><span>NMR </span><span>methods are </span><span>now </span><span>available to inform on backbone and methyl-bearing </span><span>side-chains</span><span> of proteins, </span><span>a </span><span>characterisation</span><span> of</span><span> aromatic side chains is more challenging and often require</span><span>s</span><span> specific labelling or </span></span><span><span>13</span></span><span><span>C-detection. Here we </span><span>present</span><span> a deep neural network (DNN)</span> <span>named FID-Net-2</span><span>, which transforms NMR spectra recorded on simple uniformly </span></span><span><span>13</span></span><span><span>C labelled samples to yield high-quality </span></span><span><span>1</span></span><span><span>H</span><span>-</span></span><span><span>13</span></span><span><span>C correlation spectra of the aromatic side chains. </span><span>Key to the success of the DNN is the design of a </span><span>complementary</span><span> set of</span> <span>NMR experiment</span><span>s</span><span> that </span><span>produce</span> <span>spectra with </span><span>unique </span><span>features</span><span> to aid the</span> <span>DNN </span><span>prod</span><span>uce</span><span> high-resolution aromatic </span></span><span><span>1</span></span><span><span>H-</span></span><span><span>13</span></span><span><span>C correlation spectra with </span><span>accurate</span><span> intensities. </span><span>The </span><span>reconstructed spectra can be used for quantitative </span><span>purposes as FID-Net-2</span><span> predicts uncertainties </span><span>in </span><span>the </span><span>resulting</span> <span>spectra</span><span>. </span><span>We </span><span>have </span><span>validated</span> <span>the new </span><span>methodology</span><span> experimentally</span> <span>on protein</span><span> samples</span><span> ranging from 7 to 40 </span><span>kDa</span><span> in size. We </span><span>demonstrate</span><span> that the method can</span> <span>accurately reconstruct</span> <span>high resolution </span><span>two-dimensional </span><span>aromatic </span></span><span><span>1</span></span><span><span>H-</span></span><span><span>13</span></span><span><span>C correlation </span><span>maps</span><span>,</span> <span>high resolution </span><span>three-dimensional aromatic</span><span>-</span><span>methyl NOESY spectra to </span><span>facilitate</span><span> aromatic </span></span><span><span>1</span></span><span><span>H-</span></span><span><span>13</span></span><span><span>C assignments</span><span>,</span><span> and </span><span>that the intensities of peaks from the reconstructed</span> <span>aromatic </span></span><span><span>1</span></span><span><span>H-</span></span><span><span>13</span></span><span><span>C correlation maps </span><span>can be used to </span><span>quantitatively </span><span>characterise</span><span> the kinetics of protein folding</span><span>.&nbsp; </span><span>More generally, w</span><span>e believe that this strategy of</span><span> devising new</span><span> NMR</span><span> experiments</span><span> specifically</span> <span>for analysis</span> <span>using customised </span><span>DNN</span><span>s </span><span>represents</span><span> a </span><span>substantial</span><span> advance </span><span>that</span> <span>will</span><span> have a major impact on the study of molecules </span><span>using NMR </span><span>in the years to come.</span></span><span>&nbsp;</span></p>