Metabolomics: Complete Guide for Beginners (MS, NMR & Data)

When I teach metabolomics, I start with a simple picture: genes tell you what could happen, proteins do the work, but metabolites show what is happening right now. A blood sample from someone who just drank coffee, fasted overnight or took paracetamol tells a different chemical story, even though their genome has not changed. That is why metabolomics has become one of the most exciting, and most in-demand, skills in biology, medicine and agriculture.

It is also a field where beginners get lost quickly: m/z, retention time, adducts, features, PQN, PLS-DA, MSI levels. So I built this guide from the slides and infographics I use in class and from my new Metabolomics Complete Course on YouTube. It goes from “what is a metabolite?” all the way to a working Python data-analysis pipeline, in plain English, with every important claim linked to its source.

If you are new to the field, read my article on why learning bioinformatics is essential first, then come back here. If you already work in a lab, jump straight to study design and quality control or statistics.

How small is “small”? A typical enzyme weighs around 50,000 daltons (Da); glucose weighs 180 Da. Metabolites are the inputs, intermediates and outputs of the cell’s chemical reactions: glucose is broken down to pyruvate, pyruvate becomes lactate or enters the TCA cycle, and so on.

Why the metabolome is closest to the phenotype

Think of the “omics cascade”. The genome says what can happen, the transcriptome what the cell plans to do, and the proteome what does the work. The metabolome shows what is happening now, so it tracks traits and health most closely. Oliver Fiehn called metabolomics “the link between genotypes and phenotypes” (Plant Molecular Biology, 2002).

Your metabolome mixes chemistry from five sources

Not every molecule in your blood was made by your own cells. The human metabolome combines chemistry from your cells (glucose, lactate, cholesterol), food (vitamins, plant polyphenols), gut microbes (short-chain fatty acids, TMAO), drugs (paracetamol and its breakdown products) and the environment (pollutants, pesticides, plastics). Everything outside your genes that you are exposed to over a lifetime is called the exposome.

Start the video course here. In episode 1 I explain these basics in Urdu/Hindi, with examples from biology and disease research:

Episode 1 · 21:33

↑ Back to topNext: A short history of metabolomics →

• 1999“Metabonomics” is coined for NMR studies of metabolic responses in body fluids (Nicholson, Lindon & Holmes, Xenobiotica).
• 2002The link from genotype to phenotype: Fiehn frames metabolomics as the missing layer (Plant Mol Biol).
• 2007HMDB launches with 2,180 metabolites, and the Metabolomics Standards Initiative proposes the first reporting standards (Sumner et al.).
• 2012“The apogee of the omics trilogy” (Patti, Yanes & Siuzdak, Nat Rev Mol Cell Biol).
• 2016Beyond biomarkers, towards mechanisms (Johnson et al.).
• 2021Community guide to MS best practice for annotation, quantification and reporting (Alseekh et al., Nat Methods).
• 2022HMDB 5.0 lists about 218,000 entries (Wishart et al., NAR).
• TodaySingle-cell, spatial and population-scale studies, and AI tools that predict structures from MS/MS spectra.

The human metabolome catalogue grew 100× in 15 years

Metabolite entries in the Human Metabolome Database, by release year

↑ Back to topNext: Targeted vs untargeted →

↑ Back to topNext: The eight-step workflow →

1. PlanAsk a questione.g. does drug X change liver metabolism?
2. PlanDesign the studygroups, controls, biological replicates, random run order
3. Wet labCollect and quenchfreeze fast so enzymes stop working
4. Wet labExtractsolvents pull metabolites out of the sample
5. Wet labMeasureLC-MS, GC-MS or NMR, with blanks and QCs
6. Dry labProcess dataraw signals become a feature table
7. Dry labAnalyze and identifystatistics, then name the molecules
8. Dry labInterpretmap to pathways; validate; form new hypotheses

In episode 3 of the course I walk through this whole workflow in detail, including the dry-lab pipeline step by step. If you plan to process your own raw data, this is your roadmap:

Episode 3 · 1:05:38

↑ Back to topNext: Study design and quality control →

Your study-design checklist

1. One clear question and a defined comparison (for example treated vs control, disease vs healthy, time course).
2. Biological replicates in every group, not only repeat injections of one sample. Use pilot data or a power analysis to choose the number.
3. Control confounders: match age, sex, diet, fasting state and time of sampling, or record them for statistics.
4. Randomize the run order, and the extraction order, so drift and batch effects are not confused with groups.
5. Plan QC samples: a pooled QC every 5–10 injections, internal standards in every sample, extraction and solvent blanks.
6. Write down everything: sample collection, storage, freeze–thaw cycles and instrument settings. You will need it for reporting.

