KAVNN-Dataset

Motivation: Toxicity assessment is essential for drug development and clinical safety. However, a prevalent phenomenon in computational toxicology is the trade-off between predictive performance and mechanistic interpretability, where conventional machine learning offers greater transparency but deep learning typically achieves higher accuracy. This limitation hinders the adoption of models in regulatory decision-making, where understanding the biological reason for toxicity is as vital as the prediction itself. Results: We present ToxKAN, an adverse outcome pathway (AOP) knowledge-guided Kolmogorov-Arnold framework that integrates chemical structures and gene expression profiles for mechanistic drug toxicity prediction. Extensive evaluations demonstrate that ToxKAN outperforms state-of-the-art methods in binary toxicity prediction and achieves competitive or superior, ranking among the highest reported accuracy on fine-grained multi-label pathological phenotype prediction. Critically, the model identifies biologically coherent hierarchical reasoning paths, prioritizes core toxicity genes in low-signal settings where conventional differential expression analysis fails, and learns latent representations that reflect mechanism-aware compound organization. These results indicate that mechanism-aware architecture design can simultaneously advance predictive performance and biological interpretability. Availability and implementation: The source code and processed datasets are publicly available at https://github.com/shuangquanZW/ToxKAN and Edit upload | Zenodo.