Beyond Organic Molecules: MetalLipoDB, a Lipophilicity Data Set for Metal Complexes and Machine Learning Benchmarks

Lipophilicity governs the uptake, distribution, and toxicity of drug candidates, yet the experimental log P data sets used to build predictive models contain almost exclusively organic molecules. Here we present MetalLipoDB, a manually curated data set of 5158 experimental log P and log D values of complexes of 23 metals from 1192 articles, including 556 entries for multinuclear and heterometallic complexes, annotated with the measurement conditions and counterions. The data set quantifies the interstudy reproducibility of these measurements (0.56 log units) and serves as a benchmark for machine learning, with a mean absolute error of 0.72–0.75 for complexes from unseen articles. For articles with at least four measurements, anchoring the predictions to three measured complexes of the series reduced the error on the remaining complexes from 0.73 to 0.51. MetalLipoDB, the code, and an interactive web application (https://biometaldb.streamlit.app/) are openly available to support the design of metal-based drugs.