Physiologically Based Pharmacokinetic (PBPK) Modelling have emerged as a critical method in pharmaceutical and environmental sciences. Despite its widespread adoption, inconsistent vocabulary used across models, software platforms, and literature remains a significant challenge. Such inconsistency limits the potential for data harmonization and model interoperability. This gap also hinders the adoption of machine-readable semantic technologies that could enhance PBPK model reporting, validation, and interoperability.To address this gap, we developed PBPKO (Physiologically Based Pharmacokinetic Ontology), a dedicated ontology for the kinetic modelling community. PBPKO captures mechanisms and terminology commonly used in PBPK modelling, including concepts related to physiological parameters, biochemical parameters, biological processes, enzymes, transporters, and related concepts. As of release of v2026-07-15, the ontology comprises 858 classes and 7 object properties native to PBPKO. To support semantic interoperability within the broader biomedical ontology ecosystem, PBPKO is aligned with upper-level ontologies such as the Basic Formal Ontology (BFO) and the Relation Ontology (RO), and reuses terms from established resources including Gene Ontology (GO) and the Ontology for Biomedical Investigations (OBI).In this work, we further applied PBPKO to the annotation of PFAS multi-compartment PBPK models as a case study, while also highlighting the parallel development effort of “FAIR PBPK Inspector”, which leverages the ontology for model annotation. Together, these efforts support FAIR sharing of PBPK models and facilitate their regulatory acceptance through improved standardization. PBPK ontology is accessible with CC4 licences on https://github.com/InSilicoVida-Research-Lab/pbpko .
During early pregnancy, significant perturbations to thyroid hormone activity in the mammalian foetal brain can lead to adverse neurodevelopmental outcomes. Endocrine-disrupting chemicals have the potential to cause these perturbations by interfering with normal systemic thyroid hormone regulation. This represents a complex multi-scale problem involving both whole-body (systemic) and tissue-based (local) regulation of thyroid hormones. Understanding whether biological barriers such as the blood-brain barrier and localised transcriptional feedback systems can resist systemic changes in thyroid hormones is critical. We developed a novel mathematical model describing thyroid hormone regulation dynamics in the foetal rat brain and used it to predict the effects of three chemicals (propylthiouracil, sodium phenobarbital, and methimazole) on foetal brain thyroid hormone concentrations. The model predicted changes in foetal brain thyroid hormones within two-fold accuracy based on measured changes in systemic foetal thyroid hormone levels. The model also demonstrates that maternal (dam) systemic thyroid hormone levels can be used to predict these effects during gestation. Mathematical modelling of thyroid hormone regulation in the foetal brain can improve the prediction and understanding of the dynamic effects of endocrine-disrupting chemicals on neurodevelopment and support developmental neurotoxicity risk assessment.
Essential-oil nanocapsules are bio-based antimicrobial formulations central to agri-food sustainability, yet their human and environmental hazard as formulated products cannot be assessed by conventional molecular QSAR, which ignores controlled release kinetics, nano-specific bio-interactions, and shell-material contributions. We apply a systems toxicology approach — integrating molecular QSAR (Tier 1), Korsmeyer–Peppas release kinetics (Tier 2), and a nano-specific bioavailability correction layer (Tier 3) — to produce the first quantitative integrated hazard prediction for two clove-oil advanced-material (AdMa) nanocapsule formulations: AdMa EO@PEC-GEL (pectin–gelatin shell, CaCl₂ crosslinker) and AdMa EO@Chi (chitosan shell, formaldehyde crosslinker).Eugenol (∼90 wt% core) is the principal toxicophore. Release parameters were scaled from published eugenol–chitosan kinetic data to the 1 μm target particle size; nano-correction factors were calibrated from published surface-charge and uptake relationships. The integrated model predicts AdMa EO@PEC-GEL to be 9.1× safer than free eugenol (IC₅₀ ∼3.5 mM vs. 0.38 mM), driven by anionic surface charge and 80% encapsulation efficiency. AdMa EO@Chi retains toxicity close to the free molecule (IC₅₀ ∼0.91 mM; 2.4×) because cationic surface charge offsets the encapsulation benefit. The formaldehyde crosslinker introduces independent sensitisation, genotoxicity, and IARC Group 1 carcinogenicity flags — quantified via the concentration-addition mixture model — entirely absent from PEC-GEL.Sensitivity analysis identifies zeta potential as the dominant model uncertainty driver. Tier 2 is retrospectively validated against published release kinetics (R2 = 0.997). The Korsmeyer–Peppas tier applies directly to environmental fate scenarios — pH-dependent shell dissolution in soil and aquatic compartments — providing a unified architecture for human–environment integrated hazard assessment aligned with planetary health priorities. An interactive, browser-based digital twin is provided as Supplementary Information: it recomputes all integrated predictions in real time as the user adjusts the model inputs (zeta potential, particle size, encapsulation efficiency, eugenol fraction, release time and pH/enzyme condition), enabling transparent scenario analysis and progressive refinement as experimental data become available, in line with the iterative Safe-and-Sustainable-by-Design (SSbD) workflow.
