Metabolomics Data Analysis Service — From Raw Spectra to Biological Insight

All major workflows: targeted LC-MS/MS (MRM), untargeted LC-MS (DDA/DIA), GC-MS, CE-MS, and NMR. Raw files from SCIEX, Thermo, Agilent, Waters, Bruker; processed peak tables (.csv, .txt, .xlsx); public datasets from MetaboLights, Metabolomics Workbench, and GNPS/MassIVE. Data from Creative Proteomics, your own lab, or any third party — pipeline is identical regardless of source.

At every stage: (1) QC-based feature filtration — RSD >30% in pooled QCs flagged before testing; (2) FDR correction (Benjamini-Hochberg) on all univariate tests; (3) permutation testing (n≥1,000) for PLS-DA/OPLS-DA — model only considered valid if Q²Y intercept <0.05; (4) k-fold cross-validation + independent test set for machine learning; (5) pathway enrichment FDR correction. Every threshold documented in your methods report.

Yes, fully platform-agnostic. Submit data from your instrument, a collaborator's lab, any CRO, or a public repository. Analysis quality is identical. Need data generation too? Our metabolomics platform provides complete sample-to-report workflows with a single contact for chemistry and bioinformatics.

Standard pipeline (preprocessing + statistics + pathway enrichment + figures): 2–4 weeks. Projects with machine learning, multi-omics integration, or custom development: 4–6 weeks. Expedited analysis available for manuscript revisions (1–2 weeks). We provide a detailed timeline based on your dataset size and scope during initial consultation.

Send us: (1) your data files organized by batch; (2) a sample metadata table (Sample ID, Group, Batch, Injection Order, Sample Type, covariates); (3) experimental design description (groups, pairing, replicates, hypotheses). Large datasets (≥50 GB): secure FTP available. We confirm data integrity within 2 business days.

Complete analysis package: Statistical Report (all univariate + multivariate results, model metrics); Publication-Ready Figures (PCA, PLS-DA, volcano, heatmap, pathway enrichment, box plots — 300 DPI TIFF + vector PDF/AI); Metabolite Annotation Table (MSI Level 1–4, HMDB/METLIN/KEGG IDs); Pathway Enrichment Results with KEGG maps and interpretation; Methods Documentation (ready for manuscript); Reproducible Code (R Markdown/Jupyter); Processed Data Tables (Excel + CSV).

Yes. Modules are available individually — if you have already done preprocessing and need only pathway enrichment, or have statistics done and need multi-omics integration, we pick up where you left off. We also review existing analyses before building on them to ensure the upstream work is sound.

Differential metabolites are identified using combined thresholds from univariate and multivariate statistics. The standard approach applies VIP > 1 (from PLS-DA/OPLS-DA) AND P < 0.05 (from t-test or ANOVA). For stricter screening: raise the VIP threshold (e.g., VIP > 2) or lower the P-value cutoff (e.g., P < 0.01). Fold-change (FC) can also be incorporated — |log₂FC| > 1 (i.e., FC > 2 or FC < 0.5) combined with VIP and P-value thresholds. When these criteria are too stringent for your dataset, we can widen thresholds (e.g., VIP > 1 and P < 0.1) or apply FC-only screening (FC > 1.5, P < 0.05). All thresholds are documented in your methods report for transparency and reproducibility.

This is common in datasets with high biological variability or subtle treatment effects. First, we relax the thresholds — e.g., VIP > 1 and P < 0.1 instead of P < 0.05. If still no hits, we apply univariate-only screening (FC > 1.5 or FC < 0.67, P < 0.05) and visualize results via volcano plot. We also assess whether the experimental design provides adequate statistical power — if group sizes are small (n < 5) or within-group variability is high, non-parametric tests (Mann-Whitney, Kruskal-Wallis) may be more appropriate than parametric tests. In all cases, we transparently report which criteria were applied and why.

A volcano plot displays two statistical dimensions simultaneously: log₂(fold-change) on the x-axis (magnitude of change between groups) and −log₁₀(P-value or FDR) on the y-axis (statistical significance). Metabolites with large fold-changes AND high significance appear in the upper-left (down-regulated) or upper-right (up-regulated) corners. Threshold lines (typically |log₂FC| > 0.585 and FDR < 0.05) divide the plot — points above both thresholds are considered significantly differential. The volcano plot is the central figure for differential analysis because it shows both the biological effect size (FC) and the statistical confidence (P-value) for every detected metabolite in a single view.

PLS-DA (Partial Least Squares Discriminant Analysis) finds latent variables that maximize covariance between the metabolite matrix (X) and group membership (Y) — effective for identifying which metabolites drive group separation. OPLS-DA (Orthogonal PLS-DA) goes further: it separates systematic variation into components that are predictive of group membership and components that are orthogonal (within-group variability, batch effects, noise). This filtering makes OPLS-DA VIP scores more specific to group differences and less influenced by within-group variability — the preferred method when within-group variation is large relative to between-group differences. Both models require permutation testing to validate that the observed separation is not due to chance.

R² (R²X for PCA, R²Y for PLS-DA/OPLS-DA) represents the proportion of variance in the data explained by the model — the goodness of fit. A higher R² means the model captures more of the data's structure. Q² represents the proportion of variance that can be predicted by the model, estimated through cross-validation — a measure of predictive ability. Key diagnostic: if R² is high but Q² is low (< 0.05 or negative), the model is overfitting (fitting noise rather than signal). For a valid model, Q² should be substantially greater than zero with a modest R²-to-Q² gap. Permutation testing provides an independent validation — a valid PLS-DA/OPLS-DA model should have the permuted Q² intercept below 0.05.

Permutation testing validates whether observed group separation could occur by random chance. Group labels are randomly shuffled many times (n ≥ 1,000 iterations), and the model is re-fit to each permuted dataset, generating a null distribution of R² and Q² values. Acceptance criteria: The permuted Q² intercept should be below 0.05 (negative values are ideal); the permuted R² intercept should generally be below 0.3–0.4; all permuted Q² values should be lower than the original model's Q²; and the regression line of permuted Q² values should have a positive slope. If these criteria are not met, the model's group discrimination is not statistically reliable and should not be interpreted.

Differential analysis determines whether a metabolite's abundance differs between groups — this is inherently a pairwise comparison. A metabolite is classified as increased or decreased relative to a specific reference group; the same metabolite could be up-regulated versus Group A, unchanged versus Group B, and down-regulated versus Group C. Multi-group tests (ANOVA, Kruskal-Wallis) can tell you that groups differ somewhere, but cannot identify which specific groups differ — pairwise post-hoc testing is required. For multi-group studies, we perform all pairwise comparisons of interest and apply FDR correction across the full set of tests to maintain false discovery control.

This feature-to-metabolite gap is normal. Three main reasons: (1) One metabolite generates multiple ion forms — protonated [M+H]⁺, deprotonated [M−H]⁻, sodium adducts [M+Na]⁺, ammonium adducts, isotopic peaks (¹³C, ³⁴S), dimers, and in-source fragments — a single metabolite may produce 5–15 detectable features. (2) Database coverage — MS/MS spectral libraries (HMDB, METLIN, MassBank) contain validated spectra for a fraction of the known metabolome; features without high-quality spectral matches remain unidentified. (3) Background signals — column bleed, plasticizers, and solvent contaminants produce non-metabolite features. Our pipeline applies strict criteria — retention time filtering, isotopic pattern scoring, adduct deconvolution, and MS/MS spectral matching (MSI Levels 1–4) — to maximize confident identifications while controlling false positives.