TransIA: A transformer-based few-shot learning framework for detecting PDE5 inhibitor derivatives

The proliferation of phosphodiesterase-5 inhibitor (PDE5I) derivatives in food products and environmental samples poses a growing health concern, as these illicit additives (IAs) are often designed to evade conventional targeted detection methods. A cheminformatics tool capable of rapid and accurate screening of PDE5Is is therefore urgently needed for both food safety and environmental monitoring. To address this need, we developed TransIA (Transformer-based Illicit Additives Detection Network), which identifies PDE5Is from raw mass spectra. The model achieved 98% accuracy on the testing set with a false-positive rate below 3% across diverse environmental and biological matrices. Deployment across 197 real-world samples facilitated the identification of 21 IAs, of which 16 were previously unseen by our model (5 from the test set, 7 catalogued in PubChem but absent from our database, and 4 entirely unreported). This approach may offer a theoretical framework for integrating the model with portable mass spectrometry to assist on-site environmental analysis.