Peroxisome proliferator–activated receptor gamma (PPARγ) plays a pivotal role in adipogenesis, glucose and lipid metabolism, and endocrine regulation, making it a critical molecular target in both pharmacology and toxicology. Given growing concern about industrial chemicals that can perturb PPARγ signaling, reliable computational approaches are needed to support chemical evaluation and informed regulatory decision-making. In this study, we applied a hybrid quantitative structure–activity relationship–quantitative read-across (QSAR–qRA) framework that integrates descriptor-driven modeling with similarity-based features to enhance predictive accuracy, mechanistic interpretability, and regulatory relevance. Transparent two-dimensional descriptors were employed for similarity-based feature extraction, while the contributing molecular descriptors provided a physical interpretation context, ensuring both statistical rigor and interpretability. The final structural similarity-driven univariate regression model, derived within this hybrid QSAR–qRA framework, demonstrated improved robustness compared to conventional partial least squares (PLS)-based QSAR and standalone qRA approaches, as supported by internal validation, applicability domain (AD) analysis, and descriptor ablation studies confirming the critical role of the top three features for PPARγ activity. Incorporation of RA-derived similarity-based descriptors enhanced descriptor-space coverage and reduced extrapolative uncertainty, consistent with leverage–residual diagnostics and the Laplacian kernel (LK) similarity-based closest-source analysis, with the exclusion of the lowest-similarity 10% of external (test set) compounds to ensure reliable predictions. Evaluation on a rigorously curated independent external dataset comprising 1153 unique compounds demonstrated realistic prospective performance and a marked expansion of the applicability domain compared with the global PLS model, underscoring improved transferability across heterogeneous chemical space. Importantly, descriptor interpretation provided physical insights into structural drivers of PPARγ inhibitory activity, reinforcing the biological relevance of the predictions. Overall, the similarity-driven univariate linear regression model offers improved accuracy, transparency, and adaptability compared with traditional methods, positioning it as a valuable chemoinformatics tool for advancing PPARγ modeling and broader toxicological applications.
Gentamicin is an aminoglycoside antibiotic whose clinical use is limited by nephrotoxicity, while the underlying mechanistic drivers remain incompletely understood. In this study, we reanalyze previously published transcriptomics data where RPTEC/TERT1 cells were exposed to gentamicin. We use a transcriptomics-driven coregulated gene network approach to identify key events (KEs) relevant to gentamicin-induced nephrotoxicity from a broad adverse outcome pathway (AOP) network. We subsequently employ concentration-response modeling combined with bootstrapping to generate artificial response-response data, which we use to define and calibrate a quantitative AOP (qAOP) network model specific to gentamicin exposure. This approach enables the generation of biologically meaningful hypotheses despite limited experimental data availability, for example, quantifying the contribution of different network pathways to cytotoxicity. Specifically, our analysis suggests that cytotoxicity depends more strongly on mitochondrial than on lysosomal disruption. This analysis can be validated by studying how experimental manipulation of these KEs affects cytotoxicity, and by including additional gentamicin concentrations. The qAOP model resulting from our analysis can be directly applied to other compounds sharing a similar mode of action, while the overall methodology is widely applicable to diverse biological case studies for which transcriptomics data and a broad AOP network are available. Notably, this framework does not rely on extensive experimental conditions, highlighting its potential utility for mechanistic understanding and predictive toxicology.
