Corrigendum to “Kolmogorov–Arnold networks for predicting drug–gene associations of HDAC1 inhibitors in periodontitis” [Comput. Biol. Chem. 118 (2025) 108451] (Computational Biology and Chemistry (2025) 118, (S1476927125001112), (10.1016/j.compbiolchem.2025.108451))

The authors regret that the combination of a high threshold-based accuracy (reported as 96.49 %) with a comparatively low receiver operating characteristic (ROC) area under the curve (AUC = 0.55) may be interpreted as inconsistent. We clarify that these metrics capture different evaluation views of the same Kolmogorov–Arnold Network (KAN) model. Threshold-specific classification performance. The published confusion matrix (TN = 41, FP = 2, FN = 2, TP = 69) indicates strong performance at the chosen decision threshold, yielding an accuracy of 96.49 %, a precision of 0.972, a recall of 0.972, a specificity of 0.953, an F1 of 0.972, a balanced accuracy of 0.963, and an MCC of 0.925. These metrics summarize the model's behavior for binary allocation of drug–gene pairs into non-interacting vs interacting classes, which was the practical objective of this study. Ranking performance across all thresholds. The ROC AUC of 0.55 reflects global ranking ability over varying thresholds and was computed from uncalibrated score outputs (prior to any probability calibration). While the model performs well at the chosen operating point, its discrimination across all thresholds is limited, aligning with the reported AUC. This clarification tempers the interpretation of our results: the KAN showed high accuracy at the stated threshold for classifying candidate HDAC1 drug–gene associations, but its overall ranking ability across thresholds (AUC) was modest. The biological motivation for focusing on HDAC1 (epigenetic modulation of inflammatory and osteoclastic pathways in periodontitis) remains supported by the literature cited in the article, and the study's conclusions are accordingly reframed to emphasize threshold-specific classification utility rather than global ranking.