Ensemble Machine Learning for Interpretable Prediction of Acute Toxicity in Metal-Organic Framework Linkers

Metal-organic frameworks (MOFs) are promising platforms for drug delivery due to their high porosity, tunable chemistry, and controlled release capabilities. However, their clinical translation is limited by insufficient safety data, particularly regarding the toxicity of organic linkers that may leach or degrade under physiological conditions. Experimental evaluation of linker biocompatibility across the vast MOF design space is impractical, creating a critical bottleneck for biomedical applications. Here, we present an integrated machine learning framework for systematic toxicity assessment of MOF organic linkers. Using acute toxicity data from the TOXRIC database, we trained four complementary modeling architectures: a directed message-passing graph neural network (Chemprop), a transformer-based SMILES model (ChemBERTa-2), a Random Forest using Morgan fingerprints, and a Support Vector Machine based on physicochemical descriptors. Across 5-fold stratified cross-validation, all models demonstrated strong predictive performance, achieving micro F1 scores up to ∼0.87, and microaveraged ROC-AUC values between ∼0.95 and 0.96, indicating robust discrimination despite substantial class imbalance. The trained models were applied to a large hypothetical MOF linker library, and ensemble predictions were used to improve robustness and reduce model-specific bias. Applicability-domain analysis based on chemical space overlap and nearest-neighbor similarity confirmed that predictions remained within well-supported regions of chemical space. To enable mechanistic interpretability, we integrated graph-based SHAP analysis with a SHAP-weighted Morgan fingerprint representation, allowing systematic identification of molecular substructures driving high-severity toxicity predictions. The most influential motifs showed strong agreement with independently documented toxic scaffolds, including polycyclic aromatic hydrocarbons. Together, this framework provides a scalable, accurate, and interpretable approach for prioritizing safe MOF linkers and guiding experimental screening toward candidates with favorable safety profiles.