Data based predictive models for odor perception.

Area of Science:

• Computational chemistry and cheminformatics
• Sensory science and psychophysics
• Data science and machine learning

Background:

• Traditional methods for odorant characterization are limited.
• Olfaction is a complex sense, poorly understood compared to others.
• Machine learning offers a data-driven approach for quantitative structure-property relations (QSPR).

Purpose of the Study:

• To develop and evaluate machine learning models for predicting odorant characters (OC) of "sweet" and "musky".
• To identify structural features correlating with specific odor perceptions.
• To analyze the impact of data quality and subject bias on model performance.

Main Methods:

• Analysis of a psychophysical dataset from 55 subjects rating odorant perceptions.
• Training and comparison of machine learning algorithms: Random Forest, Gradient Boosting, Support Vector Machine.
• Evaluation of model performance and identification of key structural predictors.

Main Results:

• Machine learning models successfully predicted odor characters based on chemical structures.
• Identified specific structural features significantly correlating with "sweet" and "musky" perceptions.
• Demonstrated the influence of data quality and subject bias on predictive model efficacy.

Conclusions:

• A data-driven methodology for predicting odor perception using machine learning is presented.
• Insights into human odor perception, including biases in untrained subjects, were gained.
• The developed models and methodology can predict odor characters of novel chemical compounds.