Computational Toxicology最新文献

O-QT assistant: a multi-agent AI system for streamlined chemical hazard assessment and read-across analysis using the OECD QSAR toolbox API O-QT助手：一个多代理人工智能系统，用于简化化学危害评估和使用OECD QSAR工具箱API的读取分析

IF 2.9 Q2 TOXICOLOGY

Computational Toxicology

Pub Date : 2026-03-01 Epub Date: 2025-12-08 DOI: 10.1016/j.comtox.2025.100395

Ivo Djidrovski , Raymond Pieters , Juliette Legler , Marc Teunis

The OECD QSAR Toolbox is a vital resource in regulatory toxicology for assessing chemical hazards and filling data gaps using in silico methods, supporting the move away from animal testing. However, manually interpreting its complex outputs (physicochemical properties, profiling results, experimental data) and synthesizing this information into consistent, justified assessment reports represents a significant bottleneck requiring substantial expert effort. To address this challenge, we developed the O-QT assistant: the first open-source (Apache 2.0 licensed) pipeline employing a multi-agent Large Language Model (LLM) system, featuring distinct agents for interpreting properties, environmental fate, reactivity, metabolism, QSAR predictions, experimental data, and read-across strategies. The system offers both automated analysis and a guided mode allowing user customization of scope and methods. We demonstrate the O-QT Assistant’s workflow using 1,1-diethoxyheptane (CAS 688–82-4), a fragrance ingredient, as a detailed case study, supplemented by characterization across nine additional chemicals. Its LLM agents, operating under constraints derived from structured prompts and the retrieved data, synthesized these findings into a narrative report and a comprehensive JSON log. This approach, validated across multiple chemicals demonstrating high factual accuracy (>99 %), enables full auditability of the AI interpretations against the source data.. The O-QT Assistant is freely available on GitHub at https://github.com/VHP4Safety/O-QT-OECD-QSAR-Toolbox-AI-assistant under the Apache 2.0 license. By automating key interpretation and reporting steps, the O-QT Assistant has the potential to significantly improve the efficiency and consistency of workflows involving OECD QSAR Toolbox data, promoting more standardized interpretations and potentially reducing variability in chemical safety assessments.

Scientific Contribution

An open-source multi-agent LLM assistant automating OECD QSAR Toolbox data interpretation and narrative report generation via its API for regulatory toxicology workflows.

经合组织QSAR工具箱是监管毒理学的重要资源，用于评估化学品危害和使用计算机方法填补数据空白，支持远离动物试验。然而，手动解释其复杂的输出（物理化学性质、分析结果、实验数据）并将这些信息合成为一致的、合理的评估报告是一个重大的瓶颈，需要大量的专家努力。为了应对这一挑战，我们开发了O-QT助手：第一个开源（Apache 2.0许可）管道，采用多代理大型语言模型（LLM）系统，具有不同的代理来解释属性、环境命运、反应性、代谢、QSAR预测、实验数据和读取策略。该系统提供自动分析和引导模式，允许用户自定义范围和方法。我们使用香料成分1,1-二氧基庚烷（CAS 688-82-4）作为详细的案例研究，并通过对另外九种化学物质的表征进行补充，展示了O-QT助手的工作流程。它的LLM代理在结构化提示和检索数据的约束下运行，将这些发现合成为叙述报告和全面的JSON日志。这种方法经过多种化学物质的验证，显示出较高的事实准确性（> 99%），可以针对源数据对人工智能解释进行全面审计。O-QT助手在Apache 2.0许可下可以在GitHub上免费获得https://github.com/VHP4Safety/O-QT-OECD-QSAR-Toolbox-AI-assistant。通过自动化关键解释和报告步骤，O-QT助手有可能显著提高涉及OECD QSAR工具箱数据的工作流程的效率和一致性，促进更标准化的解释，并有可能减少化学品安全评估的可变性。一个开源的多代理LLM助理，通过其API为监管毒理学工作流程自动化OECD QSAR工具箱数据解释和叙事报告生成。

{"title":"O-QT assistant: a multi-agent AI system for streamlined chemical hazard assessment and read-across analysis using the OECD QSAR toolbox API","authors":"Ivo Djidrovski , Raymond Pieters , Juliette Legler , Marc Teunis","doi":"10.1016/j.comtox.2025.100395","DOIUrl":"10.1016/j.comtox.2025.100395","url":null,"abstract":"<div><div>The OECD QSAR Toolbox is a vital resource in regulatory toxicology for assessing chemical hazards and filling data gaps using in silico methods, supporting the move away from animal testing. However, manually interpreting its complex outputs (physicochemical properties, profiling results, experimental data) and synthesizing this information into consistent, justified assessment reports represents a significant bottleneck requiring substantial expert effort. To address this challenge, we developed the O-QT assistant: the first open-source (Apache 2.0 licensed) pipeline employing a multi-agent Large Language Model (LLM) system, featuring distinct agents for interpreting properties, environmental fate, reactivity, metabolism, QSAR predictions, experimental data, and read-across strategies. The system offers both automated analysis and a guided mode allowing user customization of scope and methods. We demonstrate the O-QT Assistant’s workflow using 1,1-diethoxyheptane (CAS 688–82-4), a fragrance ingredient, as a detailed case study, supplemented by characterization across nine additional chemicals. Its LLM agents, operating under constraints derived from structured prompts and the retrieved data, synthesized these findings into a narrative report and a comprehensive JSON log. This approach, validated across multiple chemicals demonstrating high factual accuracy (>99 %), enables full auditability of the AI interpretations against the source data.. The O-QT Assistant is freely available on GitHub at https://github.com/VHP4Safety/O-QT-OECD-QSAR-Toolbox-AI-assistant under the Apache 2.0 license. By automating key interpretation and reporting steps, the O-QT Assistant has the potential to significantly improve the efficiency and consistency of workflows involving OECD QSAR Toolbox data, promoting more standardized interpretations and potentially reducing variability in chemical safety assessments.</div></div><div><h3>Scientific Contribution</h3><div>An open-source multi-agent LLM assistant automating OECD QSAR Toolbox data interpretation and narrative report generation via its API for regulatory toxicology workflows.</div></div>","PeriodicalId":37651,"journal":{"name":"Computational Toxicology","volume":"37 ","pages":"Article 100395"},"PeriodicalIF":2.9,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147396727","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

