Predicting human olfactory perception from chemical features of odor molecules

Area of Science:

• Computational chemistry
• Cheminformatics
• Sensory science
• Machine learning

Background:

• Predicting molecular odor perception from chemical structure is a significant challenge in olfaction research.
• Understanding structure-odor relationships is crucial for various scientific and industrial applications.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting olfactory percepts from molecular chemoinformatic features.
• To assess the accuracy of these models in predicting odor intensity, pleasantness, and semantic descriptors.

Main Methods:

• Organized the crowd-sourced DREAM Olfaction Prediction Challenge.
• Utilized a large olfactory psychophysical dataset.
• Trained machine learning algorithms (including regularized linear models and random forests) on chemoinformatic features to predict sensory attributes.

Main Results:

• Models accurately predicted odor intensity and pleasantness.
• Successfully predicted 8 out of 19 semantic odor descriptors (e.g., 'garlic,' 'fish,' 'sweet,' 'fruit,' 'burnt,' 'spices,' 'flower,' 'sour').
• Regularized linear models achieved performance comparable to random forest models, nearing theoretical predictive limits.

Conclusions:

• Developed accurate predictive models for molecular odor perception.
• These models offer high accuracy in predicting perceptual qualities and can aid in reverse-engineering molecular smells.
• Advances in machine learning applied to chemoinformatic data significantly improve our ability to predict olfactory experiences.