OdorNet standardizes smell labels for 9,000 molecules

Scientific Data published OdorNet on October 2. The team built a SEA taxonomy from statistical co-occurrence, perfumery expertise and AI-assisted semantic alignment to address inconsistent source labels, redundant synonyms and skewed distributions. The dataset combines academic literature and open databases from 1960-2021 into more than 20,000 entries covering 9,000 unique molecules, organizing messy descriptors into a unified hierarchy. In technical validation, the same baseline model reached a Macro F1 of 0.42 with SEA labels, outperforming commonly used odor datasets and OdorNet without the taxonomy. It creates a foundation for digital odor editing and human-computer interaction, but 0.42 also shows that learning structure-odor relationships remains far from solved.