Computer vision-based automated peak picking applied to protein NMR spectra

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.

Machine Learning in Automated Monitoring of Metabolic Changes Accompanying the Differentiation of Adipose-Tissue-Derived Human Mesenchymal Stem Cells Employing 1H-1H TOCSY NMR

The ability to monitor the dynamics of stem cell differentiation is a major goal for understanding biochemical evolution pathways. Automating the process of metabolic profiling using 2D NMR helps us to understand the various differentiation behaviors of stem cells, and therefore sheds light on the cellular pathways of development, and enhances our understanding of best practices for in vitro differentiation to guide cellular therapies. In this work, the dynamic evolution of adipose-tissue-derived human Mesenchymal stem cells (AT-derived hMSCs) after fourteen days of cultivation, adipocyte and osteocyte differentiation, was inspected based on 1H-1H TOCSY using machine learning. Multi-class classification in addition to the novelty detection of metabolites was established based on a control hMSC sample after four days' cultivation and we successively detected the changes of metabolites in differentiated MSCs following a set of 1H-1H TOCSY experiments. The classifiers Kernel Null Foley-Sammon Transform and Kernel Density Estimation achieved a total classification error between 0% and 3.6% and false positive and false negative rates of 0%. This approach was successfully able to automatically reveal metabolic changes that accompanied MSC cellular evolution starting from their undifferentiated status to their prolonged cultivation and differentiation into adipocytes and osteocytes using machine learning supporting the research in the field of metabolic pathways of stem cell differentiation.

RTExtract: time-series NMR spectra quantification based on 3D surface ridge tracking

MOTIVATION

Time-series nuclear magnetic resonance (NMR) has advanced our knowledge about metabolic dynamics. Before analyzing compounds through modeling or statistical methods, chemical features need to be tracked and quantified. However, because of peak overlap and peak shifting, the available protocols are time consuming at best or even impossible for some regions in NMR spectra.

RESULTS

We introduce Ridge Tracking-based Extract (RTExtract), a computer vision-based algorithm, to quantify time-series NMR spectra. The NMR spectra of multiple time points were formulated as a 3D surface. Candidate points were first filtered using local curvature and optima, then connected into ridges by a greedy algorithm. Interactive steps were implemented to refine results. Among 173 simulated ridges, 115 can be tracked (RMSD < 0.001). For reproducing previous results, RTExtract took less than 2 h instead of ∼48 h, and two instead of seven parameters need tuning. Multiple regions with overlapping and changing chemical shifts are accurately tracked.

AVAILABILITY AND IMPLEMENTATION

Source code is freely available within Metabolomics toolbox GitHub repository (https://github.com/artedison/Edison_Lab_Shared_Metabolomics_UGA/tree/master/metabolomics_toolbox/code/ridge_tracking) and is implemented in MATLAB and R.

SUPPLEMENTARY INFORMATION

Supplementary data are available at Bioinformatics online.

NV center based nano-NMR enhanced by deep learning

The growing field of nano nuclear magnetic resonance (nano-NMR) seeks to estimate spectra or discriminate between spectra of minuscule amounts of complex molecules. While this field holds great promise, nano-NMR experiments suffer from detrimental inherent noise. This strong noise masks to the weak signal and results in a very low signal-to-noise ratio. Moreover, the noise model is usually complex and unknown, which renders the data processing of the measurement results very complicated. Hence, spectra discrimination is hard to achieve and in particular, it is difficult to reach the optimal discrimination. In this work we present strong indications that this difficulty can be overcome by deep learning (DL) algorithms. The DL algorithms can mitigate the adversarial effects of the noise efficiently by effectively learning the noise model. We show that in the case of frequency discrimination DL algorithms reach the optimal discrimination without having any pre-knowledge of the physical model. Moreover, the DL discrimination scheme outperform Bayesian methods when verified on noisy experimental data obtained by a single Nitrogen-Vacancy (NV) center. In the case of frequency resolution we show that this approach outperforms Bayesian methods even when the latter have full pre-knowledge of the noise model and the former has none. These DL algorithms also emerge as much more efficient in terms of computational resources and run times. Since in many real-world scenarios the noise is complex and difficult to model, we argue that DL is likely to become a dominant tool in the field.