Finding optimal designs for estimating hormesis effect sizes 寻找估计激效大小的最佳设计

IF 2.9 Q2 TOXICOLOGY

Computational Toxicology

Pub Date : 2026-03-01 Epub Date: 2026-02-25 DOI: 10.1016/j.comtox.2026.100406

William Gertsch, Weng Kee Wong

There is increasing interest in hormesis as the phenomenon gains growing recognition in toxicology, public health and medicine. This paper is the first to focus on constructing model-based optimal designs for estimating the expected amount of hormesis using measures based on the area-under-the-curve (AUC). These measures can be used regardless of the choice of dose–response model. We first review optimal design techniques and propose new designs for estimating the degree or amount of hormesis with maximum precision at minimal cost when the sample size is large or small. We use a variety of state-of-the-art algorithms , including nature-inspired metaheuristic algorithms to search for the optimal designs. While the latter are not new, they are very flexible, fast and do not need technical assumptions for them to work well. Consequently, they are general purpose optimization algorithms and can be used to optimize any one or more objective functions for different models. When the experiment has two or more objectives, possibly with unequal interest, we find Pareto-optimal designs that balance the tradeoffs among the objectives. Experimental data are used to motivate and illustrate our approaches.

随着毒物学、公共卫生和医学对激效现象的认识日益加深，人们对激效的兴趣也越来越大。本文首次着重于构建基于模型的优化设计，利用基于曲线下面积（AUC）的度量来估计预期的激效量。无论选择何种剂量-反应模型，都可以使用这些措施。我们首先回顾了最佳设计技术，并提出了新的设计，用于在样本量大或小时以最小的成本以最大的精度估计激效的程度或数量。我们使用各种最先进的算法，包括自然启发的元启发式算法来搜索最优设计。虽然后者并不新鲜，但它们非常灵活、快速，并且不需要技术上的假设就能很好地工作。因此，它们是通用优化算法，可用于优化不同模型的任何一个或多个目标函数。当实验有两个或更多的目标时，可能有不相等的兴趣，我们发现帕累托最优设计平衡了目标之间的权衡。实验数据被用来激励和说明我们的方法。

{"title":"Finding optimal designs for estimating hormesis effect sizes","authors":"William Gertsch, Weng Kee Wong","doi":"10.1016/j.comtox.2026.100406","DOIUrl":"10.1016/j.comtox.2026.100406","url":null,"abstract":"<div><div>There is increasing interest in hormesis as the phenomenon gains growing recognition in toxicology, public health and medicine. This paper is the first to focus on constructing model-based optimal designs for estimating the expected amount of hormesis using measures based on the area-under-the-curve (AUC). These measures can be used regardless of the choice of dose–response model. We first review optimal design techniques and propose new designs for estimating the degree or amount of hormesis with maximum precision at minimal cost when the sample size is large or small. We use a variety of state-of-the-art algorithms , including nature-inspired metaheuristic algorithms to search for the optimal designs. While the latter are not new, they are very flexible, fast and do not need technical assumptions for them to work well. Consequently, they are general purpose optimization algorithms and can be used to optimize any one or more objective functions for different models. When the experiment has two or more objectives, possibly with unequal interest, we find Pareto-optimal designs that balance the tradeoffs among the objectives. Experimental data are used to motivate and illustrate our approaches.</div></div>","PeriodicalId":37651,"journal":{"name":"Computational Toxicology","volume":"37 ","pages":"Article 100406"},"PeriodicalIF":2.9,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147396569","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

Nanoparticle in vitro dosimetry via supervised machine learning 基于监督机器学习的纳米颗粒体外剂量测定

IF 2.9 Q2 TOXICOLOGY

Computational Toxicology

Pub Date : 2026-03-01 Epub Date: 2026-01-08 DOI: 10.1016/j.comtox.2026.100401

Linard David Hoessly , Sandor Balog

To advance the dosimetry of nanoparticles in the context of in vitro cell culture experiments (assays), we propose an inferential machine learning approach realized by supervising a deep neural network trained for function approximation as a substitute for nonparametric regression. This study explicitly addresses the limitations of current PDE-based models by introducing a supervised machine learning framework for parameter inference, ensuring predictive accuracy and interpretability for computational toxicology applications. The approach—exhaustively tested via Monte Carlo simulations—can quantitatively estimate fundamental parameters, such as the particle diffusion coefficient, particle settling velocity, and the probability of particle association with cells, directly from the temporal progression of dosimetry data. The results demonstrate that accurate analyses can be obtained through supervised machine learning, which has the capacity to define a key domain in the interpretation of in vitro assays dedicated to hazard and risk assessment of nanoparticles.

