Computational Toxicology最新文献_第4页

Machine learning-based structural analysis of OATP1B1 interactors/non-interactors: Discriminating toxic and non-toxic alerts for transporter-mediated toxicity 基于机器学习的OATP1B1相互作用物/非相互作用物的结构分析：对转运蛋白介导的毒性的毒性和无毒警报的区分

IF 2.9 Q2 TOXICOLOGY

Computational Toxicology

Pub Date : 2025-09-01 Epub Date: 2025-08-08 DOI: 10.1016/j.comtox.2025.100373

Shovanlal Gayen , Indrasis Dasgupta , Balaram Ghosh , Insaf Ahmed Qureshi , Partha Pratim Roy

This hepatic transporter, OATP1B1, plays a critical role in transporter-related toxic responses and drug-drug interactions (DDIs). Several drug-drug interactions associated with OATP1B1 are clinically reported during combination therapies of lipid-lowering statins with antihypertensive, antiviral, and antibiotic drugs.

In the present study, different molecular properties of OATP1B1-interactors and non-interactors were initially compared, and the results revealed a distinct pattern in molecular weight, hydrophobicity, and number of rotatable bonds between them. Further chemical space, scaffold content, and diversity analyses indicated that OATP1B1-interactors/non-interactors are structurally diverse. Recursive partitioning and Bayesian classification analyses, involving ECFP and FCFP fingerprints, highlighted critical structural features that may serve as alerts for toxic or non-toxic effects on OATP1B1-mediated toxicity. Other machine learning-based classification models were also constructed, where Support Vector Classifier (SVC) shows higher statistical significance and predictive ability (accuracy: 0.797; precision: 0.833, and recall: 0.758). Moreover, local and global SHAP analyses were also performed to explain the distinctive structural features of OATP1B1-interactors and non-interactors.

Overall, the study offers insights into structural determinants of OATP1B1 interactions and provides predictive models to distinguish interactors from non-interactors, which may aid in reducing transporter-related toxicity risks in drug development. The outcomes may assist in advancing the safety and performance of medicinal compounds.

这种肝脏转运蛋白OATP1B1在转运蛋白相关的毒性反应和药物-药物相互作用（ddi）中起关键作用。在降脂的他汀类药物与降压药、抗病毒药物和抗生素药物联合治疗期间，临床报道了几种与OATP1B1相关的药物-药物相互作用。在本研究中，我们首先比较了oatp1b1相互作用物和非相互作用物的不同分子性质，结果揭示了它们之间在分子量、疏水性和可旋转键数上的不同模式。进一步的化学空间、支架含量和多样性分析表明，oatp1b1相互作用物/非相互作用物具有结构多样性。涉及ECFP和FCFP指纹图谱的递归划分和贝叶斯分类分析强调了可能作为oatp1b1介导毒性毒性或无毒作用警报的关键结构特征。本文还构建了其他基于机器学习的分类模型，其中支持向量分类器（SVC）具有较高的统计显著性和预测能力(准确率：0.797；精密度：0.833，召回率：0.758)。此外，还进行了局部和全局SHAP分析，以解释oatp1b1相互作用体和非相互作用体的独特结构特征。总的来说，该研究提供了对OATP1B1相互作用的结构决定因素的见解，并提供了区分相互作用物和非相互作用物的预测模型，这可能有助于降低药物开发中与转运蛋白相关的毒性风险。这些结果可能有助于提高药用化合物的安全性和性能。

{"title":"Machine learning-based structural analysis of OATP1B1 interactors/non-interactors: Discriminating toxic and non-toxic alerts for transporter-mediated toxicity","authors":"Shovanlal Gayen , Indrasis Dasgupta , Balaram Ghosh , Insaf Ahmed Qureshi , Partha Pratim Roy","doi":"10.1016/j.comtox.2025.100373","DOIUrl":"10.1016/j.comtox.2025.100373","url":null,"abstract":"<div><div>This hepatic transporter, OATP1B1, plays a critical role in transporter-related toxic responses and drug-drug interactions (DDIs). Several drug-drug interactions associated with OATP1B1 are clinically reported during combination therapies of lipid-lowering statins with antihypertensive, antiviral, and antibiotic drugs.</div><div>In the present study, different molecular properties of OATP1B1-interactors and non-interactors were initially compared, and the results revealed a distinct pattern in molecular weight, hydrophobicity, and number of rotatable bonds between them. Further chemical space, scaffold content, and diversity analyses indicated that OATP1B1-interactors/non-interactors are structurally diverse. Recursive partitioning and Bayesian classification analyses, involving ECFP and FCFP fingerprints, highlighted critical structural features that may serve as alerts for toxic or non-toxic effects on OATP1B1-mediated toxicity. Other machine learning-based classification models were also constructed, where Support Vector Classifier (SVC) shows higher statistical significance and predictive ability (accuracy: 0.797; precision: 0.833, and recall: 0.758). Moreover, local and global SHAP analyses were also performed to explain the distinctive structural features of OATP1B1-interactors and non-interactors.</div><div>Overall, the study offers insights into structural determinants of OATP1B1 interactions and provides predictive models to distinguish interactors from non-interactors, which may aid in reducing transporter-related toxicity risks in drug development. The outcomes may assist in advancing the safety and performance of medicinal compounds.</div></div>","PeriodicalId":37651,"journal":{"name":"Computational Toxicology","volume":"35 ","pages":"Article 100373"},"PeriodicalIF":2.9,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144831426","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Development of mathematical new approach methods to assess chemical mixtures 发展新的数学方法来评估化学混合物

