XGBoost odor prediction model: finding the structure-odor relationship of odorant molecules using the extreme...

Area of Science:

• Computational Chemistry
• Cheminformatics
• Machine Learning

Background:

• Determining structure-odor relationships is challenging due to ambiguous odor descriptors.
• Machine learning (ML) offers new approaches for quantitative structure-activity relationship (QSAR) studies in chemistry.
• Olfaction remains one of the least understood human senses.

Purpose of the Study:

• To develop an ML model for predicting odor descriptors from molecular structure.
• To investigate the structure-odor relationship using computational methods.
• To classify smells of odorant molecules based on their chemical properties.

Main Methods:

• Collected a dataset of 1278 odorant molecules with seven basic odor descriptors.
• Calculated 1875 physicochemical properties for each molecule.
• Employed Principal Component Analysis (PCA) for feature reduction.
• Developed an XGBoost model for odor prediction from SMILES strings.

Main Results:

• The XGBoost-PCA model achieved high precision (>99%) and sensitivity (>99%) in predicting seven basic smells on an independent test set.
• The model demonstrated superior performance compared to other recent studies in predicting common odor descriptors.
• Accurate prediction of odor classification from molecular structure was achieved.

Conclusions:

• The developed XGBoost-PCA model effectively predicts odor descriptors from molecular structure.
• This methodology advances the understanding of the complex structure-odor relationship.
• The approach provides a valuable tool for knowledge discovery in olfaction research.