DeepCPR: UNet3+ and Multivariate Curve Resolution for Automated GC−MS Data Deconvolution

Abstract Resolving overlapping peaks in gas chromatography−mass spectrometry (GC−MS), especially for isomers and low-abundance compounds, remains challenging due to spectral similarity and noise. Traditional multivariate curve resolution (MCR) methods heavily rely on initial spectral estimates, often leading to biased results. Here we present DeepCPR, a deep learning framework integrating a UNet3+ architecture with multi-scale feature fusion and dynamic channel attention to predict chromatographic profiles as robust initializations for MCR. This design accurately resolves co-eluted peaks even at low signal-to-noise ratios (SNR of 3−10). On amino acid and fatty acid standard mixtures, DeepCPR achieves mean explained variances (R2) of 0.987 and concentration correlation coefficients (r) of 0.994, outperforming AMDIS and DeepResolution2. Crucially, DeepCPR resolves near-identical isomers (spectral similarity > 0.999) with 94.1% spectral match accuracy, an 18% improvement over spectral-based methods. DeepCPR automates untargeted GC−MS processing and peak-table generation. OPLS-DA of a fixed 53-variable matrix from 136 human plasma samples achieved 98.54% mean accuracy under repeated stratified 10-fold cross-validation (100 repetitions) and prioritized 16 variables as potential biomarkers. The open-source implementation (https://github.com/YuChuanxiu/DeepCPR) ensures reproducibility and adaptability across diverse analytical workflows for precision metabolomics and environmental analysis.