A novel Transformer-MLP fusion network for metabolite identification from mass spectra

Accurate metabolite identification remains a major challenge in untargeted metabolomics, particularly for novel compounds that are absent from existing spectral libraries. In this work, we propose a novel Feature Fusion Network (FFNet) based on Transformer and multilayer perceptron (MLP), a dual-stream architecture specifically designed to predict molecular fingerprints from mass spectra. FFNet incorporates a transformer-based pathway that captures the global context of spectral data and an MLP-based pathway that focuses on extracting local spectral features. These complementary representations are integrated through an attentive fusion mechanism to produce comprehensive molecular fingerprints. Extensive evaluations on the General Metabolite Identification Set (GMIS, 17,267 spectra), a custom-built test set designed for benchmarking metabolite identification, and the MassBank of North America (MoNA, 1243 spectra) dataset demonstrate that FFNet consistently outperforms baseline neural network models in both fingerprint prediction and metabolite identification tasks. Moreover, in structure elucidation experiments using the CASMI 2022 dataset, FFNet effectively retrieves candidate molecules with high structural similarity to unknown compounds, even in the absence of exact matches within the spectral library. These findings suggest that neural feature fusion significantly improves mass spectral analysis and enables more reliable metabolite identification in complex biological samples.