Abstract
Uveitis is a severe ocular inflammatory disease with complex immune-mediated pathogenesis, posing significant challenges for drug discovery. While artificial intelligence has accelerated virtual screening, existing models often inadequately integrate heterogeneous molecular features or address disease-specific mechanisms. To address these gaps, we propose MLGT (Multimodal Learning with Graph and molecular descriptors for Therapeutics), a novel graph attention network based on GATv2 that synergistically integrates molecular graph topology, bond attributes, and physicochemical descriptors within a unified deep learning framework. The model employs dynamic attention mechanisms to capture non-local atomic interactions and a dual-stream fusion module to combine graph embeddings with molecular descriptors. To mitigate data imbalance and overfitting, we implement label smoothing, class-balanced sampling, and SMILES randomization. Evaluated on a rigorously curated Uveitis-related compound dataset from ChEMBL, MLGT achieves state-of-the-art performance: 97.7% accuracy, 97.2% F1 score, 96.1% recall, and an AUC-ROC of 0.9156, surpassing existing graph learning and classical machine learning benchmarks. Ablation studies confirm the essential roles of multimodal fusion and attention mechanisms. This study provides an efficient, attention-based computational tool for targeted Uveitis drug screening and establishes a scalable AI-driven paradigm for precision drug discovery in complex diseases.