Sample types and how to handle them

Quenching and extraction

Enzymes keep working after sampling, so some metabolites (for example ATP or glucose-6-phosphate) change within seconds. Quenching stops metabolism with liquid nitrogen or cold solvent. Extraction then pulls the metabolites out: cold methanol, acetonitrile or methanol–water mixtures for polar metabolites, and two-phase methods (for example with chloroform or MTBE) to separate polar metabolites from lipids. Whatever you choose, extract every sample the same way, in random order.

↑ Back to topNext: How the instruments work →

m/z and adducts: why one molecule gives several signals

m/z means mass divided by charge. Most metabolite ions carry a single charge, so m/z is the molecule’s mass plus or minus a proton, or plus another ion (an adduct) such as Na⁺, K⁺ or NH₄⁺. Glucose (monoisotopic mass 180.0634 Da) appears at m/z 179.0561 as [M−H]⁻ in negative mode and at 203.0526 as [M+Na]⁺ in positive mode. This short script calculates the common ions and matches an observed m/z within a ppm tolerance:

# mz_calculator.py: which m/z values will a metabolite give in the mass spectrometer?
# No extra packages needed (standard Python 3).

PROTON = 1.007276   # mass of a proton (Da)
NA = 22.989218      # Na+ ion
K = 38.963158       # K+ ion
NH4 = 18.033823     # NH4+ ion

# Monoisotopic masses (Da) from HMDB
METABOLITES = {
    "glucose": 180.063388,
    "phenylalanine": 165.078979,
    "lactate (lactic acid)": 90.031694,
    "citrate (citric acid)": 192.027003,
    "glutamine": 146.069142,
}


def adducts(m):
    """Common single-charge ions for a neutral mass m."""
    return {
        "[M+H]+": m + PROTON,
        "[M+Na]+": m + NA,
        "[M+K]+": m + K,
        "[M+NH4]+": m + NH4,
        "[M-H]-": m - PROTON,
    }


def ppm_error(observed, theoretical):
    """Mass error in parts per million (ppm)."""
    return (observed - theoretical) / theoretical * 1e6


def match(observed_mz, tolerance_ppm=5):
    """Find which metabolite/adduct could explain an observed m/z."""
    hits = []
    for name, mass in METABOLITES.items():
        for ion, mz in adducts(mass).items():
            err = ppm_error(observed_mz, mz)
            if abs(err) <= tolerance_ppm:
                hits.append((name, ion, round(mz, 4), round(err, 2)))
    return hits


if __name__ == "__main__":
    for name, mass in METABOLITES.items():
        ions = ", ".join(f"{ion} {mz:.4f}" for ion, mz in adducts(mass).items())
        print(f"{name:22s} M = {mass:.4f} -> {ions}")
    print()
    for observed in [179.0560, 166.0864, 203.0530]:
        print(f"Observed m/z {observed}: {match(observed) or 'no match within 5 ppm'}")

python mz_calculator.py

MS/MS: fragment patterns fingerprint a molecule

In tandem mass spectrometry, the instrument picks one ion (the precursor), smashes it by collision with gas (CID) and records the fragments. Phenylalanine [M+H]⁺ at m/z 166.09 loses formic acid (−46) to give 120.08, and also produces a ring fragment at 91.05. The same fragments plus the same mass are strong evidence of identity. In untargeted work, instruments collect MS/MS automatically, either for the most intense ions (DDA, data-dependent acquisition) or for all ions in wide windows (DIA, data-independent acquisition).

MS or NMR?

