Parsing Sage and Rosemary in Time: The Machine Learning Race to Crack Olfactory Perception

Abstract

Color and pitch perception are largely understandable from characteristics of physical stimuli: the wavelengths of light and sound waves, respectively. By contrast, understanding olfactory percepts from odorous stimuli (volatile molecules) is much more challenging. No intuitive set of molecular features is up to the task. Here in Chemical Senses, the Ray lab reports using a predictive modeling framework-first breaking molecular structure into thousands of features and then using this to train a predictive statistical model on a wide range of perceptual descriptors-to create a tool for predicting the odor character of hundreds of thousands of available but previously uncharacterized molecules (Kowalewski et al. 2021). This will allow future investigators to representatively sample the space of odorous molecules as well as identify previously unknown odorants with a target odor character. Here, I put this work into the context of other modeling efforts and highlight the urgent need for large new datasets and transparent benchmarks for the field to make and evaluate modeling breakthroughs, respectively.

Keywords: computation; feature extraction; modeling; olfaction; psychophysics; smell.

Figure 1.

Channeling the ancient Roman philosopher Lucretius, John Amoore used small plastic vessels to predict the odor character of molecules. If a plastic molecular model could fit into the vessel, it was predicted to have the odor descriptor with which the vessel was labeled (reproduced with permission from Wiley from Amoore 1964).

Figure 2.

Most machine-learning breakthroughs are data-limited, not model-limited. These 4 examples show that the development of the relevant algorithms for modeling a data domain were not enough. Only when sufficiently large, labeled datasets emerged did models finally achieve human-level performance. Figure based on text from Wissner-Gross (2016).

Figure 3.

Comparison of predictive performance of Kowalewski et al. (2021) to previous work. (A) Kowalewski slightly outperforms algorithms based on 2 common feature-based decompositions of molecular structure on prediction of descriptor ratings from Keller and Vosshall (2016). Black circles indicate, for each of 21 perceptual descriptors, the area under the receiver-operating characteristic (ROC) curve, reflecting sensitivity and specificity. A value of 1 indicates that the model perfectly ranks molecules according to their descriptor ratings; 0.5 reflects chance prediction. Blue circles indicate mean and standard error across descriptors. (B) Similar to A, but for the data from Dravnieks (1985), versus the text-prediction model from Gutiérrez et al. (2018), and using the Pearson correlation between predicted and observed descriptor ratings. (C) Similar to A, but for binary descriptor labels from The Good Scents Company—Flavor, Fragrance, Food and Cosmetics Ingredients Information (no date), versus the graph convolutional network model from Sanchez-Lengeling et al. (2019). Figure based on supplementary tables in Kowalewski et al. (2021).