Bio Metabolomics Statistical Analysis

Version Compatibility

Reference examples tested with: ropls 1.34+, scipy 1.12+, statsmodels 0.14+, numpy 1.26+, pandas 2.1+, matplotlib 3.8+

Before using code patterns, verify installed versions match. If versions differ:

• Python: pip show <package> then help(module.function) to check signatures
• R: packageVersion('<pkg>') then ?function_name to verify parameters

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

Metabolomics Statistical Analysis

"Tell me which metabolites separate my groups" -> run an honest univariate test with dependence-aware FDR AND a permutation-validated multivariate model, then reconcile the two.

• R: ropls::opls() (PCA/PLS-DA/OPLS-DA + permutation), wilcox.test()/lm()/lme4::lmer(), p.adjust(method='BH')
• Python: scipy.stats.ttest_ind(equal_var=False)/mannwhitneyu, statsmodels multipletests(method='fdr_bh'), sklearn.cross_decomposition.PLSRegression + permutation_test_score

The Single Most Important Modern Insight -- A Score Plot Is a Hypothesis, Not a Result

In metabolomics the regime is p >> n (hundreds-to-thousands of features, tens of samples) with strongly correlated features. In that regime any binary labelling of n points in >= n-1 dimensions is linearly separable with probability 1, so PLS-DA and OPLS-DA produce a clean two-cluster score plot even for randomly assigned labels. A beautiful score plot is the generic output of the algorithm and carries essentially zero information. Only a cross-validated Q2 benchmarked against a permutation null distinguishes signal from geometry (Westerhuis 2008; Ruiz-Perez 2020). Three corollaries reorganize the whole skill: (1) R2 is no evidence (it can be driven to 1 by adding components); only permutation-validated Q2 licenses a claim. (2) Scaling is a hidden hypothesis -- variance-driven methods weight a feature by the variance it is allowed to contribute, so Pareto vs unit-variance hands back a different VIP list and a different biological story (van den Berg 2006). (3) Features are not independent (pathways, adducts, isotopologues), so naive BH-independence is violated and one real signal lights up its whole correlated cluster.

Transformation and Scaling -- the Decision That Changes Conclusions

Transformation (nonlinear, per value: corrects heteroscedastic multiplicative MS noise and skew) and scaling (linear, per feature: sets relative weight) are distinct. Mean-centering is the universal first step. Choosing not to scale is the strongest prior of all -- it lets the most abundant metabolite drive PC1.

van den Berg 2006: on real data autoscale and range recovered biologically meaningful loadings; Pareto is the pragmatic middle. Decision rule: transform first (if heteroscedastic -- usually yes for MS), then center, then scale; run at least Pareto AND UV, and if the top-VIP list or conclusion flips, the result is scaling-fragile and must be tempered.

Decision Tree by Scenario

Scaling + PCA

Goal: Get the honest unsupervised first look that cannot chase the labels, with QC as the primary data-quality readout.

Approach: Transform if heteroscedastic, then PCA with an explicit scaling; inspect QC clustering, batch coloring, and Hotelling T2.

library(ropls)
# scaleC default is "standard" (unit-variance/autoscale), NOT Pareto -- set explicitly
pca <- opls(t(feature_matrix), scaleC = 'pareto', fig.pdfC = 'none', info.txtC = 'none')
scores <- getScoreMN(pca)               # samples x components
getSummaryDF(pca)                       # R2X(cum) per component
# Tight pooled-QC clustering in the center = trustworthy run; QC scatter = analytical variance dominates

Permutation-Validated PLS-DA / OPLS-DA

Goal: Decide whether group separation is real, not a geometry artifact, before reading any VIP or S-plot.

Approach: Fit with an explicit scaling, raise permI far above the default of 20, and read pQ2/pR2Y -- a model whose true Q2 sits inside the permutation cloud is indistinguishable from chance.

library(ropls)
group <- factor(sample_info$group)
# OPLS-DA: 1 predictive + auto orthogonal; permI default 20 is too few for a reliable pQ2 -> >=1000
oplsda <- opls(t(feature_matrix), group, predI = 1, orthoI = NA,
               scaleC = 'pareto', permI = 1000, crossvalI = 7,
               fig.pdfC = 'none', info.txtC = 'none')
summ <- getSummaryDF(oplsda)            # R2X(cum), R2Y(cum), Q2(cum), pre, ort, pR2Y, pQ2
vip_pred <- getVipVn(oplsda)            # predictive VIP (Galindo-Prieto 2014); orthoL=TRUE for orthogonal
# Claim is licensed only if Q2 high AND pQ2 small. R2Y alone proves nothing.

