ToxSTK: A multi-target toxicity assessment utilizing molecular structure and stacking ensemble learning.

Drug registration requires risk assessment of new active pharmaceutical ingredients or excipients to ensure they are safe for human health and the environment. However, traditional risk assessment is expensive and relies heavily on animal testing. Machine learning (ML) has been used as a risk assessment tool, providing less time, money, and involved animals than in vivo experiments. Despite that, the ML models often rely on a single model, which may introduce bias and unreliable prediction. Stacking ensemble learning is an ML framework that makes predictions based on multimodal outcomes. This framework performs well in quantitative structure-activity relationship (QSAR) studies. In this study, we developed ToxSTK, a multi-target toxicity assessment using stacking ensemble learning. We aimed to create an ML tool that facilitates toxicity assessments more affordably with reduced reliance on animal models. We focused on four key targets generally assessed in early-stage drug development: hERG toxicity, mTOR toxicity, PBMCs toxicity, and mutagenicity. Our model integrated 12 molecular fingerprints with 3 ML algorithms, generating 36 novel predictive features (PFs). These PFs were then combined to construct the final meta-decision model. Our results demonstrated that the ToxSTK model surpasses standard regression and classification metrics, ensuring it is highly reliable and accurate in predicting chemical toxicities within its application domain. This model passed the y-randomization test, confirming that the identified QSAR is robust and not due to random chance. Additionally, this model outperforms the existing ML methods for these endpoints, suggesting its effectiveness for risk assessment applications. We recommend incorporating this stacking ensemble learning framework into the chemical risk assessment pipeline to improve model generalization, accuracy, robustness, and reliability.