DeepPIC

DeepPIC extracts pure ion chromatograms (PICs) directly from raw liquid chromatography–mass spectrometry (LC-MS) centroid-mode data using a customized U-Net deep learning model to automate PIC construction for metabolomics feature extraction.

Key Features:

• Automated Extraction: Automates construction of PICs from raw LC-MS centroid-mode data without requiring parameter re-optimization across datasets.
• Deep Learning Framework: Employs a customized U-Net architecture to identify and extract PICs from centroid-mode LC-MS data.
• Integration into KPIC2: Integrates with the KPIC2 framework to enable end-to-end processing from raw LC-MS data to discriminant models.
• Performance Superiority: Demonstrates higher recall rates and stronger correlation with sample concentrations than XCMS, FeatureFinderMetabo, and peakonly across MM48, simulated MM48, and quantitative datasets.
• Universal Applicability: Validated on five diverse datasets from different instruments and samples, with 95.12% of extracted PICs matching manually labeled counterparts.

Scientific Applications:

• Biomarker identification: Provides reliable PICs for downstream biomarker discovery in metabolomics studies.
• Metabolic pathway analysis: Supports elucidation of metabolic pathways by supplying accurately extracted ion chromatograms for quantitation and comparison.
• Large-scale quantitative studies: Enables large-scale and quantitative LC-MS analyses by reducing the need for dataset-specific parameter tuning and improving reproducibility across instruments and samples.

Methodology:

Processes centroid-mode LC-MS data with a customized U-Net deep learning model and was trained, validated, and tested on an Arabidopsis thaliana dataset comprising 200 input–label pairs.