Capacity Building

Meet the Next Generation of Toxicologists

Discover ONTOX’s PhD students and the research they conducted in connection with the ONTOX project, contributing to the advancement of next-generation toxicology.

Devon A. Barnes | Utrecht University

From Molecules to Mechanisms: Developing Adverse Outcome Pathways for Tubular Necrosis and Crystalline Nephropathies

“By integrating adverse outcome pathways with computational tools, human kidney cell models, and transcriptomics, my research shows how fragmented toxicity data can be transformed into mechanistic, human-relevant evidence for next-generation kidney safety assessment.”

ABSTRACT

Nephrotoxicity remains an important clinical and regulatory concern, while conventional toxicity testing often provides limited mechanistic insight and may not reliably predict human kidney injury. My thesis aimed to advance human-relevant and mechanistically informed approaches for assessing nephrotoxicity by integrating computational modelling, adverse outcome pathways, in vitro experimentation, and transcriptomics.

Computational models were first developed to predict drug-induced nephrotoxicity from chemical structure. Existing mechanistic knowledge was subsequently organised into a nephrotoxicity adverse outcome pathway network and a more specific pathway linking DNA-adduct formation to kidney failure. Using human proximal tubule cell models, adverse outcome pathway-guided assays were developed to investigate platinum-induced tubular injury, identify early cellular events, and integrate transcriptomic evidence into the mechanistic framework. The approach was then extended to crystalline nephropathies through experimental modelling of pH-dependent uric acid toxicity and the development of a pathway connecting urinary supersaturation and crystal formation with tubular injury, inflammation, fibrosis, and impaired kidney function.

Collectively, this work demonstrates how computational prediction, mechanistic knowledge, omics data, and human-relevant experimental systems can be combined to support more predictive and animal-free kidney safety assessment.

Joh H. Berkhout | National Institute for Public Health and the Environment

A new approach methodology for probabilistic assessment of neural tube closure defects

“Integrating diverse biological evidence matters more than perfecting any single testing method. By feeding in vitro perturbation data into a dynamic ‘Virtual Embryo’ simulation, we can translate molecular toxicity into precise probabilities of neural tube defect phenotypes, without animal testing.”

ABSTRACT

During the early development of vertebrate embryos, a flat sheet of tissue must elevate, bend, and fuse to form the neural tube. When chemical exposures disrupt this delicate process, severe birth defects such as spina bifida can occur. To understand and predict these disruptions without relying on traditional animal testing, we developed an advanced computational framework that simulates embryonic development. Instead of treating the tissue as a static object, we built a model where individual virtual cells react dynamically to biomechanical and genetic signals. To ground this simulation in real biology, we mapped the complex cascade of developmental toxicity using an Adverse Outcome Pathway network. Because the initial chemical triggers are crucial, we trained an artificial intelligence system to read molecular structures as spatial graphs. This allows us to rapidly predict which everyday chemicals might initiate toxic biological reactions. However, biology is not static, so we also tracked the time-resolved genetic changes in zebrafish embryos exposed to various compounds. By feeding these temporal gene expression measurements directly into our updated computational model, we can accurately simulate how the neural tissue’s behavior shifts under chemical stress. Rather than yielding a simple yes-or-no answer, this data-driven simulator calculates the precise probability that a specific developmental defect will emerge. By bridging artificial intelligence, dynamic genetic data, and computational modeling, this research provides a robust, animal-free alternative to prioritize hazardous chemicals and protect human health.

Brian Bwanya | Maastricht University

Decoding the molecular language of toxicity using omics-informed models

“The tools for animal-free, human-relevant toxicology already work. Across machine learning, single-cell sequencing, image analysis, and large language models, the methods delivered what was asked of them, and in every case the limiting factor was the data rather than the method, its scale, quality, and how well it was annotated and preserved. The clearest message of this thesis is that the next advances in safety science will come not from inventing new techniques but from building shared, well-curated, human-relevant datasets and sustaining them beyond the life of any single funded consortium. The future of toxicology depends as much on how we care for our data as on how we generate it.”

