MS²-SMILES AlignNet: A cross-modal contrastive learning framework for direct alignment of tandem mass spectra and...

Structural annotation of metabolites via tandem mass spectrometry (MS²) remains a long-standing core challenge in analytical chemistry. To address this issue, we introduce MS²-SMILES AlignNet (MSAN), a cross-modal contrastive learning framework tailored for direct alignment between MS² spectra and molecular structures. Its key innovations lie in a dual-branch Transformer-based MS² encoder and a hybrid loss function, which jointly enable effective cross-modal alignment while preserving molecular structural similarity. Trained on more than 1.6 million high-quality spectrum-structure pairs, MSAN achieves state-of-the-art performance on the unified test subset of the CASMI 2022 benchmark, attaining a Recall@1 of 54.23% in positive ion mode and 45.37% in negative ion mode. Notably, it outperforms CSU-MS² by a substantial 14.81 percentage points in the more challenging negative ion mode. Furthermore, its generalization capacity and structural isomer discrimination ability are validated on the CASMI 2016 dataset. Meanwhile, MSAN exhibits favorable robustness across diverse experimental conditions. Collectively, this work provides a precise and robust tool for high-throughput metabolite annotation in untargeted metabolomics research.