IF 2.9 Q2 TOXICOLOGY

Computational Toxicology

Pub Date : 2025-09-01 Epub Date: 2025-08-21 DOI: 10.1016/j.comtox.2025.100376

R. Broughton , M. Feshuk , Z. Stanfield , K.K. Isaacs , K. Paul Friedman

The Toxicity Forecaster (ToxCast) program contains targeted bioactivity screening data for thousands of chemicals, but chemicals are often encountered as co-exposures. This work evaluated the feasibility of using single chemical ToxCast data to predict mixture bioactivity assuming chemical additivity. Twenty-one binary mixtures and their single components, inspired by consumer product chemical exposures, were screened in concentration–response using a multidimensional in vitro assay platform for transcription factor activity. Three models were applied to simulate mixtures’ concentration-responses: concentration addition (CA), independent action (IA), and a model that treats the mixture as the most potent single chemical component (MP). Uncertainty in the modeled and observed mixture points of departure and full concentration-responses was considered using bootstrap resampling and a Bayesian statistical framework. Approximately 80 % of the predicted mixture point of departure values were within ±0.5 on a log₁₀-micromolar scale of the observed concentrations; a majority of these predicted points of departure were protective (90–96 %), whether using CA, IA, or MP derived with the screened single components, when compared to the observed mixture. For most mixtures, ≥80 % of the observed mixture concentration–response data points fell within the modeled 95 % prediction interval, suggesting it would be difficult to observe deviations from additivity when accounting for experimental and mixtures modeling uncertainties. As it is resource-prohibitive to screen all mixtures, a case study to estimate bioactivity:exposure ratios for mixtures of per- and polyfluoroalkyl chemicals demonstrated the utility of operationalizing existing ToxCast data with mixtures modeling that includes uncertainty to predict potential risk from co-exposures.

毒性预报（ToxCast）项目包含数千种化学物质的目标生物活性筛选数据，但化学物质通常是共同暴露的。本研究评估了假设化学可加性，使用单一化学ToxCast数据预测混合物生物活性的可行性。21种二元混合物及其单一成分，受到消费品化学暴露的启发，使用转录因子活性的多维体外检测平台进行浓度响应筛选。采用三个模型来模拟混合物的浓度-响应：浓度添加（CA），独立作用（IA），以及将混合物视为最有效的单一化学成分（MP）的模型。利用自举重采样和贝叶斯统计框架考虑了模型和观测的出发点和全浓度响应混合点的不确定性。大约80%的预测混合起点值在观测浓度的log10-微摩尔尺度上的±0.5以内；与观察到的混合物相比，无论是使用筛选的单一成分衍生的CA， IA还是MP，这些预测的出发点大多数都是保护性的（90 - 96%）。对于大多数混合物，≥80%的观察到的混合物浓度响应数据点落在建模的95%预测区间内，这表明当考虑实验和混合物建模的不确定性时，很难观察到可加性的偏差。由于对所有混合物进行筛选是资源限制的，一项估计全氟烷基和多氟烷基化学品混合物的生物活性暴露比的案例研究表明，利用现有ToxCast数据和包括不确定性在内的混合物建模来预测共同暴露的潜在风险是有用的。

{"title":"Development of mathematical new approach methods to assess chemical mixtures","authors":"R. Broughton , M. Feshuk , Z. Stanfield , K.K. Isaacs , K. Paul Friedman","doi":"10.1016/j.comtox.2025.100376","DOIUrl":"10.1016/j.comtox.2025.100376","url":null,"abstract":"<div><div>The Toxicity Forecaster (ToxCast) program contains targeted bioactivity screening data for thousands of chemicals, but chemicals are often encountered as co-exposures. This work evaluated the feasibility of using single chemical ToxCast data to predict mixture bioactivity assuming chemical additivity. Twenty-one binary mixtures and their single components, inspired by consumer product chemical exposures, were screened in concentration–response using a multidimensional <em>in vitro</em> assay platform for transcription factor activity. Three models were applied to simulate mixtures’ concentration-responses: concentration addition (CA), independent action (IA), and a model that treats the mixture as the most potent single chemical component (MP). Uncertainty in the modeled and observed mixture points of departure and full concentration-responses was considered using bootstrap resampling and a Bayesian statistical framework. Approximately 80 % of the predicted mixture point of departure values were within ±0.5 on a log<sub>10</sub>-micromolar scale of the observed concentrations; a majority of these predicted points of departure were protective (90–96 %), whether using CA, IA, or MP derived with the screened single components, when compared to the observed mixture. For most mixtures, ≥80 % of the observed mixture concentration–response data points fell within the modeled 95 % prediction interval, suggesting it would be difficult to observe deviations from additivity when accounting for experimental and mixtures modeling uncertainties. As it is resource-prohibitive to screen all mixtures, a case study to estimate bioactivity:exposure ratios for mixtures of per- and polyfluoroalkyl chemicals demonstrated the utility of operationalizing existing ToxCast data with mixtures modeling that includes uncertainty to predict potential risk from co-exposures.</div></div>","PeriodicalId":37651,"journal":{"name":"Computational Toxicology","volume":"35 ","pages":"Article 100376"},"PeriodicalIF":2.9,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144891954","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

On the comparability between studies in predictive ecotoxicology 预测生态毒理学研究的可比性

IF 2.9 Q2 TOXICOLOGY

Computational Toxicology

Pub Date : 2025-09-01 Epub Date: 2025-08-07 DOI: 10.1016/j.comtox.2025.100367

Christoph Schür , Kristin Schirmer , Marco Baity-Jesi

Comparability across in silico predictive ecotoxicology studies remains a significant challenge, particularly when assessing model performance. In this work, we identify key criteria necessary for meaningful comparison between independent studies: (i) the use of identical datasets that represent the same chemical and/or taxonomic space; (ii) consistent data cleaning procedures; (iii) identical train/test splits; (iv) clearly defined evaluation metrics, as subtle differences — such as alternative formulations of

R^{2}

— can lead to misleading discrepancies; and (v) transparent reporting through code and dataset sharing. Our review of recent literature on fish acute toxicity prediction reveals a critical gap: no two studies fully meet these criteria, rendering cross-study comparisons unreliable. This lack of comparability hampers scientific progress in the field. To address this, we advocate for the adoption of benchmark datasets with standardized cleaning protocols, version control, and defined data splits. We further emphasize the importance of precise metric definitions and transparent reporting practices, including code availability and the use of structured reporting or data sheets, to foster reproducibility and advance the discipline.

