Guiding the rational design of biocompatible metal-organic frameworks for drug delivery using machine learning

In drug delivery, metal-organic frameworks (MOFs) have emerged as an interesting paradigm owing to their tuneable porosity, structural diversity, and ease of external surface functionalization. Compared to other drug delivery systems, state-of-the-art MOFs can encapsulate up to 50 times higher quantities of drugs per unit of mass while showing very long release times. Nevertheless, for their translation into clinical applications, concerns regarding MOF biocompatibility necessitate comprehensive mitigation. Unfortunately, experiments are resource and time-intensive, while modeling approaches fail to capture the behavior of MOFs in intricate biological systems. Herein, we report a novel computational pipeline guided by machine learning (ML) for probing the biocompatibility of MOFs. The pipeline is designed to expedite the assessment of potential MOF toxicity based on the chemical properties of their precursors. Interpretable ML models were built on a database of over 35,000 organic molecules, predicting the potential toxicity of MOF linkers with 83% accuracy. Additionally, we established a comprehensive database cataloging the toxicity of MOF metallic centers. Leveraging these, we have screened the Cambridge Structural Database (CSD) containing 86,000 non-disordered MOFs, identifying existing and future promising candidates with minimal toxicity profiles for drug delivery applications. Beyond high-throughput screening, the developed models shed light on the chemical landscape associated with high biocompatibility precursor molecules, enabling the derivation of guidelines for the rational design of biocompatible MOFs. This framework thus expedites the identification of biocompatible materials, while facilitating deeper insights into their underlying chemistry.