PLS-DA is orthoI = 0. OPLS-DA has identical predictive power to PLS-DA -- it is a coordinate rotation, not a better model; the orthogonal block often encodes a confounder (inspect what correlates with it). DQ2 (Westerhuis 2008b) is the discriminant-appropriate figure of merit when Q2 penalizes correct-side over-predictions.

PLS-DA / OPLS-DA Validation Checklist

1. Report the transformation + scaling used (it changes the loadings, VIPs, and story).
2. Report R2X, R2Y, Q2 and the number of predictive + orthogonal components.
3. Choose the number of components inside CV, not by eye on the training fit.
4. Permutation test (>= 1000) of the full pipeline -> permutation p for Q2 (and R2Y). Permute every step that touched the labels.
5. For honest generalization error use double (cross-model) CV or an untouched test set; single CV that also tuned the model is optimistic.
6. Independent validation cohort for any biomarker claim.
7. Read VIP / S-plot only from a validated model; corroborate with univariate FDR + effect size; report ranking stability across resamples.
8. Put a PCA score plot beside the PLS-DA one -- separation only under supervision is the artifact signature.

Univariate Testing + Correct FDR

Goal: Produce an interpretable, FDR-controlled per-metabolite answer with effect sizes.

Approach: Match the test to the design, compute log2 fold change as a difference of group means on transformed data, then apply BH explicitly (defaults are not BH in either language).

import numpy as np
import pandas as pd
from scipy.stats import ttest_ind
from statsmodels.stats.multitest import multipletests

logged = np.log2(intensities.replace(0, np.nan))   # transform before testing
pvals, lfc = [], []
for feat in logged.index:
    a = logged.loc[feat, case].dropna().values
    b = logged.loc[feat, ctrl].dropna().values
    if len(a) >= 3 and len(b) >= 3:
        pvals.append(ttest_ind(a, b, equal_var=False)[1])   # Welch: scipy defaults to Student
        lfc.append(a.mean() - b.mean())                     # geometric-mean ratio on log scale
    else:
        pvals.append(np.nan); lfc.append(np.nan)
res = pd.DataFrame({'feature': logged.index, 'log2fc': lfc, 'pval': pvals}).dropna(subset=['pval'])
# statsmodels default is 'hs' (Holm-Sidak); R p.adjust default is 'holm' -- ALWAYS pass BH explicitly
res['padj'] = multipletests(res['pval'], method='fdr_bh')[1]

BH controls FDR under independence and PRDS; positively-correlated metabolomics features roughly satisfy PRDS, so BH is valid but conservative -- but closure-induced negative correlations (after total-area/PQN normalization) fall outside the clean case, where a permutation FDR sidesteps the dependence assumptions. The effective number of independent tests is far below the feature count (one compound = many adducts/isotopologues/fragments); use an effective-number-of-tests correction (Peluso 2021) rather than Bonferroni-on-features, and collapse features to compounds before counting "how many metabolites changed."

Volcano Plot

Goal: Show significance and magnitude together for all features.

Approach: Plot log2 fold change vs -log10(p), with the FDR cutoff annotated (raw p on the axis is fine only if the FDR line is drawn).

import matplotlib.pyplot as plt
hit = (res['padj'] < 0.05) & (res['log2fc'].abs() > 1)   # 2-fold + FDR 5%
plt.scatter(res['log2fc'], -np.log10(res['pval']), c=np.where(hit, 'firebrick', 'gray'), s=12, alpha=0.6)
plt.axhline(-np.log10(0.05), ls='--'); plt.axvline(1, ls='--'); plt.axvline(-1, ls='--')
plt.xlabel('log2 fold change'); plt.ylabel('-log10(p)')

Per-Method Failure Modes

Noise separation (the cardinal sin)

• Trigger: Reporting a PLS-DA/OPLS-DA score plot as evidence of a group difference.
• Mechanism: In p>>n any labelling is linearly separable; the algorithm always finds a covariance-maximizing direction, even for random labels.
• Symptom: Clean two-cluster score plot, high R2Y, but Q2 low/negative or inside the permutation cloud; PCA shows no separation.
• Fix: Permutation test (>=1000) of the full pipeline; require small pQ2; put the PCA plot beside it.

