Metabolomics Data Integration With Xcms

metabolomics-data-integration-with-xcms

Summary

Integrate peak-quality assessment into untargeted LC-MS metabolomics preprocessing pipelines by coupling XCMS chromatographic peak detection with MetaClean's machine learning classifier for flagging low-quality peaks. This skill enables quality control of extracted ion chromatograms (EICs) and peaks before downstream analysis.

When to use

You have untargeted LC-MS metabolomics data preprocessed with XCMS and need to filter out low-quality peak integrations that could introduce false positives or noise into metabolite quantification. Apply this skill when you want to automatically detect and flag EICs with poor signal-to-noise, baseline drift, or integration artifacts before statistical analysis or annotation.

When NOT to use

• Your LC-MS data has not yet been preprocessed with XCMS or an equivalent peak-detection tool; MetaClean operates on detected peaks, not raw spectra.
• You have only a small number of peaks (<100) or an unlabeled dataset with no ground truth for classifier training; MetaClean requires sufficient labeled data to train and validate robust models.
• You are working with targeted or data-independent acquisition (DIA) workflows where peak detection is not the primary quality bottleneck; MetaClean is designed for untargeted metabolomics EIC quality assessment.

Inputs

• XCMS-preprocessed peak detection results (mzML or NetCDF format LC-MS data)
• Extracted ion chromatograms (EICs) with retention time, m/z, and intensity profiles
• Labeled peak-quality ground truth (for development/training set)
• MetaCleanData example data (pqm_development, pqm_test) or user-provided peak quality metrics matrix

Outputs

• Trained MetaClean classifier model (e.g., AdaBoost_M11.model)
• Peak-quality predictions with pass/fail flags and confidence scores for each EIC
• Evaluation metrics (Pass_FScore, Fail_FScore, Accuracy, Precision, Recall) and model comparison plots (bar plots, CD plots)
• Filtered EIC list with low-quality peaks flagged or removed

How to apply

Install MetaClean alongside your existing XCMS preprocessing pipeline. After XCMS peak detection, extract EICs and apply MetaClean's getPeakQualityMetrics function to compute 11–12 quality metrics (adapted from Zhang & Zhao 2014 and Eshghi et al. 2018) on each detected peak. Use runCrossValidation with k=5 folds and multiple repetitions to identify the best-performing machine learning algorithm (e.g., AdaBoost, RandomForest, or SVM) on a labeled development set using the M11 metric set. Once the best model is selected via statistical ranking (e.g., CD plots), train a final classifier using trainClassifier, then apply getPredictions to score test or production EICs and assign each peak a quality flag. Optionally filter by RSD % using rsdFilter before metric calculation to reduce computational burden.

Related tools

• XCMS (Performs chromatographic peak detection and EIC extraction on raw LC-MS data; MetaClean integrates downstream to assess quality of XCMS-detected peaks.)
• MetaClean (Core tool for computing 11–12 peak-quality metrics and training/applying machine learning classifiers to detect low-quality peaks.) — https://github.com/KelseyChetnik/MetaClean
• MetaCleanData (Provides example development and test datasets for training and validating MetaClean classifiers.) — https://github.com/KelseyChetnik/MetaCleanData
• caret (R package underlying MetaClean's implementation of 8 classification algorithms (DecisionTree, LogisticRegression, NaiveBayes, RandomForest, SVM_Linear, AdaBoost, NeuralNetwork, ModelAveragedNeuralNetwork).)
• R (Programming environment for running MetaClean and generating evaluation visualizations.)

Examples

runCrossValidation(pqm_development, k=5, repNum=10, rand.seed=453, algorithms=c('DecisionTree','LogisticRegression','NaiveBayes','RandomForest','SVM_Linear','AdaBoost','NeuralNetwork','ModelAveragedNeuralNetwork'), metricSet='M11')

Evaluation signals

• Cross-validation fold results are reproducible: runCrossValidation with identical rand.seed, k, and repNum parameters produces identical model rankings and evaluation metrics across repeated runs.
• Statistical ranking via CD (Friedman-Demšar) plots shows clear differentiation among algorithms; the selected 'best' model has significantly lower average rank and confidence interval that does not overlap with lower-ranked competitors.
• Pass_FScore, Fail_FScore, and Accuracy metrics for the best-performing model on the held-out test set meet or exceed domain-specific thresholds (e.g., Fail_FScore ≥ 0.80 to reliably detect low-quality peaks).
• Predictions on new EICs are assigned confidence scores; predictions with high confidence (>0.9) on a validation set show >90% agreement with manual quality assessment or spike-in ground truth.
• Bar plots and CD plots generated by getBarPlots and getCDPlots show no visual anomalies (e.g., extreme outliers, missing algorithms, reversed metric scales) and confirm ranking consistency across evaluation measures.

Limitations

• MetaClean uses 8 or 9 classification algorithms (source cards mention both counts); performance is sensitive to the choice of algorithm and metric set (e.g., M11), requiring cross-validation to select the best pair.
• The 11–12 peak-quality metrics are adapted from literature but may not capture all types of integration artifacts in user-specific LC-MS methods; metric relevance depends on ionization mode, chromatographic resolution, and sample complexity.
• MetaClean requires labeled training data (ground truth peak-quality annotations); de novo quality assessment without labeled data is not supported.
• No changelog or versioning documentation is available; reproducibility across MetaClean releases or dependencies (caret, R versions) is not guaranteed.
• The skill is demonstrated on small example datasets (MetaCleanData); scalability to large untargeted metabolomics cohorts (>10,000 peaks) is not explicitly validated.

Evidence

• [readme] MetaClean utilizes 12 peak-quality metrics and 9 diverse machine learning algorithms to build a classifier to detect and flag low-quality peaks in untargeted metabolomics data.: "MetaClean utilizes 12 peak-quality metrics and 9 diverse machine learning algorithms to build a classifier to detect and flag low-quality peaks in untargeted metabolomics data"
• [intro] can be easily incorporated into existing untargeted LC-MS metabolomics pipelines that utilize the pre-processing software XCMS: "can be easily incorporated into existing untargeted LC-MS metabolomics pipelines that utilize the pre-processing software XCMS"
• [methods] The package is designed for use with the preprocessing package XCMS and can be easily integrated into existing untargeted metabolomics pipelines.: "The package is designed for use with the preprocessing package XCMS and can be easily integrated into existing untargeted metabolomics pipelines"
• [readme] The 12 peak-quality metrics used by MetaClean are adapted from the following publications: Zhang, W., & Zhao, P. X. (2014)… and Eshghi, S. T., Auger, P., & Mathews, W. R. (2018).: "The 12 peak-quality metrics used by MetaClean are adapted from the following publications… (2014)… and (2018)"
• [methods] MetaClean provides 8 classification algorithms (implemented with the R package caret): "MetaClean provides 8 classification algorithms (implemented with the R package caret)"
• [methods] runCrossValidation function is a wrapper function that uses cross-validation to train a user-selected subset of the 8 available algorithms.: "runCrossValidation function is a wrapper function that uses cross-validation to train a user-selected subset of the 8 available algorithms"
• [methods] MetaCleanData provides example data for development and test sets.: "MetaCleanData provides example data for development and test sets"
• [methods] getCDPlots… also provides a wrapper function to easily generate CD plots of the evaluation measures for each model.: "MetaClean also provides a wrapper function to easily generate CD plots of the evaluation measures for each model"