DeepNose: An Equivariant Convolutional Neural Network Predictive Of Human Olfactory Percepts

Area of Science:

• Computational neuroscience
• Chemosensation
• Artificial intelligence in sensory science

Background:

• The olfactory system uses odorant receptors (ORs) to interpret molecular signals and create smell perceptions.
• Odorant receptors (ORs) are hypothesized to function as 3D spatial filters, extracting essential molecular features for olfactory processing, analogous to filters in other sensory systems.

Purpose of the Study:

• To investigate the hypothesis that ORs act as 3D spatial filters by developing a computational model.
• To predict human olfactory percepts using 3D molecular structures and identify key molecular features driving odor perception.

Main Methods:

• A convolutional neural network (CNN) named DeepNose was trained to predict human olfactory percepts from semantic datasets.
• The network utilized an equivariant architecture for orientation invariance and processed multiple molecules simultaneously to infer mixture perception.

Main Results:

• DeepNose achieved high-fidelity perceptual predictions across various olfactory datasets.
• The model demonstrated invariance to molecular orientation and accurately predicted distinct perceptual qualities for different stereoisomers.
• The approach successfully identified molecular features contributing to specific perceptual descriptors and could infer the quality of odor mixtures.

Conclusions:

• DeepNose effectively uses 3D molecular shapes to generate accurate predictions of human olfactory percepts.
• The model provides a valuable tool for identifying molecular determinants of odor quality and advancing our understanding of the olfactory system.
• This AI-driven approach aligns with biological systems, offering insights into OR function and sensory processing.