New Approach Methodologies (NAMs) for Carcinogenicity Evaluation

European Partnership for Alternative Approaches to Animal Testing (EPAA) Initiative on NAMs for Carcinogenicity of Agrochemicals

Prof Mirjam Luijten introduced the concept of New Approach Methodologies (NAMs). Although different definitions for NAMs are being used, there is general consensus that NAMs comprise innovative methods that can be used in regulatory safety testing of chemicals and pharmaceuticals. NAMs include in silico methods, ranging from well-established tools like SARs (structure-activity relationships) to (upcoming) machine learning techniques, as well as in vitro test methods, which may range from simple 2D models to advanced, complex 3D models. Depending on the problem formulation and regulatory framework at hand, a combination of NAMs can either complement or even replace traditional animal testing. For regulatory assessment of more complex health effects such as carcinogenicity, an alternative approach that no longer relies on animal testing is more difficult to achieve. An interim solution may be to combine NAMs with short-term (ie, 90-day) toxicity studies in experimental animals encompassing traditional endpoints and, for example, omics readouts to increase mechanistic insight in the health effects induced by chemical treatment. Such an approach would already contribute to the 3Rs principle and could, in the shorter term, provide sufficient confidence to serve as the basis for deriving a point of departure (PoD). In addition, enhanced mechanistic insight could support further improvement of an NAM-based approach for carcinogenicity assessment.

The selection of methods to be used in an NAM-based approach should be guided by our knowledge of the adverse outcome pathways (AOPs) relevant for carcinogenesis. AOPs constitute a mechanistic representation of critical toxicological effects over different layers of biological organization, starting with initial interaction of a chemical with a molecular target (molecular initiating event [MIE]), leading via a series of key events (KEs) to an adverse outcome (AO) at the individual or population level.^22,26

Prof Mirjam Luijten then presented the initiative funded by the European Partnership on Alternative Approaches to Animal Testing (EPAA) aimed at the development of an NAM-based approach for assessing the carcinogenicity of agrochemicals. The traditional 2-year carcinogenicity assays in rodents are under scrutiny regarding their accuracy, cost, and relevance regarding the number of animals used.^10,12,21 A mechanism-based, weight of evidence (WoE) approach requires enhanced understanding of the AOPs relevant for carcinogenesis in humans. The EPAA project is aimed at the identification of both AOPs to be included in a WoE-based approach and possible knowledge gaps. For this, over 400 unique agrochemicals were analysed, of which 170 agrochemicals could be categorized as nongenotoxic carcinogens (NGCs). ¹⁴ In the analysis, only tumours that show a treatment-related tumour response were included. For two-thirds of the approximately 340 tumour cases, an underlying AOP or network of AOPs could be identified. Despite the wide variety of tumours in various organs distilled from the carcinogenicity studies, the number of identified AOPs was limited. This illustrates that the transition towards a WoE approach appears manageable. The AOPs involved in the remaining tumour cases remained undetermined. For these ‘unknowns’—representing 114 tumour cases related to 72 chemicals—patterns in tumour incidences and underlying AOPs were explored in a stepwise fashion. In brief, tumours suspected to have occurred due to excessive dosing were excluded and emphasis was placed on commonalities in tumour type and putative preneoplastic lesions across chemicals. The project team therefore also involved multiple experienced pathologists. The analysis yielded additional AOPs that should be included in WoE-based carcinogenicity assessment. For some tumour cases, the underlying AOPs could not yet be unravelled (Bouwmeester et al., 2026). ⁴ The resulting overview of the AOP (networks) underlying nongenotoxic carcinogenicity of agrochemicals will greatly facilitate the selection of NAMs for the development of a WoE approach to carcinogenicity assessment.

In Silico Approaches to Predict Carcinogenicity

Prof Mark Cronin provided an overview of the history and advancements in in silico methodologies for predicting the carcinogenic potential of chemicals. He highlighted both real successes and the current limitations of these approaches.

The knowledge that chemical structure is fundamental to biological activity, and most notably its toxicity, has been apparent since the mid-19th-century. This has evolved into a number of techniques that can be used to predict hazard and exposure. The most usable methods have been formalized into computational tools for toxicity prediction which include, but are not limited to, (quantitative) structure-activity relationships (QSARs), structural alerts, read-across, machine learning and artificial intelligence (AI). These varied approaches have in common that they are attempting to relate some feature(s) of chemical structure and/or physicochemical properties to toxicological activity. Such tools have been used routinely since the 1980s, with a particular growth in their use in the past three decades due to commercial pressures including the need to eliminate toxicity early, ethical concerns over the use of animals, as well as regulatory need and legislation. ¹⁹

