Metabolomics Data Analysis in R: A Practical Workflow

Metabolomics data analysis in R requires a structured pipeline that moves from raw instrument output to cleaned data matrices, statistical testing, and biological interpretation. This workflow is written for students, researchers, analysts, and life-science professionals who need concrete decisions at each step instead of abstract descriptions of available methods. The practical outcome is an R script template that you can adapt to your own datasets, with clear checkpoints for quality control, normalization choices, multivariate modeling, and visualization.

At a Glance

The table below summarizes the main stages of an R-based metabolomics workflow, the key decisions at each stage, and the common tools or packages you will encounter.

Understanding the Data Landscape

Metabolomics datasets arrive in several forms depending on the analytical platform. Nuclear magnetic resonance spectroscopy produces spectra that require binning or peak integration before statistical analysis. Liquid chromatography-mass spectrometry and gas chromatography-mass spectrometry produce raw files that must go through peak detection, deconvolution, and alignment before you have a feature table. Compound concentration tables from targeted assays may already be in a format suitable for direct import into R.

The choice of preprocessing strategy depends heavily on your data type. For LC-MS and GC-MS data, the informatics pipeline starts with preprocessing raw data, and the results of that preprocessing feed into statistical analysis and then pathway mapping for biological context [19]. The xcms R package has been one of the most widely used tools for LC-MS data preprocessing since its introduction in 2005, and recent developments have positioned it as a central component of a modular software ecosystem for metabolomics analysis [8]. These developments include enhanced scalability that enables processing of large-scale experiments with thousands of samples on standard computing hardware [8].

For researchers who prefer a graphical interface, MZmine 2 provides a cross-platform framework that incorporates the ADAP suite of computational algorithms for both LC-MS and GC-MS data [19]. The ADAP workflows handle construction of extracted ion chromatograms, detection of chromatographic peaks, spectral deconvolution, and alignment [19]. You can export the resulting feature table from MZmine and import it into R for downstream analysis.

Setting Up Your R Environment

Before you begin any analysis, establish a reproducible environment. Create a project directory with subfolders for raw data, processed data, scripts, figures, and results. Record the R version and package versions you use, because metabolomics packages evolve rapidly and results can change between versions.

Install the core packages you will need. The Bioconductor repository hosts many metabolomics packages, including xcms for preprocessing and CAMERA or annotation packages for feature annotation. The Comprehensive R Archive Network hosts packages such as MetaboDiff for differential analysis and GSimp for missing value imputation [23]. You can install packages from CRAN with the install.packages function and from Bioconductor with the BiocManager::install function.

Write your analysis as a script file instead of running commands interactively. This practice ensures that you can rerun the entire pipeline when you receive updated data or when you need to audit your methods. Include comments that explain why you made each decision, beyond what the code does.

Data Import and Initial Inspection

The first step is loading your feature table into R. A typical feature table has rows representing metabolites or features and columns representing samples. You will also need a metadata table that describes each sample, including group assignments, batch information, and any covariates such as age, sex, or treatment dose.

Use the read.csv or readr::read_csv function to import your data. Check the structure of the imported data with str and head. Verify that sample names in the feature table match sample names in the metadata table. Mismatches here are a common source of errors that are difficult to detect later in the pipeline.

Inspect the dimensions of your data. Untargeted metabolomics experiments can produce thousands of features. One study of resveratrol-treated breast cancer cells reported 4751 peaks after data preprocessing and filtering, with 312 peaks matched to an in-house standards library and 3459 peaks matched to public databases [22]. These numbers illustrate the scale of data you may need to manage.

Check for missing values early. Missing values in metabolomics data are often left-censored, meaning they fall below the limit of detection instead of being randomly absent. The GSimp package implements a Gibbs sampler based approach for left-censored missing value imputation [23]. Understanding the pattern of missingness in your data will guide your imputation strategy.

Preprocessing and Quality Control

Preprocessing transforms raw instrument data into a clean feature table suitable for statistical analysis. High-quality data preprocessing is essential for untargeted metabolomics experiments, where increasing dataset scale and complexity demand adaptable, robust, and reproducible software solutions [8].

Peak Detection and Alignment

If you are working with LC-MS or GC-MS raw data, you will need to perform peak detection and alignment. The xcms package provides functions for this purpose, including centWave for peak detection and peakGroups or obiwarp for retention time correction. The ADAP suite integrated into MZmine 2 offers an alternative workflow with graphical user interfaces [19].

