X-Align: annotation-independent cross-platform alignment of untargeted metabolomics features

Massive untargeted metabolomics datasets are rapidly accumulating, but cross-study and cross-platform reuse remain limited by poor comparability. This limitation is exacerbated by the low rate of confident structural annotation, which restricts structure-based alignment to only a small fraction of detected metabolic features. Here, X-Align is presented, which is an annotation-independent framework for cross-platform alignment of untargeted metabolomics features. Its core component, Align-ID, is a residual attention-based deep regression model that learns reference chromatographic behavior directly from MS/MS (tandem mass spectrometry) spectrum embeddings and maps heterogeneous spectra onto a common reference chromatographic scale without prior structural identification. Align-ID is trained on 804 150 high-resolution MS/MS spectra associated with reference retention values in the METLIN SMRT chromatographic system. X-Align achieved 99.6% accuracy at 52.3% coverage in the controlled benchmark and maintained high accuracy in independent cross-platform and external library validations. X-Align produces a cross-center feature repository containing 1402 aligned features and 60 688 MS/MS spectra from public plasma datasets and supports more consistent biological interpretation across independent experiments in a cross-center high-fat diet case study. This work establishes MS/MS-derived reference retention behavior as a practical basis for annotation-independent comparison and integration of cross-platform untargeted metabolomics data.