From voided sample to disease classifier — 1H-NMR spectral acquisition, bucket binning, and OPLS-DA metabolic fingerprint discrimination
Urine is an attractive biofluid for metabolomics: non-invasive, available in large volume, and rich with hundreds of small-molecule end-products of metabolism, diet, and the gut microbiome. But it is also notoriously variable — dilution, circadian rhythm, diet, and even the exact pH at measurement time all shift NMR peak positions and intensities. A rigorous preanalytical protocol is the single biggest determinant of whether a downstream metabolic fingerprint reflects disease biology or merely sample handling noise.
Standardized collection follows a strict sequence to minimize preanalytical variance across a cohort:
• Timing: second morning void after an overnight (≥8h) fast is preferred over random spot collection — it removes the confound of recent meals and matches circadian excretion phase across subjects. • Midstream, clean-catch technique: reduces contamination from urethral flora and epithelial debris. • Immediate quenching: samples are placed on ice within minutes of collection and centrifuged (2000×g, 10 min, 4°C) to pellet cells, casts, and crystals — bacterial overgrowth in unrefrigerated urine rapidly alters the metabolite profile (urease-producing bacteria hydrolyze urea within hours, spiking ammonia). • Preservatives: sodium azide (0.1%) may be added to inhibit bacterial growth in samples awaiting same-day analysis; azide-free aliquots are kept for any bioassay follow-up. • Aliquoting and freezing: samples are split into single-use aliquots and stored at −80°C to avoid repeated freeze-thaw cycling, which measurably degrades ascorbate, citrate, and certain amino acids.
Dilution correction is the central preanalytical challenge specific to urine (unlike serum/plasma, urine concentration varies 5–10× across individuals and even within the same person across a day). Three correction strategies are used in parallel: (1) creatinine normalization — dividing metabolite intensities by the co-measured creatinine peak, assuming roughly constant daily creatinine excretion; (2) osmolality-based normalization via freezing-point depression; (3) probabilistic quotient normalization (PQN) applied computationally after acquisition, which is now the field-standard because it is robust to a handful of grossly perturbed metabolites that would bias a simple total-intensity scaling.
Immediately before acquisition, 400–600 μL of urine supernatant is mixed with a 0.2M phosphate buffer (pH 7.40 ±0.05) — because carboxylic-acid and amine-bearing metabolites (citrate, creatinine, amino acids) titrate across the physiological urine pH range (4.5–8.0), and their chemical shifts move by up to 0.1–0.3 ppm per pH unit. Without tight pH buffering, the same metabolite would land in a different spectral bucket in different samples, destroying cross-sample comparability before a single spectrum is even acquired.
Two additional additives complete sample prep: • D2O (10% v/v): supplies the deuterium lock signal the spectrometer’s field-frequency lock circuit uses to maintain magnetic field stability throughout acquisition. • TSP-d4 (3-(trimethylsilyl)propionic acid-d4, 1 mM): a small, chemically inert reference molecule whose singlet is defined as exactly 0.00 ppm — every other peak in the spectrum is reported relative to this internal anchor, and its integral additionally serves as a concentration reference for absolute quantification.
Once prepared, the sample is loaded into a high-field NMR spectrometer, and a radiofrequency pulse tips the population of hydrogen nuclei into the transverse plane; as they relax back, they emit a decaying oscillating signal — the Free Induction Decay (FID) — which is Fourier-transformed into a frequency-domain spectrum. Urine's single largest signal, by orders of magnitude, is the ~110M water protons; suppressing that peak without distorting nearby metabolite resonances is the technical crux of the acquisition step.
The workhorse pulse sequence for urine metabolomics is 1D NOESY with presaturation (noesypr1d), chosen over simpler single-pulse acquisition specifically for its superior water suppression and flat baseline:
Sequence: RD – 90° – t1 – 90° – tm (mixing, ~100ms) – 90° – acquire, with continuous low-power irradiation at the water frequency during the relaxation delay (RD) and mixing time (tm).
• Presaturation: a weak, continuous RF field applied exactly at the water resonance for 1–2 seconds before the pulse sequence begins saturates (equalizes) the water proton spin populations so they produce negligible signal — while metabolite protons, at different frequencies, are unaffected. • Why NOESY over simple presaturation alone: the additional short mixing period allows residual water magnetization that has recovered via T1 relaxation to be further dephased, giving a flatter baseline around the 4.7 ppm water region — critical because sugars, amino acid α-protons, and other metabolites resonate close to water. • Trade-off: presaturation partially suppresses exchangeable protons (–OH, –NH) that resonate near water, and unavoidably attenuates some signal within ~0.1 ppm of the suppression frequency — a known and accepted limitation.
Acquisition parameters typically used in large cohort studies: 128 scans (NS), 4–8 dummy scans for equilibration, spectral width of ~20 ppm (12,000–18,000 Hz depending on field), 64k time-domain points, relaxation delay of 4 seconds (long enough for near-full T1 recovery of most urinary metabolites), total acquisition time ~4 minutes per sample — enabling autosampler-driven throughput of ~200–300 samples per instrument-day.
