Unlabeled but Not Unseen: Cytotoxicity Classification of Re(I) Tricarbonyl Complexes via K‑Means Clustering

The prediction of cytotoxicity for metal-based drug candidates remains a significant challenge due to the structural diversity and multifactorial mechanisms of action inherent to transition metal complexes. Here, we present an unsupervised machine learning approach employing K-means clustering to cluster the cytotoxicity of 225 rhenium-(I) tricarbonyl complexes based solely on molecular descriptors. After comprehensive descriptor calculation and reduction, principal component analysis was used to assess chemical space coverage and identify key variables. K-means clustering, applied without prior toxicity labels, successfully partitioned the data set into cytotoxic and noncytotoxic clusters, achieving high concordance with known biological activity and accurately clustering control compounds. Analysis of misclassified cases provided further insight into structural motifs associated with ambiguous toxicity profiles. This work demonstrates that K-means clustering, when integrated with robust descriptor selection and PCA, offers a transparent, efficient, and interpretable framework for early stage toxicity assessment in metal-based drug discovery, particularly when labeled data are limited.