CoRe-Gen: Robust Spectrum-to-Structure Generation under Imperfect Fingerprint Conditions

CoRe-Gen, an innovative technique, tackles the difficulty of determining molecular structures from tandem mass spectra (MS/MS) for de novo generation that exceeds database limitations. This method breaks down the process into two stages: predicting fingerprints from spectra and then decoding structures from those fingerprints, utilizing extensive molecular datasets. However, its implementation depends on predicted fingerprints instead of ideal ones, leading to structured errors that can affect generation. To enhance the intermediate conditions, CoRe-Gen employs synthetic-spectrum pretraining for the encoder, addresses deployment-related noise with frequency-aware fingerprint corruption during decoder training, and reduces remaining errors through structure-aware strategies.