Interpretable multitask deep learning models for odor perception based on molecular structure.

Area of Science:

• Computational chemistry
• Chemoinformatics
• Machine learning

Background:

• Predicting odor from molecular structure is crucial for industries like fragrance and food.
• Current methods relying on sensory evaluation or manual feature engineering are inefficient and not scalable.
• Understanding structure-odor relationships aids in designing novel molecules with desired olfactory properties.

Purpose of the Study:

• To develop a multitask learning model for predicting multiple odor categories from chemical structures simultaneously.
• To capture shared representations across related odors for improved prediction.
• To provide a scalable and interpretable framework for rational olfactory design.

Main Methods:

• Developed a graph neural network-based multitask learning model (kMoL).
• Trained the model on experimental data covering 14 odor categories.
• Utilized Integrated Gradients for atom-level contribution analysis and UMAP/t-SNE for structure visualization.

Main Results:

• The multitask model (kMoL) demonstrated superior accuracy and stability compared to single-task models and Random Forests.
• Label co-occurrence analysis indicated that compounds often possess multiple odor characteristics, benefiting multitask learning.
• Atom-level analysis identified chemically relevant substructures, aligning with known olfactory receptor interaction sites.

Conclusions:

• The multitask learning approach effectively predicts odor categories from chemical structures, outperforming conventional methods.
• The model captures chemically and biologically relevant features, enhancing interpretability and mechanistic understanding.
• This framework offers a scalable and interpretable solution for designing molecules with specific olfactory profiles.