Mitigating silent failures in toxicity prediction: a conformalized heteroscedastic Bayesian framework

Accurate uncertainty quantification is a prerequisite for reliable toxicity assessments in drug discovery. Traditional QSAR models provide point estimates but fail to communicate prediction reliability, particularly for structurally complex compounds. We propose Conformalized Two-Stage Heteroscedastic BART (C2S-HBART), a novel framework addressing homoscedastic modelling assumptions. A first-stage BART model estimates the mean toxicity response; a second stage explicitly estimates local predictive variance from SMILES-derived structural descriptors. Split Conformal Prediction is then integrated to provide distribution-free validity guarantees. Evaluated on the Tox21 dataset, C2S-HBART achieves near-nominal coverage (0.952 vs. 0.880 for the uncalibrated baseline) and reduces the rate of silent failures - confident but incorrect predictions - from 12.02% to 1.8%. Compared to a standard conformalized baseline requiring wide intervals (Avg Width: 1.63), the heteroscedastic approach achieves equivalent safety with sharper predictions (Avg Width: 1.20), representing a 26% improvement in information efficiency. Variable importance analysis further reveals that molecular size and topological complexity are primary drivers of predictive uncertainty. C2S-HBART provides a statistically rigorous, transparent decision-support tool for preclinical screening, enabling toxicologists to prioritize safer compounds and flag structurally complex molecules for experimental validation.