ABSTRACT

Modern toxicology is shifting away from animal testing toward human-relevant, mechanism-based approaches, a transition at the heart of the ONTOX project. This thesis asks whether omics technologies and computational methods are now capable of supporting that shift. The central obstacle is no longer generating biological data but converting increasingly complex information into mechanistically interpretable evidence for regulatory decision making. Four research chapters address this challenge across complementary human-relevant systems. The first study set out to test whether machine learning can predict drug-induced hepatic steatosis from toxicogenomic profiles, and whether the models also reveal mechanism. Supervised classifiers accurately distinguished steatogenic compounds, and the discriminating genes mapped onto lipid metabolism and established steatosis pathways, showing that the models generate mechanistic insight rather than opaque outputs. Building on this foundation, the second study sought to establish how tissue architecture shapes the cellular response to a hepatotoxicant, and which culture format better reflects human liver biology. Single-cell RNA sequencing was applied to two- and threedimensional human liver microtissues exposed to acetaminophen. Three-dimensional cultures exhibited hypoxic gradients that recapitulate liver zonation, and hepatocytes carrying a hypoxic signature raise cytochrome P450 expression while their phase II detoxification enzymes fall with dose, revealing an interplay between oxygen availability and drug metabolism. The next question was whether image-based morphology can substitute for transcriptomics in developmental toxicity screening, or whether each captures distinct aspects of the biological response. To address this, high-content imaging and transcriptomics were profiled together in three-dimensional stem-cellderived microtissues. Both readouts identified the same toxic compounds but ranked their potency independently, showing that they provide complementary rather than interchangeable evidence. Their combined points of departure also tracked in vivo teratogenic outcomes. The final study addressed a challenge that precedes biological analysis by asking whether large language models can recover the unstructured clinical metadata that leaves many archived datasets unusable. Applied to a fetal biobank, the models converted narrative records into analyzable metadata and restored an otherwise inaccessible dataset for analysis, with the choice of model determining how much information could be recovered. The reactivated dataset revealed a coordinated ciliary signature in the spinal cord of neural tube defect cases, demonstrating that recovered metadata can lead directly to new biological insight. Together, these studies answer the central question posed at the outset. Omics technologies and computational methods, from machine learning to large language models, can transform complex biological data into mechanistically interpretable, human-relevant evidence across diverse systems and scales. By defining what each approach contributes, and where they are most powerful in combination, this work provides a concrete step toward the integrated, mechanism-based framework required for next-generation risk assessment.

Marie Corradi | Hogeschool Utrecht

Small and Large Language models usages to support knowledge retrieval in toxicology

“Scientists should use (and be supported to use) new technologies to help them retrieve information from fast-growing scientific sources. However, this should not be done at the cost of their critical thinking abilities, and these technologies should always be used as a complement to scientists’ expertise rather than a replacement.”

ABSTRACT

To be able to transition to animal-free toxicology, scientists need to be able to understand the mechanisms leading to adverse outcome events in the scope of human biology. This means there is a need for quicker review of the information that is already available in various databases and scientific literature. In this thesis, we show how language models can support this goal. In the first three chapters, we demonstrate how classic, small natural language processing (NLP) models are able to extract toxicology-related information from scientific articles in a reliable and traceable way. By organizing the information obtained in a graph database, we were able to partly recreate existing adverse outcome pathways. In Chapter 4, we evaluate more modern retrieval methods with the support of embedding models to extract passages of text supporting hypothesized events leading to kidney toxicity. While these models are promising, they are not performant enough on their own and need to be supplemented with either other large languages models (LLMs) – rerankers – and/or traditional NLP. Chapter 5 shows that traditional NLP methods such as named entity recognition and keywords matching coupled with embedding and text classification models can be used beyond purely scientific text on a European project database. These allowed to recover projects relevant to modern toxicology in a faster and more effective way. Finally, in Chapter 6, we lowered the entry barrier to pathway databases by developing an agentic chatbot interfacing with MINERVA disease and physiological maps. In this user-driven effort, we also showed the importance of thorough, community-based validation, both for the adoption and for the evaluation of LLM-based tools for science.

Annika Hanna Drees | Vrije Universiteit Brussel

Development of a mechanism-based in vitro test platform to predict chemical-induced cholestatic liver injury.

“The mechanism-based in vitro detection shows promise in the improved detection of cholestatic liver injury and the consequent replacement of animal testing.”

