bio-metabolomics-statistical-analysis

What it does

The skill directs an agent to perform statistical analysis on metabolomics data, including preprocessing (log2 transformation, normalization), differential testing with limma moderated statistics, Welch's t-tests with BH correction, fold-change estimation on log2-transformed, normalized data, and multivariate approaches (PCA, PLS-DA, OPLS-DA) for identifying differentially abundant metabolites or building classification models.

How it works

• Applies a processing pipeline: zero/missing handling, log2 transformation, normalization, then statistical testing.
• For limma: constructs a design matrix from sample info, fits linear models, uses contrasts, applies empirical Bayes moderation with trend and robust options, and outputs BH-adjusted results with columns including logFC, AveExpr, t, P.Value, adj.P.Val, B. Includes guidance to add batch as a covariate in the design matrix when needed.
• For Python-based analysis: performs per-feature Welch's t-tests on log2-transformed data, then applies BH correction using multipletests with method='fdr_bh'; fold changes are computed as the difference of group means in log2 space.
• Fold change calculation emphasizes using log2-transformed, unscaled data, with notes on geometric means (difference of log means) versus arithmetic means, and provides guidance on reporting FC and applying shrinkage via ashr in R.
• Provides examples for normalization methods, including PQN normalization, and cautions about when normalization may hurt due to global shifts.
• Includes steps for creating volcano plots, PCA, and initial steps for supervised methods like PLS-DA with cross-validation and VIP scoring (in the truncated portion).

When to use it

Use when identifying differentially abundant metabolites or when building classification models from metabolomics data, particularly when you need a structured pipeline combining preprocessing, differential testing, and multivariate analyses.

What it can touch

• Tools: limma (R) for moderated testing; Python components rely on numpy, pandas, scipy.stats.ttest_ind, and statsmodels.stats.multitest.tesults for BH correction. The skill references file-based inputs like 'feature_table.tsv' and 'sample_info.tsv' and outputs results in data frames with columns such as log2fc and padj.
• Scripts present in the skill folder include executable components that implement these patterns.

Caveats

• Emphasizes that PQN normalization assumes most features are not differentially abundant; if global metabolic shifts occur (>50% of features), normalization may be inappropriate.
• Warns against imputing BPMV zeros (true absences) with half-minimum, recommending two-part tests instead.
• Notes that MetaboAnalyst uses different defaults (Student's t-test and arithmetic mean FC); users should override to align with this workflow.
• Highlights that post-hoc FC filtering can inflate FDR and recommends using treat() + topTreat() for minimum-effect testing.
• Contains explicit guidance about not pre-removing batch effects with removeBatchEffect(); instead, include batch as a covariate in the design.