Set your parameters based on your instrument and chromatography. Key parameters include the mass accuracy of your instrument, the expected peak width in seconds, and the signal-to-noise threshold. Document these parameters in your script so that another researcher can reproduce your preprocessing.

After peak detection, align features across samples. Retention time drift is common in LC-MS experiments, and alignment corrects for this drift so that the same metabolite is represented by the same feature across all samples. The output of this step is a feature table with rows for each feature and columns for each sample, where each cell contains the integrated peak area or intensity.

Quality Control Samples

Include quality control samples in your experimental design. These are typically pooled samples that represent the average composition of all study samples. Quality control samples are injected periodically throughout the analytical run to monitor instrument performance.

Plot the total ion current or base peak intensity for quality control samples over the course of the run. Drift in these values indicates instrument instability that may require correction. The TidyMS package for Python demonstrates quality control procedures for LC-MS data, including system suitability checks, signal drift evaluation, and data curation [18]. While TidyMS is a Python tool, the principles it applies are directly relevant to R-based workflows.

Feature Filtering

Not all detected features are biologically meaningful. Filter out features with a high proportion of missing values, features with low signal-to-noise ratios, and features that show excessive variation in quality control samples. A common approach is to retain features that are present in at least 80 percent of samples in at least one experimental group.

Filtering reduces the number of statistical tests you will perform, which reduces the multiple testing burden. Filtering also removes features that are likely to be noise or artifacts, improving the reliability of downstream analysis.

Normalization and Scaling

Normalization corrects for systematic variation between samples that is not biologically meaningful. Sources of such variation include differences in sample extraction efficiency, instrument sensitivity over time, and injection volume differences.

Normalization Methods

Total ion current normalization divides each feature value by the sum of all feature values in that sample. This method assumes that the total amount of detected metabolites is similar across samples. Median normalization divides each feature value by the median intensity of that sample. Quantile normalization forces the distribution of intensities to be identical across all samples.

The choice of normalization method depends on your data and experimental design. A comprehensive evaluation of metabolomics data preprocessing methods for deep learning applications found that the choice of preprocessing method affects downstream model performance [17]. The authors summarized their results into a set of best practices that researchers can use as a starting point for classification and reconstruction tasks [17].

For targeted metabolomics data, the MetaboDiff package provides a low-level entry to differential analysis that starts with the table of metabolite measurements [12]. This package is designed for translational researchers who may not have extensive mass spectrometry experience [12].

Scaling Methods

After normalization, you may need to scale your data before statistical analysis. Scaling adjusts the variance of each feature so that features with high intensity do not dominate the analysis.

Autoscaling, also known as unit variance scaling, subtracts the mean and divides by the standard deviation for each feature. This gives every feature equal weight in the analysis. Pareto scaling divides by the square root of the standard deviation, which reduces the influence of large fold changes while preserving some information about magnitude. Range scaling divides by the range of values for each feature.

The choice of scaling method affects the results of multivariate analysis. Autoscaling is appropriate when you want all features to contribute equally. Pareto scaling is useful when you want to retain some information about the magnitude of differences while reducing the dominance of high-intensity features.

Missing Value Imputation

Missing values are a persistent challenge in metabolomics data. They can arise from features that are absent in some samples, features below the limit of detection, or technical failures during data acquisition.

Understanding Missingness Patterns

Examine the pattern of missing values in your data before choosing an imputation method. Features that are missing in all samples of one group but present in another group may represent biologically meaningful differences. Features that are missing randomly across all samples may represent technical noise.

Left-censored missing values occur when the metabolite concentration is below the limit of detection. These values are not truly missing but are instead known to be below a threshold. The GSimp package implements a Gibbs sampler based approach specifically designed for left-censored missing value imputation in metabolomics studies [23].

Imputation Methods

Common imputation methods include replacing missing values with half the minimum detected value for that feature, using the k-nearest neighbors algorithm to estimate missing values from similar samples, and using model-based approaches such as GSimp [23].

The choice of imputation method can substantially affect downstream results. For features with a high proportion of missing values, consider excluding the feature entirely instead of imputing. For features with a low proportion of missing values, imputation is generally safe.

Record the imputation method you use and the proportion of missing values in your data. This information is important for reproducibility and for interpreting the reliability of your results.

Univariate Statistical Analysis

Univariate analysis examines each metabolite individually to identify features that differ between experimental groups. This approach is straightforward to interpret and provides a list of candidate biomarkers.