ABSTRACT

Cholestasis is described as the noxious accumulation of bile acids in the liver or blood circulation, and can be caused by chemicals from various applicability domains. Current animal methods applied in risk assessment and drug development poorly predict chemical-induced cholestatic liver injury in humans. This is due, at least in part, to interspecies differences and knowledge gaps regarding the mechanisms behind the adversity. The VUB host laboratory has introduced and refined an adverse outcome pathway (AOP) network to map the mechanisms of chemical-induced cholestasis. The present doctoral thesis project intended to develop and improve methods to advance the early detection of cholestatic chemicals using this AOP network as a mechanistic compass. In the first study, an existing assay examining cholestatic potential was optimised. The drug-induced cholestasis index (DICI) was used to assess the cholestatic potential by the ratio of cytotoxicity induced by a chemical in the presence and absence of bile acids, targeting the central key event of bile acid accumulation in the AOP network. This study relied on the use of 3D spheroids of human hepatoma HepaRG cells. A large array of drugs with different hepatotoxic properties were investigated to determine the predictive capabilities of the DICI. The performance of the DICI was improved by the inclusion of predictive modelling, and 2 models with a high specificity and sensitivity for cholestatic drugs, respectively, were identified. In the second study, an AOP-based integrated in vitro/in silico test battery was developed, which investigated multiple molecular initiating events specific for chemical-induced cholestasis along with generic key events. The in vitro aspect of this study was based on the use of HepaRG cells in monolayer culture. The performance of the test battery was determined by testing well-characterised cholestatic and non-cholestatic drugs and industrial chemicals. The results were analysed and integrated using an innovative weight-of-evidence approach. The test battery showed promising potential in the differentiation between cholestatic and non-cholestatic chemicals. Overall, a higher specificity than sensitivity in the detection of cholestatic chemicals was observed. The test battery was intended to be suitable for chemicals from different applicability domains and would therefore require validation for each applicability domain, but the toxic effects of industrial chemicals on the liver are usually poorly reported and their mechanisms are rarely elucidated. This hinders the choice of suitable reference chemicals for the validation. As such, the occurrence and experimental data of chemicals classified as hepatotoxic at the European Chemicals Agency were investigated in the third study. The liver was found to be the most sensitive target organ and, based on biomarker fingerprints, industrial chemicals with cholestatic, steatotic and fibrotic potential were identified. In the fourth study, the test battery developed in the second study was used to assess 3 potential cholestatic chemicals identified in the third study in order to assess the context of use of the test battery in chemical risk assessment. The test battery showed great capability in elucidating the cholestatic mechanism behind the hepatotoxicity of the chemicals and the ability to rank them based on their cholestatic potential. Overall, the present doctoral thesis project was conducted within the Flemish VLAIO HOBOTA and European ONTOX project. The results provide the basis for a new approach methodology applicable in chemical risk assessment and safety evaluation, and thus contribute to reduction of animal tests and improvement of human health by the prioritisation of chemicals.

René Geci | EsqLABS GmbH

High-throughput PBK Modelling for the Prediction of Systemic Exposure and Effects of Chemicals on Humans

“Our research shows that the concentrations chemicals reach inside the human body can be predicted with reasonable reliability, without the need for animal data. This works by combining machine learning models that predict chemical properties from molecular structure with mechanistic PBK models that describe human physiology. Knowing these concentrations is essential for safety assessment, because laboratory tests on cells only tell us at which concentration a chemical becomes harmful. Bringing the two together offers a practical route to assessing chemical safety without animal testing.”

ABSTRACT

Modern toxicology and pharmacology are undergoing a fundamental transformation. Ethical considerations and advances in computational methods are increasingly enabling the replacement of animal testing with more human-relevant, non-animal assessment approaches. This thesis contributes to this transition by developing and validating an approach termed high-throughput physiologically based kinetic (HT-PBK) modelling. This new method integrates mathematical PBK models with modern machine learning (ML) tools that predict key physicochemical and absorption, distribution, metabolism and excretion (ADME) properties of molecules directly from their chemical structure. Thereby, it enables rapid, automated model parameterisation while preserving mechanistic interpretability and removing bottlenecks in traditional PBK model development. Relying solely on in vitro and computational input data, HT-PBK modelling enables the estimation of human exposure to chemicals, and consequently the effects of chemicals on humans.

The overarching aim of this thesis is to quantify the predictive performance of HT-PBK-based pharmacokinetic (PK) simulations against in vivo PK data observed in humans, and to showcase the utility of such simulations for real-world application in safety assessment. To enable a comprehensive evaluation, the work assembled and curated one of the largest available datasets of human plasma concentration-time profiles, covering over 2,000 human PK curves of more than 200 compounds following intravenous and oral administration. Across this dataset, HT-PBK simulations reproduced the PK of most compounds sufficiently reliably, with 87% of Cmax and 84% of AUC values within tenfold of the observed data.