↑ Back to topNext: Data processing →

The processing pipeline, step by step

1. Convert files from the vendor format to open mzML with ProteoWizard msConvert (MS-DIAL also reads many vendor files directly).
2. Peak picking: detect chromatographic peaks for every m/z in every sample.
3. Alignment: match the same feature across samples and correct retention-time shifts.
4. Gap filling: go back to the raw data to recover peaks missed in some samples.
5. Grouping: link isotopes, adducts and in-source fragments that belong to one compound (e.g. CAMERA, MS-DIAL, MZmine ion identity networking).
6. Filtering: remove features found in blanks (e.g. sample/blank ratio below 3), missing in most samples, or unstable in pooled QCs (RSD above 20–30%).
7. Drift and batch correction: model the QC signal over the run order (e.g. LOESS-based QC-RSC) and correct every feature.
8. Normalization, transformation and scaling: correct dilution (PQN, internal standards, creatinine for urine), log-transform, then scale (Pareto or autoscaling) for multivariate statistics.

In episode 4, a one-minute summary, I list the core tasks a bioinformatician owns in a metabolomics project:

Episode 4 · 0:59

↑ Back to topNext: Statistics with Python →

A complete pipeline in Python

These two scripts let you practise without your own data. The first makes a realistic example feature table (22 samples: 8 control, 8 treated, 6 pooled QCs; 300 features, of which 30 really change). The second runs the full pipeline: missing-value filter, imputation, QC-RSD filter, PQN normalization, log transformation, Pareto scaling, PCA, Welch t-tests with FDR, and a volcano plot. To use your own data, export a CSV with the same layout from MZmine, MS-DIAL or XCMS.

# make_example_data.py: create a realistic example feature table to practise on
# pip install numpy pandas
import numpy as np
import pandas as pd

rng = np.random.default_rng(42)
n_features, n_per_group, n_qc = 300, 8, 6

# each feature gets a typical intensity (log-normal, like real LC-MS data)
base = rng.lognormal(mean=10, sigma=1.2, size=n_features)
effect = np.ones(n_features)
effect[:15] = rng.uniform(2.0, 4.0, 15)     # 15 features go UP in treated samples
effect[15:30] = rng.uniform(0.25, 0.5, 15)  # 15 features go DOWN

def samples(n, fold, noise=0.25):
    return base * fold * rng.lognormal(0, noise, size=(n, n_features))

control = samples(n_per_group, 1.0)
treated = samples(n_per_group, effect)
qc = samples(n_qc, (1 + effect) / 2, noise=0.08)           # pooled QC = mix of all samples
dilution = rng.uniform(0.7, 1.3, size=(2 * n_per_group + n_qc, 1))  # unequal sample amounts
X = np.vstack([control, treated, qc]) * dilution
X[rng.random(X.shape) < 0.03] = np.nan                      # a few missing values, as in real data

names = [f"C{i+1}" for i in range(n_per_group)] + [f"T{i+1}" for i in range(n_per_group)] + [f"QC{i+1}" for i in range(n_qc)]
groups = ["Control"] * n_per_group + ["Treated"] * n_per_group + ["QC"] * n_qc
mz = np.round(rng.uniform(70, 900, n_features), 4)
rt = np.round(rng.uniform(0.5, 15, n_features), 2)
cols = [f"{m}_{r}" for m, r in zip(mz, rt)]                 # feature name = m/z_retention time

df = pd.DataFrame(X, columns=cols)
df.insert(0, "group", groups)
df.insert(0, "sample", names)
df.to_csv("feature_table.csv", index=False)
print(f"Saved feature_table.csv: {df.shape[0]} samples x {n_features} features")

# metabolomics_stats.py: from a feature table to PCA, statistics and a volcano plot
# pip install numpy pandas scipy scikit-learn matplotlib
# Input: feature_table.csv with columns sample, group (Control / Treated / QC), then one column per feature
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from scipy import stats
from sklearn.decomposition import PCA

df = pd.read_csv("feature_table.csv")
meta, X = df[["sample", "group"]], df.drop(columns=["sample", "group"])
print("Start:", X.shape[1], "features")

# 1. Remove features missing in more than half of the samples
X = X.loc[:, X.isna().mean() <= 0.5]

# 2. Fill the remaining gaps with 1/5 of the feature's smallest value (a common, simple choice)
X = X.fillna(X.min() / 5)

