EUVIMEDEUVIMED entdecken
Crossref

A methodological audit of validation design, label provenance, and dataset lineage in machine learning for low-molecular-weight respiratory sensitization

Amr Saber, Abdelrahman Hassan Ibrahim

Computational Toxicology · 2026-12 · 40 · 100467

DOI: 10.1016/j.comtox.2026.100467 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Computational Toxicology, 2026-12. Verfasst von Amr Saber und 1 weiteren Autor:innen. Die Arbeit nennt 53 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Reliable machine-learning evaluation for low-molecular-weight respiratory sensitization is limited by inherited endpoints and structurally related validation compounds. Of 654 source records, two inherited-negative records failed parsing, leaving 652 canonical molecular-graph representations (136 positive, 516 negative). Three baseline models and their equal-weight consensus were evaluated internally. Across 20 complete split–fit–predict repetitions, mean consensus PR AUC was 0.6819 for random stratified and 0.6245 for Butina-grouped validation; the difference between unrounded procedure means was 0.0574. A duplicate-safe full-refit cluster bootstrap, refitting all preprocessing and models in 1000 resampled datasets, gave a mean difference of 0.0695; 92.4% of differences were positive, but the percentile interval (−0.0207, 0.1644) included zero. This differs from the historical conditional stored-prediction mean of 0.0564 (−0.0092, 0.1251). Composition matching attenuated the gap to 0.0440; 30- and 50-family sensitivities gave 0.0411 and 0.0408. Nearest-training similarity was associated with lower log loss, but these analyses do not causally isolate analogue sharing. Five repeated nested partitions gave a mean constrained-optimization change of −0.0071. At the primary 15% inner-error target, selective prediction achieved 71.3% coverage and retained balanced accuracy of 0.802 versus 0.751 for all predictions and 0.748 for fold-matched random retention. All models and inner-selected operating points were reported. OECD and ToxCML analyses assess provenance only. All predictive evaluations are internal; no endpoint-matched external validation, deployment, or regulatory generalization is claimed.

Zitieren

APA

Saber, A., & Ibrahim, A. H. (2026). A methodological audit of validation design, label provenance, and dataset lineage in machine learning for low-molecular-weight respiratory sensitization. Computational Toxicology, 40, 100467. https://doi.org/10.1016/j.comtox.2026.100467

Vancouver

Saber A, Ibrahim AH. A methodological audit of validation design, label provenance, and dataset lineage in machine learning for low-molecular-weight respiratory sensitization. Computational Toxicology. 2026;40:100467. doi:10.1016/j.comtox.2026.100467

BibTeX
@article{Saber2026,
  author = {Saber, Amr and Ibrahim, Abdelrahman Hassan},
  title = {A methodological audit of validation design, label provenance, and dataset lineage in machine learning for low-molecular-weight respiratory sensitization},
  journal = {Computational Toxicology},
  year = {2026},
  volume = {40},
  pages = {100467},
  doi = {10.1016/j.comtox.2026.100467}
}
Kostenlos starten →