为了在体外细胞培养实验（测定）的背景下推进纳米颗粒的剂量测定，我们提出了一种推理机器学习方法，该方法通过监督用于函数近似训练的深度神经网络来实现，以替代非参数回归。本研究通过引入有监督的机器学习框架进行参数推断，明确解决了当前基于pde的模型的局限性，确保了计算毒理学应用的预测准确性和可解释性。该方法通过蒙特卡罗模拟进行了详尽的测试，可以直接从剂量学数据的时间进展中定量估计基本参数，如颗粒扩散系数、颗粒沉降速度和颗粒与细胞关联的概率。结果表明，通过监督机器学习可以获得准确的分析，这有能力定义一个关键领域的解释，致力于纳米颗粒的危害和风险评估的体外分析。

引用次数: 0

A weighted-likelihood framework for class imbalance in Bayesian prediction models 贝叶斯预测模型中阶级不平衡的加权似然框架

IF 2.9 Q2 TOXICOLOGY

Computational Toxicology

Pub Date : 2026-03-01 Epub Date: 2026-01-07 DOI: 10.1016/j.comtox.2026.100400

Stanley E. Lazic

Class imbalance is a pervasive problem in predictive toxicology, where the number of non-toxic compounds often exceeds the number of toxic ones. Models trained on such data often perform well on the majority class but poorly on the minority class, which is most relevant for safety assessment. We propose a simple and general Bayesian framework that addresses class imbalance by modifying the likelihood function. Each observation’s likelihood is raised to a power inversely proportional to its class proportion, with the weights normalised to preserve the overall information content. This weighted-likelihood (or power-likelihood) approach embeds cost-sensitive learning directly into Bayesian updating. The method is demonstrated using simulated binary data and an ordered logistic model for drug-induced liver injury (DILI). Weighting alters parameter estimates and decision boundaries, improving balanced accuracy and sensitivity for the minority (toxic) class. The approach can be implemented with minimal changes in standard probabilistic programming languages such as Stan, PyMC, and Turing.jl. This framework provides an easily extensible foundation for developing Bayesian prediction models that better reflect the asymmetric costs of safety-critical decisions.

在预测毒理学中，类不平衡是一个普遍存在的问题，即无毒化合物的数量往往超过有毒化合物的数量。在这些数据上训练的模型通常在多数类别上表现良好，但在与安全评估最相关的少数类别上表现不佳。我们提出了一个简单而通用的贝叶斯框架，通过修改似然函数来解决类不平衡问题。每个观测值的似然被提高到与其类比例成反比的幂，权重归一化以保持整体信息内容。这种加权似然（或幂似然）方法将代价敏感学习直接嵌入到贝叶斯更新中。用模拟二值数据和药物性肝损伤的有序logistic模型对该方法进行了验证。加权改变了参数估计和决策边界，提高了少数（有毒）类的平衡准确性和灵敏度。这种方法可以在Stan、PyMC和Turing.jl等标准概率编程语言中进行最小的修改来实现。该框架为开发贝叶斯预测模型提供了一个易于扩展的基础，该模型可以更好地反映安全关键决策的不对称成本。

{"title":"A weighted-likelihood framework for class imbalance in Bayesian prediction models","authors":"Stanley E. Lazic","doi":"10.1016/j.comtox.2026.100400","DOIUrl":"10.1016/j.comtox.2026.100400","url":null,"abstract":"<div><div>Class imbalance is a pervasive problem in predictive toxicology, where the number of non-toxic compounds often exceeds the number of toxic ones. Models trained on such data often perform well on the majority class but poorly on the minority class, which is most relevant for safety assessment. We propose a simple and general Bayesian framework that addresses class imbalance by modifying the likelihood function. Each observation’s likelihood is raised to a power inversely proportional to its class proportion, with the weights normalised to preserve the overall information content. This weighted-likelihood (or power-likelihood) approach embeds cost-sensitive learning directly into Bayesian updating. The method is demonstrated using simulated binary data and an ordered logistic model for drug-induced liver injury (DILI). Weighting alters parameter estimates and decision boundaries, improving balanced accuracy and sensitivity for the minority (toxic) class. The approach can be implemented with minimal changes in standard probabilistic programming languages such as <span>Stan</span>, <span>PyMC</span>, and <span>Turing.jl</span>. This framework provides an easily extensible foundation for developing Bayesian prediction models that better reflect the asymmetric costs of safety-critical decisions.</div></div>","PeriodicalId":37651,"journal":{"name":"Computational Toxicology","volume":"37 ","pages":"Article 100400"},"PeriodicalIF":2.9,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145939538","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

