Adverse Outcome Pathway and Machine Learning to Predict Drug Induced Seizure Liability

Abstract

Central nervous system (CNS) drugs have the highest clinical attrition, often due to CNS-related toxicities such as drug-induced seizures (DIS). Early prediction of DIS risk could reduce failure rates and optimize drug development by prioritizing testing in experimental models of DIS. Using seizure-relevant Adverse Outcome Pathways (AOPs) from various sources, we identified 67 seizure-associated protein targets. Biological activity data (EC₅₀, IC₅₀, K_i) for these targets were curated from ChEMBL, enabling development of ∼2000 regression and classification (random forest, support vector, XGBoost) models. Support vector regression (SVR) models achieved an average MAE of 0.54 ± 0.09 (-log M), while random forest classifiers yielded mean ROC AUC, accuracy, and recall of 0.88, 0.85, and 0.70, respectively (5-fold CV) across all targets. Multitarget XGBoost models concatenating ECFP6 fingerprints and target encodings (one-hot or ProtBERT) also demonstrated excellent overall performance, although their predictive accuracy was notably lower for leave-out sets compared to individual target-specific models. These models were used to predict activity for a seizure-liability data set with target-annotated DIS risk predictions. Overall, our findings support the utility of using target-specific machine-learning models for DIS prediction to aid in early toxicity testing prioritization and reduce CNS drug attrition.

Keywords: ProtBERT; adverse outcome pathway; central nervous system; drug induced seizures; machine learning; one-hot encoding.