Emzing: A Deep Learning-Enabled Software Suite for End-to-End Analysis of High-Resolution LC-MS Metabolomics Data

Abstract

Abstract High-resolution LC-MS-based metabolomics requires transforming noise-dominated raw data into quantitative feature tables and, ultimately, biologically meaningful insights. In practice, this process is fragmented across multiple tools and involves extensive parameter tuning and manual curation. Emzing is a web-based, vendor-agnostic platform that integrates end-to-end analysis, from raw data interrogation to statistical interpretation, within a unified environment. Emzing introduces data-driven, deep learning-based methods for signal extraction and feature quantification, formulating peak detection as a supervised learning problem in both spectral and chromatographic domains. A data-driven noise estimation strategy replaces fixed intensity thresholds, improving sensitivity to low-intensity signals. Feature correspondence across samples is established using robust retention-time alignment and graph-based grouping, while correlation-driven feature tagging assigns isotopes, adducts, and in-source fragments. Metabolite annotation (MSI level 2) combines spectral denoising with composite scoring of MS/MS spectra against reference libraries. Emzing further integrates processing with downstream statistical and machine learning-based analyses to facilitate reliable biological interpretation. Interactive visualization across all stages enables transparent curation and rapid iteration from initial results to publication-quality outputs. The use of pretrained models in inference mode, together with a reduced parameter space, supports reproducible analysis. Emzing improves true-feature recovery, quantitation accuracy, and reproducibility on benchmark data sets while reducing reliance on manual intervention. Scalability is demonstrated through Emzing Vault, a repository containing over 110 reprocessed public domain data sets.