Higher static magnetic field strength (expressed by convention as the proton resonance frequency in MHz) delivers two compounding benefits for a complex biofluid mixture like urine:
• Increased spectral dispersion: chemical shift differences (in Hz) scale linearly with field strength, so peaks that overlap at 400 MHz can resolve into distinct, integrable resonances at 800–900 MHz — directly reducing the ambiguity of assigning a bucket to a single metabolite versus a co-eluting mixture of several. • Increased sensitivity: signal-to-noise scales roughly with B0^(3/2) to B0^(7/4), so an 800 MHz instrument can detect metabolites present at severalfold lower concentration than a 400 MHz instrument in the same acquisition time.
Cryogenically cooled probes (cryoprobes), which reduce electronic thermal noise in the receiver coil, provide an additional 3–4× sensitivity gain independent of field strength and are now standard in metabolomics core facilities. Combined 800 MHz + cryoprobe configurations can detect urinary metabolites down to low-micromolar concentration in a 4-minute acquisition, which is what makes population-scale (n>1,000) urine metabolomic cohorts practically achievable.
A raw NMR spectrum is not yet data a statistical model can use — it is a continuous curve of ~65,000 intensity points per sample, subject to small instrumental drifts in phase, baseline, and exact peak position between runs. Spectral processing converts this into a clean, aligned, and dimensionally reduced numerical matrix: one row per sample, one column per spectral "bucket," ready for multivariate analysis.
After Fourier transformation of the FID, several corrections are applied automatically in software (TopSpin, MNova, or the R package "speaq"):
• Phase correction: zero- and first-order phase errors (from imperfect pulse timing) are corrected so all peaks appear as pure absorption-mode lineshapes rather than mixed absorption/dispersion — essential for accurate integration. • Baseline correction: a smooth polynomial or spline baseline is subtracted to remove broad rolling distortions from probe ringdown and imperfect shimming, so peak areas reflect true metabolite concentration rather than baseline offset. • Chemical-shift referencing: every spectrum is calibrated so the TSP singlet sits at exactly 0.00 ppm, correcting for small run-to-run magnetic field drift. • Spectral alignment (icoshift, recursive segment-wise peak alignment): even after referencing, real metabolite peaks can still drift by a few Hz between samples due to residual pH/ionic-strength micro-differences; alignment algorithms warp small spectral segments to maximize cross-sample correlation before binning, preventing the same metabolite peak from splitting across two adjacent buckets in different samples.
Two competing philosophies exist for reducing a spectrum to a feature vector:
1. Uniform bucketing (bulk region binning): the spectrum from 0.5–9.5 ppm is divided into equal-width bins (commonly 0.04 ppm, ~220 buckets after exclusions), and the integral within each bin becomes one feature. Simple, robust to minor peak drift, and the field-standard for large-cohort screening — but a single bucket can contain overlapping signal from more than one metabolite.
2. Targeted / deconvoluted quantification: software such as Chenomx NMR Suite fits reference metabolite spectral libraries (from the Human Metabolome Database, HMDB) directly onto the observed spectrum, yielding absolute concentrations (μM) for ~40–60 confidently identified metabolites per sample. Slower and requiring more expert curation, but produces biologically interpretable, directly comparable concentration values rather than arbitrary bucket intensities.
Excluded regions: the residual water resonance (4.5–5.1 ppm) and, in urine specifically, the very broad urea singlet (5.5–6.0 ppm, driven by exchange with water and pH) are excluded from binning — both carry essentially no discriminative metabolic information and their large, variable intensity would otherwise dominate any downstream normalization.
Normalization: probabilistic quotient normalization (PQN) computes, for each sample, the ratio of its intensity to a reference spectrum (the cohort median spectrum, bin-by-bin), then takes the median of all these bin-wise quotients as a single most-probable dilution factor for that sample — a substantial improvement over creatinine normalization alone, since creatinine excretion itself varies with muscle mass, age, and renal function and is not a perfectly stable denominator, particularly in kidney disease cohorts where it is often the very variable under study.
With a clean sample-by-bucket matrix in hand, multivariate statistics extract the pattern of covarying metabolite changes that separates disease groups — a task impossible by inspecting univariate peaks one at a time, since the diagnostic signal in metabolomics is almost always a coordinated shift across a panel of metabolites rather than any single dramatic outlier.
Principal Component Analysis is always run first, before any group labels are used, as an unbiased quality-control step:
• PCA finds orthogonal linear combinations of the ~220 bucket variables (principal components) that successively capture the maximum remaining variance in the dataset. • Score plots (samples projected onto PC1 vs. PC2) reveal gross outliers — a sample far outside the main cloud usually indicates a technical failure (poor water suppression, contamination, degraded sample) rather than true biology, and is flagged for exclusion or re-run before proceeding. • Because PCA is unsupervised, any separation by disease class that emerges spontaneously on PC1/PC2 is a strong, unbiased signal that metabolic differences are a dominant source of variance in the dataset — though in practice, dietary and hydration variance often dominate PC1, and disease-related separation frequently only emerges on later components or after supervised modeling.
