Trialblazer: A chemistry-focused predictor of toxicity risks in late-stage drug development

Failures in drug discovery due to intolerable levels of adverse effects remain a significant concern. These failures can profoundly impact drug development, especially when identified during the later stages. Despite the challenges and risks posed by unexpected late-stage toxicity, publicly available data and predictive methods for these events remain limited. To address the need for improved predictive methods, we compiled a new, high-quality data set comprising 1603 benign drugs and 238 drug candidates that failed during clinical studies or were withdrawn from the market due to toxicity reasons. This enabled the development of classifiers based on multilayer perceptrons (MLPs). The most suitable model ("Trialblazer") is trained on Morgan fingerprints combined with bioactivity profiles predicted based on molecular similarity. Consequently, applying the model does not require prior knowledge of the compound's specific biological target(s), making it particularly useful for profiling innovative compounds for which such information is typically unavailable. Trialblazer achieved ROC-AUC and MCC values of 0.87 and 0.47 during cross-validation. When applied to external data, the model effectively distinguished drugs consistent with their safety profiles, as reflected in pharmacovigilance data from the European Medicines Agency (EMA). These results suggest that the model's predictions may serve as an indicator tool to flag compounds with a potentially increased risk of toxicity. However, similar to most theoretical and experimental models, the approach should not be used as a strict filter for rejecting compounds. Trialblazer is publicly accessible via PyPI and an API-powered public web service.