Deep Learning for Digital Olfaction: Graph-Based Self-Supervised Learning

Exploring the relationship between a molecule’s structure and its olfactory characteristics is a longstanding challenge addressed through Quantitative Structure-Odor Relationship (QSOR) modeling. This research holds significant implications for diverse sectors, including food production, perfume formulation, environmental conservation, and sensory perception studies. Our study introduces a novel approach utilizing a Graph Neural Network (GNN) with contrastive self-supervised learning to predict molecule odor descriptions. The integration of GNN enhances the extraction of molecular structure embeddings, contributing to superior performance when incorporated into a fully connected neural network. Machine learning has already made a big difference in how we see and hear things with graph neural networks in molecular science, we aim to achieve similar advancements in understanding smells.