Choosing the Appropriate Test

For two-group comparisons, use a Student t-test if the data are approximately normally distributed and variances are similar between groups. Use a Welch t-test if variances differ. Use a Mann-Whitney U test if the data are not normally distributed.

For multi-group comparisons, use ANOVA if the data meet the assumptions of normality and homogeneity of variance. Use the Kruskal-Wallis test as a nonparametric alternative.

A study of serum and tissue NMR metabolomics in gliomas used univariate and multivariate statistical analysis to identify metabolites that differentiated control from disease groups [14]. Eight metabolites were statistically significant in differentiating control from disease groups in serum samples, and five metabolites were significant in ANOVA across different tissue grades [14].

Multiple Testing Correction

When you test thousands of features, some will appear significant by chance alone. Multiple testing correction controls the rate of false positives.

The Bonferroni correction is the most conservative approach and controls the family-wise error rate. The Benjamini-Hochberg procedure controls the false discovery rate and is less conservative, making it more powerful for metabolomics data where you expect many true positives.

Report both the raw p-values and the adjusted p-values in your results. Use the adjusted p-values to determine significance.

Effect Size and Fold Change

Statistical significance does not necessarily imply biological importance. A metabolite can show a statistically significant difference with a very small effect size. Report effect sizes such as fold change or Cohen d alongside p-values.

Fold change is calculated as the ratio of the mean intensity in one group to the mean intensity in another group. A fold change threshold of 1.5 or 2.0 is commonly used to identify metabolites with meaningful differences, but the appropriate threshold depends on your biological question and the technical variability of your assay.

Multivariate Statistical Analysis

Multivariate analysis examines all metabolites simultaneously to identify patterns that distinguish experimental groups. This approach accounts for correlations between metabolites and can reveal differences that are not apparent in univariate analysis.

Principal Component Analysis

Principal component analysis is an unsupervised method that reduces the dimensionality of your data while preserving the maximum amount of variation. The first principal component captures the largest amount of variation, the second captures the next largest, and so on.

Plot the scores of the first two or three principal components to visualize the overall structure of your data. Samples that cluster together are similar in their metabolic profiles. Samples that separate by experimental group suggest that the groups have distinct metabolic signatures.

Principal component analysis is also useful for detecting outliers and batch effects. Samples that fall far from the main cluster may represent technical failures or mislabeled samples.

Partial Least Squares Discriminant Analysis

Partial least squares discriminant analysis is a supervised method that finds the linear combinations of features that best separate known groups. This method is commonly used in metabolomics because it can achieve good separation even when the number of features greatly exceeds the number of samples.

A study of lactic acid bacteria in food fermentation used PCA and PLS-DA to identify differential metabolites and perturbed pathways [15]. The PLS-DA model delivered robust group discrimination with R2X of 0.89, R2Y of 0.95, and Q2 of 0.91 [15]. These metrics indicate that the model explained a high proportion of the variation and had good predictive ability.

Model Validation

Supervised methods such as PLS-DA can overfit the data, meaning the model performs well on the training data but poorly on new data. Cross-validation assesses the predictive ability of your model by holding out a portion of the samples during model fitting and then testing the model on the held-out samples.

Report the Q2 value from cross-validation, which estimates the predictive ability of the model. A Q2 value above 0.5 is generally considered good, while a Q2 value near zero or negative indicates that the model has no predictive ability.

Permutation testing provides an additional validation step. In permutation testing, the group labels are randomly shuffled and the model is refit many times. The performance of the model on the permuted data is compared to the performance on the real data. If the real model performs much better than the permuted models, the separation is likely to be real instead of an artifact of overfitting.

Visualization Strategies

Visualization is essential for exploring your data, communicating your results, and identifying patterns that statistical tests may miss.

Principal Component Analysis Scores Plot

The PCA scores plot is the most common visualization in metabolomics. Each point represents a sample, and the position of the point reflects the metabolic profile of that sample. Color the points by experimental group to visualize separation between groups.

Add ellipses to the plot to show the confidence region for each group. Ellipses that do not overlap suggest that the groups are metabolically distinct.

Heatmaps

Heatmaps display the intensity of each feature across all samples. Rows represent features and columns represent samples. The color of each cell represents the intensity of that feature in that sample.

Cluster the rows and columns to reveal patterns in the data. Features that cluster together may be co-regulated or may belong to the same metabolic pathway. Samples that cluster together are similar in their metabolic profiles.

