A semantic-based community model for high-fidelity tuning of olfactory mixture distances

Area of Science:

• Sensory Science
• Olfactory Neuroscience
• Computational Chemistry

Background:

• Establishing quantitative links between physical stimuli and perception is a key goal in sensory science.
• Such mappings are well-defined for vision and audition but remain elusive for olfaction, especially for complex odor mixtures.
• Current understanding of olfactory perception struggles with predicting the experience of combined odors.

Purpose of the Study:

• To develop a predictive model for perceptual distances between complex odor mixtures.
• To investigate if a compact semantic space derived from single-molecule properties can capture mixture perception.
• To benchmark prediction models in the Dialogue for Reverse Engineering Assessment and Methods Olfactory Mixtures Prediction Challenge.

Main Methods:

• Assembled a unified dataset of odor-mixture pairs for the Olfactory Mixtures Prediction Challenge.
• Benchmarked prediction models on a hidden test set and integrated top performers into an ensemble model.
• Utilized single-molecule representations to derive a compact semantic space for odor prediction.

Main Results:

• The developed ensemble model significantly outperformed existing state-of-the-art methods on a hidden test set.
• Achieved a 33% reduction in RMSE (to 0.08) and a 53% increase in Pearson correlation (to 0.57) on the test set.
• An ensemble using only olfactory semantic features further improved predictions, showing strong performance on both test and validation sets.

Conclusions:

• Perceptual distances in odor mixtures can be accurately predicted using a compact semantic space derived from single-molecule properties.
• Mixture perception may rely on similar representational principles as single-molecule olfaction.
• Established a reproducible quantitative framework for olfactory mixture perception, aiding future research and applications in smell engineering.