Metabolite Phenotype Association Partial Correlation

metabolite-phenotype-association-partial-correlation

Summary

Compute partial Spearman correlations between metabolic features and a phenotype of interest while simultaneously adjusting for epidemiological confounders (age, gender, disease status). This generates metabolome-wide association estimates, raw p-values, and multiple-testing corrected p-values suitable for hypothesis testing in metabolomic cohorts.

When to use

You have a SummarizedExperiment object containing NMR or MS metabolomic data with aligned phenotype information (BMI, disease status, age, gender), and you need to identify metabolites associated with a continuous or categorical outcome while controlling for known confounders that might distort the observed associations.

When NOT to use

• You have already computed univariate metabolite associations without confounder adjustment; partial correlation adds computational cost when confounding is negligible.
• Your phenotype data is missing confounder values for a large fraction of samples; partial correlation requires complete cases for all variables.
• Your outcome is binary and sample counts are extremely imbalanced; generalized linear models (logistic regression) may be more appropriate than Spearman correlation.

Inputs

• SummarizedExperiment object with metabolomic data (rows=metabolic features, columns=samples)
• Phenotype data frame with outcome variable and confounder columns (age, gender, disease status)
• Cross-validation filter metadata (CV assignments)

Outputs

• MWAS_BMI matrix: 3-column data frame with metabolite identifiers (ppm) as row names
• Column 1: r (Spearman partial correlation coefficient)
• Column 2: p (raw p-value from partial correlation test)
• Column 3: pFDR (Benjamini-Hochberg corrected p-value)

How to apply

Load the cross-validated metabolomic SummarizedExperiment object and ensure the phenotype slot contains the outcome variable (e.g., BMI) and all confounders (age, gender, T2D status). Apply the MWAS_stats function from MWASTools with disease_id set to your outcome, specify confounders as a list, and set correlation method to 'Spearman'. The function computes partial correlations for each metabolite against the outcome, adjusting for confounders simultaneously, then calculates raw p-values from the partial correlation test statistics. Apply Benjamini-Hochberg correction to generate false discovery rate–adjusted p-values (pFDR). Assemble results into a three-column matrix with metabolite identifiers (ppm values) as row names and columns for correlation coefficient (r), raw p-value (p), and BH-corrected p-value (pFDR).

Related tools

• MWASTools (R package providing MWAS_stats function to compute partial correlations and p-values with confounder adjustment) — github.com/AndreaRMICL/MWASTools
• R (Statistical computing environment; required version ≥3.3)
• Bioconductor (Infrastructure for SummarizedExperiment object classes and statistical methods)

Examples

MWAS_stats(metabo_SE, disease_id='BMI', confounder=c('Age','Gender','T2D'), corr_method='Spearman')

Evaluation signals

• Output matrix has exactly 3 columns (r, p, pFDR) with row names matching metabolite identifiers from input SummarizedExperiment
• Raw p-values are uniformly distributed on [0,1] under null hypothesis; pFDR values are ≥ corresponding raw p-values
• Number of rows equals number of metabolic features in input; no metabolites are dropped unless filtered by prior QC
• pFDR values satisfy monotonicity: features ranked by raw p-value also rank identically by pFDR
• Partial correlation coefficients r lie in [-1, 1] range; magnitude should reflect biological signal after confounder removal

Limitations

• Assumes linearity of relationships between metabolites, outcome, and confounders; non-linear associations may be missed.
• Spearman correlation is non-parametric but less efficient than Pearson if data are jointly normal; choose based on NMR spectral preprocessing and normality assessment.
• Multiple testing correction (Benjamini-Hochberg) controls false discovery rate but not family-wise error; at FDR threshold α=0.05 expect ~5% of significant findings to be false positives.
• Requires complete case deletion if confounders are missing; large amounts of missing confounder data may substantially reduce effective sample size.
• Confounder specification is user-determined; omitting true confounders or including colliders can bias estimates or inflate type I error.

Evidence

• [other] metabolite-phenotype-association-partial-correlation via partial Spearman correlations: "Reproduce the BMI metabolome-wide association matrix (MWAS_BMI) via partial Spearman correlations"
• [intro] confounder adjustment in MWAS: "a major limitation of these multivariate models from the epidemiological perspective is that they do not properly account for cofounding factors (e.g. age, gender), which might distort the observed"
• [intro] partial correlation and GLM methods: "metabolite-phenotype association models (partial correlations, generalized linear models) adjusted for epidemiological confounders"
• [intro] MWASTools MWAS_stats function: "we present a package to perform MWAS using univariate hypothesis testing with efficient handling of epidemiological confounders"
• [other] output format: r, p, pFDR columns: "MWAS_BMI is a 3-column matrix containing metabolic features (ppm values) in rows with columns for estimates (r coefficients), raw p-values, and BH-corrected p-values (pFDR)"
• [other] Benjamini-Hochberg multiple testing correction: "Apply Benjamini-Hochberg correction to generate adjusted p-values (pFDR)"