The xOmicsShiny application offers heatmap visualization along with PCA, volcano plots, Venn diagrams, WGCNA, and advanced clustering analyses [7]. While xOmicsShiny is a web-based application, the visualization principles it implements are directly applicable to R-based workflows.

Volcano Plots

Volcano plots combine fold change and statistical significance in a single visualization. The x-axis represents the log2 fold change and the y-axis represents the negative log10 p-value. Features in the upper left and upper right corners of the plot have both large fold changes and high statistical significance.

Set thresholds for fold change and p-value to highlight features of interest. Features that pass both thresholds are typically considered candidate biomarkers.

Pathway Visualization

Mapping significant metabolites to metabolic pathways provides biological context for your results. The xOmicsShiny application covers a broad range of pathway databases, including WikiPathways, Reactome, and KEGG pathways [7]. Pathway mapping can reveal which biological processes are most affected by your experimental condition.

A study of resveratrol-treated breast cancer cells used pathway analysis to identify significant metabolic pathways affected by treatment, particularly those involving steroid, fatty acid, amino acid, and nucleotide metabolism [22]. This type of analysis moves beyond listing significant metabolites to understanding the biological implications of your results.

Annotation and Identification

Annotation assigns chemical identities to the features in your dataset. This step is critical for biological interpretation but remains a major bottleneck in untargeted metabolomics.

Annotation Approaches

The most reliable annotation approach matches features to an in-house standards library using retention time and mass spectrometry data. When standards are not available, match features to public databases using accurate mass and predicted retention time.

Poor chemical annotation of high-resolution mass spectrometry data limits the applications of untargeted metabolomics datasets [9]. The IDSL.CSA R package generates composite spectra libraries from MS1-only data, enabling chemical annotation of high-resolution mass spectrometry coupled with liquid chromatography peaks regardless of the availability of MS2 fragmentation spectra [9]. This approach demonstrated comparable annotation rates for commonly detected endogenous metabolites in human blood samples when compared to MS/MS libraries in validation tests [9].

Reporting Annotation Confidence

Report the level of confidence for each annotation. Metabolomics Standards Initiative levels range from level 1, where the identification is confirmed with a reference standard, to level 4, where the feature is unidentified but reproducible.

Be transparent about the limitations of your annotations. Features that are annotated based on accurate mass alone should be clearly distinguished from features that are confirmed with standards.

Reproducibility and Documentation

Reproducibility is a core requirement for metabolomics research. Another researcher should be able to take your raw data and your analysis script and reproduce your results.

Version Control

Record the versions of R and all packages used in your analysis. The sessionInfo function in R provides this information. Save the output of sessionInfo to a text file and include it with your analysis.

Use a package management tool such as renv to create a project-specific library of packages. This ensures that your analysis uses the same package versions even if the packages are updated on CRAN or Bioconductor.

Script Organization

Organize your analysis into sections that correspond to the workflow stages described in this article. Use comments to explain the purpose of each section and the rationale for each decision.

Save intermediate outputs at each stage of the pipeline. If you need to change a parameter in the preprocessing step, you should not need to rerun the entire pipeline from the beginning.

Data Sharing

Consider sharing your data and analysis code when you publish your results. The FAIR Guiding Principles provide a framework for making data findable, accessible, interoperable, and reusable [4]. Following these principles increases the impact of your research and enables others to build on your work.

The National Institutes of Health Genomic Data Sharing Policy provides guidance on data sharing for genomic research [3]. While this policy is specific to genomic data, the principles of responsible data sharing apply to metabolomics data as well.

Common Failure Patterns

Several recurring problems can compromise metabolomics analysis in R. Recognizing these patterns early can save substantial time and prevent incorrect conclusions.

Sample Name Mismatches

The most common and most damaging error is a mismatch between sample names in the feature table and sample names in the metadata table. This error can silently reorder your data, causing all downstream results to be wrong. Always verify that sample names match exactly before proceeding with any analysis.

Inappropriate Normalization

Applying the wrong normalization method can introduce artifacts or obscure real biological differences. Total ion current normalization fails when samples have substantially different total metabolite concentrations. Quantile normalization can distort the relationships between features.

Overfitting in Multivariate Models

PLS-DA models can achieve perfect separation on training data even when there is no real difference between groups. Always validate supervised models with cross-validation and permutation testing. Report the Q2 value and the results of permutation testing.

