Towards autonomous analysis of chemical exchange saturation transfer experiments using deep neural networks

Deep learning and its applications in nuclear magnetic resonance spectroscopy

Nuclear Magnetic Resonance (NMR), as an advanced technology, has widespread applications in various fields like chemistry, biology, and medicine. However, issues such as long acquisition times for multidimensional spectra and low sensitivity limit the broader application of NMR. Traditional algorithms aim to address these issues but have limitations in speed and accuracy. Deep Learning (DL), a branch of Artificial Intelligence (AI) technology, has shown remarkable success in many fields including NMR. This paper presents an overview of the basics of DL and current applications of DL in NMR, highlights existing challenges, and suggests potential directions for improvement.

Improving quantification accuracy of a nuclear Overhauser enhancement signal at -1.6 ppm at 4.7 T using a machine learning approach

Objective.A new nuclear Overhauser enhancement (NOE)-mediated saturation transfer MRI signal at -1.6 ppm, potentially from choline phospholipids and termed NOE(-1.6), has been reported in biological tissues at high magnetic fields. This signal shows promise for detecting brain tumors and strokes. However, its proximity to the water peak and low signal-to-noise ratio makes accurate quantification challenging, especially at low fields, due to the difficulty in separating it from direct water saturation and other confounding signals. This study proposes using a machine learning (ML) method to address this challenge.Approach.The ML model was trained on a partially synthetic chemical exchange saturation transfer dataset with a curriculum learning denoising approach. The accuracy of our method in quantifying NOE(-1.6) was validated using tissue-mimicking data from Bloch simulations providing ground truth, with subsequent application to an animal tumor model at 4.7 T. The predictions from the proposed ML method were compared with outcomes from traditional Lorentzian fit and ML models trained on other data types, including measured and fully simulated data.Main results.Our tissue-mimicking validation suggests that our method offers superior accuracy compared to all other methods. The results from animal experiments show that our method, despite variations in training data size or simulation models, produces predictions within a narrower range than the ML method trained on other data types.Significance.The ML method proposed in this work significantly enhances the accuracy and robustness of quantifying NOE(-1.6), thereby expanding the potential for applications of this novel molecular imaging mechanism in low-field environments.

Protein NMR assignment by isotope pattern recognition

The current standard method for amino acid signal identification in protein NMR spectra is sequential assignment using triple-resonance experiments. Good software and elaborate heuristics exist, but the process remains laboriously manual. Machine learning does help, but its training databases need millions of samples that cover all relevant physics and every kind of instrumental artifact. In this communication, we offer a solution to this problem. We propose polyadic decompositions to store millions of simulated three-dimensional NMR spectra, on-the-fly generation of artifacts during training, a probabilistic way to incorporate prior and posterior information, and integration with the industry standard CcpNmr software framework. The resulting neural nets take [1H,13C] slices of mixed pyruvate-labeled HNCA spectra (different CA signal shapes for different residue types) and return an amino acid probability table. In combination with primary sequence information, backbones of common proteins (GB1, MBP, and INMT) are rapidly assigned from just the HNCA spectrum.

Solution-state methyl NMR spectroscopy of large non-deuterated proteins enabled by deep neural networks

Methyl-TROSY nuclear magnetic resonance (NMR) spectroscopy is a powerful technique for characterising large biomolecules in solution. However, preparing samples for these experiments is demanding 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 information for proteins that cannot be produced in bacterial systems. We validate the methodology experimentally on three proteins with molecular weights in the range 42-360 kDa. We further demonstrate the applicability of our methodology to 3D NOESY spectra of Escherichia coli Malate Synthase G (81 kDa), where observed NOE cross-peaks are in good agreement with the available structure. The method represents an advance in the field of using deep learning to analyse complex magnetic resonance data and could have an impact on the study of large biomolecules in years to come.

Biomolecular NMR spectroscopy in the era of artificial intelligence

Biomolecular nuclear magnetic resonance (NMR) spectroscopy and artificial intelligence (AI) have a burgeoning synergy. Deep learning-based structural predictors have forever changed structural biology, yet these tools currently face limitations in accurately characterizing protein dynamics, allostery, and conformational heterogeneity. We begin by highlighting the unique abilities of biomolecular NMR spectroscopy to complement AI-based structural predictions toward addressing these knowledge gaps. We then highlight the direct integration of deep learning approaches into biomolecular NMR methods. AI-based tools can dramatically improve the acquisition and analysis of NMR spectra, enhancing the accuracy and reliability of NMR measurements, thus streamlining experimental processes. Additionally, deep learning enables the development of novel types of NMR experiments that were previously unattainable, expanding the scope and potential of biomolecular NMR spectroscopy. Ultimately, a combination of AI and NMR promises to further revolutionize structural biology on several levels, advance our understanding of complex biomolecular systems, and accelerate drug discovery efforts.