In next-generation risk assessment (NGRA), there is a growing need for physiologically-based kinetic (PBK) models to extrapolate external doses to internal concentrations in biological matrices and vice versa. PBK models are used in different scientific contexts, including chemical risk assessment. As they are commonly developed for specific chemicals, effects, and populations, a wide range of PBK models is available. However, because PBK models are developed and implemented by different researchers using varying concepts, programming languages, and coding standards, they are often difficult to find, access and reuse across platforms or contexts. Improving their findability, accessibility, interoperability, and reusability (i.e., FAIR) is essential to gain acceptance of PBK models in risk assessment for regulatory purposes. Although many elements to enhance the FAIRness of PBK models are available, a standardised good modelling practice has yet to be established. This paper presents a practical strategy for implementing the FAIR principles for PBK models through a proposed FAIR PBK standard designed to promote interoperability and reusability. The standard builds upon the existing Systems Biology Markup Language (SBML), the dedicated physiologically-based pharmacokinetic ontology (PBPKO) for semantic annotation, and guidelines and tools to combine these to create FAIR PBK models. The standard aids both model developers and (re-)users. It is accompanied by a dedicated software tool, the FAIR PBK inspector, to aid model developers in creating PBK models compliant with the standard. Also, support for PBK models compliant with the standard was implemented in the Monte Carlo Risk Assessment (MCRA) software, enabling their reuse in chemical risk assessment. More broadly, the SBML basis enables reuse in other environments, including R, MATLAB, and Python. A proof-of-principle application within the EU Partnership for the Assessment of Risks from Chemicals (PARC) demonstrates the use and potential of the standard.
Maximizing the use of already generated in vivo ecotoxicological data is essential, to develop and validate alternative in vitro and in silico methods, establish relative species sensitivity, and assess the role of fish acute and chronic studies in hazard assessment. We constructed high-quality reference datasets for fish toxicity by extracting experimental studies from REACH registration dossiers and rigorously curating them using AI-assisted algorithms. Strict filtering criteria were applied to ensure study reliability, which include verification of the exposure concentrations, overall compliance with test guidelines, and accurate molecular structure assignment. The resulting dataset includes acute toxicity data for 2315 structures and chronic toxicity data for 418 structures, encompassing a broad spectrum of industrial chemistries, test designs, and regulatory guidelines. The datasets contain short-term fish toxicity LC50 values for 610 standardised structures and fish earlyto life stage toxicity NOEC/EC10 values for 135 standardised structures amenable to modelling, when considering only bounded effects levels derived from concentration-response modelling using analytically verified exposure concentrations. These were used to quantitatively assess the accuracy and applicability domain coverage of widely used predictive models. We benchmarked the models for their ability to predict effect concentrations and regulatory classification thresholds both globally and within chemical subspaces defined with a range of clustering methods. For acute toxicity, up to 94% of LC50 values can be predicted within 1 log unit at an applicability domain coverage of 30%; this coverage increases to nearly 80% when 85% of the predictions fall within 1 log unit. Absolute error increases with ionisation potential, and local predictive performance varies substantially even within the applicability domain. For chronic toxicity, performance was considerably lower, particularly for highly toxic substances, with much smaller applicability domain coverage due to the scarcity of experimental data, which also makes the model assessment less conclusive. Overall, our analysis demonstrates the potential of modelling for aquatic toxicity hazard assessment, highlights current limitations, and provides curated data and suggestions to support the further development and promotion of predictive methods as alternatives to in vivo testing.
Traditional aggregate risk models rely on the availability of toxicological parameters for each substance in the mixture and assume a linear relationship between the substances. However, these models are limited by incomplete toxicological data and the potential for non-linear (antagonistic or synergistic) interactions among substances. To address these limitations, this work presents a framework for characterizing and interpreting toxicological data to support health risk prediction. Toxicological parameters were obtained from the United States Environmental Protection Agency's Integrated Risk Information System database, and exposure data from an artificially generated dataset based on the properties of exposure data published by Heck et al. The latter dataset is highly imbalanced, with a ratio of negative to positive individuals of 0.0024 of total samples, thus posing modelling challenges. Specifically, we apply Principal Component Analysis (PCA) to infer relationships among exposure substances, identify clusters in the principal component space to estimate missing toxicological parameters, and use the resulting PCA loadings to cross-check linear aggregation weights learned via gradient descent. We then extend the linear risk aggregation model to include pairwise interaction terms to capture non-linear dependencies and compare its interpretability and predictive performance against three tree-based models: Random Forest (RF), Balanced Random Forest (BRF), and eXtreme Gradient Boosting (XGBoost). Our results show that PCA reveals groupings of high-risk pollutants that align with both the established literature and the gradient-learned linear aggregation weights. Interaction-based extensions to the linear model produced unstable and uninterpretable results, likely due to extreme data imbalance and the limitations of using pairwise exposure products to represent interactions. Among tree-based methods, BRF achieved the best performance on imbalanced data, and using a minority oversampling technique to balance the data was insufficient to achieve strong predictive performance on the extremely imbalanced datasets typically encountered in toxicological risk prediction applications.
