Aligning Mass Spectra with Molecular Structure for Fast Computational Olfaction

Understanding human olfaction from a computational point of view is an under-explored frontier within the machine learning community. Previous works leverage the chemical structure of odorant molecules for olfactory prediction tasks. However, obtaining the structure of unknown odorants requires significant time- and labor-intensive techniques, limiting their suitability for fast computational olfaction. In this work, we explore the use of direct electron ionization mass spectrometry (direct EI-MS), a fast sensing technique (in the order of seconds), for olfactory prediction tasks. We contribute Spectrum-to-Chemical Embedding alignmeNT (SCENT), a multi-modal contrastive learning framework to align mass spectra with explicit chemical information. In particular, we augment the mass spectrum representation with a chemical structure prior during training, requiring only mass spectrum information at test time. Our approach performs on par with state-of-the-art approaches that require chemical information explicitly during inference, and significantly better than previous mass-spectrum-only baselines.