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

Area of Science:

• Materials Science
• Computational Chemistry
• Toxicology

Background:

• Metal-organic frameworks (MOFs) show great potential for drug delivery due to their unique properties.
• Clinical application of MOFs is hindered by the unknown toxicity of organic linkers under physiological conditions.
• Experimental toxicity assessment for the vast number of MOF linkers is infeasible.

Purpose of the Study:

• To develop an integrated machine learning framework for assessing the toxicity of MOF organic linkers.
• To enable systematic and scalable evaluation of linker safety for biomedical applications.
• To guide the design and selection of safer MOF linkers for drug delivery.

Main Methods:

• Trained four machine learning models (Graph Neural Network, Transformer, Random Forest, SVM) on acute toxicity data.
• Utilized ensemble predictions and applicability domain analysis for robust and reliable assessments.
• Integrated graph-based SHAP analysis for mechanistic interpretability and identification of toxic substructures.

Main Results:

• All models achieved high predictive performance (micro F1 up to ~0.87, ROC-AUC ~0.95-0.96) in cross-validation.
• Ensemble predictions on a large linker library demonstrated robustness and reduced bias.
• SHAP analysis identified known toxic molecular motifs, validating the model's interpretability.

Conclusions:

• The developed framework offers a scalable, accurate, and interpretable method for MOF linker toxicity assessment.
• This approach can significantly accelerate the screening of safe MOF candidates for drug delivery.
• Facilitates prioritizing linkers with favorable safety profiles for experimental validation.