Metabolomics training set preparation

metabolomics-training-set-preparation

Summary

Preparation of annotated MS/MS spectra with purity labels from LC-MS/MS Data-Dependent Acquisition (DDA) mode for training customized chimeric spectra purification models. This skill bridges raw mass spectrometry data collection and machine learning model training within the DNMS2Purifier workflow.

When to use

When you have LC-MS/MS acquisitions in DDA mode and need to train a customized DNMS2Purifier model to purify chimeric MS/MS spectra specific to your experimental conditions, metabolite classes, or ionization settings. Use this skill if the default model does not adequately separate pure from chimeric spectra in your dataset.

When NOT to use

• Input spectra are already pre-filtered or purified by instrument-level methods and no chimeric contamination is suspected.
• Training data contains fewer than 50–100 reliably annotated spectra (insufficient for stable model training).
• MS/MS data are acquired in targeted or parallel reaction monitoring (PRM/MRM) mode rather than DDA, as chimeric spectra are an artifact specific to DDA.

Inputs

• LC-MS/MS DDA-mode spectra (MS/MS peak lists with m/z and intensity values)
• Purity annotations or labels for each spectrum (binary or quantitative)
• Precursor m/z and isolation window metadata
• Feature vectors derived from MS/MS spectral properties

Outputs

• Serialized trained model object (.rds file)
• Model performance metrics (accuracy, precision, recall, or equivalent)
• Feature importance rankings (if provided by the model)

How to apply

Collect LC-MS/MS data in DDA (Data-Dependent Acquisition) mode to generate MS/MS spectra. Annotate each MS/MS spectrum with a purity label (e.g., binary: pure vs. chimeric, or quantitative purity score) based on manual inspection, fragmentation consistency, or precursor isolation quality. Organize the annotated spectra and labels into a structured training dataset with explicit feature vectors and target labels. Load this prepared training dataset into the R 4.2.1 environment, then execute the DNMS2Purifier_model_generation.r script to train the customized model. The trained model object is serialized and saved to an .rds file for downstream use in the DNMS2Purifier.r purification workflow.

Related tools

• R (Execution environment for DNMS2Purifier_model_generation.r script; required version 4.2.1)
• DNMS2Purifier_model_generation.r (R script that trains customized model on annotated MS/MS spectra and purity labels) — https://github.com/HuanLab/DNMS2Purifier
• DNMS2Purifier.r (Main purification script that consumes the trained .rds model for MS/MS purification) — https://github.com/HuanLab/DNMS2Purifier

Examples

source('DNMS2Purifier_model_generation.r'); trainedModel <- trainDNMS2PurifierModel(spectraData = annotated_spectra, purityLabels = purity_labels); saveRDS(trainedModel, file = 'custom_purification_model.rds')

Evaluation signals

• Trained model .rds file is successfully serialized and can be loaded without errors into the R environment.
• Model achieves acceptable classification performance on a held-out test set (e.g., >80% accuracy or AUC >0.85 for pure vs. chimeric discrimination).
• When the trained model is applied within DNMS2Purifier.r purification workflow, the number of flagged chimeric spectra matches manual inspection or known chimera rate in the sample.
• Feature importance plot or coefficients show that model relies on plausible spectral properties (e.g., precursor isolation width, spectral entropy, fragment intensity variance) rather than random features.
• Training dataset contains sufficient annotation diversity (e.g., representation across precursor mass ranges, charge states, and fragmentation patterns) to avoid overfitting to specific metabolite classes.

Limitations

• Model training requires manual or semi-automated annotation of MS/MS purity, which is labor-intensive and may introduce subjective bias.
• Trained model is specific to the LC-MS/MS platform, ionization mode, and instrument settings used during training; transfer to different instruments or conditions may require retraining.
• No changelog or versioning guidance is provided in the repository, making it difficult to track which training datasets correspond to which model versions.
• The DNMS2Purifier_model_generation.r script documentation is limited; exact statistical method, feature engineering, and hyperparameter tuning are not fully detailed.

Evidence

• [other] Prepare training data containing annotated MS/MS spectra and purity labels from LC-MS/MS acquisitions in DDA mode.: "Prepare training data containing annotated MS/MS spectra and purity labels from LC-MS/MS acquisitions in DDA mode."
• [other] Execute the model generation script to train the customized purification model on the prepared training dataset.: "Execute the model generation script to train the customized purification model on the prepared training dataset."
• [other] Serialize and save the trained model object to an .rds file for downstream use in the DNMS2Purifier.r purification workflow.: "Serialize and save the trained model object to an .rds file for downstream use in the DNMS2Purifier.r purification workflow."
• [readme] we also provide the script DNMS2Purifier_model_generation.r for customized model training: "we also provide the script DNMS2Purifier_model_generation.r for customized model training"
• [readme] The program is written in R (ver 4.2.1).: "The program is written in R (ver 4.2.1)."