计算机预测生态毒理学研究的可比性仍然是一个重大挑战，特别是在评估模型性能时。在这项工作中，我们确定了在独立研究之间进行有意义比较所需的关键标准：(i)使用代表相同化学和/或分类空间的相同数据集；（ii）一致的数据清理程序；（iii）相同的训练/测试分割；（iv）明确定义的评估指标，因为细微的差异（例如R2的不同公式）可能导致误导性的差异；(v)通过代码和数据集共享进行透明报告。我们回顾了最近关于鱼类急性毒性预测的文献，发现了一个关键的差距：没有两项研究完全符合这些标准，使得交叉研究比较不可靠。这种可比性的缺乏阻碍了该领域的科学进步。为了解决这个问题，我们提倡采用带有标准化清理协议、版本控制和定义数据分割的基准数据集。我们进一步强调精确的度量定义和透明的报告实践的重要性，包括代码可用性和使用结构化报告或数据表，以促进可重复性和推进该学科。

{"title":"On the comparability between studies in predictive ecotoxicology","authors":"Christoph Schür , Kristin Schirmer , Marco Baity-Jesi","doi":"10.1016/j.comtox.2025.100367","DOIUrl":"10.1016/j.comtox.2025.100367","url":null,"abstract":"<div><div>Comparability across <em>in silico</em> predictive ecotoxicology studies remains a significant challenge, particularly when assessing model performance. In this work, we identify key criteria necessary for meaningful comparison between independent studies: (i) the use of identical datasets that represent the same chemical and/or taxonomic space; (ii) consistent data cleaning procedures; (iii) identical train/test splits; (iv) clearly defined evaluation metrics, as subtle differences — such as alternative formulations of <span><math><msup><mrow><mi>R</mi></mrow><mrow><mn>2</mn></mrow></msup></math></span> — can lead to misleading discrepancies; and (v) transparent reporting through code and dataset sharing. Our review of recent literature on fish acute toxicity prediction reveals a critical gap: no two studies fully meet these criteria, rendering cross-study comparisons unreliable. This lack of comparability hampers scientific progress in the field. To address this, we advocate for the adoption of benchmark datasets with standardized cleaning protocols, version control, and defined data splits. We further emphasize the importance of precise metric definitions and transparent reporting practices, including code availability and the use of structured reporting or data sheets, to foster reproducibility and advance the discipline.</div></div>","PeriodicalId":37651,"journal":{"name":"Computational Toxicology","volume":"35 ","pages":"Article 100367"},"PeriodicalIF":2.9,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144866339","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

In silico analyses as a tool for regulatory assessment of protein digestibility: Where are we? 计算机分析作为蛋白质消化率调节评估的工具：我们在哪里？

IF 2.9 Q2 TOXICOLOGY

Computational Toxicology

Pub Date : 2025-09-01 Epub Date: 2025-08-05 DOI: 10.1016/j.comtox.2025.100372

Fernando Rivero-Pino, Caroline Idowu, Hannes Malfroy, Diana Rueda, Hannah Lester

In silico tools are emerging as a valuable resource for predicting the behaviour of proteins, not only for the assessment of toxicity and allergenicity, but also for modelling digestion to study protein digestibility. These methods offer cost-effective, high-throughput alternatives to traditional in vitro and in vivo methods. Computational models simulate enzymatic digestion, allowing the analysis of protein cleavage and peptide release. Complementary tools such as molecular docking have also been proposed as part of the in silico battery of tests. Given their efficiency, in silico approaches could ultimately be proposed to support regulated product applications, particularly in assessing protein digestibility for novel foods. However, their acceptance and use in risk assessment remains uncertain due to a lack of validation in part due to conflicting findings cited in the literature − while some studies report strong correlations between in silico and in vitro digestibility results, others indicate significant discrepancies. This review critically evaluates the potential regulatory application of in silico protein digestibility models for use in novel food risk assessment, highlighting key challenges such as model standardization, validation against experimental data, and the influence of protein structure and digestion conditions. Future research should focus on refining model accuracy and establishing clear validation frameworks to enhance regulatory confidence in in silico digestion tools.

计算机工具正在成为预测蛋白质行为的宝贵资源，不仅用于评估毒性和过敏原，而且用于模拟消化以研究蛋白质消化率。这些方法为传统的体外和体内方法提供了成本效益高、高通量的替代方法。计算模型模拟酶消化，允许分析蛋白质裂解和肽释放。分子对接等补充工具也被提议作为硅电池测试的一部分。鉴于它们的效率，计算机方法最终可能会被提议用于支持受监管的产品应用，特别是在评估新食品的蛋白质消化率方面。然而，由于文献中引用的相互矛盾的研究结果缺乏验证，它们在风险评估中的接受和使用仍然不确定——尽管一些研究报告了体内消化率和体外消化率结果之间的强相关性，但其他研究表明存在显著差异。这篇综述批判性地评估了硅蛋白质消化率模型在新型食品风险评估中的潜在监管应用，强调了模型标准化、实验数据验证以及蛋白质结构和消化条件的影响等关键挑战。未来的研究应侧重于提高模型的准确性和建立明确的验证框架，以提高对硅消化工具的监管信心。

{"title":"In silico analyses as a tool for regulatory assessment of protein digestibility: Where are we?","authors":"Fernando Rivero-Pino, Caroline Idowu, Hannes Malfroy, Diana Rueda, Hannah Lester","doi":"10.1016/j.comtox.2025.100372","DOIUrl":"10.1016/j.comtox.2025.100372","url":null,"abstract":"<div><div><em>In silico</em> tools are emerging as a valuable resource for predicting the behaviour of proteins, not only for the assessment of toxicity and allergenicity, but also for modelling digestion to study protein digestibility. These methods offer cost-effective, high-throughput alternatives to traditional <em>in vitro</em> and <em>in vivo</em> methods. Computational models simulate enzymatic digestion, allowing the analysis of protein cleavage and peptide release. Complementary tools such as molecular docking have also been proposed as part of the <em>in silico</em> battery of tests. Given their efficiency, <em>in silico</em> approaches could ultimately be proposed to support regulated product applications, particularly in assessing protein digestibility for novel foods. However, their acceptance and use in risk assessment remains uncertain due to a lack of validation in part due to conflicting findings cited in the literature − while some studies report strong correlations between <em>in silico</em> and <em>in vitro</em> digestibility results, others indicate significant discrepancies. This review critically evaluates the potential regulatory application of <em>in silico</em> protein digestibility models for use in novel food risk assessment, highlighting key challenges such as model standardization, validation against experimental data, and the influence of protein structure and digestion conditions. Future research should focus on refining model accuracy and establishing clear validation frameworks to enhance regulatory confidence in <em>in silico</em> digestion tools.</div></div>","PeriodicalId":37651,"journal":{"name":"Computational Toxicology","volume":"35 ","pages":"Article 100372"},"PeriodicalIF":2.9,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144780871","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

