SRMMP-CharQM, Physics-Informed Deep Learning for Toxicity Prediction: Quantum Mechanical Descriptors Enable Scaffold Hopping in Mitochondrial Membrane Potential Assays

Disruption of mitochondrial membrane potential (SR-MMP, Stress Response─Mitochondrial Membrane Potential) is a critical toxicity end point in early drug safety assessment. However, traditional prediction models based on molecular topology or sequences often struggle to resolve Activity Cliffs and suffer from limited generalization capabilities when facing unseen molecular scaffolds (i.e., scaffold-hopping scenarios), largely due to overfitting to specific structural fragments. To address this challenge, we propose SRMMP-CharQM, a physics-informed dual-branch deep learning framework. This model integrates BiGRU-based SMILES sequence encoding with global electronic descriptors (HOMO-LUMO gap and Total Energy) calculated via semiempirical quantum mechanical methods (xTB). Validated on a rigorous scaffold-split data set exhibiting significant label distribution shift (toxicity ratio: 6.400% in validation vs 28.9% in testing), SRMMP-CharQM demonstrated superior generalization performance. Experimental results show that the incorporation of quantum features boosted the model's performance on the external test set to an Area Under the Receiver Operating Characteristic Curve (AUC) of 0.854 and an Area Under the Precision-Recall Curve (AUPRC) of 0.693, achieving a significant improvement (approximately 16% relative gain in AUPRC) compared to the structure-only baseline. Mechanistic analysis reveals that quantum descriptors act as a critical physical inductive bias, empowering the model with reasoning capabilities beyond structural memorization. First, the model successfully resolved activity cliff cases where structurally similar compounds exhibit divergent toxicities, accurately identifying unstable, high-reactivity molecules via energy gap differences. Second, the predicted probability landscape reveals that the model captures complex nonlinear interactions: compounds with narrow gaps and moderate sizes are assigned the highest toxicity risk (probability > 0.600). Conversely, for ultralarge molecules with similarly narrow gaps, the model exhibits a learned "Steric Cutoff", correctly classifying them as safe.