Abstract
<jats:p>Automating scientific discovery has been a long-standing objective in the scientific community, and the recent advances in artificial intelligence, particularly large language models and agentic workflows, have made this goal increasingly feasible. In chemistry, however, many problems remain limited by sparse data and fragmentary prior knowledge, leading standard data-driven approaches to struggle. In these low-data settings, progress should instead be accelerated by autonomous systems that strategically generate new informative data. Here, we introduce Chelatron, an agentic platform for autonomous metal-ligand design that leverages large language model-based reasoning to coordinate an iterative, hypothesis-driven framework of molecular discovery via effective chemical space exploration. The platform emulates key elements of the scientific method by proposing chemically meaningful hypotheses, generating candidate molecules, evaluating molecular properties, and using accumulated evidence to guide subsequent exploration. As an initial demonstration, we apply Chelatron to the design of chelators for actinium, an element whose scarcity and radioactivity have left its coordination chemistry underexplored. Across two campaigns, Chelatron proposed 119 hypotheses for improving chelator design and screened nearly 65,000 candidate molecules. 3D metal-chelator structures were constructed and evaluated at the xTB-level of theory for around 2,500 chelators, of which 679 were predicted to bind actinium comparable to or more favorable than the reference chelator, macropa, and are under further investigation. These results demonstrate Chelatron as an operational framework to accelerate hypothesis-driven discovery in data-limited metal-ligand chemistry.</jats:p>