Convolutional Neural Network Analysis of Two-Dimensional Hyperfine Sublevel Correlation Electron Paramagnetic Resonance Spectra

A machine learning approach is presented for analyzing complex two-dimensional hyperfine sublevel correlation electron paramagnetic resonance (HYSCORE EPR) spectra with the proficiency of an expert spectroscopist. The computer vision algorithm requires no training on experimental data; rather, all of the spin physics required to interpret the spectra are learned from simulations alone. This approach is therefore applicable even when insufficient experimental data exist to train the algorithm. The neural network is demonstrated to be capable of utilizing the full information content of two-dimensional ¹⁴N HYSCORE spectra to predict the magnetic coupling parameters and their underlying probability distributions that were previously inaccessible. The predicted hyperfine ( a, T) and ¹⁴N quadrupole ( K, η) coupling constants deviate from the previous manual analyses of the experimental spectra on average by 0.11 MHz, 0.09 MHz, 0.19 MHz, and 0.09, respectively.

Combining NMR Spectroscopy and Molecular Dynamic Simulations to Solve and Analyze the Structure of Protein-RNA Complexes

Understanding the RNA binding specificity of protein is of primary interest to decipher their function in the cell. Here, we review the methodology used to solve the structures of protein-RNA complexes using solution-state NMR spectroscopy: from sample preparation to structure calculation procedures. We also describe how molecular dynamics simulations can help providing additional information on the role of key amino acid side chains and of water molecules in protein-RNA recognition.

Robust and transferable quantification of NMR spectral quality using IROC analysis

Non-Fourier methods are increasingly utilized in NMR spectroscopy because of their ability to handle nonuniformly-sampled data. However, non-Fourier methods present unique challenges due to their nonlinearity, which can produce nonrandom noise and render conventional metrics for spectral quality such as signal-to-noise ratio unreliable. The lack of robust and transferable metrics (i.e. applicable to methods exhibiting different nonlinearities) has hampered comparison of non-Fourier methods and nonuniform sampling schemes, preventing the identification of best practices. We describe a novel method, in situ receiver operating characteristic analysis (IROC), for characterizing spectral quality based on the Receiver Operating Characteristic curve. IROC utilizes synthetic signals added to empirical data as "ground truth", and provides several robust scalar-valued metrics for spectral quality. This approach avoids problems posed by nonlinear spectral estimates, and provides a versatile quantitative means of characterizing many aspects of spectral quality. We demonstrate applications to parameter optimization in Fourier and non-Fourier spectral estimation, critical comparison of different methods for spectrum analysis, and optimization of nonuniform sampling schemes. The approach will accelerate the discovery of optimal approaches to nonuniform sampling experiment design and non-Fourier spectrum analysis for multidimensional NMR.

Facilitating the performance of qNMR analysis using automated quantification and results verification

Quantitative nuclear magnetic resonance (qNMR) is considered as a powerful tool for measuring the absolute amount of small molecules in complex mixtures. However, verification of the accuracy of such quantification is not a trivial task. In particular, preprocessing and integration steps are challenging and potentially erroneous. A script was developed in Matlab environment to automate qNMR analysis. Verification of the results is based on two evolving integration profiles. The analysis of binary mixtures of internal standards as well as pharmaceutical products has shown that all common artifacts (phase and baseline distortion, impurities) can be easily recognized in routine qNMR experiments. In the absence of distortion, deviation between automatically (mean value of several integrals) and manually calculated values was generally below 0.1%. The routine is independent of multiplet pattern, solvent, spectrometer, nuclei type and pulse sequence used. In general, the usage of the developed script can facilitate and verify results of routine qNMR analysis in an automatic manner. Copyright © 2017 John Wiley & Sons, Ltd.

