Abstract
<jats:p>Metal-organic compounds have unique structural and electronic properties that make them attractive candidates for protein inhibition. However, the rational design and computational prioritization of such compounds remain challenging due to the limited availability of experimental bioactivity data. In contrast, large experimental databases have enabled substantial advances in data-driven approaches for purely organic inhibitors. Here, we investigated whether knowledge learned from organic inhibitor datasets can be transferred to the metal-organic domain. For this purpose, we curated two datasets of 134 ruthenium complexes associated with 62 distinct proteins, resulting in 201 pIC50 and 219 pKi values. We then developed a transfer-learning pipeline that combines self-supervised pre-training on organic inhibition data with contrastive semantic alignment to enable domain adaptation to metal-organic compounds. This strategy significantly improved discrimination between potent and weak inhibitors, as measured by the Matthews correlation coefficient (MCC), relative to baseline models without transfer learning. At the pharmacologically relevant threshold of 1 µM, performance increased from 0.48 to 0.62 on the pIC50 dataset and from 0.50 to 0.58 on the pKi dataset relative to the baseline model, approaching the upper limit imposed by experimental measurement variability of 0.71 and 0.68, respectively.</jats:p>