The FAIR AOP roadmap for 2025: Advancing findability, accessibility, interoperability, and re-usability of adverse outcome pathways 2025年的FAIR AOP路线图：提高不利结果路径的可查找性、可访问性、互操作性和可重用性

IF 2.9 Q2 TOXICOLOGY

Computational Toxicology

Pub Date : 2025-09-01 Epub Date: 2025-07-21 DOI: 10.1016/j.comtox.2025.100368

Holly M. Mortensen , Maciej Gromelski , Ginnie Hench , Marvin Martens , Clemens Wittwehr , Saurav Kumar , Vikas Kumar , Karine Audouze , Vassilis Virvilis , Penny Nymark , Michelle Angrish , Iseult Lynch , Stephen Edwards , Barbara Magagna , Marcin W. Wojewodzic , The FAIR AOP Cluster Working Group

Adverse Outcome Pathways (AOPs) describe the mechanistic interactions of biological entities with a stressor (chemical, nanomaterial, radiation, virus, etc.) that produce an adverse response. How these interactions and associations are catalogued contributes to our ability to understand mechanistic effects and apply this knowledge to New Approach Methods (NAMs) that have the potential to reduce animal testing in chemical, biological, and material safety assessments. Making AOP data align with FAIR (Findable, Accessible, Interoperable, and Reusable) metadata standards relies on technical tools that implement and process AOP data and related metadata, and the establishment of coordinated and consensus computational bioinformatic methods. Herein current efforts in addressing the FAIRification of AOP mechanistic data and metadata, as well as the international, collaborative efforts to document, and improve the (re)-use and reliability of AOP information will be described. These coordinated efforts contribute to the establishment of a directive for the processing and storing of standardized AOP mechanistic data in the AOP-Wiki repository, and application of these data to next generation risk assessment.

不良后果途径（AOPs）描述了生物实体与应激源（化学物质、纳米材料、辐射、病毒等）产生不良反应的机制相互作用。如何对这些相互作用和关联进行分类有助于我们理解机制效应，并将这些知识应用于新方法（NAMs），这些方法有可能减少化学、生物和材料安全评估中的动物试验。使AOP数据与FAIR（可查找、可访问、可互操作和可重用）元数据标准保持一致，依赖于实现和处理AOP数据和相关元数据的技术工具，以及建立协调一致的计算生物信息学方法。本文将描述当前在处理AOP机制数据和元数据的标准化方面的努力，以及在记录和改进AOP信息的（再）使用和可靠性方面的国际协作努力。这些协调的工作有助于在AOP- wiki存储库中建立处理和存储标准化AOP机制数据的指令，并将这些数据应用于下一代风险评估。

{"title":"The FAIR AOP roadmap for 2025: Advancing findability, accessibility, interoperability, and re-usability of adverse outcome pathways","authors":"Holly M. Mortensen , Maciej Gromelski , Ginnie Hench , Marvin Martens , Clemens Wittwehr , Saurav Kumar , Vikas Kumar , Karine Audouze , Vassilis Virvilis , Penny Nymark , Michelle Angrish , Iseult Lynch , Stephen Edwards , Barbara Magagna , Marcin W. Wojewodzic , The FAIR AOP Cluster Working Group","doi":"10.1016/j.comtox.2025.100368","DOIUrl":"10.1016/j.comtox.2025.100368","url":null,"abstract":"<div><div>Adverse Outcome Pathways (AOPs) describe the mechanistic interactions of biological entities with a stressor (chemical, nanomaterial, radiation, virus, etc.) that produce an adverse response. How these interactions and associations are catalogued contributes to our ability to understand mechanistic effects and apply this knowledge to New Approach Methods (NAMs) that have the potential to reduce animal testing in chemical, biological, and material safety assessments. Making AOP data align with FAIR (Findable, Accessible, Interoperable, and Reusable) metadata standards relies on technical tools that implement and process AOP data and related metadata, and the establishment of coordinated and consensus computational bioinformatic methods. Herein current efforts in addressing the FAIRification of AOP mechanistic data and metadata, as well as the international, collaborative efforts to document, and improve the (re)-use and reliability of AOP information will be described. These coordinated efforts contribute to the establishment of a directive for the processing and storing of standardized AOP mechanistic data in the AOP-Wiki repository, and application of these data to next generation risk assessment.</div></div>","PeriodicalId":37651,"journal":{"name":"Computational Toxicology","volume":"35 ","pages":"Article 100368"},"PeriodicalIF":2.9,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144748772","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Screening for genotoxicants in food: A data-driven approach using food composition data and machine learning based in silico models 筛选食品中的基因毒物：使用食品成分数据和基于计算机模型的机器学习的数据驱动方法

IF 2.9 Q2 TOXICOLOGY

Computational Toxicology

Pub Date : 2025-09-01 Epub Date: 2025-07-24 DOI: 10.1016/j.comtox.2025.100370

Jakob Menz, Bernd Schäfer

Foods represent complex mixtures of constituents and contaminants, some of which may pose risks to health through genotoxic effects. We investigated the current capabilities and limitations of a data-driven approach for the systematic identification of genotoxic substances in food. To this end, we used machine learning to develop quantitative structure–activity relationship (QSAR) models aimed at predicting outcomes for three in vitro genotoxicity assays: the bacterial reverse mutation assay (Ames test), the in vitro chromosomal aberration test (CAvit) and the in vitro micronucleus test (MNvit). These models were applied to screen for putative dietary genotoxicants using the FooDB compound dataset (n = 70,477) as a case study. Overall, 6.6 % of the FooDB compounds were predicted as positive by at least one in silico model, while 7.1 % were predicted as negative by all three models. Depending on the predicted endpoint, between 77 % and 82 % of the FooDB compounds fell outside the model’s applicability domain or gave an equivocal prediction. Interestingly, of the 4,683 FooDB compounds predicted to be positive in at least one in vitro assay, only 491 could be mapped to an experimental data point. As a strategy to progress from in silico screening to risk assessment, we propose a tiered approach that integrates in silico modelling, exposure assessment and experimental testing. While it has to be acknowledged that current food composition databases and in silico models still have limitations, this work illustrates that data-driven approaches hold great promise for enhancing the identification of genotoxicants in foods.

