Physically Constrained Mass Spectrometry Data Binning for Multi-Platform Untargeted Metabolomics Data Integration

Machine learning (ML) modeling on mass spectrometry (MS)-based untargeted metabolomics data facilitates cancer diagnosis, such as hepatocellular carcinoma (HCC). Cohort-specific models often suffer from limited transferability and generalizability, necessitating the integration of multi-center and multi-platform data. MS data are typically deconvoluted through binning; however, conventional methods can either disintegrate or aggregate m/z features, making the combination of external data unreliable for model improvement. Here, we introduce “(mass) resolution-adaptive MS data integration (raMSIn)” to address these limitations. Compared to traditional methods, raMSIn achieved the highest proportion (>94%) of perfectly binned features and produced a superior HCC detection model. Using a mouse model, 10 generic metabolites were identified as significant for HCC diagnosis across multiple MS platforms, including desorption electrospray MS imaging on liver cryosections (F1 score = 0.87) and glass smears from fine-needle aspirates (F1 score = 0.72), as well as direct infusion analysis (recall = 0.67). raMSIn represents a significant advancement in MS data processing, offering a robust framework for improved biomarker discovery and disease monitoring by integrating data from various sources and institutions. It is potentially applicable to clinical MS-based metabolomics.