Molecular dynamics insights into antibiotic–microplastic interactions: mechanisms, environmental risks, and predictive perspectives 分子动力学洞察抗生素-微塑料相互作用：机制，环境风险和预测观点

IF 2.9 Q2 TOXICOLOGY

Computational Toxicology

Pub Date : 2026-03-01 Epub Date: 2026-01-09 DOI: 10.1016/j.comtox.2026.100402

K. Shashikala , V. Sudheesh , Deepa Janardanan , Suja P Devipriya

Microplastics (MPs) have evolved from being viewed as inert pollutants to dynamic vectors that alter the environmental behaviour of antibiotics, intensifying their persistence, transport, and ecotoxicological impact. Despite a surge of experimental and computational studies, inconsistencies in methodology, polymer selection, and environmental realism continue to obscure the mechanistic understanding of antibiotic-MP interactions. This critical review re-evaluates the current evidence, contrasting adsorption kinetics, isotherm models, and desorption dynamics reported across different microplastic-antibiotic systems. We examine how polymer composition, environmental ageing, and biofilm colonisation jointly modulate the strength and nature of antibiotic adsorption, while also addressing inconsistencies in reported adsorption behaviours that stem from overly simplified laboratory conditions. Molecular dynamics (MD) and quantum–mechanical (DFT) simulations have provided unprecedented atomistic insights into these interactions; yet, their predictive potential remains underexploited due to inconsistent parameterisation, limited simulation time scales, and weak integration with environmental data. By synthesising empirical observations with simulation results, this review identifies dominant interaction pathways, including hydrophobic, electrostatic, hydrogen bonding, and π–π stacking, and examines how these mechanisms are modulated by environmental variables such as pH, salinity, and natural organic matter. We further assess the emerging role of machine-learning-accelerated MD, hybrid QM/MM approaches, and multiscale digital-twin frameworks that aim to bridge molecular-scale processes with ecosystem-level behaviour. Finally, this review proposes a unified framework for standardising simulation protocols, integrating MD-derived energetics into environmental fate and transport models, and translating atomistic insights into regulatory and risk-assessment contexts. Collectively, these critical perspectives reposition MD simulations not merely as interpretive tools but as predictive engines essential for managing the intertwined challenges of microplastic pollution and antimicrobial resistance.

微塑料（MPs）已从被视为惰性污染物演变为改变抗生素环境行为的动态载体，增强其持久性、运输和生态毒理学影响。尽管实验和计算研究激增，但方法、聚合物选择和环境现实主义方面的不一致继续模糊了抗生素- mp相互作用的机制理解。这篇重要的综述重新评估了目前的证据，对比了不同微塑料-抗生素系统的吸附动力学、等温线模型和解吸动力学。我们研究了聚合物组成、环境老化和生物膜定植如何共同调节抗生素吸附的强度和性质，同时也解决了由于过度简化的实验室条件而导致的报告中吸附行为的不一致。分子动力学（MD）和量子力学（DFT）模拟为这些相互作用提供了前所未有的原子性见解；然而，由于不一致的参数化、有限的模拟时间尺度以及与环境数据的弱集成，它们的预测潜力仍未得到充分利用。通过综合经验观察和模拟结果，本综述确定了主要的相互作用途径，包括疏水、静电、氢键和π -π堆叠，并研究了这些机制如何受到环境变量（如pH、盐度和天然有机物）的调节。我们进一步评估了机器学习加速MD、混合QM/MM方法和多尺度数字孪生框架的新兴作用，这些框架旨在将分子尺度过程与生态系统级行为联系起来。最后，本文提出了一个统一的框架，用于标准化模拟协议，将md衍生的能量学整合到环境命运和运输模型中，并将原子见解转化为监管和风险评估背景。总的来说，这些关键的观点重新定位了MD模拟，不仅是解释工具，而且是管理微塑料污染和抗菌素耐药性交织挑战所必需的预测引擎。

{"title":"Molecular dynamics insights into antibiotic–microplastic interactions: mechanisms, environmental risks, and predictive perspectives","authors":"K. Shashikala , V. Sudheesh , Deepa Janardanan , Suja P Devipriya","doi":"10.1016/j.comtox.2026.100402","DOIUrl":"10.1016/j.comtox.2026.100402","url":null,"abstract":"<div><div>Microplastics (MPs) have evolved from being viewed as inert pollutants to dynamic vectors that alter the environmental behaviour of antibiotics, intensifying their persistence, transport, and ecotoxicological impact. Despite a surge of experimental and computational studies, inconsistencies in methodology, polymer selection, and environmental realism continue to obscure the mechanistic understanding of antibiotic-MP interactions. This critical review re-evaluates the current evidence, contrasting adsorption kinetics, isotherm models, and desorption dynamics reported across different microplastic-antibiotic systems. We examine how polymer composition, environmental ageing, and biofilm colonisation jointly modulate the strength and nature of antibiotic adsorption, while also addressing inconsistencies in reported adsorption behaviours that stem from overly simplified laboratory conditions. Molecular dynamics (MD) and quantum–mechanical (DFT) simulations have provided unprecedented atomistic insights into these interactions; yet, their predictive potential remains underexploited due to inconsistent parameterisation, limited simulation time scales, and weak integration with environmental data. By synthesising empirical observations with simulation results, this review identifies dominant interaction pathways, including hydrophobic, electrostatic, hydrogen bonding, and π–π stacking, and examines how these mechanisms are modulated by environmental variables such as pH, salinity, and natural organic matter. We further assess the emerging role of machine-learning-accelerated MD, hybrid QM/MM approaches, and multiscale digital-twin frameworks that aim to bridge molecular-scale processes with ecosystem-level behaviour. Finally, this review proposes a unified framework for standardising simulation protocols, integrating MD-derived energetics into environmental fate and transport models, and translating atomistic insights into regulatory and risk-assessment contexts. Collectively, these critical perspectives reposition MD simulations not merely as interpretive tools but as predictive engines essential for managing the intertwined challenges of microplastic pollution and antimicrobial resistance.</div></div>","PeriodicalId":37651,"journal":{"name":"Computational Toxicology","volume":"37 ","pages":"Article 100402"},"PeriodicalIF":2.9,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145978245","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