Ignoring Batch Effects

Samples analyzed on different days or in different batches can show systematic differences that are unrelated to biology. If your samples are spread across multiple batches, include batch as a covariate in your analysis or use a batch correction method.

Incomplete Documentation

Failing to document preprocessing parameters, imputation methods, and statistical thresholds makes your analysis impossible to reproduce. Write your analysis as a script with detailed comments, and record the versions of all software used.

Limitations and Interpretation

Metabolomics data analysis has inherent limitations that you must consider when interpreting your results.

Correlation Does Not Imply Causation

Metabolomics identifies associations between metabolite levels and experimental conditions. These associations do not prove that the metabolites cause the observed effects. Follow up metabolomics findings with targeted experiments to establish causality.

Annotation Uncertainty

Many features in untargeted metabolomics datasets remain unidentified. The IDSL.CSA approach can improve annotation rates by generating composite spectra libraries from MS1-only data [9], but even this approach has limitations. Features that cannot be annotated cannot be mapped to biological pathways.

Technical Variability

Metabolomics measurements are subject to technical variability from sample preparation, chromatography, and mass spectrometry. Quality control samples help monitor this variability, but they cannot eliminate it. Interpret small differences between groups with caution.

Sample Size Considerations

A priori sample size determination is essential to ensure the validity of findings and optimize resource allocation [13]. However, sample size calculation in metabolomics remains challenging due to the high dimensionality and variability inherent in the data [13]. A systematic review of the literature identified twenty relevant studies on sample size calculation and power analysis in metabolomics, highlighting ongoing challenges and directions for future research [13].

Professional Escalation Criteria

Some problems in metabolomics analysis require expertise beyond the typical R user. Recognize when to escalate to a bioinformatics specialist, a statistician, or the instrument facility manager.

When to Consult a Bioinformatics Specialist

Consult a bioinformatics specialist when your preprocessing pipeline produces inconsistent results across batches, when you need to integrate multiple omics data types, or when you are developing custom analysis methods. Multi-omics integration approaches are vital for understanding the hierarchical complexity of human biology and have proven valuable in cancer research and precision medicine [11].

When to Consult a Statistician

Consult a statistician when your experimental design is complex, when you need to account for multiple covariates, or when you are analyzing survival outcomes. A general workflow for survival analysis applicable to high-dimensional omics data has been published, focusing on Cox-type penalized regressions and hierarchical Bayesian models for feature selection [10]. A step-by-step R tutorial using The Cancer Genome Atlas survival and omics data is available for this workflow [10].

When to Consult the Instrument Facility

Consult the instrument facility when quality control samples show excessive drift, when retention times are unstable, or when you suspect instrument contamination. These problems are best addressed at the instrument level instead of through computational correction.

Frequently Asked Questions

What is the minimum R experience needed to analyze metabolomics data?

You need working knowledge of data frames, basic plotting with ggplot2, and the ability to install and load packages from CRAN and Bioconductor. Familiarity with the tidyverse family of packages is helpful for data manipulation. If you can import a CSV file, filter rows, and create a scatter plot, you have the foundational skills to begin metabolomics analysis in R.

How do I choose between xcms and MZmine for preprocessing?

Both tools are valid for LC-MS data preprocessing. The xcms package is integrated into the R and Bioconductor ecosystem, which means you can move directly from preprocessing to statistical analysis without exporting and importing data [8]. MZmine 2 provides a graphical interface that may be easier for beginners, and it incorporates the ADAP algorithms for peak detection and deconvolution [19]. Choose xcms if you want a fully scripted workflow in R. Choose MZmine if you prefer a graphical interface and are comfortable exporting the feature table for import into R.

What should I do when my PCA plot shows separation by batch instead of by experimental group?

Batch separation indicates that technical variation is dominating your data. First, verify that the batch effect is real by checking whether samples from the same batch cluster together. If batch effects are present, include batch as a covariate in your statistical model or apply a batch correction method. Consider whether the batch effect is confounded with your experimental groups, because confounding cannot be corrected computationally.

How many quality control samples should I include in my experiment?

The number of quality control samples depends on the size of your study and the stability of your instrument. A common practice is to inject quality control samples at regular intervals throughout the analytical run, such as one quality control sample for every five to ten study samples. The quality control samples should be pooled from all study samples to represent the average composition.

What is the difference between normalization and scaling?

Normalization corrects for systematic differences in total metabolite concentration between samples, such as differences in sample amount or instrument sensitivity. Scaling adjusts the variance of each feature so that features with different intensity ranges contribute equally to the analysis. Normalization is applied before scaling, and both steps are needed for most metabolomics datasets.