After evaluating intravenous and oral exposure, the thesis extends the assessment to dermal exposure, which is a major route for many consumer and cosmetic products and for which animal testing is increasingly restricted by regulation. Although the prediction of systemic PK after dermal exposure is more challenging, 75% of Cmax and AUC values remained within a tenfold range of observed human data.

The evaluation is then extended to real-world risk assessment tasks. First, within the European Partnership for Alternative Approaches to Animal Testing (EPAA) NAM Designathon, the thesis demonstrates how HT-PBK predictions alone can support the classification of chemicals based on their potential for systemic availability when toxic potency information is unavailable. Finally, HT-PBK predictions are combined with in vitro toxicity assay data to assess the risk of drugs causing drug-induced liver injury (DILI). The integration of in vitro assays and HT-PBK simulations yields strong alignment with clinical DILI outcomes, with a ROC AUC of 91%.

By combining mechanistic simulation with ML-driven parameter prediction, this thesis advances the scientific basis for next-generation risk assessment. The resulting tools are transparent, independent of animal data and designed for efficient evaluation across large chemical sets. They support growing scientific and regulatory efforts to integrate ML-enabled modelling into toxicology and pharmacology.

Saad Lodhi | Maastricht University

Advancing Translational Genomics in Toxicology to Improve Human Relevance

“The central theme of my thesis is translation. Transcriptomic data can provide a common molecular layer for connecting toxicological findings across species, experimental models, organs, and levels of biological interpretation. By combining standardized and reproducible analysis with comparative toxicogenomics, machine learning, and AOP-oriented interpretation, the thesis explores different ways of making molecular responses more comparable and interpretable for human-relevant toxicology.”

ABSTRACT

This thesis investigates how transcriptomics and computational approaches can support more human-relevant and mechanistically informed toxicology. The core theme of the work is translation: using transcriptomic data to connect findings across experimental models, species, organs, and different levels of biological interpretation. Within this translational genomics framework, the thesis addresses challenges in analytical standardization, comparison of toxicological responses across biological systems, prediction of toxicological outcomes, and mechanistic interpretation of transcriptomic data.

The thesis first introduces R-ODAF-Shiny, a graphical implementation of the R-ODAF differential-expression workflow designed to make standardized and reproducible RNA-seq analysis more accessible. This analytical framework is then applied across several organ-specific toxicology case studies.

In the intestinal toxicity work, doxorubicin-induced responses are investigated across human colonoids, mouse colonoids, and mouse colon in vivo. This comparative transcriptomic study examines conserved and model-specific molecular responses across species and experimental systems and explores how findings from different models relate to one another and to human-relevant toxicity. A second intestinal case study focuses on gefitinib and examines regional toxicity in the jejunum and colon, together with their corresponding organoid models. This work investigates how transcriptomic responses differ between intestinal regions and how these responses compare between in vivo and organoid systems.

The thesis also includes a machine-learning study of drug-induced hepatic steatosis using transcriptomic data from primary human hepatocytes, primary rat hepatocytes, and rat liver tissue. Multiple machine-learning approaches are evaluated to classify compounds according to steatogenic potential and to identify predictive transcriptomic signatures across experimental models.

Finally, the thesis integrates transcriptomic analysis with Adverse Outcome Pathway (AOP)-based tools in a platinum-induced kidney toxicity case study involving cisplatin, carboplatin, and oxaliplatin. Differentially expressed genes are linked to AOP Key Event-associated gene sets to support structured mechanistic interpretation of the transcriptomic responses.

Together, these studies approach translation from several complementary perspectives: standardizing transcriptomic analysis, comparing responses across models and species, examining organ- and region-specific toxicity, applying machine learning for prediction, and linking molecular responses to AOP-based mechanistic frameworks. The broader aim is to improve how toxicogenomic data are analyzed and interpreted in support of more human-relevant toxicological research. Several of these research activities were conducted within or in connection with the ONTOX/ASPIS and VHP4Safety initiatives.

Rita Ortega Vallbona | ProtoQSAR

Computational approaches for next-generation risk assessment of hepatic toxicity

“The key message of my thesis is that integrating complementary, interpretable computational methods within the AOP framework can provide mechanistically grounded predictions of drug-induced hepatic steatosis, supporting the development and regulatory uptake of NAMs.”

