KAVNN-Dataset

\abstract{Adverse drug reactions pose significant risks to human health; thus, accurate assessment of Drug-Induced Liver and Kidney Injury (DILKI) is critical for drug safety and adhering to the 3R principles (Replacement, Reduction, Refinement). However, current computational strategies often encounter a trade-off: traditional machine learning offers feature-based interpretability but frequently demonstrates constrained predictive capability, whereas deep learning achieves higher performance but typically remains a mechanistic “black box”. To address this challenge, we present the Kolmogorov-Arnold Visible Neural Network (KAVNN), a model designed to accurately predict DILKI and pathological phenotypes while inferring biological mechanisms. KAVNN integrates chemical structures with multi-scale biological priors and employs Fourier series-based Kolmogorov-Arnold Networks (KANs) to replace standard Multi-Layer Perceptrons. This architecture enhances non-linear expressiveness with a compact parameter space. In the drug toxicity prediction task, KAVNN achieves superior average performance over the state-of-the-art baselines, with an approximate 15.5\% improvement observed in the pathological phenotype prediction task. Moreover, it exhibits robust interpretability by identifying toxicity-associated genes, tracing potential causal trajectories from molecular perturbations to tissue injury, and uncovering targets consistent with established toxicology. KAVNN thus bridges the gap between accuracy and interpretability, offering a transparent, mechanism-driven tool for early-stage safety assessment.}\keywords{Drug-Induced Liver and Kidney Injury (DILKI), Kolmogorov-Arnold Networks (KANs), Toxicity prediction, Pathological phenotype prediction}