BRACE: Bayesian Residual-covariance Adaptive Calibration for Multi-Endpoint ADMET Prediction
Year: 2026
Type: preprint
Abstract: Deciding which ADMET assays to run on a compound requires a calibrated probability that it will pass each assay not yet measured, given those already run. A multi-task predictor supplies these via the cross-assay covariance of its prediction errors (residuals), but this covariance is usually estimated once, from historical data. It therefore misses correlations specific to a program’s chemical series (scaffolds) and knows nothing about assays unique to the program. With only hundreds to thousands of program compounds, we adapt this covariance rather than retrain the predictor. BRACE merges historical and program residuals by expectation–maximization into one covariance covering every assay, then corrects it per scaffold series with an Inverse-Wishart hierarchical model; an optional copula handles non-Gaussian data. A closed-form identity makes its posterior-mean covariance equal the Ledoit–Wolf shrinkage estimator exactly, so its prior strength comes from the data, without cross-validation. On a semi-synthetic 36-assay benchmark built from the residual covariance of about 780,000 real ADMET compounds, BRACE cuts the pass-probability mean absolute error by up to 76% versus the QComp multi-task baseline. Its always-on merging step matches or beats QComp in all 29 settings of a robustness sweep; the optional per-scaffold step helps when series-specific structure exists and hurts when it does not, and we map when to use it (chosen with hindsight here; an automatic rule is future work). On a real temporal split, the per-scaffold steps raise AUC-ROC from 0.789 to 0.860 and MCC from 0.37 to 0.53, and lower Brier from 0.187 to 0.152. However, about 96% of that gain comes from recalibrating each series’ mean and variance, which simple untuned shrinkage also achieves; BRACE’s distinctive cross-assay term adds about 4% here and matters mainly under strong cross-assay structure.
Source: ChemRxiv
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7221703901