A review of machine learning in toxicology: current practices and reporting gaps

Clinical Snapshot

Key Findings

• P Value: Not applicable; descriptive review without hypothesis testing

• Effect Size: Approximately 50% of reviewed papers employed model interpretation methods to address 'black box' predictions; frequent gaps identified in reporting of missing value handling and availability of data and code (specific proportions not extractable from abstract alone)

• Primary Outcome: Characterisation of ML reporting practices in toxicology literature (2022–2024): prevalence of internal validation, model interpretability methods, missing data handling, and data/code availability

• Nnt Or Sensitivity: Not applicable; this is a methodological review, not a clinical intervention or diagnostic accuracy study. No NNT, sensitivity, specificity, or hazard ratio is calculable.

• Confidence Interval: Not reported; no confidence intervals provided for descriptive frequency estimates

Clinical Application

The review's findings are directly actionable for researchers and journal editors seeking to improve ML reporting standards. Implementation of recommended reporting practices (e.g., mandatory data/code availability, explicit missing data handling, use of interpretability methods) is feasible at low cost. Adoption would require editorial policy changes at toxicology journals and updated author guidelines. In Australia, the Australian Industrial Chemicals Introduction Scheme (AICIS, formerly NICNAS) and the Therapeutic Goods Administration (TGA) increasingly encounter ML-based toxicological assessments in chemical notification and preclinical drug dossiers. The reporting gaps identified — particularly around model interpretability and data availability — are directly relevant to regulatory submission quality. The RACGP and clinical pharmacology community would benefit from improved transparency in ML-derived toxicity predictions used to inform prescribing safety data. Australia's participation in OECD chemical safety frameworks means that any international movement toward standardised ML reporting (analogous to OECD QSAR guidance documents) would have direct regulatory implications domestically. The TGA's adoption of ICH guidelines for preclinical safety (ICH S7A/S7B) may need to be updated to address ML-specific validation and reporting requirements. Toxicologists, computational chemists, regulatory scientists, and clinical pharmacologists who develop, evaluate, or apply ML-based predictive models for toxicological risk assessment. Relevant to preclinical drug development teams, chemical safety assessors, and academic researchers publishing in the field.

Abstract

In recent years, machine learning and artificial intelligence approaches have been increasingly applied in the context of toxicological risk assessment. Many published overview, review, and comment papers discuss advantages, disadvantages, success stories, and open challenges for the application of machine learning models in toxicology. Machine learning methods using information from in vitro experiments can help to avoid animal experiments, thus allowing for larger numbers of experiments to be conducted. Drawbacks of machine learning models are the lack of mechanistic interpretability and the need for large amounts of high-quality data. In this work, we present a literature review of papers indexed in PubMed or published in the journal Computational Toxicology in the years 2022 to 2024, to assess the usage of machine learning methods in toxicology as well as the practices in reporting of methods and corresponding results. We do not address the suitability or the performance of methods, which is impossible to assess objectively without reanalysis on raw data, but focus on common practices and gaps in reporting. Major results are that many different machine learning methods are used in toxicology, often with appropriate internal validation. However, in only half of the cases, interpretation methods are used to address the problem that these models often make predictions as a black box. Moreover, there are very frequent gaps in reporting, in particular related to handling of missing values, and availability of data and code. Thus, this review can serve as a starting point for further tailored methodological research and guidance.

References

1. 1.Kappenberg, F., Stolte, M., Sauer, L., Duda, J. C., Lau, M., Schürmeyer, L., Zhou, H., Schwender, H., Schorning, K., & Rahnenführer, J. (2026). A review of machine learning in toxicology: current practices and reporting gaps. Archives of Toxicology. https://doi.org/10.1007/s00204-025-03987-2