VIP misuse

• Trigger: Selecting biomarkers by VIP > 1 from a single model fit.
• Mechanism: VIPs are normalized so the mean squared VIP = 1 -- roughly half the features exceed 1 by construction; there is no null, no p-value, and the ranking is unstable under resampling in p>>n.
• Symptom: Top-20 VIP list reshuffles when the model is re-bootstrapped; VIP-only hits fail to replicate.
• Fix: Use VIP only from a permutation-validated model; require univariate FDR + effect-size concordance and resampling stability; use the OPLS-specific VIP so a high orthogonal-block VIP (the confounder) is not credited to disease.

Naive FDR under correlation

• Trigger: BH or Bonferroni applied as if the features were independent.
• Mechanism: Pathway co-regulation plus adducts/isotopologues/fragments make features strongly correlated; one signal lights up its whole cluster, and closure (after sample-wise normalization) injects negative correlations.
• Symptom: A "200 significant metabolites" list that encodes a handful of independent signals; over-conservative threshold from Bonferroni-on-features.
• Fix: Effective-number-of-tests or permutation FDR (Peluso 2021); collapse features to compounds before counting hits; report independent-signal counts.

Log with zeros / detection-rate confound

• Trigger: Half-min (or zero) imputation followed by log, especially when detection rate differs between groups.
• Mechanism: "Missing" is left-censored (MNAR); a constant imputed at the LOD then logged spikes the censored region, and a detection-rate difference masquerades as a concentration difference.
• Symptom: Fake bimodality; a low-abundance "hit" that is really a difference in how often the metabolite was detected.
• Fix: Report per-group detection rates with any low-abundance hit; prefer glog or a left-censored imputer (QRILC/GSimp) over impute-constant-then-log when detection differs (see metabolomics/normalization-qc).

Quantitative Thresholds

Common Errors

References

• van den Berg RA, Hoefsloot HCJ, Westerhuis JA, Smilde AK, van der Werf MJ. 2006. Centering, scaling, and transformations: improving the biological information content of metabolomics data. BMC Genomics 7:142.
• Westerhuis JA, Hoefsloot HCJ, Smit S, Vis DJ, Smilde AK, et al. 2008. Assessment of PLSDA cross validation. Metabolomics 4:81-89.
• Westerhuis JA, van Velzen EJJ, Hoefsloot HCJ, Smilde AK. 2008. Discriminant Q2 (DQ2) for improved discrimination in PLSDA models. Metabolomics 4:293-296.
• Saccenti E, Hoefsloot HCJ, Smilde AK, Westerhuis JA, Hendriks MMWB. 2014. Reflections on univariate and multivariate analysis of metabolomics data. Metabolomics 10:361-374.
• Broadhurst DI, Kell DB. 2006. Statistical strategies for avoiding false discoveries in metabolomics and related experiments. Metabolomics 2:171-196.
• Thevenot EA, Roux A, Xu Y, Ezan E, Junot C. 2015. Analysis of the human adult urinary metabolome variations with age, body mass index, and gender by implementing a comprehensive workflow for univariate and OPLS statistical analyses. J Proteome Res 14:3322-3335.
• Triba MN, Le Moyec L, Amathieu R, Goossens C, Bouchemal N, et al. 2015. PLS/OPLS models in metabolomics: the impact of permutation of dataset rows on the K-fold cross-validation quality parameters. Mol BioSyst 11:13-19.
• Szymanska E, Saccenti E, Smilde AK, Westerhuis JA. 2012. Double-check: validation of diagnostic statistics for PLS-DA models in metabolomics studies. Metabolomics 8(Suppl 1):3-16.
• Galindo-Prieto B, Eriksson L, Trygg J. 2014. Variable influence on projection (VIP) for orthogonal projections to latent structures (OPLS). J Chemometr 28:623-632.
• Ruiz-Perez D, Guan H, Madhivanan P, Mathee K, Narasimhan G. 2020. So you think you can PLS-DA? BMC Bioinformatics 21(Suppl 1):2.
• Peluso A, Glen R, Ebbels TMD. 2021. Multiple-testing correction in metabolome-wide association studies. BMC Bioinformatics 22:67.
• Storey JD, Tibshirani R. 2003. Statistical significance for genomewide studies. Proc Natl Acad Sci USA 100:9440-9445.

Related Skills

• metabolomics/normalization-qc - Sample-wise normalization, drift/batch correction, missing-value imputation upstream of testing
• metabolomics/pathway-mapping - Functional interpretation of differential metabolites
• machine-learning/biomarker-discovery - Feature selection inside CV, stability, minimal-optimal vs all-relevant
• machine-learning/model-validation - Leakage taxonomy, nested CV, calibration vs discrimination
• experimental-design/multiple-testing - FDR vs FWER regime, discovery vs confirmatory
• data-visualization/volcano-and-ma-plots - Volcano plot recipes