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

Area of Science:

• Computational toxicology
• Cheminformatics
• Medicinal chemistry

Background:

• Halogens are crucial in drug design, affecting bioactivity and stability.
• Assessing halogen impact on drug toxicity (genotoxicity, cardiotoxicity, hepatotoxicity) is a significant challenge.
• Existing models often assume halogenation increases toxicity.

Purpose of the Study:

• To develop an AI framework (HD-GEM) for predicting toxicity of halogenated compounds.
• To investigate the specific contributions of halogens versus scaffold atoms to toxicity.
• To challenge conventional assumptions about halogenation and drug toxicity.

Main Methods:

• Hybrid Dynamic Graph-based Ensemble Model (HD-GEM) integrating graph neural networks, molecular fingerprints, and ensemble meta-learning.
• Node perturbation analysis to identify key atoms influencing toxicity predictions.
• Comparative analysis against conventional ML models and existing toxicity prediction tools (ProTox, ADMETlab, admetSAR).

Main Results:

• HD-GEM significantly outperforms existing models in toxicity prediction accuracy and ROC-AUC scores.
• Core scaffold atoms (carbon, nitrogen, oxygen) were found to be the primary drivers of toxicity.
• Halogen contributions to toxicity were minimal; iodine-substituted compounds showed the lowest toxicity.
• Polyhalogenated scaffolds exhibited reduced toxicity, potentially due to stabilized reactive metabolite formation.

Conclusions:

• HD-GEM provides an interpretable AI-driven framework for accurate toxicity prediction.
• The study challenges the paradigm that halogenation inherently increases drug toxicity.
• Scaffold composition and specific features are more critical determinants of toxicity than halogenation status.