Residual diagnostics and structural support in machine-learning prediction of mouse intraperitoneal acute toxicity

We investigated mouse intraperitoneal acute-toxicity prediction using source-traceable public records, residual diagnostics, and evaluation stratified by structural similarity. Rule-based curation and compound aggregation of TOXRIC and PubChem LD₅₀ records produced a 34,959-compound development pool and a reserved 1068-compound PubChem-only holdout. The studied 20× OOF discrepancy criterion excluded 448 compounds and retained 34,511 for ECFP4-count/RDKit-2D modeling with multi-seed XGBoost averaging. Database checks reproduced all 448 TOXRIC source values and modeling targets, establishing record traceability without identifying label error. Random and scaffold-based nested evaluations did not show improved R², RMSE, or MAE from exclusion on fixed outer-test populations. The final ensemble achieved a holdout R² of 0.7440, RMSE of 0.4703, and MAE of 0.2445 on the -log₁₀(LD₅₀,mol/kg) scale. Structural decomposition showed R² = 0.9794 for 641 compounds with fingerprint-identical development-set neighbors and R² = 0.4290 for the remaining 427 fingerprint-nonidentical compounds. Upstream-record overlap was identified within the fingerprint-identical group. The combined in-domain R² of 0.9136 included all 641 fingerprint-identical compounds and 92 similar but nonidentical compounds. These findings distinguish source-record traceability from model agreement and characterize prediction accuracy in relation to structural and provenance support. The released compound-level data, predictions, source checks, and code support reproducible endpoint-specific modeling and examination of its prediction limits.