# 3. Keep features that are stable in the pooled QC samples (RSD below 30%)
qc = X[meta.group == "QC"]
rsd = qc.std() / qc.mean() * 100
X = X.loc[:, rsd < 30]
print("After missing-value and QC-RSD filters:", X.shape[1], "features")

# 4. PQN normalization: corrects for samples that are more or less concentrated
reference = X[meta.group == "QC"].median()
quotients = X.div(reference, axis=1)
X = X.div(quotients.median(axis=1), axis=0)

# 5. Log-transform (makes intensities more normal), then Pareto-scale for PCA
L = np.log2(X)
P = (L - L.mean()) / np.sqrt(L.std())

# 6. PCA: do the groups separate? Are the QCs tightly clustered?
pca = PCA(n_components=2)
scores = pca.fit_transform(P)
var = pca.explained_variance_ratio_ * 100
colours = {"Control": "#0072B2", "Treated": "#D55E00", "QC": "#999999"}
fig, ax = plt.subplots(figsize=(6, 5))
for g, c in colours.items():
    m = (meta.group == g).values
    ax.scatter(scores[m, 0], scores[m, 1], c=c, label=g, s=60, edgecolor="white")
ax.set_xlabel(f"PC1 ({var[0]:.1f}%)")
ax.set_ylabel(f"PC2 ({var[1]:.1f}%)")
ax.legend(frameon=False)
ax.set_title("PCA score plot")
fig.tight_layout()
fig.savefig("pca.png", dpi=200)

# 7. Univariate test per feature: Welch t-test on log data, then FDR correction
c, t = L[meta.group == "Control"], L[meta.group == "Treated"]
p = stats.ttest_ind(t, c, equal_var=False).pvalue
q = stats.false_discovery_control(p, method="bh")          # Benjamini-Hochberg
log2fc = t.mean() - c.mean()                               # already in log2 units
res = pd.DataFrame({"feature": L.columns, "log2FC": log2fc.values, "p": p, "q_FDR": q})
res["significant"] = (res.q_FDR < 0.05) & (res.log2FC.abs() >= 1)
res = res.sort_values("q_FDR")
res.to_csv("results.csv", index=False)
print("Significant features (FDR < 0.05 and at least 2-fold):", int(res.significant.sum()))
print(res.head(5).to_string(index=False))

# 8. Volcano plot
fig, ax = plt.subplots(figsize=(6, 5))
col = np.where(~res.significant, "#BBBBBB", np.where(res.log2FC > 0, "#D55E00", "#0072B2"))
ax.scatter(res.log2FC, -np.log10(res.p), c=col, s=18)
ax.axvline(-1, ls="--", c="grey", lw=0.8)
ax.axvline(1, ls="--", c="grey", lw=0.8)
ax.set_xlabel("log2 fold change (Treated / Control)")
ax.set_ylabel("-log10 p-value")
ax.set_title("Volcano plot")
fig.tight_layout()
fig.savefig("volcano.png", dpi=200)
print("Saved pca.png, volcano.png and results.csv")

pip install numpy pandas scipy scikit-learn matplotlib
python make_example_data.py
python metabolomics_stats.py

↑ Back to topNext: Metabolite identification →

For high-resolution MS, many labs also use the five-level scale of Schymanski et al. (2014), which splits putative structures into finer levels.

Databases and tools for annotation

• HMDB: human metabolites with masses, spectra and biology
• MassBank and MoNA: open MS/MS reference spectra
• GNPS: library search and molecular networking that groups related molecules by similar spectra
• SIRIUS with CSI:FingerID: predicts molecular formulas and structures from MS/MS, even for compounds not in libraries
• LIPID MAPS: lipid structures and classification
• PubChem and ChEBI: general chemical databases

↑ Back to topNext: Pathway analysis →

• Over-representation analysis (ORA): is the list of significant metabolites enriched in a pathway?
• Quantitative enrichment (MSEA): uses all measured values, not only a cut-off list
• Topology analysis: gives more weight to metabolites at important positions in a pathway
• mummichog / functional analysis: works from m/z features before identification (Li et al., 2013); available in MetaboAnalyst

↑ Back to topNext: Applications →

↑ Back to topNext: Challenges and frontiers →

Multi-omics integration

Metabolomics becomes even more powerful when combined with other layers: a gene variant (genomics) that changes an enzyme’s expression (transcriptomics and proteomics) and so a metabolite level (metabolomics) tells a complete causal story. Methods include correlation networks, joint dimensionality reduction (for example MOFA or DIABLO in mixOmics) and pathway-level integration. For protein structures behind enzymes, see my AlphaFold tutorial.