Federation of toxicological data resources for in silico new approach methodologies (NAMs) 计算机新方法（NAMs）毒理学数据资源联合会

IF 2.9 Q2 TOXICOLOGY

Computational Toxicology

Pub Date : 2026-03-01 Epub Date: 2026-01-29 DOI: 10.1016/j.comtox.2026.100404

Nicoleta Spînu , Dimitris Stripelis , Mark T.D. Cronin , Gregory L. Warren , Andrew P. Worth

Next Generation Risk Assessment (NGRA) promotes animal-free, exposure-informed, and hypothesis-driven approaches to chemical safety assessment. In silico tools, such as quantitative structure–activity relationship (QSAR) models, are valuable new approach methodologies (NAMs) for use in NGRA. However, the practical implementation of in silico NAMs remains limited by challenges in data availability, heterogeneity, and regulatory acceptance. In this study, federated learning is introduced to advance chemical safety assessment while leveraging proprietary data domains. Federated learning is a decentralised machine learning approach where multiple organisations, devices or servers collaboratively train a model while keeping their data locally, sharing only model updates to preserve confidentiality and privacy. Three use cases were simulated with the Flower open-source federated learning framework, namely (i) federated analytics for dermal permeability (log Kp) screening; (ii) federated convolutional neural networks (CNNs) for mutagenicity prediction from SMILES strings, and (iii) federated eXtreme Gradient Boosting (XGBoost) models for predicting skin sensitisation potential using molecular fingerprints and descriptors. The results show that federated learning approaches can yield predictive performance comparable to centralised models while mitigating concerns over the visibility of, and access to, commercially sensitive data. Open challenges related to data curation, interpretability, and model governance, as well as future directions, are discussed. This work demonstrates that federated learning can facilitate secure collaboration across organisations, enhance the utility of distributed chemical datasets, and accelerate the adoption of in silico NAMs.

下一代风险评估（NGRA）促进无动物、接触知情和假设驱动的化学品安全评估方法。计算机工具，如定量构效关系（QSAR）模型，是NGRA中有价值的新方法方法（NAMs）。然而，计算机NAMs的实际实施仍然受到数据可用性、异构性和监管接受方面的挑战的限制。在本研究中，引入了联邦学习来推进化学品安全评估，同时利用专有数据域。联邦学习是一种分散的机器学习方法，其中多个组织、设备或服务器协同训练模型，同时将数据保存在本地，仅共享模型更新以保护机密性和隐私。使用Flower开源联邦学习框架模拟了三个用例，即(i)皮肤渗透性（log Kp）筛选的联邦分析；（ii）联合卷积神经网络（cnn）用于从SMILES字符串中预测突变性；（iii）联合极端梯度增强（XGBoost）模型用于使用分子指纹和描述符预测皮肤致敏潜力。结果表明，联邦学习方法可以产生与集中式模型相当的预测性能，同时减轻对商业敏感数据可见性和访问的担忧。讨论了与数据管理、可解释性和模型治理以及未来方向相关的公开挑战。这项工作表明，联邦学习可以促进跨组织的安全协作，增强分布式化学数据集的效用，并加速采用计算机NAMs。

{"title":"Federation of toxicological data resources for in silico new approach methodologies (NAMs)","authors":"Nicoleta Spînu , Dimitris Stripelis , Mark T.D. Cronin , Gregory L. Warren , Andrew P. Worth","doi":"10.1016/j.comtox.2026.100404","DOIUrl":"10.1016/j.comtox.2026.100404","url":null,"abstract":"<div><div>Next Generation Risk Assessment (NGRA) promotes animal-free, exposure-informed, and hypothesis-driven approaches to chemical safety assessment. <em>In silico</em> tools, such as quantitative structure–activity relationship (QSAR) models, are valuable new approach methodologies (NAMs) for use in NGRA. However, the practical implementation of <em>in silico</em> NAMs remains limited by challenges in data availability, heterogeneity, and regulatory acceptance. In this study, federated learning is introduced to advance chemical safety assessment while leveraging proprietary data domains. Federated learning is a decentralised machine learning approach where multiple organisations, devices or servers collaboratively train a model while keeping their data locally, sharing only model updates to preserve confidentiality and privacy. Three use cases were simulated with the Flower open-source federated learning framework, namely (i) federated analytics for dermal permeability (log Kp) screening; (ii) federated convolutional neural networks (CNNs) for mutagenicity prediction from SMILES strings, and (iii) federated eXtreme Gradient Boosting (XGBoost) models for predicting skin sensitisation potential using molecular fingerprints and descriptors. The results show that federated learning approaches can yield predictive performance comparable to centralised models while mitigating concerns over the visibility of, and access to, commercially sensitive data. Open challenges related to data curation, interpretability, and model governance, as well as future directions, are discussed. This work demonstrates that federated learning can facilitate secure collaboration across organisations, enhance the utility of distributed chemical datasets, and accelerate the adoption of <em>in silico</em> NAMs.</div></div>","PeriodicalId":37651,"journal":{"name":"Computational Toxicology","volume":"37 ","pages":"Article 100404"},"PeriodicalIF":2.9,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146188953","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 machine learning-based multi-task quantitative structure–activity relationship models for predicting toxicities in six human organ systems 基于机器学习的多任务定量构效关系模型的开发，用于预测六种人体器官系统的毒性

