Application of machine learning methods in toxicology for transcriptomic data analysis

Introduction. The etiology of acute toxic hepatitis induces a multifaceted and time-dependent profile of transcriptomic remodeling, simultaneously engaging several pathogenic axes. For accurate etiological classification, classical differential expression analysis should be complemented by methods focused on recognizing multidimensional feature combinations and quantifying their contributions. The aim of this study was to evaluate the expression profile of a gene panel reflecting the antioxidant response, glutathione-dependent detoxification, cell cycle control, and programmed cell death for differentiating the etiology of acute toxic hepatitis and identifying the contribution of these markers to class discrimination.

Material and methods. The experiment was performed on 210 male rats. Models of carbon tetrachloride-induced, paracetamol-induced, and ethanol-induced toxic hepatitis were established. The expression of Nfe2l2, Nqo1, Hmox1, Sod1, Gclc, Gstm1, Gstp1, Gstt1, Ripk1, Chek1, Casp7 genes in liver tissue was assessed 24 and 72 hours after exposure. The level of statistical significance was evaluated using the Mann-Whitney U test followed by the Benjamini-Hochberg correction for multiple comparisons. For integrative analysis, a multiclass classification task was performed on tabular features using XGBoost, LightGBM, and CatBoost algorithms. Performance was assessed by ROC/AUC in a one-vs-rest scheme with a 75/25 hold-out split and hyperparameter tuning via stratified cross-validation. Interpretation was performed using SHAP and catboost-evaluation methods.

Results. Transcriptomic responses depended on the toxicant and time point, involving bidirectional shifts within functional pathways. All three algorithms demonstrated very high class discrimination (AUC 0.987–0.9998). For carbon tetrachloride, XGBoost achieved the highest AUC (0.9998), while for paracetamol and ethanol, CatBoost showed the most balanced performance (0.9967 and 0.9960, respectively). SHAP analysis revealed that discrimination was primarily driven by markers of oxidative stress and detoxification, foremost Hmox1, followed by Gstm1, Sod1, and Gstt1. CatBoostEvaluation confirmed a robust performance gain upon inclusion of Hmox1 (Score 23.145%; p=0.000089), Gclc (13.627%; p=0.000089), and Gstm1 (12.898%; p=0.000089), whereas Gstp1 and Chek1 did not demonstrate significant incremental contribution in this setting.

Limitations. The study was conducted on a single in vivo model, within an acute experimental design, and using specific toxicant doses. This may limit the extrapolation of the results to chronic exposure scenarios or other dose regimens, animals of a different species and sex.

Conclusion. The expression profile of a compact gene panel provides reliable etiological classification of acute toxic hepatitis. An approach combining classical statistics and interpretable machine learning establishes a foundation for subsequent validation on independent series and expansion to multimodal features in tasks of early diagnosis and assessment of corrective interventions.

Compliance with ethical standards. The study was approved by the Bioethics Committee of the Ufa Research Institute of Occupational Medicine and Human Ecology (protocol No. 06-09 dated 05.09.2024). The study was conducted in accordance with the European Convention for the Protection of Vertebrate Animals Used for Experimental and Other Scientific Purposes (ETS N 123) and with the Directive 2010/63/EC of the European Parliament and of the Council of 22 September 2010 on the Protection of Animals Used for Scientific Purposes.

Authors’ contribution.
Karimov D.O. – study concept and design, material processing, statistical analysis, application of machine learning methods, manuscript writing, editing, approval of the final version of the article, responsibility for the integrity of all its parts.

Conflict of interests. The authors declare no apparent and potential conflicts of interest in relation to the publication of this article.

Funding. The work was carried out as part of the state assignment for the industry research program of Rospotrebnadzor “Scientific substantiation of the national system for ensuring sanitary and epidemiological welfare, managing health risks and improving the quality of life of the population of Russia” for 2021–2025. clause 6.1.8, state registration number 121062100058-8.

Received: January 26, 2026 / Accepted: February 2, 2026 / Published: March 18, 2026

1. Ajisafe O.M., Adekunle Y.A., Egbon E., Ogbonna C.E., Olawade D.B. The role of machine learning in predictive toxicology: A review of current trends and future perspectives. Life Sciences. 2025; 378: 123821. https://doi.org/10.1016/j.lfs.2025.123821

2. Jia X., Wang T., Zhu H. Advancing computational toxicology by interpretable machine learning. Environ. Sci. Technol. 2023; 57(46): 17690–706. https://doi.org/10.1021/acs.est.3c00653

3. Bergen V., Kodella K., Srikrishnan S., Barrandon O., Anderson S., Rogers-Grazado M., et al. A large-scale human toxicogenomics resource for drug-induced liver injury prediction. Nat. Commun. 2025; 16(1): 9860. https://doi.org/10.1038/s41467-025-65690-3

4. Shin S., Lee C., Park T. Comprehensive analysis of high-throughput transcriptomics to distinguish drug-induced liver injury (DILI) phenotypes. Arch. Toxicol. 2025; 99(9): 3721–34. https://doi.org/10.1007/s00204-025-04089-x

5. Li T., Tong W., Roberts R., Liu Z., Thakkar S. Deep learning on high-throughput transcriptomics to predict drug-induced liver injury. Front. Bioeng. Biotechnol. 2020; 8: 562677. https://doi.org/10.3389/fbioe.2020.562677

6. Jiang J., van Ertvelde J., Ertaylan G., Peeters R., Jennen D., de Kok T.M., et al. Unraveling the mechanisms underlying drug-induced cholestatic liver injury: identifying key genes using machine learning techniques on human in vitro data sets. Arch. Toxicol. 2023; 97(11): 2969–81. https://doi.org/10.1007/s00204-023-03583-4

7. Ikoma K., Hosaka T., Ooka A., Shizu R., Yoshinari K. Machine learning on toxicogenomic data reveals a strong association between the induction of drug-metabolizing enzymes and centrilobular hepatocyte hypertrophy in rats. Int. J. Mol. Sci. 2025; 26(10): 4886. https://doi.org/10.3390/ijms26104886

8. Bifarin O.O. Interpretable machine learning with tree-based shapley additive explanations: Application to metabolomics datasets for binary classification. PLoS One. 2023; 18(5): e0284315. https://doi.org/10.1371/journal.pone.0284315

9. Hu Q., Chen Y., Zou D., He Z., Xu T. Predicting adverse drug event using machine learning based on electronic health records: a systematic review and meta-analysis. Front. Pharmacol. 2024; 15: 1497397. https://doi.org/10.3389/fphar.2024.1497397

10. Chenga H., Garg A., Das S.S., Ramamurthi N. Prediction of drug-induced nephrotoxicity using chemical information and transcriptomics data. J. Chem. Inf. Model. 2025; 65(10): 5139–51. https://doi.org/10.1021/acs.jcim.5c00141