Reporting and sharing your data

Good science is reproducible. Deposit raw data and metadata in MetaboLights or the Metabolomics Workbench, describe your QC procedures, and report identification levels. These repositories are also free sources of real datasets for practice.

↑ Back to topNext: Your learning roadmap →

1. Week 1: concepts. Read Sections 1–4 of this guide and watch episodes 1 and 2 of the video course.
2. Week 2: instruments and workflow. Sections 5–6 and episode 3; read the primer by Liu & Locasale (2017).
3. Week 3: programming and statistics. Run the Python scripts in Section 8; learn pandas, PCA and FDR.
4. Week 4: real data. Download a small study from MetaboLights, process it with MZmine or MS-DIAL, and analyse it in MetaboAnalyst and Python.
5. Week 5: annotation and pathways. Annotate top features with GNPS or SIRIUS, report MSI levels, run pathway analysis.
6. Week 6: portfolio. Write up your project as a GitHub repository and a short report or poster.

Episode 2 is the best place to start building your vocabulary before you open any software:

Episode 2 · 6:41

The free Metabolomics Complete Course (Urdu/Hindi)

4 videos so far, in order. New episodes are added to the same playlist.

Key papers to read

Peer-reviewed reviews, standards and databases cited in this guide

Start here: reviews

• 1
  Liu & Locasale (2017) Metabolomics: a primerTrends in Biochemical Sciences 42:274–284
• 2
  Patti, Yanes & Siuzdak (2012) Metabolomics: the apogee of the omics trilogyNat Rev Mol Cell Biol 13:263–269
• 3
  Johnson, Ivanisevic & Siuzdak (2016) Metabolomics: beyond biomarkers and towards mechanismsNat Rev Mol Cell Biol 17:451–459
• 4
  Rinschen et al. (2019) Identification of bioactive metabolites using activity metabolomicsNat Rev Mol Cell Biol 20:353–367
• 5
  Wishart (2016) Emerging applications of metabolomics in drug discovery and precision medicineNat Rev Drug Discov 15:473–484

Methods, standards and databases

• 1
  Alseekh et al. (2021) Mass spectrometry-based metabolomics: a guide for annotation, quantification and best reporting practicesNature Methods 18:747–756
• 2
  Sumner et al. (2007) Proposed minimum reporting standards for chemical analysisMetabolomics 3:211–221
• 3
  Schymanski et al. (2014) Identifying small molecules via high resolution mass spectrometry: communicating confidenceEnviron Sci Technol 48:2097–2098
• 4
  Emwas et al. (2019) NMR spectroscopy for metabolomics researchMetabolites 9:123
• 5
  Wishart et al. (2022) HMDB 5.0: the Human Metabolome Database for 2022Nucleic Acids Research 50:D622–D631
• 6
  Pang et al. (2024) MetaboAnalyst 6.0Nucleic Acids Research 52:W398–W406
• 7
  Dührkop et al. (2019) SIRIUS 4: turning tandem mass spectra into metabolite structure informationNature Methods 16:299–302
• 8
  Li et al. (2013) Predicting network activity from high throughput metabolomics (mummichog)PLoS Comput Biol 9:e1003123

Free tools, databases and data repositories

• 1
  HMDBHuman Metabolome Database
• 2
  MetaboAnalystStatistics and pathway analysis online
• 3
  MZmineOpen-source LC-MS/GC-MS processing
• 4
  MS-DIALProcessing with built-in libraries
• 5
  XCMS (R)Peak picking and alignment in R
• 6
  GNPSSpectral libraries and molecular networking
• 7
  SIRIUSFormula and structure prediction from MS/MS
• 8
  MoNAMassBank of North America spectra
• 9
  KEGG PathwayMetabolic pathway maps
• 10
  LIPID MAPSLipid structures and tools
• 11
  MetaboLightsPublic metabolomics datasets (EMBL-EBI)
• 12
  Metabolomics WorkbenchPublic datasets and tools (NIH)