IF 2.9 Q2 TOXICOLOGY

Computational Toxicology

Pub Date : 2026-03-01 Epub Date: 2025-12-24 DOI: 10.1016/j.comtox.2025.100399

Pei-Yu Wu , Wei-Chun Chou , Venkata N. Kamineni , Chi-Yun Chen , Jui-Hua Hsieh , Chris D. Vulpe , Zhoumeng Lin

Traditional toxicity assessment relies heavily on animal testing, particularly for chemicals lacking toxicity data. This study developed machine learning (ML)-driven quantitative structure–activity relationship (QSAR) models to predict human organ-specific toxicities, including cardiotoxicity, developmental toxicity, hepatotoxicity, neurotoxicity, nephrotoxicity, and reproductive toxicity. We collected in vivo data for 2,389 chemicals and Tox21 high-throughput screening data for 1,746 chemicals, resulting in 1,743 chemicals with matched datasets. Eighty-eight ML-based QSAR models were developed using three feature scenarios: (1) Tox21 data alone, (2) molecular descriptors alone, and (3) combined features. Five descriptor types and four ML algorithms (random forests, decision trees, support vector machines, and deep neural network [DNN]) were applied, with and without chi-square-based feature selection. Performance was evaluated using nested cross-validation and five metrics (recall, precision, balanced accuracy, F1 score, and ROC-AUC). DNN models in Scenario 2 performed best for developmental and neurotoxicity, while those in Scenario 3 outperformed others for the remaining toxicities. ROC-AUC values approached 0.8 across endpoints, and models without feature selection generally performed better. SHAP and contribution maps enhanced interpretability, highlighting key structural features of toxicity. This study demonstrates the potential of ML-assisted QSAR models for accurate multi-organ toxicity prediction, supporting drug development and chemical risk assessment.

传统的毒性评估严重依赖于动物试验，特别是对于缺乏毒性数据的化学品。本研究开发了机器学习（ML）驱动的定量构效关系（QSAR）模型来预测人类器官特异性毒性，包括心脏毒性、发育毒性、肝毒性、神经毒性、肾毒性和生殖毒性。我们收集了2389种化学物质的体内数据和1746种化学物质的Tox21高通量筛选数据，得到了1743种化学物质的匹配数据集。采用三种特征场景(1)单独使用Tox21数据，(2)单独使用分子描述符，(3)组合特征，开发了88个基于ml的QSAR模型。五种描述符类型和四种机器学习算法（随机森林、决策树、支持向量机和深度神经网络[DNN]）被应用，有和没有基于卡方的特征选择。使用嵌套交叉验证和五个指标（召回率、精度、平衡准确度、F1分数和ROC-AUC）评估性能。情景2中的DNN模型在发育和神经毒性方面表现最佳，而情景3中的DNN模型在其余毒性方面表现优于其他模型。各个端点的ROC-AUC值接近0.8，没有特征选择的模型通常表现更好。SHAP和贡献图增强了可解释性，突出了毒性的关键结构特征。该研究证明了ml辅助QSAR模型在准确预测多器官毒性、支持药物开发和化学品风险评估方面的潜力。

{"title":"Development of machine learning-based multi-task quantitative structure–activity relationship models for predicting toxicities in six human organ systems","authors":"Pei-Yu Wu , Wei-Chun Chou , Venkata N. Kamineni , Chi-Yun Chen , Jui-Hua Hsieh , Chris D. Vulpe , Zhoumeng Lin","doi":"10.1016/j.comtox.2025.100399","DOIUrl":"10.1016/j.comtox.2025.100399","url":null,"abstract":"<div><div>Traditional toxicity assessment relies heavily on animal testing, particularly for chemicals lacking toxicity data. This study developed machine learning (ML)-driven quantitative structure–activity relationship (QSAR) models to predict human organ-specific toxicities, including cardiotoxicity, developmental toxicity, hepatotoxicity, neurotoxicity, nephrotoxicity, and reproductive toxicity. We collected in vivo data for 2,389 chemicals and Tox21 high-throughput screening data for 1,746 chemicals, resulting in 1,743 chemicals with matched datasets. Eighty-eight ML-based QSAR models were developed using three feature scenarios: (1) Tox21 data alone, (2) molecular descriptors alone, and (3) combined features. Five descriptor types and four ML algorithms (random forests, decision trees, support vector machines, and deep neural network [DNN]) were applied, with and without chi-square-based feature selection. Performance was evaluated using nested cross-validation and five metrics (recall, precision, balanced accuracy, F1 score, and ROC-AUC). DNN models in Scenario 2 performed best for developmental and neurotoxicity, while those in Scenario 3 outperformed others for the remaining toxicities. ROC-AUC values approached 0.8 across endpoints, and models without feature selection generally performed better. SHAP and contribution maps enhanced interpretability, highlighting key structural features of toxicity. This study demonstrates the potential of ML-assisted QSAR models for accurate multi-organ toxicity prediction, supporting drug development and chemical risk assessment.</div></div>","PeriodicalId":37651,"journal":{"name":"Computational Toxicology","volume":"37 ","pages":"Article 100399"},"PeriodicalIF":2.9,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145884618","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