ABSTRACT

Drug-induced liver injury (DILI), understood as liver damage caused by exposure to drugs or other xenobiotics, is one of the main reasons for drug attrition. The clinical manifestations of DILI range from hepatocellular and cholestatic damage to steatosis and fibrosis. While most cases of DILI resolve after discontinuation, it can progress to life-threatening conditions, such as acute liver failure. The network of biological pathways that can lead to the appearance of DILI is very complex and involves numerous mechanisms, making the prediction of DILI extremely difficult. In the context of next-generation risk assessment, the use of human-relevant approaches offers insights into the mechanisms of toxicity. The knowledge about these mechanisms is ideally gathered into Adverse Outcome Pathway (AOP) networks that connect Molecular Initiating Events (MIEs) and Key Events (KEs) to an adverse outcome. In this thesis, we explore several computational methodologies that can be used to input data predictions into an AOP network to add to the weight of evidence during its development and transition to a quantitative AOP. In the first chapter, we review examples of in silico approaches that have been used to address different building blocks in the steatosis AOP, including (Quantitative) Structure-Activity Relationships ((Q)SAR) models, read-across, omics analysis, and structure-based methods. In the second chapter, we develop (Q)SAR models for two types of chemical-induced lipid accumulation, taking a small group of carboxylic acids that present different effects in HepG2 cells with only small variations in their chemical structure. In the third chapter, we switch to the use of structure-based methods for the prediction of MIE alteration by exposure to compounds. We explain how we have developed an automated docking server with prepared protein structures that allow for the binding of query compounds to the set of proteins of interest and not only offer the binding energy as a result but also analyse the protein-ligand interactions and use that to add information that could lead to a better classification of the query compounds as binders or non-binders. Overall, the work presented here intends to bridge the application of computational approaches for next-generation risk assessment. By emphasizing the importance of mechanistic information and tool transparency, this work aims to aid in advancing the regulatory acceptance of these approaches to be used in AOP-informed hazard assessment pipelines.

Anouk Verhoeven | vrije universiteit brussel

DEVELOPMENT OF A MECHANISM-BASED TEST PLATFORM TO DETECT CHEMICAL-INDUCED STEATOTIC LIVER INJURY

“Through these different studies, this doctoral thesis project has provided an important contribution to the establishment of a NAM to study the potential of chemicals to cause steatosis in the liver. This NAM is not only of interest to pharmaceutical industry, but equally to food and biocide industries, aiding in the progression towards a future with less reliance on the use of animals for chemical safety evaluations.”

ABSTRACT

Steatosis constitutes a major manifestation of liver injury. Steatotic liver insults result from the accumulation of fatty acids and can be caused by chemicals from various applicability domains. Current animal models poorly detect human chemical-induced steatotic liver injury, which is mainly due to interspecies differences and gaps in the mechanistic understanding of this type of hepatotoxicity. This doctoral thesis project aimed to develop a new approach methodology (NAM) that focuses on key molecular and cellular events in toxicity pathways suitable for integration into a testing strategy to assess the steatogenic potential of various chemicals. In the first study, an existing adverse outcome pathway (AOP) network mechanistically describing the development of liver steatosis was optimized and analyzed using semi-automated approaches to determine key events (KE) relevant for the prediction of liver steatosis. Subsequently, an 2-step in vitro test battery was developed to target these KEs, consisting of a human-relevant cell-based model linked to assays that quantify KE alterations at transcriptional and/or functional levels. The test battery was set up in a human stem cell–based liver model (second study) and a human HepaRG liver cell model (third study), and the capability to detect steatogenic hazard properties was evaluated for chemicals upon repeated dose exposure. The results of the second study revealed the power of the in vitro system to assess anabolic pathways associated with liver steatosis that contribute to a chemical’s steatogenic properties. In the third study, the in vitro test battery was complemented with in silico methods, enhancing the mechanistic coverage of KEs within the AOP network. The NAM was evaluated for its ability to predict steatogenic potential using a weight-of-evidence (WoE) scoring framework. This study demonstrated that combining in silico and in vitro mechanistic data enabled the identification of chemicals possessing a steatogenic hazard. In the fourth study, in vitro distribution kinetic profiles of chemicals were evaluated in the human HepaRG liver cell model, demonstrating that a chemical’s potency estimate is derived both by its intrinsic bioactivity as well as by its in vitro biokinetics. Therefore, cell-associated concentrations are a more appropriate metric for hazard characterization. Overall, this doctoral thesis project, conducted within the European ONTOX project, provides an important contribution to the establishment of a NAM to assess liver steatogenic potential of chemicals and thus to the operationalization of next generation risk assessment (NGRA).