Studying micro to millisecond protein dynamics using simple amide 15N CEST experiments supplemented with major-state R2 and visible peak-position constraints

Over the last decade amide 15N CEST experiments have emerged as a popular tool to study protein dynamics that involves exchange between a 'visible' major state and sparsely populated 'invisible' minor states. Although initially introduced to study exchange between states that are in slow exchange with each other (typical exchange rates of, 10 to 400 s-1), they are now used to study interconversion between states on the intermediate to fast exchange timescale while still using low to moderate (5 to 350 Hz) 'saturating' B1 fields. The 15N CEST experiment is very sensitive to exchange as the exchange delay TEX can be quite long (~0.5 s) allowing for a large number of exchange events to occur making it a very powerful tool to detect minor sates populated ([Formula: see text]) to as low as 1%. When systems are in fast exchange and the 15N CEST data has to be described using a model that contains exchange, the exchange parameters are often poorly defined because the [Formula: see text] versus [Formula: see text] and [Formula: see text] versus exchange rate ([Formula: see text]) plots can be quite flat with shallow or no minima and the analysis of such 15N CEST data can lead to wrong estimates of the exchange parameters due to the presence of 'spurious' minima. Here we show that the inclusion of experimentally derived constraints on the intrinsic transverse relaxation rates and the inclusion of visible state peak-positions during the analysis of amide 15N CEST data acquired with moderate B1 values (~50 to ~350 Hz) results in convincing minima in the [Formula: see text] versus [Formula: see text] and the [Formula: see text] versus [Formula: see text] plots even when exchange occurs on the 100 μs timescale. The utility of this strategy is demonstrated on the fast-folding Bacillus stearothermophilus peripheral subunit binding domain that folds with a rate constant ~104 s-1. Here the analysis of 15N CEST data alone results in [Formula: see text] versus [Formula: see text] and [Formula: see text] versus [Formula: see text] plots that contain shallow minima, but the inclusion of visible-state peak positions and restraints on the intrinsic transverse relaxation rates of both states during the analysis of the 15N CEST data results in pronounced minima in the [Formula: see text] versus [Formula: see text] and [Formula: see text] versus [Formula: see text] plots and precise exchange parameters even in the fast exchange regime ([Formula: see text]~5). Using this strategy we find that the folding rate constant of PSBD is invariant (~10,500 s-1) from 33.2 to 42.9 °C while the unfolding rates (~70 to ~500 s-1) and unfolded state populations (~0.7 to ~4.3%) increase with temperature. The results presented here show that protein dynamics occurring on the 10 to 104 s-1 timescale can be studied using amide 15N CEST experiments.

Towards autonomous analysis of chemical exchange saturation transfer experiments using deep neural networks

Macromolecules often exchange between functional states on timescales that can be accessed with NMR spectroscopy and many NMR tools have been developed to characterise the kinetics and thermodynamics of the exchange processes, as well as the structure of the conformers that are involved. However, analysis of the NMR data that report on exchanging macromolecules often hinges on complex least-squares fitting procedures as well as human experience and intuition, which, in some cases, limits the widespread use of the methods. The applications of deep neural networks (DNNs) and artificial intelligence have increased significantly in the sciences, and recently, specifically, within the field of biomolecular NMR, where DNNs are now available for tasks such as the reconstruction of sparsely sampled spectra, peak picking, and virtual decoupling. Here we present a DNN for the analysis of chemical exchange saturation transfer (CEST) data reporting on two- or three-site chemical exchange involving sparse state lifetimes of between approximately 3-60 ms, the range most frequently observed via experiment. The work presented here focuses on the ¹H CEST class of methods that are further complicated, in relation to applications to other nuclei, by anti-phase features. The developed DNNs accurately predict the chemical shifts of nuclei in the exchanging species directly from anti-phase ¹H^N CEST profiles, along with an uncertainty associated with the predictions. The performance of the DNN was quantitatively assessed using both synthetic and experimental anti-phase CEST profiles. The assessments show that the DNN accurately determines chemical shifts and their associated uncertainties. The DNNs developed here do not contain any parameters for the end-user to adjust and the method therefore allows for autonomous analysis of complex NMR data that report on conformational exchange.