Accurate prediction of personalized olfactory perception from large-scale chemoinformatic features

Area of Science:

• Neuroscience
• Cheminformatics
• Computational Biology

Background:

• The olfactory stimulus-percept problem remains challenging, with difficulty predicting odor from molecular features due to individual variability and complex receptor interactions.
• Existing models struggle to predict personalized multi-odor attributes, highlighting a gap in understanding structure-odor relationships for intensity and pleasantness.

Purpose of the Study:

• To develop a predictive model for individual and population perceptual responses to odorants.
• To identify key chemical features influencing olfactory perception.

Main Methods:

• Utilized a random forest model, a machine learning approach employing multiple decision trees.
• Integrated both population-level and individual-specific perceptual data to enhance prediction accuracy.
• Analyzed feature importance to identify critical chemical descriptors for odor prediction.

Main Results:

• The random forest model demonstrated success in predicting personalized odor attributes for diverse molecules.
• Integration of diverse perceptual data reduced noise and outlier influence, improving model robustness.
• A small subset of low- and non-degenerative chemical features proved sufficient for accurate olfactory predictions.

Conclusions:

• The developed random forest model accurately predicts personalized odor attributes, advancing the field of olfactory perception.
• The identified discriminative features offer insights into olfactory mechanisms and facilitate rational odorant design.