Abstract
<jats:p>Molecular property prediction is a key computational tool for drug design, allowing researchers to predict several key parameters such as Absorption, Distribution, Metabolism, Excretion and Toxicity (ADMET) properties of candidate molecules. While the field has undergone rapid development in recent years, for example with the introduction of graph neural networks and chemical foundation models, the tool landscape is highly fragmented across diDerent packages, making the adoption of state-of-the-art approaches particularly cumbersome. To address this issue, we introduce MATCHA (Modelling and AI Toolkit for Chemistry in Healthcare Applications), an open-source Python package for various molecular property prediction algorithms, which we used to rank 3rd out of 370 participants in the recent OpenADMET-ExpansionRx molecular property prediction blind challenge. We provide three diDerent case studies tackling large-scale benchmarking on the OpenADMET-ExpansionRx challenge dataset, estimating prediction confidence via uncertainty quantification approaches and explaining black-box model decisions via medicinal chemistry inspired Local Interpretable Modelagnostic Explanations (LIME). By providing multiple programming interfaces and agentic skills for AI-assisted workflow automation, the package is suited for experts and non-experts alike to develop complex molecular property prediction workflows and integrate new algorithms. The package is available at https://github.com/emdgroup/matcha.</jats:p>