食品是成分和污染物的复杂混合物，其中一些可能通过基因毒性作用对健康构成风险。我们调查了目前的能力和局限性的数据驱动的方法系统识别食品中的遗传毒性物质。为此，我们利用机器学习开发了定量结构-活性关系（QSAR）模型，旨在预测三种体外遗传毒性试验的结果：细菌反向突变试验（Ames试验）、体外染色体畸变试验（CAvit）和体外微核试验（MNvit）。这些模型被应用于筛选假定的膳食基因毒物，并以FooDB化合物数据集（n = 70,477）作为案例研究。总体而言，6.6%的FooDB化合物被至少一个硅模型预测为阳性，而7.1%的化合物被所有三个模型预测为阴性。根据预测端点的不同，77%到82%的FooDB化合物超出了模型的适用范围，或者给出了模棱两可的预测。有趣的是，在4683种FooDB化合物中，预计至少有一种在体外试验中呈阳性，但只有491种可以映射到实验数据点。作为一种从硅筛选到风险评估的策略，我们提出了一种集成硅建模、暴露评估和实验测试的分层方法。虽然必须承认，目前的食品成分数据库和计算机模型仍然有局限性，但这项工作表明，数据驱动的方法对加强食品中基因毒物的识别具有很大的希望。

{"title":"Screening for genotoxicants in food: A data-driven approach using food composition data and machine learning based in silico models","authors":"Jakob Menz, Bernd Schäfer","doi":"10.1016/j.comtox.2025.100370","DOIUrl":"10.1016/j.comtox.2025.100370","url":null,"abstract":"<div><div>Foods represent complex mixtures of constituents and contaminants, some of which may pose risks to health through genotoxic effects. We investigated the current capabilities and limitations of a data-driven approach for the systematic identification of genotoxic substances in food. To this end, we used machine learning to develop quantitative structure–activity relationship (QSAR) models aimed at predicting outcomes for three <em>in vitro</em> genotoxicity assays: the bacterial reverse mutation assay (Ames test), the <em>in vitro</em> chromosomal aberration test (CAvit) and the <em>in vitro</em> micronucleus test (MNvit). These models were applied to screen for putative dietary genotoxicants using the FooDB compound dataset (n = 70,477) as a case study. Overall, 6.6 % of the FooDB compounds were predicted as positive by at least one <em>in silico</em> model, while 7.1 % were predicted as negative by all three models. Depending on the predicted endpoint, between 77 % and 82 % of the FooDB compounds fell outside the model’s applicability domain or gave an equivocal prediction. Interestingly, of the 4,683 FooDB compounds predicted to be positive in at least one <em>in vitro</em> assay, only 491 could be mapped to an experimental data point. As a strategy to progress from <em>in silico</em> screening to risk assessment, we propose a tiered approach that integrates <em>in silico</em> modelling, exposure assessment and experimental testing. While it has to be acknowledged that current food composition databases and <em>in silico</em> models still have limitations, this work illustrates that data-driven approaches hold great promise for enhancing the identification of genotoxicants in foods.</div></div>","PeriodicalId":37651,"journal":{"name":"Computational Toxicology","volume":"35 ","pages":"Article 100370"},"PeriodicalIF":2.9,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144757488","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

An in silico protocol for endocrine activity assessment: Integrating predictions, experimental evidence, and expert reviews across estrogen, androgen, thyroid, and steroidogenesis modalities 内分泌活动评估的计算机程序：整合雌激素、雄激素、甲状腺和类固醇生成模式的预测、实验证据和专家评论

IF 3.1 Q2 TOXICOLOGY

Computational Toxicology

Pub Date : 2025-09-01 Epub Date: 2025-06-25 DOI: 10.1016/j.comtox.2025.100364

Candice Johnson , Sue Marty , Marlene Kim , Kevin Crofton , Alessandra Roncaglioni , Arianna Bassan , Tara Barton-Maclaren , Ana Domingues , Markus Frericks , Agnes Karmaus , Sunil Kulkarni , Elena Lo Piparo , Stephanie Melching-Kollmuss , Ray Tice , David Woolley , Kevin Cross

Endocrine disruption (ED) has been introduced as a new classification, labelling and packaging (CLP) hazard category under Regulation (EC) No 1272/2008. Additionally, consideration of endocrine-disrupting properties and endocrine-related effects continues to be an important aspect of chemicals management under the Canadian Environmental Protection Act (CEPA) 1999 for the prioritization and hazard characterization of potential hormone disrupting substances. To support chemical prioritization and hazard assessment, this study presents a structured in silico protocol for assessing endocrine activity across the estrogen (E), androgen (A), thyroid (T), and steroidogenesis (S) (EATS) modalities. The protocol integrates (Quantitative) Structure–Activity Relationship ((Q)SAR) predictions with experimental data using a structured approach grounded in a hazard assessment framework (HAF) and defines principles for evaluating the reliability and confidence of predictions. Key endpoints and model development opportunities are identified for each modality. Two case studies are presented to demonstrate the application of the protocol. In the assessment of 4-Chloro-1-[2,2-dichloro-1-(4-chlorophenyl)ethenyl]-2-(methylsulfonyl)benzene, structurally similar analogs supported a medium-confidence assessment of estrogen and androgen activity. Whereas, in the assessment of chloroprene, uncertainties due to potential metabolic transformation limited confidence in negative assessments. These case studies illustrate how model outputs, experimental evidence, an analysis of analogs, and expert review can be integrated to produce transparent and reproducible assessments. The framework supports a weight-of-evidence (WOE) non-testing approach for identifying endocrine-active substances.

根据法规（EC） No 1272/2008，内分泌干扰（ED）作为一个新的分类、标签和包装（CLP）危害类别被引入。此外，根据1999年加拿大环境保护法（CEPA），考虑内分泌干扰特性和内分泌相关影响仍然是化学品管理的一个重要方面，以确定潜在激素干扰物质的优先级和危害特征。为了支持化学物质优先级和危害评估，本研究提出了一种结构化的计算机程序，用于评估雌激素(E)、雄激素(a)、甲状腺(T)和类固醇生成(S)模式的内分泌活性。该方案使用基于危害评估框架（HAF）的结构化方法，将（定量）结构-活性关系（(Q)SAR）预测与实验数据相结合，并定义了评估预测可靠性和置信度的原则。为每个模态确定关键端点和模型开发机会。给出了两个案例研究来演示该协议的应用。在对4-氯-1-[2,2-二氯-1-（4-氯苯基）乙基]-2-（甲基磺酰基）苯的评估中，结构相似的类似物支持了对雌激素和雄激素活性的中等置信度评估。然而，在氯丁二烯的评估中，由于潜在代谢转化的不确定性限制了负面评估的可信度。这些案例研究说明了如何将模型输出、实验证据、类似物分析和专家评审结合起来，以产生透明和可重复的评估。该框架支持证据权重（WOE）非测试方法来识别内分泌活性物质。

