Beyond QSARs – Quantitative Knowledge-Activity Relationships (QKARs) for Toxicity Prediction

Computational toxicology is vital for drug safety and risk assessment, yet traditional Quantitative Structure–Activity Relationship (QSAR) models rely solely on chemical structure, limiting their ability to predict toxicity when small structural variations lead to major biological differences. Recent advances in artificial intelligence (AI), including text embeddings and generative AI, offer new opportunities to enhance toxicity prediction by integrating broader chemical and biological knowledge. In this study, we present a novel framework, Quantitative Knowledge–Activity Relationships (QKARs), which leverage domain-specific knowledge to predict toxicity. QKAR models were developed for two endpoints, drug-induced liver injury (DILI) and drug-induced cardiotoxicity (DICT), using three levels of knowledge representation. Models incorporating comprehensive knowledge consistently outperformed those based on simpler representations, and performance was not strongly correlated with algorithmic complexity across five machine learning methods. Compared with QSARs built on identical datasets, QKARs achieved superior predictive performance for both DILI and DICT. Notably, QKARs more effectively differentiated structurally similar drugs with distinct liver toxicity profiles. We further explored hybrid models that combined knowledge-based and structure-based features, referred to as Q(K+S)ARs, which yielded additional accuracy improvements. Overall, our results demonstrate that QKARs provide a robust alternative to QSARs by using domain-specific knowledge, offering enhanced capabilities for predicting and understanding drug toxicity.