Orthogonal Projections to Latent Structures Discriminant Analysis (OPLS-DA) is the standard supervised classifier in metabolomics because it explicitly separates variance into two blocks:
• Predictive component(s): variance correlated with the class label (e.g., healthy vs. CKD) — this is the axis of biological interest. • Orthogonal components: variance uncorrelated with class (dietary variation, hydration, batch effects) — mathematically removed from the predictive axis, dramatically sharpening the group separation compared to plain PLS-DA or PCA.
Model validity is never assessed by R2Y (goodness-of-fit) alone, since a model with enough components can always fit training data perfectly — instead: • Q2Y (goodness-of-prediction) from 7-fold cross-validation estimates how well the model predicts held-out samples; Q2Y >0.4 is generally considered a valid, generalizable model in metabolomics. • Permutation testing (200+ iterations of randomly reshuffled class labels, rebuilding the model each time) generates a null distribution of R2/Q2 values; the real model's Q2Y must exceed essentially all permuted values (empirical p<0.01–0.005) to rule out overfitting to noise — a mandatory reviewer-requested control in any published metabolomics classifier.
Variable Importance in Projection (VIP) scores rank each of the ~220 spectral buckets by its contribution to the OPLS-DA separation; buckets with VIP >1 (combined with univariate significance, e.g. FDR-corrected p<0.05) are carried forward as the candidate discriminant panel, which is then manually cross-referenced against HMDB and Chenomx spectral libraries to assign a metabolite identity to each significant bucket.
A discriminant model built and cross-validated on a single cohort is a hypothesis, not a diagnostic test. The final and most rigorous stage of the pipeline freezes the metabolite panel and its coefficients exactly as trained, then evaluates it — with no further tuning — against an entirely independent patient cohort, quantifying real-world diagnostic performance with ROC curves and comparing it against the clinical chemistry tests already in routine use.
The locked OPLS-DA predictive score (or a simplified linear combination of the top VIP metabolites) is applied to every sample in the external cohort, and Receiver Operating Characteristic (ROC) analysis plots true positive rate against false positive rate across all possible decision thresholds:
• Area Under the Curve (AUC) summarizes overall discriminative power in a single number: 0.5 = no better than chance, 1.0 = perfect separation. Published urinary NMR panels for early diabetic nephropathy and CKD staging typically report external AUC of 0.85–0.92 — substantially better than urinary albumin-to-creatinine ratio alone in early-stage disease, where albuminuria has not yet risen above detection thresholds. • The operating threshold is chosen based on clinical context: a screening application prioritizes sensitivity (catch every true case, tolerate more false positives for confirmatory follow-up testing), while a test used to trigger an invasive biopsy prioritizes specificity. • DeLong's test statistically compares the NMR panel's ROC curve against a reference clinical model (e.g., eGFR + albuminuria) to establish that the metabolomic panel adds genuine incremental diagnostic value rather than merely reproducing existing clinical markers by proxy.
Translating a validated multivariate NMR signature into a deployable clinical test requires several additional steps beyond statistical validation:
• Assay simplification: the full ~220-bucket OPLS-DA model is distilled to a short panel of 4–8 named, individually quantifiable metabolites (e.g., citrate, hippurate, TMAO, glucose, alanine, creatinine-normalized ratios) that can be reported and interpreted by a clinician, rather than requiring the original multivariate black-box model. • Reference range establishment: each metabolite in the final panel needs population-specific reference intervals (age, sex, and often ethnicity-stratified), since urinary excretion baselines vary systematically across these strata. • Analytical validation: inter-day and inter-instrument reproducibility (CV <10–15% is the typical target for quantitative NMR metabolomics), and cross-platform concordance if the assay will also be offered on other spectrometers or via LC-MS as an orthogonal method. • Regulatory pathway: as a laboratory-developed test (LDT) it can be offered under CLIA certification in the US, but broader distribution as an IVD product requires FDA clearance with a defined intended-use population and formal clinical-utility evidence (that using the test actually changes management and improves outcomes, not just diagnostic accuracy in isolation).
The largest published urinary 1H-NMR metabolomics validation to date profiled >5,000 individuals across multiple cohorts for chronic kidney disease staging, identifying a panel anchored on citrate, hippurate, and trimethylamine-N-oxide (TMAO) that achieved external-cohort AUC of 0.88 for detecting stage-2 CKD — outperforming eGFR-based staging alone and detectable years before conventional albuminuria crossed diagnostic thresholds. This illustrates the core promise of the technique: a 4-minute, non-invasive, reagent-light NMR scan can surface a systems-level metabolic signature of organ dysfunction well before single-marker clinical chemistry does.