Modernizing pharmacopoeias with artificial intelligence 用人工智能实现药典现代化

IF 2.9 Q2 TOXICOLOGY

Computational Toxicology

Pub Date : 2026-03-01 Epub Date: 2026-02-28 DOI: 10.1016/j.comtox.2026.100407

Pawan Kumar , Saurabh Sahoo , Gaurav Pratap Singh Jadaun , Rajeev Singh Raghuvanshi

Pharmacopoeias are official compendia that provide legally binding specifications for drug substances, products, and excipients, ensuring quality, safety, and efficacy in pharmaceutical manufacturing. They define standards for identity, purity, potency, impurity limits, assays, storage, and labeling. Artificial intelligence (AI) and machine learning (ML) are emerging as transformative tools in pharmaceuticals and regulatory sciences, offering opportunities to modernize compendial processes. Their integration into pharmacopoeial science can accelerate monograph development, streamline revisions, enhance quality control, and support global harmonization. AI-driven technologies such as automated data processing, predictive analytics, and natural language processing can reduce revision timelines, improve transparency, and strengthen stakeholder engagement. These innovations support evidence-based regulatory governance by enabling co-production among regulators, manufacturers, and public health stakeholders. This contemporary review examines current and potential applications of AI and ML in compendial science, including monograph drafting, public comment processing, compliance analytics, and harmonization across pharmacopoeias, providing a regulatory- and policy-oriented perspective that prioritizes governance and implementation considerations. It also presents a pilot study on AI-assisted life cycle management of Indian Pharmacopoeia monographs. The pilot study findings indicated substantial reductions in initial drafting and review turnaround times, enhanced internal consistency across monograph sections, and more efficient categorization of stakeholder feedback during public consultation. Key challenges related to data availability, regulatory compliance, and system integration are discussed, alongside future directions for embedding AI into pharmacopoeial systems, improving accountability, efficiency, and access to quality-assured medicines.

药典是官方的药典，为原料药、制剂和辅料提供具有法律约束力的规范，确保药品生产的质量、安全性和有效性。它们定义了鉴别、纯度、效价、杂质限度、测定、储存和标记的标准。人工智能（AI）和机器学习（ML）正在成为制药和监管科学的变革性工具，为药典流程的现代化提供了机会。将它们整合到药典科学中可以加速各论的开发，简化修订，加强质量控制，并支持全球统一。人工智能驱动的技术，如自动数据处理、预测分析和自然语言处理，可以缩短修订时间，提高透明度，并加强利益相关者的参与。这些创新使监管机构、制造商和公共卫生利益攸关方能够合作生产，从而支持基于证据的监管治理。本当代综述研究了人工智能和机器学习在药典科学中的当前和潜在应用，包括专著起草、公众意见处理、合规分析和药典协调，提供了一个以监管和政策为导向的视角，优先考虑治理和实施方面的考虑。它还提出了一项关于人工智能辅助印度药典专著生命周期管理的试点研究。试点研究结果表明，初步起草和审查周转时间大幅减少，各专著部分的内部一致性增强，公众咨询期间利益相关者反馈的分类更有效。讨论了与数据可用性、法规遵从性和系统集成相关的主要挑战，以及将人工智能嵌入药典系统、改善问责制、效率和获得有质量保证的药物的未来方向。

