A comparative study of machine learning models on molecular fingerprints for odor decoding.

Area of Science:

• Computational chemistry
• Cheminformatics
• Sensory science

Background:

• The relationship between molecular structure and odor perception is crucial for fragrance development.
• Accurate in silico odor prediction remains a challenge in sensory science.

Purpose of the Study:

• To compare machine learning approaches for predicting fragrance odors.
• To evaluate the effectiveness of molecular fingerprints versus classical descriptors.

Main Methods:

• Benchmarking Random Forest, eXtreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine.
• Utilizing functional group fingerprints, molecular descriptors, and Morgan structural fingerprints on a dataset of 8681 compounds.
• Assessing model performance using AUROC and AUPRC metrics.

Main Results:

• The Morgan-fingerprint-based XGBoost model achieved the highest discrimination performance (AUROC 0.828, AUPRC 0.237).
• Molecular fingerprints demonstrated superior representational capacity for olfactory cues compared to descriptor-based models.
• A continuous, interpretable scent space was revealed, aligning with perceptual and chemical relationships.

Conclusions:

• Molecular fingerprints, particularly Morgan fingerprints, are highly effective for in silico odor prediction.
• Machine learning, especially XGBoost, can significantly advance data-driven olfactory research.
• This work enables the next generation of predictive models for fragrance and sensory science.