{"title":"An in silico protocol for endocrine activity assessment: Integrating predictions, experimental evidence, and expert reviews across estrogen, androgen, thyroid, and steroidogenesis modalities","authors":"Candice Johnson , Sue Marty , Marlene Kim , Kevin Crofton , Alessandra Roncaglioni , Arianna Bassan , Tara Barton-Maclaren , Ana Domingues , Markus Frericks , Agnes Karmaus , Sunil Kulkarni , Elena Lo Piparo , Stephanie Melching-Kollmuss , Ray Tice , David Woolley , Kevin Cross","doi":"10.1016/j.comtox.2025.100364","DOIUrl":"10.1016/j.comtox.2025.100364","url":null,"abstract":"<div><div>Endocrine disruption (ED) has been introduced as a new classification, labelling and packaging (CLP) hazard category under Regulation (EC) No 1272/2008. Additionally, consideration of endocrine-disrupting properties and endocrine-related effects continues to be an important aspect of chemicals management under the Canadian Environmental Protection Act (CEPA) 1999 for the prioritization and hazard characterization of potential hormone disrupting substances. To support chemical prioritization and hazard assessment, this study presents a structured <em>in silico</em> protocol for assessing endocrine activity across the estrogen (E), androgen (A), thyroid (T), and steroidogenesis (S) (EATS) modalities. The protocol integrates (Quantitative) Structure–Activity Relationship ((Q)SAR) predictions with experimental data using a structured approach grounded in a hazard assessment framework (HAF) and defines principles for evaluating the reliability and confidence of predictions. Key endpoints and model development opportunities are identified for each modality. Two case studies are presented to demonstrate the application of the protocol. In the assessment of 4-Chloro-1-[2,2-dichloro-1-(4-chlorophenyl)ethenyl]-2-(methylsulfonyl)benzene, structurally similar analogs supported a medium-confidence assessment of estrogen and androgen activity. Whereas, in the assessment of chloroprene, uncertainties due to potential metabolic transformation limited confidence in negative assessments. These case studies illustrate how model outputs, experimental evidence, an analysis of analogs, and expert review can be integrated to produce transparent and reproducible assessments. The framework supports a weight-of-evidence (WOE) non-testing approach for identifying endocrine-active substances.</div></div>","PeriodicalId":37651,"journal":{"name":"Computational Toxicology","volume":"35 ","pages":"Article 100364"},"PeriodicalIF":3.1,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144569991","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

“RapidTox”: A decision-support workflow to inform rapid toxicity and human health assessment “ RapidTox ”：为快速毒性和人体健康评估提供信息的决策支持工作流

IF 2.9 Q2 TOXICOLOGY

Computational Toxicology

Pub Date : 2025-09-01 Epub Date: 2025-08-29 DOI: 10.1016/j.comtox.2025.100377

Jason C. Lambert , Jason Brown , Hui Gong , Curtis Kilburn , Jan Krysa , Brad Kuntzelman , Janet Lee , April Luke , Joshua Powell , Asif Rashid , James Renner , Risa Sayre , Jyothi Tumkur , Carl F. Valone , Chelsea Weitekamp , Russell S. Thomas

Regulatory bodies such as the U.S. Environmental Protection Agency are consistently faced with decisions pertaining to potential human health impacts of a diverse landscape of chemicals encountered in exposure matrices such as water, air, and soil. For legacy chemicals or those currently in commerce, decision contexts may range from emergency response to disasters where evaluation of potential threats to human health occurs on the order of hours to days, up to site- or media-specific assessment and remediation over the course of months to years. In addition, screening and prioritization of new chemicals or emerging contaminants represents an ever-present focus area for the regulatory community. A common theme across these overarching decision contexts is the need for assembling and integrating human health relevant data such as toxicity values and associated effects information. Various activities ranging from screening and prioritization to human health risk assessment of chemicals have historically been time and resource intensive, often requiring that practitioners consult and review a variety of disparate data streams to inform a given decision. In addition, many environmental chemicals are ‘data-poor’, lacking sufficient hazard data or toxicity values applicable to a given exposure scenario. In response, decision-based workflows have been developed and deployed in the RapidTox online platform wherein available toxicity values, hazard/effects data, physicochemical properties, and new approach methods-based data (e.g., read-across; cell-based bioactivity) have been assembled into data delivery modules. To date, the user interface design and expertly scoped content have been integrated in ‘screening human health assessment’ or ‘emergency response’ workflows to support decision-making.

美国环境保护署等监管机构一直面临着与水、空气和土壤等接触基质中所遇到的各种化学品对人类健康的潜在影响有关的决定。对于遗留化学品或目前在商业上的化学品，决策背景可能从紧急反应到灾害，其中对人类健康的潜在威胁的评估需要数小时到数天的时间，到针对特定场所或媒介的评估和补救需要数月到数年的时间。此外，新化学品或新出现污染物的筛选和优先级是监管界始终关注的重点领域。在这些总体决策背景下的一个共同主题是需要收集和整合与人类健康有关的数据，如毒性值和相关影响信息。从筛选和确定优先次序到化学品的人类健康风险评估，各种活动历来都是时间和资源密集型的，往往要求从业人员咨询和审查各种不同的数据流，以便为给定的决定提供信息。此外，许多环境化学品“缺乏数据”，缺乏足够的危害数据或适用于特定暴露情景的毒性值。作为回应，RapidTox在线平台开发并部署了基于决策的工作流程，其中可用的毒性值、危害/效应数据、物理化学特性和基于新方法的数据（例如，读取、基于细胞的生物活性）已组装成数据传递模块。迄今为止，用户界面设计和专业界定的内容已纳入“筛选人体健康评估”或“应急反应”工作流程，以支持决策。

