In July 2025, AI-driven drug discovery company CALICI and a joint research team from Yeungnam University(YU) announced groundbreaking results that could shift the paradigm in drug toxicity prediction. Their latest paper, published in Briefings in Bioinformatics, demonstrates that certain halogenated structures—long avoided in drug design—can, under specific conditions, actually reduce toxicity.

Key Findings

• Conventional belief: Halogen substitutions (fluorine, chlorine, bromine, iodine) increase drug toxicity.
• New discovery: Some halogens, especially iodine, can reduce hepatotoxicity and cardiotoxicity.
• Polyhalogenation: Multiple halogen substitutions can enhance bioactivity, improve metabolic stability, and lower toxicity.
• AI prediction scope: Thousands of compounds and real drug structures containing 1–3 aromatic scaffolds were analyzed.

Technical Highlights
The HD-GEM (Hybrid Dynamic Graph-based Ensemble Model) integrates graph neural network (GNN) structural learning with chemical descriptor-based feature recognition.

• Outperforms existing prediction tools (ProTox, ADMETlab, etc.) in both accuracy and interpretability
• Incorporates advanced techniques such as SHAP-based feature selection, SMOTE imbalance correction, and Optuna optimization

Read the full paper: Briefings in Bioinformatics – Article Link