{"title":"Modernizing pharmacopoeias with artificial intelligence","authors":"Pawan Kumar , Saurabh Sahoo , Gaurav Pratap Singh Jadaun , Rajeev Singh Raghuvanshi","doi":"10.1016/j.comtox.2026.100407","DOIUrl":"10.1016/j.comtox.2026.100407","url":null,"abstract":"<div><div>Pharmacopoeias are official compendia that provide legally binding specifications for drug substances, products, and excipients, ensuring quality, safety, and efficacy in pharmaceutical manufacturing. They define standards for identity, purity, potency, impurity limits, assays, storage, and labeling. Artificial intelligence (AI) and machine learning (ML) are emerging as transformative tools in pharmaceuticals and regulatory sciences, offering opportunities to modernize compendial processes. Their integration into pharmacopoeial science can accelerate monograph development, streamline revisions, enhance quality control, and support global harmonization. AI-driven technologies such as automated data processing, predictive analytics, and natural language processing can reduce revision timelines, improve transparency, and strengthen stakeholder engagement. These innovations support evidence-based regulatory governance by enabling co-production among regulators, manufacturers, and public health stakeholders. This contemporary review examines current and potential applications of AI and ML in compendial science, including monograph drafting, public comment processing, compliance analytics, and harmonization across pharmacopoeias, providing a regulatory- and policy-oriented perspective that prioritizes governance and implementation considerations. It also presents a pilot study on AI-assisted life cycle management of Indian Pharmacopoeia monographs. The pilot study findings indicated substantial reductions in initial drafting and review turnaround times, enhanced internal consistency across monograph sections, and more efficient categorization of stakeholder feedback during public consultation. Key challenges related to data availability, regulatory compliance, and system integration are discussed, alongside future directions for embedding AI into pharmacopoeial systems, improving accountability, efficiency, and access to quality-assured medicines.</div></div>","PeriodicalId":37651,"journal":{"name":"Computational Toxicology","volume":"37 ","pages":"Article 100407"},"PeriodicalIF":2.9,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147396568","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

Corrigendum to “A computational framework for modeling VX skin penetration and RSDL-based neutralization” [Comput. Toxicol. 36 (2025) 100393] “模拟VX皮肤穿透和基于rsdl的中和的计算框架”的勘误表。毒物，36 (2025)100393 [j]

IF 2.9 Q2 TOXICOLOGY

Computational Toxicology

Pub Date : 2026-03-01 Epub Date: 2025-12-10 DOI: 10.1016/j.comtox.2025.100396

Laurent Simon

引用次数: 0

Modeling metabolism: Evolution of toxicodynamic and toxicokinetic considerations. Adding a new kinetics layer 代谢建模：毒性动力学和毒性动力学考虑的演变。增加一个新的动力学层

IF 2.9 Q2 TOXICOLOGY

Computational Toxicology

Pub Date : 2026-03-01 Epub Date: 2025-12-02 DOI: 10.1016/j.comtox.2025.100394

A. Chapkanov , H. Ivanova , G. Poryazova , I. Todorova , T.W. Schultz , O.G. Mekenyan

Modern metabolic simulation encompasses five key attributes that align with a typical data matrix, enabling accurate predictions of metabolism. These attributes are 1) the structural features of the parent molecule (S), 2) the metabolic transformations, both individual and grouped standard types (T), 3) the probability that a specific reaction will occur (P), especially if a particular structural fragment is present nearby, 4) reaction rate (R) (such as the depletion rate of a parent structure), and 5) the quantity of reaction products generated at a given time (Q). The thermodynamically informed phase of metabolism includes STP. Here, the previously described kinetic phase is expanded to include the R and Q attributes. Specifically, a proof-of-concept is described that shows how 2D, 3D, or local parameters can be aligned through regression analysis with hydroxylation and hydrolysis to explicitly simulate metabolic kinetics. In this approach, the amount of metabolite formed depends on the substrate reaction rate via chemical half-lives.

现代代谢模拟包含与典型数据矩阵一致的五个关键属性，从而能够准确预测代谢。这些属性是1)母体分子的结构特征（S), 2)个体和分组标准类型的代谢转化（T), 3）特定反应发生的概率(P)，特别是如果附近存在特定的结构片段，4）反应速率（R)（如母体结构的损耗率），以及5）在给定时间产生的反应生成物的数量(Q)。代谢的热力学信息阶段包括STP。在这里，先前描述的动力学相被扩展到包括R和Q属性。具体来说，描述了一个概念验证，显示了如何通过羟基化和水解的回归分析来对齐2D， 3D或局部参数，以明确模拟代谢动力学。在这种方法中，形成的代谢物的数量取决于通过化学半衰期的底物反应速率。

{"title":"Modeling metabolism: Evolution of toxicodynamic and toxicokinetic considerations. Adding a new kinetics layer","authors":"A. Chapkanov , H. Ivanova , G. Poryazova , I. Todorova , T.W. Schultz , O.G. Mekenyan","doi":"10.1016/j.comtox.2025.100394","DOIUrl":"10.1016/j.comtox.2025.100394","url":null,"abstract":"<div><div>Modern metabolic simulation encompasses five key attributes that align with a typical data matrix, enabling accurate predictions of metabolism. These attributes are 1) the structural features of the parent molecule (S), 2) the metabolic transformations, both individual and grouped standard types (T), 3) the probability that a specific reaction will occur (P), especially if a particular structural fragment is present nearby, 4) reaction rate (R) (such as the depletion rate of a parent structure), and 5) the quantity of reaction products generated at a given time (Q). The thermodynamically informed phase of metabolism includes STP. Here, the previously described kinetic phase is expanded to include the R and Q attributes. Specifically, a proof-of-concept is described that shows how 2D, 3D, or local parameters can be aligned through regression analysis with hydroxylation and hydrolysis to explicitly simulate metabolic kinetics. In this approach, the amount of metabolite formed depends on the substrate reaction rate via chemical half-lives.</div></div>","PeriodicalId":37651,"journal":{"name":"Computational Toxicology","volume":"37 ","pages":"Article 100394"},"PeriodicalIF":2.9,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145749932","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