Cnn Inference On Metabolomics Data

CNN inference on metabolomics data

Summary

Apply a pre-trained convolutional neural network to evaluate chromatographic peak quality and assess feature fidelity in LC-MS-based metabolomics datasets. This skill automates the classification of extracted ion chromatogram (EIC) plots as true or false features, replacing manual inspection.

When to use

Use this skill when you have LC-MS feature tables (with m/z and retention time columns) paired with raw .mzXML or .mzML files, and you need to automatically validate which detected features correspond to genuine chromatographic peaks versus false positives or noise. Apply it as a quality-control gate before downstream metabolite annotation or statistical analysis.

When NOT to use

• Input is already a validated, manually curated feature list with confirmed peak quality
• Raw mass spectrometry files (.mzXML/.mzML) are unavailable or not paired with the feature table
• Feature table lacks m/z and retention time columns in the required format (first two columns)

Inputs

• Feature table (.csv): m/z, retention time (rt), and feature intensity columns
• Raw mass spectrometry data files (.mzXML or .mzML format)
• Smoothing parameter selection (0, 1, or 2)

Outputs

• FinalTable.csv: feature table with CNN-derived peak quality scores
• PredictionOutcomes.csv: Boolean classification (true/false feature) for each input feature
• EIC plots (.png): extracted ion chromatograms visualized for quality inspection

How to apply

Load the pre-trained CNN model from pyEVA and initialize it with the trained checkpoint weights. Prepare your feature table as a .csv file with m/z and rt as the first two columns, and co-locate the corresponding .mzXML or .mzML files in the input directory. Run pyEVA's main pipeline (python main.py), selecting a smoothing level (0, 1, or 2) to match your EIC preprocessing preference. The CNN performs forward inference on each feature's chromatographic signature, outputting a quality score and Boolean classification (true/false feature) for each input. Retrieve the final predictions from FinalTable.csv and PredictionOutcomes.csv in the output folder, which contain per-feature CNN scores and fidelity assignments.

Related tools

• pyEVA (Python-based deep learning framework that wraps the pre-trained CNN model and orchestrates EIC extraction, preprocessing, inference, and result aggregation) — https://github.com/HuanLab/pyEVA
• XCMS (R package used (via RStudio) for upstream mass spectrometry feature detection and EIC plot generation)
• ggplot2 (R package for visualizing EIC plots and diagnostic graphics)
• Anaconda (Python environment manager for isolating and activating the pyEVA conda environment (pyEVA))

Examples

python main.py

Evaluation signals

• Output files (FinalTable.csv, PredictionOutcomes.csv) are generated without runtime errors and contain the same number of rows as the input feature table
• Boolean predictions in PredictionOutcomes.csv are binary (True/False) and not null for any feature
• Peak quality scores in FinalTable.csv fall within a bounded range (e.g., [0, 1] or log-likelihood scale) and show variance across features
• EIC plots are generated and visually correspond to the input m/z and retention time values, confirming correct feature extraction
• Manual spot-check of 10–20 EIC plots and their corresponding CNN classifications agrees with domain expertise (high-confidence features show characteristic peak shapes; low-confidence features show baseline noise, shoulder peaks, or multi-modal distributions)

Limitations

• CNN model was trained on a specific LC-MS dataset and instrument type; performance may degrade on significantly different instrument platforms, ionization modes, or chromatographic methods
• Smoothing parameter (0, 1, 2) choice affects EIC appearance and downstream CNN inference; no automated selection is provided; user must empirically validate which level is appropriate for their data
• Requires installation of external dependencies (Anaconda, R, R packages, Python packages), which may create environment conflicts on some systems
• CNN classifies peaks as binary (true/false) without confidence intervals or soft probability scores; borderline cases may not be distinguishable

Evidence

• [intro] EVA is a convolutional neural network-based platform designed to evaluate chromatographic peak quality for feature fidelity in LC-MS-based metabolomics: "A Convolutional Neural Network-Based Platform to Evaluate Chromatographic Peak Quality for Feature Fidelity in LC-MS-based Metabolomics"
• [readme] pyEVA automates EIC classification through deep learning: "Through pyEVA's use of deep learning tools, the process of identifying EIC plots as true or false can be automated, greatly speeding up the mass spectrometry pipeline of work."
• [readme] Input format specification for pyEVA: "In input place a .csv file with the first two columns being m/z and rt (followed by the relevant data in successive columns), along with the related .mzXML or .mzML files."
• [readme] Output files and their contents: "In output there should be a .csv named FinalTable.csv with the results of the analysis, and a .csv called PredictionOutcomes.csv with Boolean values for the data input."
• [readme] Smoothing parameter choice and inference workflow: "In the PyCharm terminal, type: python main.py. This should start your program. Choose the level of smoothing (0, 1, 2), and wait for the EICs to be output to the EICplots folder"
• [intro] EVA Python availability: "EVA now has a python version that can be easily installed and used on Mac, PC, and Linux systems. Please check out pyEVA at https://github.com/HuanLab/pyEVA"