Beyond Deep Learning Dominance: A Scaffold-Aware Hybrid Framework for Robust Toxicity Prediction in Data-Scarce Regimes

Computational toxicology plays a pivotal role in modern drug discovery and environmental risk assessment; however, the reliability of predictive models on unseen chemical scaffolds remains a critical bottleneck. Deep learning architectures, despite their prevalence, are susceptible to ’silent failures’—yielding high-confidence errors on out-of-distribution data—which poses significant risks in safety-critical applications. In this study, we propose a ’Safe-byDesign’ hybrid framework to mitigate this epistemic uncertainty. Rather than relying on a single algorithmic paradigm, we integrate the explicit structural ’memory’ of classical Random Forests with the topological ’intuition’ of Graph Neural Networks (GNNs) via a transparent Stacking Ensemble (Logistic Regression) functioning as a Mixture of Experts (MoE). Evaluated across 12 Tox21 endpoints using a rigorous 5-seed benchmarking protocol under Nested Scaffold Split, our analysis demonstrates that classical models serve as an essential robustness layer in data-scarce regimes. Crucially, the Hybrid MoE functions as a fail-safe mechanism, preventing performance degradation in tasks where pure GNNs struggle to generalize (e.g., NR-AR-LBD), while leveraging complementary strengths in others (e.g., SR-MMP). Benchmarking against the industry-standard DeepChem GraphConv model reveals statistical equivalence (p > 0.05) in overall performance, with the Hybrid framework exhibiting superior predictive stability in 7 out of 12 assays, particularly in complex endpoints such as SR-p53 and Androgen Receptor targets. These findings suggest that the proposed architecture offers a viable pathway toward achieving state-of-the-art predictive power without compromising the interpretability required for regulatory acceptance.