Multi-task learning (MTL) is a machine learning strategy that is investigated for its potential to address the inherent complexity and variability in pharmacokinetic and pharmacodynamic profiling. Recent studies underscore MTL utility in improving ADMET (Absorption, Distribution, Metabolism, Elimination, and Toxicity) and potency predictions, foundational for efficient drug design.

Building on our observations, we introduce the OneADMET dataset—a comprehensive curated resource from public datasets ready to train and validated MTL models simultaneously processing hundreds of continuous tasks relevant for drug design and bioactivities predictions.

We observed that multi-task models are as predictive and sometime outperforms single-task models. Comparatively, MTL model deployment and maintenance are simpler and computationally more efficient for profiling. Based on the OneADMET dataset, we confirm the robustness of large-scale MTL for detailed pharmacokinetics profiling.