Abstract
There has been recent progress in predicting whether common verbal descriptors such as "fishy", "floral" or "fruity" apply to the smell of odorous molecules. However, accurate predictions have been achieved only for a small number of descriptors. Here, we show that applying natural-language semantic representations on a small set of general olfactory perceptual descriptors allows for the accurate inference of perceptual ratings for mono-molecular odorants over a large and potentially arbitrary set of descriptors. This is noteworthy given that the prevailing view is that humans' capacity to identify or characterize odors by name is poor. We successfully apply our semantics-based approach to predict perceptual ratings with an accuracy higher than 0.5 for up to 70 olfactory perceptual descriptors, a ten-fold increase in the number of descriptors from previous attempts. These results imply that the semantic distance between descriptors defines the equivalent of an odorwheel.
Introduction
Quantifying smell is notoriously difficult. Unlike vision, which can be broken down into RGB values, or hearing, which relies on frequency, olfaction lacks a universal coordinate system. While humans are excellent at discriminating between millions of different scents, we are famously poor at naming them. This paper, published in Nature Communications, presents a breakthrough in "digitizing" smell. By leveraging natural language processing (NLP), the authors demonstrate that the semantic relationships between words in everyday language can be used to predict how a molecule will smell across a vast array of descriptors—effectively mapping the "odorwheel" of human perception.
Methods and Datasets
The authors’ core innovation is the use of distributional semantic models to bridge the gap between chemistry and language. Instead of trying to predict 146 different descriptors directly from chemistry, they first predicted a small set of 19 "anchor" descriptors (the DREAM dataset) using 4,884 physicochemical features of the molecules.
The workflow involves:
- Semantic Mapping: Using fastText, an NLP algorithm trained on a 16-billion-word corpus, to represent descriptors as 300-dimensional vectors.
- Cross-Lexicon Prediction: Developing a transformation matrix that maps the semantic vectors of the small anchor set (DREAM) to a much larger target set of descriptors (Dravnieks).
- Zero-Shot Learning: This allowed the model to perform "zero-shot" inference—predicting ratings for molecules or descriptors it had never seen in an olfactory training context.
The study utilized two primary datasets: the DREAM Dataset (ratings for 476 molecules across 19 descriptors by 49 laypeople) and the Dravnieks Dataset (ratings for 128 molecules across 146 descriptors by 500 experts).
Key Results and Insights
The model successfully predicted perceptual ratings with an accuracy (correlation) higher than 0.5 for up to 70 descriptors. This is a 10x increase over previous models that typically capped out at 5–7 descriptors.
A profound insight is that the "Odorwheel" is semantic. The structure of our language contains a latent map of our sensory experiences. Prediction performance is "smooth" across the semantic space; if the model is good at predicting "lemon," it is likely also good at predicting "orange" because they are semantically close. Combining a "ratings-only" model with a "semantics-only" model yielded the best results, suggesting that linguistic data can fill the gaps where experimental data is sparse.
Limitations and Future Work
While promising, the study focused exclusively on mono-molecular odorants. Real-world scents are complex mixtures involving non-linear interactions not yet accounted for. Additionally, the current model struggles with linguistic nuances like antonyms (e.g., "sweet" vs "sour") and may contain cultural biases since it was trained on English text. Future work could incorporate structured lexicons like WordNet to handle these relationships more effectively.
Summary updated