Day 3: Metabolomic Data Analysis Workshop

Summary

The session recaps statistical and data-quality considerations for metabolomics, then demonstrates analysis in MetaboAnalyst and basic machine-learning workflows in Orange. The central message is that reliable biological interpretation depends on identifying and controlling technical variation before selecting metabolites or building classifiers.

1. Sources of Variation and Experimental Design

Metabolomic measurements can vary for biological and technical reasons:

• Biological variation: Differences between people, health conditions, age, treatment, medications, diet, and sample-collection timing. Metabolites can change quickly—for example, depending on recent food or drink intake.
• Sample handling: Transportation, storage conditions, storage duration, and repeated freeze–thaw cycles can affect metabolite stability.
• Reagents and contamination: Some plastics may leach compounds into samples or contribute to mass-spectrometry interference. The instructor recommends careful reagent handling and notes that glassware is often preferred.
• Instrument and chromatography effects: LC columns and their condition can affect metabolite separation. A column change or blockage may introduce batch effects.

Recommended practices include randomizing the order of samples rather than running cohorts one by one, documenting participant and collection characteristics, and using pooled QC samples and internal standards to monitor analytical performance. QC samples should be run throughout batches and maintained under consistent conditions.

2. QC, Outliers, and Data Quality

An ideal dataset would have comparable measurements across samples, but real metabolomics data often contain substantial variation. PCA, sample correlations, and feature-level variability can help identify potential problems, but they do not by themselves determine whether a difference is biological or technical.

• Pooled QC samples should generally cluster together and show limited variability. Their behavior helps assess whether observed sample differences may be due to instrument or processing effects.
• Internal standards can help monitor extraction and run-to-run variability.
• Coefficient of variation (CV) in pooled samples is calculated as standard deviation divided by the mean, multiplied by 100. The instructor suggests checking features against roughly 20–30% CV thresholds; features with high variability in QC samples may be unreliable.
• Outliers should be assessed carefully. An unusual sample could reflect a technical problem or a genuinely different biological profile. Decisions to exclude samples should be based on the evidence and study context.
• The instructor also highlights blanks, signal-to-noise, mass accuracy (including a cited 5-ppm check), and MS/MS evidence when curating metabolite features.

3. Distribution, Missing Values, and Preprocessing

Measured relative ion intensities are continuous values, so their distributions may approximate a Gaussian pattern, though real data are rarely perfectly normal. The instructor stresses that preprocessing should reduce noise and technical variation without distorting the original data structure.

Missing Values

• Distinguish missing completely at random from values missing systematically—for example, because a metabolite is below detection or has low abundance.
• Features with a large proportion of missing values may be removed. The instructor presents 30% missingness as a common working threshold, while emphasizing that the decision depends on data quality and context.
• If imputing, choose a method appropriate to the missingness mechanism. Options discussed include mean or low-value/LOD-based imputation and K-nearest-neighbor (KNN) imputation.
• Imputation should account for cohort characteristics so that values from one group do not improperly influence another. It should also avoid substantially changing the data’s distribution or shape.

Normalization and Scaling

• Normalization aims to make samples comparable and reduce technical effects while preserving biological signal.
• Methods discussed include mean or median centering, normalization using pooled QC samples, and use of an internal standard as a reference feature.
• Scaling can follow centering. Standard-deviation scaling (described as z-scaling in the session) and median-absolute-deviation-based scaling were discussed.
• The choice should depend on the data and analysis. Preprocessing choices should be documented, especially for publications or theses.
• The instructor recommends checking the data after preprocessing rather than assuming a method improved it.

4. Statistical Comparisons and Multiple Testing

The session revisits PCA, correlation, t-tests, and heat maps:

• PCA provides a visual check for sample grouping and possible outliers or batch effects.
• Sample-wise correlation can help identify samples that behave unusually. The instructor suggests Pearson correlation for suitable normalized data and Spearman correlation for rank-based assessment, including when data are not normalized or transformed.
• t-tests can identify differences between two groups, and volcano plots can display fold change alongside statistical significance.
• Because omics studies test many features, raw p-values can produce many false positives. The instructor recommends using adjusted p-values, particularly false discovery rate (FDR) correction such as Benjamini–Hochberg. More stringent family-wise corrections may remove more candidate features.

5. MetaboAnalyst Demonstration

Using a small dataset, the instructor demonstrates a workflow in MetaboAnalyst:

1. Upload the dataset in the required format, with sample names, group labels, and metabolite-feature columns correctly specified.
2. Review the dataset and assess it without additional filtering first.
3. Apply a low-variance filter when appropriate, especially for datasets with many thousands of features.
4. Apply a suitable transformation and normalization; the demonstration includes log transformation and median-based normalization.
5. Consider normalization against pooled QC samples or an internal-standard reference feature.
6. Examine PCA and sample correlations, checking whether pooled samples behave consistently.
7. Compare groups using a t-test and inspect fold changes, adjusted significance, and heat maps of selected features.

The demonstration dataset showed substantial variability in some pooled QC features, illustrating why QC-based feature review is important before interpreting group differences.

6. Machine Learning in Orange

The session introduces Orange’s visual, widget-based interface for basic classification without coding.

Workflow

1. Load the prepared pre-diabetic versus diabetic dataset.
2. Use Select Columns to identify metabolite features, the class label, and sample-name metadata.
3. Optionally visualize PCA with a scatter plot.
4. Add classification methods such as K-nearest neighbors, SVM, random forest, Naive Bayes, and logistic regression.
5. Connect the data and models to Test & Score.
6. Evaluate models using cross-validation. The instructor stresses that cross-validation is essential and mentions stratified folds and leave-one-out cross-validation for smaller datasets.
7. Inspect the confusion matrix to review correct and incorrect classifications and related measures such as sensitivity, specificity, precision, and recall.
8. Use feature-ranking tools or model-based rankings to identify features that may distinguish groups.
9. Compare machine-learning-selected features with statistically significant features, and test whether using only top-ranked features changes model performance.

The instructor warns about underfitting (poor predictive performance) and overfitting (a model fits the training data but does not generalize). Very high scores in the demonstration should not be treated as proof of a reliable model. The session recommends separating training and test data and, for a larger study, evaluating the final model on an independent blinded dataset from another sample source. A commonly mentioned split is 70% training and 30% testing, alongside a separate blinded validation set.

Speakers and Sources Featured

• Main workshop instructor: Unnamed; presented the statistical recap, data-quality guidance, and software demonstrations.
• Workshop coordinator/host: Unnamed; shared information about the workshop page and supplementary materials.
• Participants: Several unnamed attendees asked questions about CV, KNN imputation, and scaling. One participant was addressed as Ashu while troubleshooting dataset upload.
• Software demonstrated: MetaboAnalyst and Orange.
• Data sources discussed: A small subset dataset and a pre-diabetic/diabetic comparison dataset, described during the session as originating from published data.