Machine Learning Modeling and Prediction of Transition-Metal Complex Properties

To accelerate the discovery of bioactive transition-metal complexes, we integrated combinatorial synthesis, automated structure generation, and machine learning to map a library of over 2,000 Re, Mn, Ir, and Ru complexes. Biological activities (including S. aureus and Leishmania inhibition) and physicochemical properties were experimentally determined. We developed MeCoBi (Metal Complex Builder), a novel framework for large-scale automated 3D structure generation, enabling a systematic comparison between 1D/2D, graph-based, and 3D-aware representations. Our machine learning models successfully classified biological activity and predicted HPLC retention times, demonstrating that 3D geometric descriptors offer distinct advantages over 2D representations in critical structural scenarios. This work underscores how combining combinatorial synthesis with data-driven modeling significantly accelerates the discovery and optimization of functional bioinorganic candidates.