{"title":"“RapidTox”: A decision-support workflow to inform rapid toxicity and human health assessment","authors":"Jason C. Lambert , Jason Brown , Hui Gong , Curtis Kilburn , Jan Krysa , Brad Kuntzelman , Janet Lee , April Luke , Joshua Powell , Asif Rashid , James Renner , Risa Sayre , Jyothi Tumkur , Carl F. Valone , Chelsea Weitekamp , Russell S. Thomas","doi":"10.1016/j.comtox.2025.100377","DOIUrl":"10.1016/j.comtox.2025.100377","url":null,"abstract":"<div><div>Regulatory bodies such as the U.S. Environmental Protection Agency are consistently faced with decisions pertaining to potential human health impacts of a diverse landscape of chemicals encountered in exposure matrices such as water, air, and soil. For legacy chemicals or those currently in commerce, decision contexts may range from emergency response to disasters where evaluation of potential threats to human health occurs on the order of hours to days, up to site- or media-specific assessment and remediation over the course of months to years. In addition, screening and prioritization of new chemicals or emerging contaminants represents an ever-present focus area for the regulatory community. A common theme across these overarching decision contexts is the need for assembling and integrating human health relevant data such as toxicity values and associated effects information. Various activities ranging from screening and prioritization to human health risk assessment of chemicals have historically been time and resource intensive, often requiring that practitioners consult and review a variety of disparate data streams to inform a given decision. In addition, many environmental chemicals are ‘data-poor’, lacking sufficient hazard data or toxicity values applicable to a given exposure scenario. In response, decision-based workflows have been developed and deployed in the RapidTox online platform wherein available toxicity values, hazard/effects data, physicochemical properties, and new approach methods-based data (e.g., read-across; cell-based bioactivity) have been assembled into data delivery modules. To date, the user interface design and expertly scoped content have been integrated in ‘screening human health assessment’ or ‘emergency response’ workflows to support decision-making.</div></div>","PeriodicalId":37651,"journal":{"name":"Computational Toxicology","volume":"35 ","pages":"Article 100377"},"PeriodicalIF":2.9,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144932314","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Part I. Systematic development of machine learning models for predicting mechanism-based toxicity from in vitro ToxCast bioassay data 第一部分系统开发机器学习模型，用于从体外ToxCast生物测定数据预测基于机制的毒性

IF 2.9 Q2 TOXICOLOGY

Computational Toxicology

Pub Date : 2025-09-01 Epub Date: 2025-07-25 DOI: 10.1016/j.comtox.2025.100371

Donghyeon Kim , Siyeol Ahn , Jiyong Jeong, Jinhee Choi

Artificial intelligence (AI) for toxicity prediction has gained significant attention as a potential new approach methodologies (NAMs) for next-generation risk assessment (NGRA). Among the various large toxicity data sources, the ToxCast database represents a valuable resource that is frequently used to develop AI models. To facilitate the regulatory adoption of such models, it is essential to identify those that offer both suitable predictive performance and clear relevance to regulatory endpoints. In this study, we systematically developed mechanism-based toxicity-prediction models using ToxCast bioassay data and sought to identify machine-learning models applicable to NGRA. We collected 1,485 bioassay datasets from InvitroDB v4.1 and pre-processed them for model training. Five types of molecular fingerprints (MACCS, Morgan, RDKit, Layered, and Pattern) and five machine-learning algorithms (logistic regression, decision tree, random forest, gradient boosting tree, and XGBoost) were applied to 980 bioassays, yielding 24,500 models. The best-performing model for each assay was selected according to the F1 score. Using annotations from the NTP ICE database, we ultimately selected 311 models trained on bioactivity data relevant to regulatory endpoints—including acute toxicity, developmental and reproductive toxicity, carcinogenicity, and endocrine disruption—that achieved acceptable performance (F1 score ≥ 0.5). Overall, this study provides a cornerstone for incorporating ToxCast-based AI models into NGRA.

人工智能（AI）毒性预测作为下一代风险评估（NGRA）的潜在新方法方法（NAMs）受到了广泛关注。在各种大型毒性数据源中，ToxCast数据库是经常用于开发人工智能模型的宝贵资源。为了促进此类模型的监管采用，必须确定那些既提供合适的预测性能又与监管端点明确相关的模型。在这项研究中，我们利用ToxCast生物测定数据系统地开发了基于机制的毒性预测模型，并试图确定适用于NGRA的机器学习模型。我们从InvitroDB v4.1中收集了1485个生物测定数据集，并对其进行预处理以进行模型训练。五种类型的分子指纹（MACCS、Morgan、RDKit、Layered和Pattern）和五种机器学习算法（逻辑回归、决策树、随机森林、梯度增强树和XGBoost）应用于980种生物分析，产生24,500个模型。根据F1评分选择各试验中表现最好的模型。利用NTP ICE数据库的注解，我们最终选择了311个模型，这些模型训练了与调节终点相关的生物活性数据，包括急性毒性、发育和生殖毒性、致癌性和内分泌干扰，这些模型达到了可接受的性能（F1评分≥0.5）。总的来说，本研究为将基于toxcast的人工智能模型纳入NGRA提供了基础。

{"title":"Part I. Systematic development of machine learning models for predicting mechanism-based toxicity from in vitro ToxCast bioassay data","authors":"Donghyeon Kim , Siyeol Ahn , Jiyong Jeong, Jinhee Choi","doi":"10.1016/j.comtox.2025.100371","DOIUrl":"10.1016/j.comtox.2025.100371","url":null,"abstract":"<div><div>Artificial intelligence (AI) for toxicity prediction has gained significant attention as a potential new approach methodologies (NAMs) for next-generation risk assessment (NGRA). Among the various large toxicity data sources, the ToxCast database represents a valuable resource that is frequently used to develop AI models. To facilitate the regulatory adoption of such models, it is essential to identify those that offer both suitable predictive performance and clear relevance to regulatory endpoints. In this study, we systematically developed mechanism-based toxicity-prediction models using ToxCast bioassay data and sought to identify machine-learning models applicable to NGRA. We collected 1,485 bioassay datasets from InvitroDB v4.1 and pre-processed them for model training. Five types of molecular fingerprints (MACCS, Morgan, RDKit, Layered, and Pattern) and five machine-learning algorithms (logistic regression, decision tree, random forest, gradient boosting tree, and XGBoost) were applied to 980 bioassays, yielding 24,500 models. The best-performing model for each assay was selected according to the F1 score. Using annotations from the NTP ICE database, we ultimately selected 311 models trained on bioactivity data relevant to regulatory endpoints—including acute toxicity, developmental and reproductive toxicity, carcinogenicity, and endocrine disruption—that achieved acceptable performance (F1 score ≥ 0.5). Overall, this study provides a cornerstone for incorporating ToxCast-based AI models into NGRA.</div></div>","PeriodicalId":37651,"journal":{"name":"Computational Toxicology","volume":"35 ","pages":"Article 100371"},"PeriodicalIF":2.9,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144748773","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Part II. Systematic development of machine learning models for predicting human and ecotoxicity from in vivo OECD test guideline data 第二部分。系统开发机器学习模型，用于从体内OECD测试指南数据预测人类和生态毒性

