Feature Table Construction For Multivariate Analysis

feature-table-construction-for-multivariate-analysis

Summary

Construct aligned feature tables from mass-spectrometry breath data by extracting volatile organic compound (VOC) features from individual samples, filtering by quality metrics (RSD), and merging across samples to produce a matrix suitable for multivariate statistical analysis. This skill bridges raw MS data through feature extraction and alignment to a standardized CSV format where rows are samples and columns are m/z features with intensity values.

When to use

When you have high-resolution mass-spectrometry (HRMS) breath data in mzML or mzXML format from multiple subjects or conditions and need to prepare a sample-by-feature intensity matrix for downstream statistical comparison, biomarker discovery, or classification tasks. Use this skill when raw peak data must be harmonized across samples to account for retention-time drift and instrument variation.

When NOT to use

• Input is already an aligned feature table or pre-processed CSV—skip directly to statistical analysis.
• Data is from non-breath or non-HRMS instruments (e.g., GC-FID, targeted assays) where peak detection workflows differ.
• Sample sizes are so small (<3 samples) that alignment becomes unreliable or pooling across replicates is impossible.

Inputs

• mzML or mzXML files (raw HRMS breath data)
• List of file paths (one per sample)
• Sample name identifiers (strings)

Outputs

• Aligned feature table CSV (rows: m/z features; columns: sample identifiers; values: total intensities)
• FeatureSet objects (intermediate; for quality control inspection)

How to apply

First, call find_feature() on each mzML/mzXML file with appropriate algorithm choice (Topological or Gaussian) and quality threshold (typically 0.8) to extract a FeatureSet object containing m/z values, scan times, and integrated intensities. Apply RSD (relative standard deviation) filtering—either at a fixed threshold or using a quantile (e.g., 0.1 quantile)—to remove noise features with inconsistent breath-peak intensity patterns. Next, invoke merge_result() on the list of filtered FeatureSets with custom sample names to align features across samples by m/z value and calculate per-sample intensity totals. Finally, export the resulting Sample object to CSV using to_csv(), which produces a table with m/z as the index, sample names as columns, and total intensity per feature per sample as cell values. Optional: annotate adducts and isotopes by passing adduct=True and isotope=True to to_csv().

Related tools

• BreathXplorer (Python package providing find_feature(), merge_result(), and to_csv() functions for feature extraction, alignment, and table export) — https://github.com/wykswr/breathXplorer

Examples

from breathXplorer import find_feature, merge_result
fss = [find_feature(f, False, .8, "Topological", 6) for f in ["sample1.mzML", "sample2.mzML", "sample3.mzML"]]
fss = [fs.rsd_control(fs.rsd.quantile(0.1)) for fs in fss]
sample = merge_result(fss, ["sample1", "sample2", "sample3"])
sample.to_csv("aligned_feature_table.csv")

Evaluation signals

• Output CSV has consistent dimensions: number of rows equals number of unique m/z features, number of columns equals number of input samples plus 2 (ID and m/z columns).
• All intensity values are non-negative and within the expected dynamic range for the instrument (no negative or implausibly large values).
• No m/z duplicates in the output; each feature is represented exactly once (merge_result aligns by m/z and aggregates intensities).
• RSD-filtered FeatureSets have demonstrably lower noise: median RSD of retained features is below the applied threshold (e.g., median RSD < 0.1 if threshold = 0.1 quantile).
• Cross-sample alignment produces expected feature presence/absence patterns; features detected in multiple samples show consistent m/z values (within typical MS tolerance, ~5 ppm for HRMS).

Limitations

• Peak recognition sensitivity depends on user-selected quality threshold (0.8 default); lower thresholds may introduce noise, higher thresholds may miss weak but real VOCs.
• RSD-based filtering assumes breath peaks have consistent intensity across the breath cycle; metabolically variable or transient features may be incorrectly removed.
• Feature alignment relies on m/z matching without chromatographic retention-time information in the final output; co-eluting isomers with identical m/z cannot be distinguished.
• The Gaussian algorithm requires prior knowledge (number of breaths in experiment) which may not always be available; Topological algorithm is more general but may perform differently on certain breath profiles.
• Supported input formats are limited to mzML and mzXML; other vendor formats (e.g., .raw) require conversion before processing.

Evidence

• [readme] BreathXplorer is a bioinformatic solution to process breath data generated from HRMS analysis. It contains a suite of functions, including feature extraction, feature alignment, and breath recognition.: "BreathXplorer is a bioinformatic solution to process breath data generated from HRMS analysis. It contains a suite of functions, including feature extraction, feature alignment, and breath"
• [readme] The input file should be in mzML or mzXML format (Perhaps more in the future). For the output, the single or aligned feature table can be exported as csv file.: "The input file should be in mzML or mzXML format (Perhaps more in the future). For the output, the single or aligned feature table can be exported as csv file."
• [readme] The index of the table is the m/z value of the features, and each column is the total intensity of the feature in a sample (experiment of a subject). The name of the column is the sample name.: "The index of the table is the m/z value of the features, and each column is the total intensity of the feature in a sample (experiment of a subject). The name of the column is the sample name."
• [readme] Feature extraction is used to find the volatile organic compound (VOC) in the breath sample. The feature extraction is performed using the find_feature function.: "Feature extraction is used to find the volatile organic compound (VOC) in the breath sample. The feature extraction is performed using the find_feature function."
• [readme] In practice, the relative standard deviation (RSD) can be used to filter out the noise which doesn't have a consistent intensity with breath peaks.: "In practice, the relative standard deviation (RSD) can be used to filter out the noise which doesn't have a consistent intensity with breath peaks."
• [readme] The merge_result function takes as input a list of FeatureSet objects, and returns a Sample object. It aligns the features with the similar m/z value from different samples, and calculate the total intensity of each feature in each sample.: "The merge_result function takes as input a list of FeatureSet objects, and returns a Sample object. It aligns the features with the similar m/z value from different samples, and calculate the total"
• [other] BreathXplorer implements feature extraction as a workflow step that processes mass-spectrometry input data and produces a feature table CSV output in a standardized format.: "BreathXplorer implements feature extraction as a workflow step that processes mass-spectrometry input data and produces a feature table CSV output in a standardized format."