Artificial intelligence could help fragrance and flavor developers explore a much larger pool of potential aroma molecules, including structures that conventional discovery methods might overlook. New research published in Scientific Reports demonstrates a generative AI framework designed to create novel odorant candidates while accounting for whether those molecules could realistically be synthesized.
The study, published November 2, 2026, by researchers Tim C. Pearce at the University of Leicester and Ahmed Ibrahim of the School of Clinical Medicine, University of Cambridge, combines two computational approaches: a variational autoencoder (VAE), which learns patterns in molecular structures and generates new ones, and a quantitative structure-activity relationship (QSAR) model, which predicts the likelihood that a molecule has odor-related properties.
The system generated chemically valid structures, with 93.9% of candidates being unique, per the authors. More importantly, it produced 4.3 times as many distinct, novel compounds around a starting odorant structure as a comparable model without the QSAR component. This suggests the approach can help researchers explore more potential molecular variations from a known odorant rather than simply reproduce familiar structures.
The researchers also tested whether the proposed molecules could work outside a computer model. Their analysis found:
Of 6,765 ring-containing candidates, 2,499 had core molecular frameworks absent from the training datasets.
Automated retrosynthesis identified valid synthetic routes for 100% of candidates, averaging just 2.89 steps from commercially available starting materials.
Quantum-mechanical calculations indicated that the generated molecules had thermodynamic stability profiles comparable to known volatile compounds.
Relevant properties: Candidates displayed molecular weights and predicted physicochemical properties broadly consistent with established odorants.
The research points to a potentially valuable new tool for expanding the palette of molecules available to perfumers and flavorists. Rather than relying exclusively on screening existing chemical libraries or modifying known ingredients, researchers could use generative AI to propose entirely new molecular structures, then prioritize candidates based on odor-related predictions, chemical stability and potential manufacturability.
The ability to generate multiple novel structures around a known odorant could also help ingredient developers investigate new molecular variations and potentially identify alternatives with different performance characteristics.
There is, however, an important distinction between predicting odor potential and demonstrating actual olfactory performance. The study establishes a computational framework for generating and evaluating promising candidates; it does not demonstrate that the molecules have been synthesized, smelled or validated in finished fragrance or flavor applications.