IF 3.1 Q2 TOXICOLOGY

Computational Toxicology

Pub Date : 2025-09-01 Epub Date: 2025-07-17 DOI: 10.1016/j.comtox.2025.100369

Donghyeon Kim , Jiyong Jeong , Siyeol Ahn, Jinhee Choi

Artificial intelligence (AI)-based toxicity prediction models have emerged as promising new approach methodologies (NAMs) to reduce reliance on traditional in vivo testing in chemical risk assessment. In this study, we systematically developed machine learning models using toxicity data generated in accordance with OECD Test Guidelines (TG), available in the eChemPortal database. The models targeted endpoints regulated under major chemical frameworks, including Korea’s Act on the Registration and Evaluation of Chemical Substances (K-REACH) and the Consumer Chemical Products and Biocides Safety Control Act (K-BPR), as well as the European Union’s Registration, Evaluation, Authorization and Restriction of Chemicals (EU REACH) and Biocidal Products Regulation (EU BPR). A comprehensive training dataset was curated by harmonizing dose descriptors, effect levels, and exposure routes. Model features were generated using four types of molecular fingerprints (MACCS, Morgan, RDKit, and Layered), and five machine learning algorithms—Logistic Regression, Decision Tree, Random Forest, Gradient Boosting Tree, and XGBoost—were trained. Model performance was evaluated using standard metrics, including F1 score, precision, recall, accuracy, AUC-ROC. In total, 680 models were developed for 17 TG-based endpoints. The best-performing model for each endpoint was selected based on its F1 score. Machine learning models predicting acute toxicity (TG 420, 402, 403), developmental toxicity (TG 414), carcinogenicity (TG 453), and ecotoxicity (TG 201, 202, 203, 210, 211) demonstrated acceptable performance (F1 score ≥ 0.5), whereas models for other endpoints require further improvement. Based on these findings, we suggest key challenges and considerations for applying machine learning models trained on OECD TG data to support next generation chemical risk assessment (NGRA).

基于人工智能（AI）的毒性预测模型已经成为一种有前途的新方法方法（NAMs），以减少对传统体内测试在化学品风险评估中的依赖。在这项研究中，我们根据eChemPortal数据库中提供的OECD测试指南（TG）生成的毒性数据，系统地开发了机器学习模型。这些模型针对的是受主要化学框架监管的终端，包括韩国的《化学物质注册和评价法》（K-REACH）和《消费化学产品和杀菌剂安全控制法》（K-BPR），以及欧盟的《化学品注册、评价、授权和限制法》（EU REACH）和《杀菌剂条例》（EU BPR）。通过协调剂量描述符、效应水平和暴露途径，编制了一个全面的训练数据集。使用四种类型的分子指纹（MACCS、Morgan、RDKit和Layered）生成模型特征，并训练五种机器学习算法（logistic Regression、Decision Tree、Random Forest、Gradient Boosting Tree和xgboost）。采用标准指标评估模型性能，包括F1评分、精度、召回率、准确率、AUC-ROC。总共为17个基于tg的终点开发了680个模型。根据每个端点的F1分数选择表现最佳的模型。预测急性毒性（TG 420, 402, 403），发育毒性（TG 414），致癌性（TG 453）和生态毒性（TG 201, 202, 203, 210, 211）的机器学习模型表现出可接受的性能（F1评分≥0.5），而其他终点的模型需要进一步改进。基于这些发现，我们提出了应用OECD TG数据训练的机器学习模型来支持下一代化学品风险评估（NGRA）的主要挑战和考虑因素。

{"title":"Part II. Systematic development of machine learning models for predicting human and ecotoxicity from in vivo OECD test guideline data","authors":"Donghyeon Kim , Jiyong Jeong , Siyeol Ahn, Jinhee Choi","doi":"10.1016/j.comtox.2025.100369","DOIUrl":"10.1016/j.comtox.2025.100369","url":null,"abstract":"<div><div>Artificial intelligence (AI)-based toxicity prediction models have emerged as promising new approach methodologies (NAMs) to reduce reliance on traditional in vivo testing in chemical risk assessment. In this study, we systematically developed machine learning models using toxicity data generated in accordance with OECD Test Guidelines (TG), available in the eChemPortal database. The models targeted endpoints regulated under major chemical frameworks, including Korea’s Act on the Registration and Evaluation of Chemical Substances (K-REACH) and the Consumer Chemical Products and Biocides Safety Control Act (K-BPR), as well as the European Union’s Registration, Evaluation, Authorization and Restriction of Chemicals (EU REACH) and Biocidal Products Regulation (EU BPR). A comprehensive training dataset was curated by harmonizing dose descriptors, effect levels, and exposure routes. Model features were generated using four types of molecular fingerprints (MACCS, Morgan, RDKit, and Layered), and five machine learning algorithms—Logistic Regression, Decision Tree, Random Forest, Gradient Boosting Tree, and XGBoost—were trained. Model performance was evaluated using standard metrics, including F1 score, precision, recall, accuracy, AUC-ROC. In total, 680 models were developed for 17 TG-based endpoints. The best-performing model for each endpoint was selected based on its F1 score. Machine learning models predicting acute toxicity (TG 420, 402, 403), developmental toxicity (TG 414), carcinogenicity (TG 453), and ecotoxicity (TG 201, 202, 203, 210, 211) demonstrated acceptable performance (F1 score ≥ 0.5), whereas models for other endpoints require further improvement. Based on these findings, we suggest key challenges and considerations for applying machine learning models trained on OECD TG data to support next generation chemical risk assessment (NGRA).</div></div>","PeriodicalId":37651,"journal":{"name":"Computational Toxicology","volume":"35 ","pages":"Article 100369"},"PeriodicalIF":3.1,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144714500","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0