Concentration–response curves model the relationship between a concentration of a compound and the response it elicits in a biological system, such as the viability of cells. A primary objective is to estimate an alert concentration, meaning the concentration at which a predefined effect level is reached. These values are typically derived from parametric models fitted to the data. Therefore, accurately modeling the concentration–response relationship is essential for understanding the safety and potency of compounds. However, due to errors or biological variability, it is often observed that measurements at individual concentration levels deviate from the typically assumed sigmoidal models. Such deviations in the data can compromise the accuracy of the estimation of alert concentrations. Due to the usually very low number of measured concentrations, identifying these deviations is challenging. Here, we simulated data for different scenarios of possible deviations at each individual concentration level. We propose several statistical methods that can mitigate the impact of such deviations. Two methods identify the deviations and eliminate the data of the corresponding concentrations from the curve fitting process. Three methods use different weighting approaches to assign lower weights to deviating values. All methods are compared with a baseline method that disregards deviations on how accurately they estimate the true EC20 value. Methods that remove concentrations with deviations perform best on average, but they perform notably worse if correct data are mistakenly excluded.
Adverse outcome pathways (AOP) capture sequences of key events (KE) relating toxicant exposures to adverse outcomes (AO) observed in exposed populations. Quantitative AOPs (qAOP) enable AO risk assessment based on KE quantification and modeling of their relationships. This work aimed to develop a qAOP model for AOP 411 “Oxidative Stress Leading to Decreased Lung Function”. As a proof-of-concept, the model was applied to predict the medium-term lung function evolution after switching from combustible cigarettes (CC) to heated tobacco products (HTPs). The qAOP model enables the evaluation of the HTP “risk reduction potential” in the absence of epidemiological evidence.We designed our qAOP to evaluate the relative change (RC) of the AO risk for HTP switching compared to either continuing or quitting CC consumption. Using publicly available data for KE quantification, we developed both data- and physiology-based KE relationship (KER) models.We used the resulting qAOP to evaluate the effects of switching from CC smoking to HTPs in terms of the RC of the AO “Decreased Lung Function”. To support the acceptance of our approach, we followed “best practice” recommendations for building and reporting the three KER models. We also examined several scenarios for an optimal combination of in vitro testing and in silico predictions in quantifying AOP 411.In summary, qAOP provides a multiscale mechanistic approach to evaluate medium-term lung function changes when switching from CC smoking to HTPs. It demonstrates how in vitro and in silico approaches can be combined for a pragmatic and scientifically sound risk assessment.
Next Generation Risk Assessment (NGRA) increasingly relies on quantitative measures of compound potency derived from in vitro systems, commonly expressed as points-of-departure (PoDs). High-throughput transcriptomics (HTTr) platforms may feature in NGRA workflows, generating concentration-response data for thousands of genes following chemical exposure. We present BIPODE (Bayesian Inference for Point-of-Departure Estimation), a non-parametric Bayesian workflow designed to estimate gene-level minimum effect concentrations from HTTr data. The workflow is implemented as a Python package with an accompanying Nextflow pipeline for scalable deployment. Use of the pipeline is discussed via case study applications to publicly available HTTr datasets. BIPODE offers a transparent, statistically rigorous approach to HTTr analysis, supporting NGRA by providing reliable estimates of compound potency in in vitro systems.
Pollution is a major but under-evaluated driver of biodiversity loss. While methods are being developed to predict genotype-environment (GxE) interactions for humans, such approaches are under-developed for nonhuman populations. We built a computational approach to predict GxE interactions for Mus musculus, an ecologically relevant species. We integrate disconnected datasets to form Chemical-Pathway-Gene-Variant Phenotype linkages for three chemical classes (pharmaceuticals, pesticides, and cosmetics) with two sets of genetic variants: from primarily lab mice in Ensembl (>15,000,000 variants) and from wild mice (>50,000 variants) to identify genes and pathways that may underlie genetic susceptibility or adaptation to chemicals in wild mouse populations. By comparing the Ensembl and wild mouse linkages and the three chemical classes, we 1) characterize pathways implicated in GxE interactions (e.g., PI3K-Akt signaling pathway), 2) identify genes with high variability in the wild mouse linkages that may be important in GxE interactions (e.g., IL6, NFKB1), and 3) highlight genes from Ensembl linkages in the same pathways as the wild linkages, but absent from the wild mouse data (e.g., TLR4, TNF). While some variants implicated in Ensembl may differ from what would be in wild populations, we propose that the genes may still be important given that they are in the same pathways implicated in the wild linkages. Future studies can prioritize the genes and pathways we highlight to characterize genetic susceptibility or adaptation to chemicals for wild mouse populations and expand our approach of repurposing lab data to inform GxE interactions in other ecologically relevant species.
