SRMMP-CharQM, Physics-Informed Deep Learning for Toxicity Prediction: Quantum Mechanical Descriptors Enable Scaffold...

Area of Science:

• Computational chemistry and cheminformatics
• Drug discovery and development
• Toxicology and safety assessment

Background:

• Disruption of mitochondrial membrane potential (SR-MMP) is a key toxicity endpoint in early drug safety assessment.
• Traditional models struggle with activity cliffs and generalization to new molecular scaffolds due to overfitting.
• Existing methods lack the ability to capture complex structure-toxicity relationships beyond simple structural features.

Purpose of the Study:

• To develop a physics-informed deep learning framework for accurate prediction of SR-MMP.
• To improve generalization capabilities of toxicity prediction models, especially in scaffold-hopping scenarios.
• To address limitations of traditional models in resolving activity cliffs and handling diverse molecular structures.

Main Methods:

• Proposed a dual-branch deep learning framework, SRMMP-CharQM, integrating BiGRU-based SMILES encoding with quantum mechanical descriptors (HOMO-LUMO gap, Total Energy) using xTB.
• Validated the model on a scaffold-split dataset with significant label distribution shift.
• Employed mechanistic analysis to understand the role of quantum descriptors in model reasoning.

Main Results:

• SRMMP-CharQM demonstrated superior generalization performance on an external test set, achieving an AUC of 0.854 and AUPRC of 0.693.
• Incorporation of quantum features led to a significant improvement (approx. 16% relative gain in AUPRC) over structure-only baselines.
• The model successfully resolved activity cliffs by identifying high-reactivity molecules via energy gap differences and captured complex nonlinear interactions, including a learned 'Steric Cutoff' for large molecules.

Conclusions:

• Physics-informed deep learning, integrating quantum mechanical descriptors, significantly enhances toxicity prediction accuracy and generalization.
• Quantum features provide crucial inductive bias, enabling models to reason beyond structural memorization and resolve complex chemical phenomena like activity cliffs.
• SRMMP-CharQM offers a robust framework for early drug safety assessment, improving prediction of mitochondrial toxicity and identifying potential safety liabilities in novel chemical entities.