A deep position-encoding model for predicting olfactory perception from molecular structures and electrostatics

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning

Background:

• Olfactory perception prediction from molecular structures is complex due to the discontinuous nature of smell perception.
• Existing methods often struggle to capture the nuances of odorant-receptor interactions.

Purpose of the Study:

• To introduce Mol-PECO, a deep learning model for predicting olfactory perceptions from molecular structures and electrostatics.
• To demonstrate Mol-PECO's superiority over traditional machine learning and graph neural network approaches.

Main Methods:

• Developed Mol-PECO, a deep learning model utilizing the Coulomb matrix for molecular representation and positional encoding.
• Trained and evaluated Mol-PECO on a comprehensive dataset of odor molecules and their descriptors.
• Compared Mol-PECO's performance against traditional machine learning methods and graph neural networks.

Main Results:

• Mol-PECO significantly outperforms traditional machine learning methods and graph neural networks in predicting olfactory perceptions.
• The learned molecular embeddings by Mol-PECO effectively capture the odor space.
• Mol-PECO enables global clustering of odor descriptors and local retrieval of similar odorant molecules.

Conclusions:

• Mol-PECO provides an effective deep learning framework for predicting olfactory perceptions.
• The Coulomb matrix offers a powerful alternative for molecular representation in olfactory prediction tasks.
• This research advances the understanding of olfactory mechanisms and molecular interactions.