Insight into the Structure-Odor Relationship of Molecules: A Computational Study Based on Deep Learning

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in chemistry

Background:

• Molecular odors significantly impact human life, but experimental identification is impractical for large datasets, especially for hazardous compounds.
• Predicting molecular odors computationally is gaining traction as an efficient alternative to experimental methods.
• Understanding the structure-odor relationship is crucial for identifying and designing molecules with desired olfactory properties.

Purpose of the Study:

• To develop and evaluate computer-aided models for predicting molecular odors.
• To perform two-class (odor/no odor, fruity/no odor, floral/no odor, woody/no odor) and multi-class (fruity/flowery/woody/no odor) odor predictions.
• To assess the performance of multilayer perceptron (MLP) with physicochemical descriptors (MLP-Des), MLP with molecular fingerprints, and convolutional neural networks (CNN).

Main Methods:

• Utilized three machine learning models: MLP-Des, MLP with molecular fingerprints, and CNN.
• Trained and validated models on newly refined molecular odor datasets.
• Employed two-class and multi-class classification strategies for odor prediction.

Main Results:

• All three models demonstrated robust prediction capabilities for molecular odors.
• The MLP-Des model achieved the highest accuracy, with AUC values of 0.99 for two-class and 0.86 for multi-class predictions.
• The CNN model effectively extracted structural features related to odor from 2D molecular images.

Conclusions:

• The developed computational models provide accurate predictions of molecular odors.
• MLP-Des models excel in prediction and elucidating structure-odor relationships.
• These models facilitate the discovery of novel odorant and potentially hazardous molecules and enhance understanding of chemical structure-odor perception.