How do I handle features that are missing in all samples of one group?

Features that are missing in all samples of one group but present in another group may represent biologically meaningful differences. A metabolite that is present in control samples but absent in treated samples could indicate that the treatment suppresses production of that metabolite. Do not impute these features. Instead, test whether the presence or absence pattern is statistically significant using a Fisher exact test.

Can I use the same R workflow for NMR and mass spectrometry data?

The statistical analysis steps are similar for both platforms, but the preprocessing steps differ. NMR data typically requires binning or peak integration before you have a feature table. Mass spectrometry data requires peak detection, deconvolution, and alignment. Once you have a feature table, the normalization, imputation, univariate analysis, multivariate analysis, and visualization steps are largely platform-independent.

How should I report my metabolomics analysis methods in a publication?

Report the software and versions used for each step of the pipeline, including preprocessing, normalization, imputation, and statistical analysis. Report the parameters used for peak detection and alignment. Report the normalization and scaling methods. Report the multiple testing correction method and the significance thresholds. Provide access to your analysis script and processed data so that other researchers can reproduce your results.

Related Bioinformatics Guides

• Alternative Splicing Analysis from RNA-Seq Data
• Gene Ontology (GO) and Enrichment Analysis
• The KEGG Database and Pathway Analysis
• Predicting AMR from Genomic Data
• Circular RNAs: Computational Identification and Analysis

References and Further Reading

• EMBL-EBI Training. European Bioinformatics Institute.
• NCBI Data Resources. National Center for Biotechnology Information.
• Genomic Data Sharing Policy. National Institutes of Health.
• The FAIR Guiding Principles. Scientific Data.
• Miso: an R package for multiple isotope labeling assisted metabolomics data analysis.. Bioinformatics (Oxford, England), 2019.
• WebSpecmine: A Website for Metabolomics Data Analysis and Mining.. Metabolites, 2019.
• xOmicsShiny: an R Shiny application for cross-omics data analysis and pathway mapping.. Bioinformatics advances, 2025.
• xcms in Peak Form: Now Anchoring a Complete Metabolomics Data Preprocessing and Analysis Software Ecosystem.. Analytical chemistry, 2025.
• IDSL.CSA: Composite Spectra Analysis for Chemical Annotation of Untargeted Metabolomics Datasets.. Analytical chemistry, 2023.
• Tutorial on survival modeling with applications to omics data.. Bioinformatics (Oxford, England), 2024.
• Transcriptomics and epigenetic data integration learning module on Google Cloud.. Briefings in bioinformatics, 2024.
• MetaboDiff: an R package for differential metabolomic analysis.. Bioinformatics (Oxford, England), 2018.
• A priori sample size determination and power analysis in metabolic phenotyping and integrative metabolomics: an application framework based on a systematic review of literature.. 2026.
• Integrated serum and tissue NMR metabolomics with correlation analysis reveals metabolic alterations in gliomas.. 2026.
• Untargeted Metabolomics Reveals Distinct Metabolic Signatures of Lactic Acid Bacteria in Food Fermentation and the Same Pipeline Applied to Foodborne Pathogen Detection.. 2026.
• eCOMET: An R package for evaluating metabolic diversity and enrichment from LC-MS/MS data to test ecological hypotheses from individuals to ecosystems. 2026.
• A Comprehensive Evaluation of Metabolomics Data Preprocessing Methods for Deep Learning. Metabolites, 2022.
• A Python-Based Pipeline for Preprocessing LC-MS Data for Untargeted Metabolomics Workflows. Metabolites, 2020.
• Metabolomics Data Preprocessing using ADAP and MZmine 2. Methods in molecular biology, 2020.
• Preprocessing and Pretreatment of Metabolomics Data for Statistical Analysis.. Advances in Experimental Medicine and Biology, 2017.
• Metabolomics Data Preprocessing: From Raw Data to Features for Statistical Analysis. 2018.
• Untargeted Metabolomics Reveals Acylcarnitines as Major Metabolic Targets of Resveratrol in Breast Cancer Cells. Metabolites, 2025.
• GSimp: A Gibbs sampler based left-censored missing value imputation approach for metabolomics studies. Plos Computational Biology, 2018.
• Visualization and interpretation of multivariate associations with disease risk markers and disease risk-The triplot. Metabolites, 2019.

This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.