Deep cross-modal learning between tandem mass spectrometry and molecular fingerprints for metabolite identification

Metabolite annotation plays a key role in metabolomics. To enable structural annotation of unknown tandem mass spectra, the prediction of molecular fingerprints using MS/MS is currently of great interest. However, current methods still present challenges in terms of redundancy and high dimensionality of fingerprint features, which can affect the accuracy and speed of annotation results. Therefore, we propose a dual-tower model structure consisting of an MS/MS feature extractor and a fingerprint feature extractor, which can directly compute the correlation between MS/MS and molecular fingerprints without needing to predict molecular fingerprints. Moreover, the fingerprint feature extractor, consisting of two MLPs, effectively reduces fingerprint redundancy. Both feature extractors are simultaneously optimized by contrastive learning. We trained and tested our method using data downloaded from the GNPS. The trained model was then used to search molecular structure databases such as PubChem. Experimental results show that our method outperforms MetFID, FingerScorer, MatFrag, DeepMass and CFM-ID in top-k evaluation.