A New AI Study Reveals Machine-Learning Model Achieves Human-Level Skill in Describing Chemical Smell

The fundamental challenge in neuroscience is understanding how physical properties in stimuli are associated with perceptual characteristics. While there are well-established mappings between physical properties and perceptual qualities in other senses, such as color in vision and pitch in audition, the study highlights that mapping between chemical structures and olfactory percepts remains properly understood. 

To address these concerns, researchers developed a neural network-based model to map chemical structures to odor perceptions, creating a Principal odor map (POM) that captures perceptual distances and hierarchies. They experimented with a dataset of 5,000 molecules with odor labels, trained the model, and conducted a prospective validation challenge, showing that the model’s prediction closely matched human ratings for novel odorants. The POM preserved the perceptual relationships, outperforming traditional structure-based maps. The work emphasizes the potential of machine learning to map odor space and understand olfactory perceptions. 

They have compared the graph neural network (GNN) model to a traditional count-based fingerprint model for predicting odor preferences of various models. The GNN model outperformed the cFP-based model, matching or surpassing human panelists’ ratings for 55% for odor labels. Impurities in chemical reactions were identified as potential contributors to odor perceptions, with a 31.5% rate of significant odorous contamination in the stimulus set. The GNN model performed best for labels with clear structural determinants and with many training examples, while panelists’ performance varied based on familiarity with the labels.

The Principal odor map (POM) was tested for its robustness in handling discontinuities in mapping molecular structure and odor perception. The researchers obtained the result that POM correctly predicted the counterintuitive structure odor relationship in 50% of the cases, while a baseline model performed much worse at 90%. A linear model based on POM coordinates outperformed cheminformatics models in predicting odor applicability, odor detection thresholds, and perceptual similarity across multiple datasets. 

This driven map of human olfaction provides a foundation for further explorations of complex relationships between molecular structure and odor perception. It opens up new avenues for finding the nature of olfactory sensation and promises to advance the fields of chemistry, olfactory neuroscience, and psychophysics.

Check out the Research Article and Reference Article. All Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to join our 30k+ ML SubReddit, 40k+ Facebook Community, Discord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.

If you like our work, you will love our newsletter..