Predicting a Molecular Fingerprint from an Electron Ionization Mass Spectrum with Spec2Vec

Mass spectrometry is a technique widely used for the identification of compounds in biological systems. However, traditional mass spectrometry matching methods can only identify compounds that are already present in the mass spectrometry library. One solution is to predict the mass spectra by molecular fingerprint and expand the mass spectral library by the predicted mass spectra. Another method is to predict the molecular fingerprint by mass spectra and retrieve unknown compounds by molecular fingerprint. Since deep learning networks are difficult to train sparse mass spectra. To address this problem, a fingerprint prediction method based on Spec2Vec is proposed. The proposed method uses mass spectrum embedding to transform the sparse mass spectrum vector into a dense feature vector. The experimental results show that the performance of fingerprint prediction using the mass spectral embedding is better than molecular fingerprint prediction directly using mass spectra as features. In addition, the proposed method can be linked with the fingerprint prediction mass spectrometry method to further improve the recognition accuracy.