DEPI-MS: Density-Equalized Pseudo-Image Learning for Drift-Tolerant LC−MS Metabolomics Classification

Liquid chromatography−mass spectrometry (LC−MS)-based metabolomics captures system-level biochemical variation, but classification remains difficult under realistic acquisition variability. Conventional peak-table workflows reduce the continuous retention time (RT)−m/z intensity field to discrete features, thereby losing structural information such as peak morphology, co-elution patterns, and local background. Pseudo-image learning provides a way to preserve this signal field, yet current approaches often allocate image resolution inefficiently, rely on generic image backbones, and remain sensitive to chromatographic drift. Here, we present DEPI-MS, a Density-Equalized Pseudo-Image learning for LC-MS-based metabolomics. DEPI-MS combines three practical design choices tailored to LC−MS data. First, it preserves uniform RT binning while using QC-derived nonuniform m/z binning, assigning more pixels to signal-dense mass regions without disrupting chromatographic interpretation. Second, a compact multi-scale classifier aggregates local peak-like responses and broader RT−m/z context through channel-wise global feature reweighting. Third, differentiable RT-axis warping is applied during training to reduce sensitivity to smooth chromatographic displacement without regenerating pseudo-images from raw files. In a nasopharyngeal carcinoma serum cohort, DEPI-MS achieved an AUC of 0.984 on an independent external validation set and remained stable under a 60 s RT-drift stress test, outperforming representative peak-table and pseudo-image baselines. Cohort-level attribution further localized recurrent model evidence to RT−m/z regions, providing candidate signal loci for biochemical follow-up. These results support DEPI-MS as a practical route toward transferable LC−MS metabolomics classification under acquisition variability.