alpha,beta-Unsaturated carbonyl compounds are frequently associated with Michael-type electrophilic reactivity toward biological nucleophiles. However, structurally similar electrophiles within the same chemical class may exhibit markedly different mutagenic responses, making mechanistically interpretable mutagenicity prediction challenging. In this study, we examined a curated dataset of 34 alpha,beta-unsaturated acids, a chemically coherent subclass of Michael acceptors, to evaluate whether electronically derived reactivity descriptors could discriminate mutagenic from non-mutagenic compounds. Descriptors derived from conceptual density functional theory (CDFT) and electrotopological state (E-State) indices were reduced using correlation filtering, ANOVA, and recursive feature elimination. Among the evaluated classifiers, Logistic Regression provided the most consistent balance between predictive agreement and interpretability (kappa = 0.72). The final two-descriptor model, based on the global electrophilicity index (omega) and the beta-carbon E-State descriptor (SC beta), indicates that higher electrophilicity and increased electronic activation at the beta-carbon are associated with a greater probability of a mutagenic Ames response within this restricted chemical domain. The proposed model is intended as a category-specific binary hazard classifier for alpha,beta-unsaturated acids rather than as a generalized mutagenicity predictor.
Accurate in silico toxicity prediction is essential for chemical safety assessment and computational toxicology. This study presents a novel Cliff method that leverages structure-activity relationship (SAR) information from activity cliff pairs to enhance machine learning models for toxicity prediction. We analyzed 249 bioassays from the Tox21 dataset and identified activity cliff pairs based on Tanimoto similarity and activity difference analysis. A total of 1613 molecular descriptors were calculated using Python Mordred, and single-feature decision tree models evaluated the predictive power of each descriptor across the cliff pair subsets of each bioassay dataset. Through construction and evaluation of multiple machine learning models across all datasets, we demonstrated that Cliff outperformed conventional feature engineering approaches, including variance thresholding, correlation analysis, RFE, and LASSO regression. Cliff achieved a superior median AUC of 0.736, with particularly notable improvements observed in non-tree-based machine learning models such as logistic regression and SVM (average AUC gains up to 0.071). Further analysis revealed that cliff pairs showed significantly higher descriptor value differences than non-cliff pairs for seven key descriptors, including d_ATS8Z, d_AATSC1d, d_AATSC5c, d_SLogP, d_ATSC2d, d_WPath, and d_ATS8m. These findings highlight the potential of activity cliff-guided feature selection as a valuable computational tool for improving the predictive accuracy and interpretability of in silico toxicity models.
Aquatic toxicity testing, which traditionally involves in-vivo experiments on organisms across three trophic levels (algae, invertebrates, and fish), is both resource-intensive and ethically challenging. As an alternative, insilico methods such as Quantitative Structure-Toxicity Relationship (QSTR) modeling have emerged as promising tools for predicting chemical toxicity. In this research, we present ProtoAquaTox (R), a software platform developed to predict acute and chronic aquatic toxicity using three multi-tasking QSTR models. The platform employs multi-tasking QSTR models built on a dataset of 2119 unique chemicals and 11,860 toxicity values, representing acute and chronic toxicity data for different trophic levels, namely, fish, crustaceans, and algae. These models encompass multiple test organisms (total 87 species) from all three trophic levels and are designed to account for variations in experimental parameters such as test organisms, media type, exposure type, duration, etc. Employing the random forest machine learning technique, we developed three binary classification models (Models1-3) that categorize chemicals into distinct toxicity classes, achieving optimal performance based on validation metrics. Our multi-tasking models demonstrate superior predictive performance, achieving accuracies between 0.85 and 0.90, precision values ranging from 0.85 to 0.91, MCC scores between 0.68 and 0.79, and AUC values in the range of 0.92 to 0.96 when evaluated on an external test set. Experimental validation against test organisms has demonstrated the models' efficacy, with areas for further refinement identified. ProtoAquaTox (R) provides an intuitive interface, enabling users to predict the environmental risks posed by diverse chemicals, thereby aiding hazard mitigation and contributing to the protection of aquatic ecosystems. The platform is freely available through the Chemopredictionsuite web platform, making it a valuable tool for environmental risk assessment.
