Improving ADME Property Predictions by Integrating Public and Proprietary Data

Accurate predictions of compound properties are crucial for enhancing drug discov- ery by expediting processes and increasing success rates. This study focuses on predict- ing key pharmacokinetic endpoints related to Absorption, Distribution, Metabolism, and Excretion (ADME), leveraging extensive internal and newly available public hiqh- quality ADME data. We assess data-integration strategies for ADME prediction across six endpoints using single-source (internal or public), pooled single-task, and multi-task learning models. Models trained on combined data — especially multi-task models — generally outperform single-source baselines, with consistent gains on public tests and frequent gains on internal tests when public data complement and are proportionally balanced with in-house data size. Applicability domain analyses show that multi-task learning reduces error for compounds with higher similarity to the training space, indicating better generalization across combined spaces. Analysis of prediction uncer- tainties mirrors these observations. Our study underscores that curated integration of high-quality public datasets with proprietary data can deliver more accurate and better-calibrated in silico ADME models to support computational compound design in drug discovery.