HolobiomicsLab Multivariate Statistical Analysis of Feature-Based Molecular Networks from Non-Targeted Metabolomics
Summary
Apply multivariate statistical analysis methods (e.g., PCA, PLS-DA, clustering) to feature tables derived from non-targeted LC-MS/MS data to identify patterns, classify samples, and associate metabolomic signatures with experimental conditions or phenotypes in feature-based molecular networks.
When to use
When you have preprocessed non-targeted LC-MS/MS feature tables (post-merging, cleanup, blank removal, and batch correction) and seek to uncover multivariate patterns across samples, discriminate between experimental groups, or reduce dimensionality of high-dimensional metabolomic data before downstream interpretation.
When NOT to use
- Input feature table has not undergone data cleanup, blank removal, or batch correction; apply those preprocessing steps first.
- Sample size is extremely small (n < 3 per group); multivariate methods may overfit or lack statistical power.
- Feature table contains raw, non-log-transformed abundances without appropriate normalization; preprocessing must precede multivariate analysis.
Inputs
- Preprocessed feature table (TSV, CSV, or Excel format) with features as columns and samples as rows
- Sample metadata or design matrix indicating experimental groups or conditions
- Cleaned, batch-corrected LC-MS/MS feature abundance data
Outputs
- PCA scores and loadings plots
- PLS-DA classification model and predictions
- Heatmaps of clustered features and samples
- Statistical summary tables (e.g., model performance, feature contributions)
- R objects (.RData) or Python pickled objects for downstream analysis
How to apply
Load the cleaned and batch-corrected feature table (typically in tabular format with features as columns and samples as rows) into a Jupyter Notebook environment (R or Python) from the FBMN-STATS repository. Execute the multivariate statistical analysis notebook, which automates data scaling, dimensionality reduction (PCA), supervised classification (PLS-DA), and clustering. The notebook applies standard preprocessing steps (e.g., log transformation, normalization) and generates publication-ready visualizations (scores plots, loadings, heatmaps) and statistical outputs (model performance metrics). Verify that all cells run without errors and that output files are generated in the working directory.
Related tools
- Jupyter Notebook (Interactive environment for executing R or Python multivariate statistical analysis workflows on feature tables) — https://github.com/Functional-Metabolomics-Lab/FBMN-STATS
- R (Statistical computing language used in FBMN-STATS notebooks for PCA, PLS-DA, and clustering analysis) — https://github.com/Functional-Metabolomics-Lab/FBMN-STATS
- Google Colab (Cloud-based alternative environment for running FBMN-STATS notebooks without local R installation) — https://colab.research.google.com/github/Functional-Metabolomics-Lab/FBMN-STATS/blob/main/R/Stats_Untargeted_Metabolomics.ipynb
- MZmine3 (Preprocessing tool for generating feature tables from raw LC-MS/MS data prior to multivariate analysis)
Evaluation signals
- All notebook cells execute without errors and complete in reasonable time (<30 min for test datasets MSV000082312 or MSV000085786).
- Output files (plots, tables, model objects) are generated in the working directory and match the structure and content of reference files in the associated Google Drive.
- PCA scores plot shows clear separation or clustering consistent with experimental design; PLS-DA model shows non-trivial cross-validation accuracy (>60% for multi-class, >70% for binary).
- Generated visualizations (heatmaps, loadings plots) are readable and reproducible across independent runs with the same input data.
- Dimensions of output matrices and feature contributions are consistent with input feature table dimensions (e.g., loadings matrix = features × n_components).
Limitations
- Google Colab runtime automatically disconnects after 90 minutes of idleness or 12 hours of continuous use, requiring notebook re-execution and redownload of results.
- Colab disk space is limited to 77 GB, which may constrain analysis of very large datasets or high-dimensional feature tables.
- Quickstart GNPS (older version) does not generate reformatted output compatible with the notebook; users must use GNPS 2 to ensure correct feature table format.
- When copying notebook code to other environments (e.g., RStudio), GitHub rendering limitations may affect symbol transfer; use Google Colab or locally opened Jupyter for accurate code transfer.
- Package installation in Colab must be performed every session (unlike local Jupyter), which adds runtime overhead.
Evidence
- [readme] perform data merging, data cleanup, blank removal, batch correction, and univariate and multivariate statistical analyses on their non-targeted LC-MS/MS data: "perform data merging, data cleanup, blank removal, batch correction, and univariate and multivariate statistical analyses on their non-targeted LC-MS/MS data"
- [other] Open and execute the multivariate statistical analysis notebook in Jupyter, ensuring all cells run without errors.: "Open and execute the multivariate statistical analysis notebook in Jupyter, ensuring all cells run without errors."
- [other] Verify that all expected output files are generated and match the reported results stored in the project Google Drive.: "Verify that all expected output files are generated and match the reported results stored in the project Google Drive."
- [readme] Since Colab does not come pre-installed with R packages (or libraries) when running our R Notebook in Colab, we need to install the packages every time we run the notebook: "Since Colab does not come pre-installed with R packages (or libraries) when running our R Notebook in Colab, we need to install the packages every time we run the notebook"
- [readme] Although Colab is easier to use and is all Cloud-based, the main problem with the Colab environment is when you leave the Colab notebook idle for 90 mins or continuously used it for 12 hours, the runtime will automatically disconnect.: "when you leave the Colab notebook idle for 90 mins or continuously used it for 12 hours, the runtime will automatically disconnect"
- [readme] We advise Quickstart GNPS users to switch to the latest GNPS 2 for FBMN-STATS and accessing the Notebooks. The previous version of Quickstart GNPS does not generate the reformatted output needed for Notebook/Web app integration: "The previous version of Quickstart GNPS does not generate the reformatted output needed for Notebook/Web app integration"