Machine learning modeling of acute inhalation toxicity using an RFA-RFR framework, supported by explainable AI and SALI-based activity cliff analysis

Abstract

The present study develops machine learning models to predict the acute inhalation toxicity (pLC₅₀) of a diverse set of 552 molecules. Following dataset curation from the Integrated Chemical Environment (ICE) database, 3D molecular structures were generated and molecular descriptors (∼58,000) were calculated. After unsupervised pruning and redundancy reduction, 853 descriptors remained, which were further refined to 47 descriptors. Final feature selection was carried out exclusively on the training set using the Red Fox Algorithm (RFA). Random Forest Regressor (RFR) models were then developed and interpreted using Shapley additive explanations (SHAP) values. The resulting RFA–RFR model demonstrated robust statistical performance (R²_tr = 0.907, R²_L10%O = 0.902, and Q₂F₃ = 0.887). Model robustness was confirmed through leave-10%-out cross-validation and external set validation method. Importantly, benchmarking against previous QSAR models for inhalation toxicity shows improved predictive accuracy and broader chemical applicability. Mechanistic interpretation revealed that toxicity is strongly influenced by molecular weight, hydrogen bond acceptor count, lipophilic-polar balance, and sp³-hybrid carbon with oxygen arrangements. This study highlights the value of integrating feature optimization with interpretable machine learning approaches for deriving predictive and mechanistically meaningful models of inhalation toxicity.