Source-linked AI summary
Accurate, fully-automated NMR spectral profiling for metabolomics
Siamak Ravanbakhsh, Philip Liu, Trent Bjorndahl, Rupasri Mandal, Jason R. Grant, Michael Wilson, Roman Eisner, Igor Sinelnikov, Xiaoyu Hu, Claudio Luchinat, Russell Greiner, David S. Wishart
TL;DR
Manual NMR spectral profiling is slow, error-prone, and limits metabolomics throughput. The paper presents BAYESIL, which automates spectral processing and probabilistic deconvolution using reference libraries; across diverse mixtures and biological samples, it achieves high identification and quantification performance quickly, while depending on appropriate biofluid information and NMR detection limits.
Problem
Manual NMR spectral profiling requires trained experts and can be slow, error-prone, inconsistent, and difficult to scale.
Method
BAYESIL automates spectral processing and uses factor-graph-based probabilistic deconvolution with biofluid-specific reference libraries.
Results
98% identification and 90% quantification accuracy were achieved in 5 minutes for a serum spectrum, compared with BATMAN’s 85% identification and 8% quantification accuracy in 19 hours.
Takeaways & Limitations
BAYESIL provides a fully automated, quantitative, publicly available system intended to remove the manual profiling barrier in NMR-based metabolomics.
Takeaways & Limitations
BAYESIL depends on correctly identifying the biofluid and is constrained by NMR detection limits and sample-size requirements.
Abstract
from arXiv · showhide
Many diseases cause significant changes to the concentrations of small molecules (aka metabolites) that appear in a person's biofluids, which means such diseases can often be readily detected from a person's "metabolic profile". This information can be extracted from a biofluid's NMR spectrum. Today, this is often done manually by trained human experts, which means this process is relatively slow, expensive and error-prone. This paper presents a tool, Bayesil, that can quickly, accurately and autonomously produce a complex biofluid's (e.g., serum or CSF) metabolic profile from a 1D1H NMR spectrum. This requires first performing several spectral processing steps then matching the resulting spectrum against a reference compound library, which contains the "signatures" of each relevant metabolite. Many of these steps are novel algorithms and our matching step views spectral matching as an inference problem within a probabilistic graphical model that rapidly approximates the most probable metabolic profile. Our extensive studies on a diverse set of complex mixtures, show that Bayesil can autonomously find the concentration of all NMR-detectable metabolites accurately (~90% correct identification and ~10% quantification error), in <5minutes on a single CPU. These results demonstrate that Bayesil is the first fully-automatic publicly-accessible system that provides quantitative NMR spectral profiling effectively -- with an accuracy that meets or exceeds the performance of trained experts. We anticipate this tool will usher in high-throughput metabolomics and enable a wealth of new applications of NMR in clinical settings. Available at http://www.bayesil.ca.
I. METABOLIC PROFILING PIPELINE
BAYESIL automates both processing of raw 1D 1H NMR spectra and spectral deconvolution to produce compound identities and concentrations.
- BAYESIL processes raw 1D NMR spectra from Agilent/Varian or Bruker instruments at several frequencies.It applies automated processing before deconvolution.
- BAYESIL uses intelligent phasing and baseline correction methods to process free induction decays automatically.
- Approximate inference rapidly performs spectral deconvolution, yielding both compound identities and their concentrations.
A. Spectral Processing in BAYESIL
BAYESIL automates spectral processing from the FID to create reproducible, uniform input spectra, while addressing challenging phasing and baseline-correction steps.
- Spectral processing: BAYESIL performs zero-filling, Fourier and Hilbert transformation, phasing, baseline correction, chemical shift referencing, reference deconvolution, and smoothing automatically.Automating these functions standardizes data before spectral deconvolution.
- Phasing: Phasing maximizes peak symmetry by reducing zero-order and first-order phase mismatch.BAYESIL combines spectral norm minimization with cross entropy optimization for isolated peaks.
- Baseline correction: Baseline correction removes hardware- or solvent-related distortions while preserving desirable signal.BAYESIL detects baseline points using iterative thresholding and signal-to-noise estimation, then models the baseline with interpolation and smoothing.
- Optional processing: Smoothing and linebroadening are optional because smoothing is mostly cosmetic and may remove low-amplitude, narrow peaks.Reference deconvolution is also mostly cosmetic and can adversely affect the spectrum when distortions are not caused by poor shimming.
