Odor-Sight: An Interpretable Graph Neural Network Web Platform for Prediction of Odorant Activity

Abstract

Understanding Structure–Odor Relationships remains a significant challenge due to the intrinsic noise and subjectivity of human olfactory data. Here, we present Odor-Sight 1.0, an open-source, graph-based deep learning web platform for binary classification of odorant versus odorless molecules. Trained on a rigorously curated data set of 4,201 compounds derived from OlfactionBase, the underlying Graph Neural Network achieved a Balanced Accuracy of 0.89 ± 0.011 and a Matthews Correlation Coefficient of 0.75 ± 0.022 across 15 repeated stratified splits, demonstrating strong and reproducible performance under class imbalance. Although classical machine learning models (Random Forest, Support Vector Machine, and XGBoost) trained on Morgan fingerprints achieved comparable aggregate accuracy, only the graph-based model supports bond-level explainability and an Applicability Domain (AD) defined directly in its own learned representation. To bridge the gap between statistical learning and chemical intuition, Odor-Sight integrates EdgeSHAPer, a bond-centric Explainable AI strategy that highlights molecular substructures contributing to model predictions. Furthermore, an embedding-based AD framework is implemented to assess the prediction reliability. The platform is supported by a scalable microservices architecture, enabling accessible and high-throughput analysis. Also, case studies across four distinct odorant classes, plus a borderline odorless compound, show that the model localizes each prediction on the expected osmophore. We also benchmarked Odor-Sight against a literature platform, Odorify, using the same validation metrics. Odor-Sight provides a transparent, reproducible, and practical tool for fragrance and flavor design. The platform is freely available at https://odorsight.labmol.com.br/odorsight, and the data and scripts can be found at https://github.com/LabMolUFG/OdorSight.