Predictive Modelling Enables Exploration of the Anticancer Potential of Ruthenium(II) Polypyridyl Complexes

The rational design of metal-based anticancer agents is challenged by vast chemical diversity, limited experimental throughput, and the scarcity of standardized structure-activity relationship data.Here, we investigate the extent to which predictive machine learning models can be applied productively to ruthenium(II) polypyridyl complexes in the low-data regime.A dataset of 718 cytotoxicity measurements was curated from the literature and used to benchmark multiple molecular representations and regression architectures under increasingly stringent datasplitting strategies.We observe that both predictive performance and apparent chemical-space extrapolation depend strongly on the chosen representation, with diberent encodings yielding complementary predictions.Guided by an ensemble-based modeling strategy, we screened a large virtual library of ruthenium polypyridyl complexes and selected a small, chemically diverse subset for synthesis and biological evaluation.The experimentally measured cytotoxicities are in excellent agreement with model predictions, including the discrimination of pronounced activity diberences between closely related complexes.These results demonstrate that data-driven prioritization can support the discovery and fine-tuning of cytotoxic ruthenium polypyridyl complexes despite limited and heterogeneous training data, and provide a general framework for integrating predictive modeling into medicinal inorganic chemistry drug discovery.