B. Spectral Deconvolution
BAYESIL models mixtures as sums of compound-specific Lorentzian peak clusters, partitions spectra into locally interacting regions, and solves the resulting inference problem with a factor graph and sequential Monte Carlo.
- Spectral model: Each compound spectrum is represented by Lorentzian peaks grouped into clusters, with peak height, center, and width parameters.
- Spectral model: A mixture spectrum is reconstructed by summing compound contributions, and deconvolution finds concentrations and shifts that best fit the observed spectrum.The initial loss is the squared difference between observed and reconstructed spectral heights.
- Spectral deconvolution: BAYESIL factors the large nonlinear search over cluster shifts and compound concentrations into smaller interrelated tasks.Small shift windows and localized Lorentzian influence make this factorization possible.
- Spectral deconvolution: The spectrum is partitioned into contiguous regions whose points are influenced by exactly the same subset of clusters.A typical CSF spectrum is partitioned into approximately 350 regions, each involving 1 to approximately 25 clusters.
- Spectral deconvolution: The loss is decomposed across regions, but regional solutions are combined probabilistically because solving each region independently does not produce the overall correct answer.
- Probabilistic inference: BAYESIL represents deconvolution as MAP inference in a factor graph with regional factors and concentration or chemical-shift variables.A non-parametric sequential Monte Carlo method narrows kernel-density distributions as temperature decreases until convergence.
C. BAYESIL’s Spectral Library
BAYESIL relies on reference 1D 1H NMR spectra organized into compound-specific peak clusters and uses biofluid-specific sublibraries to support spectral fitting.
- Library design: The quality and size of the spectral library are critical to spectral deconvolution performance.
- Library contents: The spectral library contains reference information for each compound, including peak clusters, peak amplitudes, positions, widths, and allowable chemical-shift windows.
- Biofluid-specific libraries: BAYESIL uses separate spectral sublibraries for serum and CSF.The serum library contains 50 NMR-detectable compounds, while the CSF library contains 48.
- Biofluid-specific libraries: Biofluid-specific or organism-specific libraries improve spectral fitting by reducing the number of possible compounds.
II. ASSESSMENT
BAYESIL was assessed on computer-generated, laboratory-generated, and biological spectra from serum and CSF, using identification and concentration-quantification measures. Figure 5 further examines compound-level decisions and concentration differences against expert estimates and detection thresholds.
- Assessment: BAYESIL was assessed on three spectral-data types spanning computer-generated mixtures, laboratory-generated mixtures, and real biological samples.The computer-generated mixtures included serum and CSF spectra sampled from measured concentration and chemical-shift distributions.
- Assessment: Identification accuracy counts true positives and true negatives relative to the detection threshold, while quantitative accuracy measures whether estimated concentrations are close enough to true values.The quantitative measure uses a median-based percentage-error criterion rather than relying only on absolute concentration differences.
- Assessment: Figure 5 reports compound-level false-positive, true-positive, false-negative, and true-negative frequencies across 50 serum samples.Its right panel compares average concentrations for correctly identified compounds with expert values and compound-specific detection thresholds.
- Assessment: BAYESIL’s results on diverse test data were often within 10% of expert concentration estimates.For spectra with known ground truth, its metabolic profiles were often more accurate than those produced by experts.
A. Efficiency
BAYESIL profiles serum or CSF spectra in less than five minutes on a single CPU, supporting substantially higher sustained processing throughput than a human expert.
- Efficiency: Less than 5 minutes is the typical runtime for profiling one serum or CSF spectrum on a single 2.8 GHz CPU processor.Over 24 hours, the system should process more than 200 spectra versus approximately 20 spectra per day for a human expert.
- Efficiency: Reducing the number of deconvolution particles by a factor of 10 lowers runtime to less than one minute, often reducing accuracy by only 5–10%.This speed-up trades computational time against identification and quantification accuracy.
B. Limitations
BAYESIL’s scope is constrained by NMR’s intrinsic sensitivity, sample-volume requirements, and dependence on correctly specified biofluid-specific spectral libraries.
- B. Limitations: NMR has a lower detection limit of approximately 1–5 µM and requires relatively large samples of about 500 µL.High-quality, high-SNR spectra such as the CSF samples studied can still achieve high accuracy within these detection limits.
- B. Limitations: BAYESIL requires users to provide the biofluid being analyzed because biofluid-specific spectral libraries support its identification and quantification performance.Misidentifying or mislabelling the biofluid could lead to poorer results.