Related guides and courses on Codanics

• 1
  Bioinformatics ka ChillaThe complete bioinformatics course
• 2
  Why learning bioinformatics is essentialStart here if you are new
• 3
  AlphaFold tutorialProtein structure prediction
• 4
  Python ka Chilla for data science40 days of Python for analysis
• 5
  R with AammarR for statistics and figures
• 6
  Pandas mastery guideHandle feature tables in Python
• 7
  Data visualisation guideChoose the right plot
• 8
  ABC of statisticsStatistics from the basics
• 9
  Handling outliersSpot and treat outlier samples
• 10
  Systematic review and meta-analysisSynthesize published evidence
• 11
  Conference posters with AI toolsPresent your metabolomics results
• 12
  Using AI tools in scientific writingWrite up your study

More research, AI and bioinformatics guides on the Codanics blog.

Metabolomics glossary

30 key terms in plain English

Metabolite

A small molecule (usually below about 1,500 Da) that a cell makes, uses or takes in, such as glucose, lactate or an amino acid.

Metabolome

The complete set of metabolites in a biological sample at a given time.

Metabolomics

The large-scale measurement and study of metabolites in biological samples.

Metabonomics

An older, closely related term coined for NMR studies of metabolic responses in body fluids; today used almost interchangeably with metabolomics.

Lipidomics

A branch of metabolomics that focuses on lipids, such as fatty acids, phospholipids and cholesterol.

Exposome

Everything outside your genes that you are exposed to over a lifetime: food, drugs, microbes and pollutants.

Xenobiotic

A foreign chemical in the body, such as a drug, food additive or pollutant.

Targeted metabolomics

Measuring a predefined panel of known metabolites, usually with pure standards for exact amounts.

Untargeted metabolomics

Profiling as many signals as possible to discover differences, usually as relative amounts.

Quenching

Stopping metabolism immediately after sampling, for example with liquid nitrogen or cold solvent.

Pooled QC

A mixture of all study samples, injected throughout the run to monitor and correct instrument drift.

Internal standard

A known compound (often isotope-labelled) added to every sample to check recovery and correct signal variation.

Blank

A sample without biological material, used to detect contamination from tubes, solvents and the instrument.

m/z

Mass-to-charge ratio of an ion, the value a mass spectrometer measures.

Adduct

The ion form of a molecule, such as [M+H]+, [M+Na]+ or [M−H]−.

Retention time

When a compound comes out (elutes) of the chromatography column.

ESI

Electrospray ionization: a soft method that turns molecules in liquid into ions for LC-MS.

MS/MS

Tandem mass spectrometry: an ion is selected and broken into fragments whose masses help identify it.

Feature

A detected signal (one m/z at one retention time), not yet a confirmed metabolite.

Feature table

A matrix with samples in rows, features in columns and intensities in the cells.

ppm error

The difference between measured and theoretical mass in parts per million; high-resolution instruments reach below 5 ppm.

PQN

Probabilistic quotient normalization: corrects for samples that are more or less diluted.

Pareto scaling

Dividing each centred feature by the square root of its standard deviation, so large and small signals both contribute.

PCA

Principal component analysis: an unsupervised map of samples that shows groups, trends and outliers.

PLS-DA

Partial least squares discriminant analysis: a supervised model that separates known groups; it must be validated because it overfits easily.

VIP score

Variable importance in projection: ranks features by their contribution to a PLS-DA model.

FDR

False discovery rate: the expected share of false positives among significant results, controlled with the Benjamini–Hochberg method.

Volcano plot

A plot of fold change against statistical significance for every feature.

Annotation

Assigning a putative identity to a feature, with a stated confidence level.

Pathway enrichment

Testing whether a metabolic pathway contains more changed metabolites than expected by chance.

Frequently asked questions