Structure-grouped QSAR validation reveals modest ranking performance for ToxRefDB-recorded lung neoplastic findings.

Organ-specific quantitative structure-activity relationship (QSAR) modeling is constrained by sparse, heterogeneous animal pathology data. We evaluated molecular structure and bioactivity data from the Toxicity Forecaster (ToxCast) program for a database-derived lung neoplastic pathology endpoint using a dataset of 560 standardized parent representations (53 positive and 507 database-supported otherwise-negative) in 328 structure-validation groups. Of 2189 contributing studies, 95.7% used oral administration and 1.6% inhalation; the inhalation-specific neoplastic subset contained one positive representation. Nested group-aware validation, whole-group bootstrap, full-pipeline label permutation and a design-qualified otherwise-negative sensitivity analysis were applied. Morgan ExtraTrees yielded a pooled out-of-fold area under the receiver operating characteristic curve (ROC-AUC) of 0.643 (95% confidence interval, 0.566-0.719) and average precision (AP) of 0.175 (0.119-0.282). Penalized Morgan logistic regression yielded ROC-AUC of 0.669 and AP of 0.202; paired confidence intervals did not establish ExtraTrees superiority. Restricting the otherwise-negative class to 427 representations meeting stricter evidence criteria yielded ExtraTrees ROC-AUC of 0.620 and AP of 0.198. ToxCast integration did not establish incremental discrimination. Screening 7216 non-training structures produced model-dependent rankings: ExtraTrees and logistic regression shared nine Top-25 entries, whereas only 4-methylbenzophenone from the nine-structure targeted review set appeared in both lists. The results support modest structure-associated ranking information for this heterogeneous database endpoint, but not pulmonary or inhalation carcinogenicity classification or validated prioritization utility. This study provides a structure-grouped benchmark for organ-specific computational toxicology and informs future comparisons of biological descriptors and the design of independent validation studies.