Impact of halogenation on scaffold toxicity assessed using HD-GEM machine learning model.

Abstract

Halogens play a fundamental role in drug design, influencing bioactivity, stability, and selectivity. However, their impact on toxicity, particularly genotoxicity, cardiotoxicity, and hepatotoxicity, remains a critical challenge in drug discovery. This study presents HD-GEM (Hybrid Dynamic Graph-based Ensemble Model), a novel machine learning framework integrating graph neural networks, descriptor-based molecular fingerprints, and ensemble meta-learning to predict the toxicity of halogenated aromatic compounds and drug scaffolds. HD-GEM demonstrates superior predictive power compared to conventional machine learning (ML) models and popular toxicity web applications like ProTox, ADMETlab, and admetSAR, achieving high accuracy and Receiver Operating Characteristic-Area Under Curve scores across diverse datasets. Importantly, a node perturbation analysis revealed that carbon, nitrogen, and oxygen atoms within the scaffold dominate toxicity predictions, whereas halogen contributions were minimal, challenging the conventional assumption that halogenation inherently increases toxicity in many pharmacological contexts. Among halogens, iodine-substituted compounds exhibit the lowest toxicity, a trend corroborated across single-, double-, and triple-ring scaffolds. Notably, polyhalogenated scaffolds show reduced toxicity, suggesting a stabilizing effect that mitigates reactive metabolite formation. This study presents an interpretable artificial intelligence-driven framework for toxicity prediction in the context of computational toxicology and cheminformatics. Atom-level and descriptor-based analyses reveal scaffold- and feature-specific contributions to toxicity.