Metabolomics Panel

A metabolomics panel measures the small molecules circulating in your blood or urine, typically a few hundred to a few thousand of them, and returns concentrations or relative intensities per sample. It sits downstream of the genome, the transcriptome, and the proteome, and it changes on the timescale of hours, which makes it the layer most sensitive to what you ate, when you slept, and whether you exercised before the draw.

First, a naming collision worth clearing up. A comprehensive metabolic panel (CMP) is a 14-analyte clinical chemistry test: glucose, calcium, albumin, total protein, sodium, potassium, bicarbonate, chloride, BUN, creatinine, ALP, ALT, AST, bilirubin. A metabolomics panel is a research-grade assay measuring hundreds to thousands of compounds. If your CMP flags out of range, that is a question for the clinician who ordered it, and a single out-of-range value on a 14-test panel is common by chance alone. Metabolomics does not replace it, and nothing here interprets it for you.

How it works

Two platforms dominate. NMR (Nightingale, Bruker) quantifies roughly 250 analytes in absolute units, mostly lipoprotein subclasses, fatty acids, a handful of amino acids, glycolysis intermediates, and GlycA. It is highly reproducible, needs little sample prep, and has a hard sensitivity floor around micromolar, so it misses most of the interesting low-abundance chemistry.

LC-MS is the other route. Reverse-phase C18 captures lipids and mid-polarity compounds, HILIC captures amino acids, sugars, nucleotides, and organic acids. Targeted panels (Biocrates MxP Quant 500, for example) use isotope-labeled internal standards and report absolute concentrations for a fixed list. Untargeted runs (Metabolon HD4 and similar) report thousands of features, of which perhaps 30-50% are identified compounds and the rest are named by mass-to-charge ratio and retention time.

We would take a targeted panel with internal standards over untargeted breadth for a personal longitudinal profile. Absolute concentrations are comparable across runs and across years. Untargeted relative intensities are not, unless every timepoint runs in the same batch, which defeats the purpose of longitudinal sampling.

In your own data

Expect a wide CSV or Excel workbook: rows are samples, columns are metabolites, plus a metadata sheet with batch, run order, injection index, and QC flags. From Metabolon you also get ScaledImpData (median-scaled, missing values imputed to the minimum) and OrigScale. Use OrigScale and do your own handling. Their imputation to the observed minimum will manufacture structure if the limit of detection differs by batch.

A working pipeline:

1. Drop features missing in more than 20-30% of samples. Everything else, impute with half the minimum observed value per feature, or use imputeLCMD::impute.QRILC if you want something less crude.
2. Log-transform (log2(x + 1)), then probabilistic quotient normalization to correct for dilution, which matters enormously in urine and still matters in plasma.
3. Fit and remove run order and batch. sva::ComBat works if batch is balanced against your variable of interest. If it is not, you are stuck, and no correction will save you.
4. Plot PCA colored by batch before and after. If PC1 still tracks batch, stop and investigate rather than proceeding to statistics.

The single most common mistake in personal metabolomics is drawing blood in a non-standardized state. Fasting duration, time of day, and recent exercise move amino acids, acylcarnitines, and bile acids by amounts that swamp most biological signals of interest. Fix the protocol: 12 hours fasted, same clock time, no exercise in the prior 24 hours, and record all three anyway. Coefficients of variation on the same pooled sample across a run typically land at 5-15% for well-behaved targeted analytes and considerably worse for low-abundance untargeted features. Ask for the pooled QC data. If the vendor will not provide it, treat the numbers as directional.

Limitations

Published metabolite panels replicate poorly. Parkinson’s disease work has repeatedly surfaced small candidate panels, but the specific metabolites differ across studies and platforms ¹². Alzheimer’s disease work implicating bile acids and sphingolipids in brain and blood is mechanistically interesting, and the effect sizes are population-level averages, not individual classifiers ³⁴. Pediatric NAFLD panels reached useful discrimination in a cohort, and none of that transfers to a single adult sample without validation in a comparable population ⁵. Metabolomics also does not replace genetic testing for inborn errors of metabolism, where gene panels and sequencing carry the diagnosis ⁶⁷.

The layer’s strength is its own history. One metabolomics panel is a snapshot with a wide confidence interval. Twelve of them, same protocol, same platform, gives you a personal reference range where a two-standard-deviation move means something. Interpretation of any abnormal result belongs with a clinician.

Woolf Software builds longitudinal molecular profiles of individuals: whole-genome sequencing, RNA sequencing, proteomics, blood biomarkers, and continuous glucose data, integrated into one model of you. Build your profile.