MFG-GIN: A Multi-Fingerprint Graph Isomorphism Network for Allergenicity Prediction

Predicting chemical allergenicity is crucial for chemical safety and drug development, yet existing computational methods have limitations; molecular fingerprints overlook intermolecular relationships, whereas atom-based graphs capture intramolecular topology. We developed MFG-GIN, which integrates 12 types of RDKit fingerprints and k-nearest neighbor similarities, and utilizes GIN with Jumping Knowledge to capture molecular and neighborhood features through message passing. Self-supervised edge prediction preserves topology, and Focal Loss handles class imbalance. On the benchmark, MCC is 0.790 (0.742 ± 0.032 across four independent end-to-end seeds) and sensitivity 0.892, competitive with published traditional and graph-based models.