Encoding LC–MS-Based Untargeted Metabolomics Data into Images toward AI-Based Clinical Diagnosis

Liquid chromatography–mass spectrometry (LC–MS)-based untargeted metabolomics provides comprehensive and quantitative profiling of metabolites in clinical investigations. The use of whole metabolome profiles is a promising strategy for disease diagnosis but technically challenging. Here, we developed an approach, namely MetImage, to encode LC–MS-based untargeted metabolomics data into multi-channel digital images. Then, the images that represent the comprehensive metabolome profiles can be employed for developing deep learning-based AI models toward clinical diagnosis. In this work, we demonstrated the application of MetImage for clinical screening of esophageal squamous cell carcinoma (ESCC) in a clinical cohort with 1104 participants. A convolutional neuronal network-based AI model was trained to distinguish ESCC screening positive and negative subjects using their serum metabolomics data. Superior performances such as sensitivity (85%), specificity (92%), and area under curve (0.95) were validated in an independent testing cohort (N = 442). Importantly, we demonstrated that our AI-based ESCC screening model is not a "black box". The encoded images reserved the characteristics of mass spectra from the raw LC–MS data; therefore, metabolite identifications in key image features were readily achieved. Altogether, MetImage is a unique approach that encodes raw LC–MS-based untargeted metabolomics data into images and facilitates the utilization of whole metabolome profiles for AI-based clinical applications with improved interpretability.