Apoptosis signal-regulating kinase 1 (ASK1/MAP3K5) is a serine/threonine kinase regulating JNK/p38 signaling under oxidative and endoplasmic reticulum stress and is a therapeutic target in multiple diseases. We developed pIC50 prediction models for ASK1 inhibitors and evaluated them under three data-partitioning schemes. From 6,280 ChEMBL and PubChem BioAssay records, 5,671 IC50 records were retained. After unit harmonization, salt removal, canonical SMILES conversion, and median aggregation of identical structures, 2,371 compounds remained. ECFP4 fingerprints and 217 RDKit descriptors were used with LightGBM, SVR, and random forest models optimized by Optuna under random, Bemis-Murcko scaffold, and Butina cluster 5-fold splits. LightGBM pooled out-of-fold R² values were 0.818, 0.796, and 0.688. Mean nearest-neighbor Tanimoto similarities were 0.806, 0.745, and 0.668, decreasing with performance (r=0.956). The R² range across splits (0.130) exceeded within-split model differences (0.011-0.040). Of the compounds, 52.9% had ≥5 IC50 records within the same assay. Among 704 corresponding ChEMBL records, 665 (94.5%) originated from 18 patents and 39 from five articles. Removing compounds with repeated IC50 records reduced Murcko-split R² from 0.795 to 0.771. These findings suggest that random splitting may provide optimistic estimates of generalization to structurally novel compounds and support evaluation using chemically informed splits.