A Spectrum-to-Structure Diffusion Model for De Novo Small Molecule Generation

Elucidating the structures of unknown small molecules from mass spectra is a fundamental yet highly challenging task in domains such as metabolomics, proteomics, and drug discovery. Although mass spectrometry provides rich fragmentation patterns critical for structural identification, interpreting these spectra to accurately reconstruct molecular structures remains a complex problem. To address this, we propose DiffNovo, a spectrum-to-structure diffusion framework designed to enhance small molecule identification. DiffNovo integrates a BART-based SMILES encoder-decoder, a high-resolution spectrum encoder, and a conditional diffusion model. The BART-based component captures intricate dependencies among SMILES tokens, while the spectrum encoder effectively leverages fine-grained spectral features. The conditional diffusion model includes a forward process that incrementally adds Gaussian noise and a reverse denoising process that reconstructs molecular structures, guided by fragmentation patterns from the spectra. This conditioning ensures that generated molecules remain faithful to the spectral data. Experimental results on the NIST20 dataset show that DiffNovo outperforms existing methods across multiple evaluation metrics, including SMILES-level accuracy, fingerprint-based similarity, and physicochemical property alignment. The code is available at https://github.com/GaoHoufu/DiffNovo.