- B. Limitations: The method depends on prior knowledge of the typical composition of each biofluid mixture.This requirement has motivated efforts to characterize NMR-detectable metabolomes across fluids including serum, CSF, saliva, urine, milk, rumen, and cell extracts.
C. Other systems
Compared with prior software, BAYESIL combines public availability with fully automated profiling of mixtures containing more than 50 compounds. In the reported serum comparison, it was faster and more accurate than BATMAN.
- C. Other systems: Existing NMR profiling packages were often inaccurate on real biofluids or not publicly available, while BATMAN had automatically handled up to 26 publicly available compounds.These limitations motivated comparison with BAYESIL’s larger-scale automated profiling.
- C. Other systems: BAYESIL analyzed the serum spectrum in 5 minutes with 98% identification accuracy and 90% quantification accuracy, whereas BATMAN took 19 hours with 85% and 8%, respectively.The comparison used a 50-compound BAYESIL library and a 40-compound BATMAN subset.
- C. Other systems: BAYESIL is freely available for profiling 1D ^1H NMR spectra of serum, CSF, and other biofluid mixtures at 500 and 600 MHz, excluding urine.The paper presents removal of manual spectral-processing and deconvolution requirements as its central practical contribution.
III. MATERIALS AND METHODS
Reference spectra are prepared from metabolite stock solutions for NMR library construction, with CSF samples diluted when insufficient volume is available.
- Reference-sample preparation: Metabolite stock solutions are prepared at 1 mM to 100 mM in 20 mM NaHPO4 at pH 7.0.These solutions are diluted as needed to obtain a final stock concentration of 1 mM.
- Reference-sample preparation: 350 µL of prepared sample is transferred to a suitable Sigma tube for NMR data acquisition.
- Biological CSF handling: CSF samples are diluted with sufficient H2O when less than 285 µL is available.
I. OTHER NMR-ANALYSIS SOFTWARE TOOLS
Existing NMR-analysis tools provide incremental or partial automation, but the paper identifies a gap in fully automated, quantitative profiling of complex mixtures.
- Desired capabilities: Fully automated processing, identification, and quantification across diverse fluids, instruments, and frequencies are specified as ideal tool features.
- Existing tools: Only a handful of 19 identified software packages provide some automated identification or quantification.
- Existing tools: Many tools require users to select compound subsets, lack quantification, or specialize in particular mixtures.
- Evaluation gaps: Accuracy is difficult to assess because many tools provide no evaluation or are tested only on simple mixtures or spike-in experiments.
- BAYESIL: BAYESIL is described as completely automating preprocessing and deconvolution for wide-ranging complex mixtures involving more than 60 compounds.
III. DETAILS ABOUT THE CONSTRUCTION OF NMR
BAYESIL constructs localized spectral regions and performs approximate factor-graph inference over particle-based distributions to estimate compound concentrations and chemical shifts.
- Spectral regions: BAYESIL bounds each cluster’s influence to a relatively small spectral region because each peak significantly affects only a limited range.
- Spectral regions: Cluster upper bounds are estimated by shifting each library signature under the spectrum and finding its maximum possible value with other compounds absent.
- Factor-graph model: The loss is decomposed over spectral regions, allowing a factor-graph representation whose factors connect regions to associated shift and concentration variables.
- Factor-graph model: Inference is difficult because factors involve many continuous variables and may be multimodal, preventing basic sampling and several parametric methods.
- Approximate inference: BAYESIL represents variable distributions with particles, repeatedly samples joint assignments, evaluates regional losses, and updates distributions toward lower-loss values.The experiments use N = 10,000 particles, while more particles can increase accuracy at the cost of runtime.
- Approximate inference: Kernel density estimation produces updated variable distributions, and convergence returns the mode as the approximate MAP assignment.
- Approximate inference: Importance sampling weights are ignored in practice to reduce variance and computation time because the mode, rather than the marginals, is the target.
V. LIST OF NMR-DETECTABLE COMPOUNDS IN SERUM AND CSF
BAYESIL maintains separate NMR-detectable compound libraries for serum and CSF, each containing dozens of named metabolites plus DSS.
- Serum library: The serum library includes metabolites spanning amino acids, organic acids, sugars, and other compounds, plus DSS.
- CSF library: The CSF library includes metabolites such as acetate, citrate, glucose, lactate, amino acids, ketone bodies, and other compounds, plus DSS.