Translational framework combining machine learning and in vivo screening for aminoglycoside ototoxicity prevention.

Aminoglycosides are potent antibiotics, but their clinical use is limited by ototoxicity. Traditional drug screening is labor-intensive and inefficient in identifying safer alternatives. To address this, we developed a rapid screening framework combining a graph attention-based transformer model with zebrafish-based validation. Due to the limited availability of ototoxicity-specific datasets, the model was trained using handcrafted labels derived from extensive literature review, along with transfer learning techniques. It was then used to screen FDA-approved drug libraries, prioritizing candidates with potential protective effects for in vivo validation with zebrafish. Among 28 compounds selected based on the model prediction, six-L-Glutamine, Ammonium Lactate solution, Malic Acid, Dexpanthenol, Calcium Citrate Tetrahydrate, and Strontium Ranelate-significantly protected zebrafish neuromast hair cells from gentamicin-induced damage. Structural analysis revealed that carboxyl and hydroxyl groups enhanced membrane permeability and antioxidant activity, contributing to efficacy, while hydrophobic bulky substituents reduced effectiveness. These findings demonstrate the utility of integrating data-driven prediction with in vivo validation and offer mechanistic insights for the rational design of otoprotective agents. The generated dataset also provides a valuable resource for future drug repurposing targeting aminoglycoside-induced ototoxicity.