NMR-based automated protein structure determination

NMR spectra analysis for protein structure determination can now in many cases be performed by automated computational methods. This overview of the computational methods for NMR protein structure analysis presents recent automated methods for signal identification in multidimensional NMR spectra, sequence-specific resonance assignment, collection of conformational restraints, and structure calculation, as implemented in the CYANA software package. These algorithms are sufficiently reliable and integrated into one software package to enable the fully automated structure determination of proteins starting from NMR spectra without manual interventions or corrections at intermediate steps, with an accuracy of 1-2 Å backbone RMSD in comparison with manually solved reference structures.

Automation of peak-tracking analysis of stepwise perturbed NMR spectra

We describe a new algorithmic approach able to automatically pick and track the NMR resonances of a large number of 2D NMR spectra acquired during a stepwise variation of a physical parameter. The method has been named Trace in Track (TINT), referring to the idea that a gaussian decomposition traces peaks within the tracks recognised through 3D mathematical morphology. It is capable of determining the evolution of the chemical shifts, intensity and linewidths of each tracked peak.The performances obtained in term of track reconstruction and correct assignment on realistic synthetic spectra were high above 90% when a noise level similar to that of experimental data were considered. TINT was applied successfully to several protein systems during a temperature ramp in isotope exchange experiments. A comparison with a state-of-the-art algorithm showed promising results for great numbers of spectra and low signal to noise ratios, when the graduality of the perturbation is appropriate. TINT can be applied to different kinds of high throughput chemical shift mapping experiments, with quasi-continuous variations, in which a quantitative automated recognition is crucial.

Peak picking multidimensional NMR spectra with the contour geometry based algorithm CYPICK

The automated identification of signals in multidimensional NMR spectra is a challenging task, complicated by signal overlap, noise, and spectral artifacts, for which no universally accepted method is available. Here, we present a new peak picking algorithm, CYPICK, that follows, as far as possible, the manual approach taken by a spectroscopist who analyzes peak patterns in contour plots of the spectrum, but is fully automated. Human visual inspection is replaced by the evaluation of geometric criteria applied to contour lines, such as local extremality, approximate circularity (after appropriate scaling of the spectrum axes), and convexity. The performance of CYPICK was evaluated for a variety of spectra from different proteins by systematic comparison with peak lists obtained by other, manual or automated, peak picking methods, as well as by analyzing the results of automated chemical shift assignment and structure calculation based on input peak lists from CYPICK. The results show that CYPICK yielded peak lists that compare in most cases favorably to those obtained by other automated peak pickers with respect to the criteria of finding a maximal number of real signals, a minimal number of artifact peaks, and maximal correctness of the chemical shift assignments and the three-dimensional structure obtained by fully automated assignment and structure calculation.

NMR structure calculation for all small molecule ligands and non-standard residues from the PDB Chemical Component Dictionary

An algorithm, CYLIB, is presented for converting molecular topology descriptions from the PDB Chemical Component Dictionary into CYANA residue library entries. The CYANA structure calculation algorithm uses torsion angle molecular dynamics for the efficient computation of three-dimensional structures from NMR-derived restraints. For this, the molecules have to be represented in torsion angle space with rotations around covalent single bonds as the only degrees of freedom. The molecule must be given a tree structure of torsion angles connecting rigid units composed of one or several atoms with fixed relative positions. Setting up CYANA residue library entries therefore involves, besides straightforward format conversion, the non-trivial step of defining a suitable tree structure of torsion angles, and to re-order the atoms in a way that is compatible with this tree structure. This can be done manually for small numbers of ligands but the process is time-consuming and error-prone. An automated method is necessary in order to handle the large number of different potential ligand molecules to be studied in drug design projects. Here, we present an algorithm for this purpose, and show that CYANA structure calculations can be performed with almost all small molecule ligands and non-standard amino acid residues in the PDB Chemical Component Dictionary.