Investigation of the structure-odor relationship using a Transformer model

Area of Science:

• Computational Chemistry
• Cheminformatics
• Machine Learning

Background:

• The relationship between molecular structure and odor is complex and not fully understood.
• Traditional methods struggle to consistently predict odor from molecular features.
• Advanced machine learning models offer new avenues for exploring structure-odor relationships.

Purpose of the Study:

• To develop and evaluate a Transformer model for predicting molecular properties, specifically substructures and odor descriptors.
• To interpret the model's predictions using attention mechanisms to understand structure-odor correlations.
• To identify key molecular substructures associated with specific odor qualities.

Main Methods:

• Collected SMILES data for 100,000 molecules to train a Transformer model for substructure prediction.
• Utilized attention matrices to visualize and interpret the model's focus on specific atoms and substructures.
• Applied the trained model to a dataset of 4462 molecules with odor descriptors to predict 98 odor qualities.

Main Results:

• Achieved a high F1 score of 0.98 in predicting molecular substructures.
• The attention mechanism accurately identified atoms within target substructures.
• The model inferred 98 odor descriptors with an average F1 score of 0.33.
• Identified potential structure-odor relationships for 19 descriptors with F1 scores > 0.45.

Conclusions:

• The Transformer model demonstrates strong performance in predicting molecular substructures.
• Attention visualization provides valuable insights into the model's decision-making process for substructure annotation.
• The study offers a data-driven approach to explore the intricate molecular structure-odor relationship, paving the way for predicting and designing molecules with desired olfactory properties.