In silico, or computational, tools have a number of uses, ranging from identifying structures with potential toxicity, for example, for screening and prioritization of large inventories, or as part of a testing strategy, or WoE to perform hazard identification of a substance. They are applied in specific regulatory situations such as ICH M7, EU REACH, Classification, Labelling and Packaging (CLP), as well as the prioritization and screening of large inventory. ²⁵ Computational models are also instrumental in dealing with new problems that may arise in toxicology, with the ongoing issue of N-nitrosamines being one such example. ²⁴

Computational tools are used widely to assist in the prediction of effects relating to carcinogenicity. The full range of techniques can be used, starting with a knowledge of molecular (sub-)structures that are associated with electrophilic interactions with DNA. These so-called structural alerts build on the fundamental basis and investigation of carcinogenicity data by Ashby and Tennant. ¹ More quantitative predictions may be achieved with quantitative structure-activity relationships (QSARs). The methodology underpinning QSARs has been extended to include machine learning methods which can relate the hazardous effects of a large range of substances to aspects of their physicochemical properties or structural attributes.

There are a variety of places in the drug development pipeline where computational tools may be applied. They have found particular use in screening of candidate molecules early in drug development. There are especially useful and widely applied in the identification of potent compounds, being largely successful at identifying DNA-reactive compounds such that they may be eliminated early. The reason for the success is due to the high level of understanding in the molecular initiating event (MIE) and the possibility of modelling interactions with DNA.

In contrast, modelling nongenotoxic carcinogenicity has resulted in much less robust predictions. There are a number of reasons for the lower quality of noncarcinogenic models, but key amongst them are the number and complexity of mechanisms, lack of definitive MIEs on which to base a model, and the paucity of suitable data for modelling. Integrating such computational tools into carcinogenicity assessment should include an assessment of the uncertainty in the approaches and whether a prediction is acceptable on its own or will require other information, which may include “big” data from New Approach Methodologies (NAMs) and in vivo bioassays.

A potential solution to prediction of complex endpoints is the process of read-across, whereby similar structures are assumed to have similar activities. The utility of read-across has been demonstrated for carcinogenic substances, for instance where substances have common metabolic processes. ² Justification may require information and variety of lines of evidence to provide an overall WoE.

At the current time, there is no, or little, use of AI beyond the development of global QSARs utilizing machine learning (as noted above). There are a number of possibilities, however, for instance probabilistic approaches could integrate information from in vitro data that have been derived from Integrated Approaches to Testing and Assessment (IATA). This would be particularly valuable for some of the nongenotoxic mechanisms of carcinogenicity (an example of such an IATA was described by Jacobs et al ¹⁵ for breast cancer. Although little explored, large language models (LLMs) may ultimately assist in compiling information from pathology reports, forming summaries and identifying otherwise undiscovered trends and relationships. ¹⁷

In summary, in silico approaches have the potential to assist in the identification of hazards relating to carcinogenicity. The approaches range from structural alerts, relating aspects of chemistry to toxicity up to machine learning approaches developing models from large datasets. Computational models work optimally when there is a mechanistic basis and have found success in predicting DNA-reactive genotoxicity. The complexity of nongenotoxic mechanisms of carcinogenicity has hindered their use to definitively replace the long-term carcinogenicity assay, but the possibility of integrating multiple lines of evidence should be investigated further to achieve this.

Genomics-Based NAMs for Classification of Carcinogens in Pharma

Dr Heidrun Ellinger-Ziegelbauer discussed the development of the use of genomics data for the prediction of carcinogenicity of pharmaceuticals. She outlined the evolution of transcriptomics data over the years and its potential for predicting long-term carcinogenicity from short-term animal studies. It is the objective of the Health and Environmental Sciences Institute (HESI) Carcinogenomics Working Group to use genomic signature and biomarkers for assessing tumorigenic potential, especially for rodent liver cancer. ⁶

The primary outcomes of this long-term, ongoing effort were presented. Carcinogens can generally be assigned to two major categories, based on their overall mode of actions (MOAs), which has implications for their regulation:

A major target organ for chemical-induced cancerogenesis is the liver, which is well described in rats due to the performance of the so-called 2-year cancer bioassay in the context of chemical risk assessment. For an initial investigation whether short-term studies combined with molecular omics methods may be able to classify compounds with respect to later liver tumour induction, liver gene expression profiles were analysed after up to 14-day treatment of rats with representative rodent hepatocarcinogens, including GCs and NGCs at carcinogenic doses, in comparison with noncarcinogens. ⁹ Histopathological findings supported the dose selection, featuring weak apoptosis and/or necrosis and inflammation for GCs, and early or regenerative hyperplasia for NGCs. For functional interpretation of the deregulated genes to increase mechanistic insight, the genes were assigned to a biochemical category describing its main biochemical, cellular or systemic function. Then, a “Tox” category for each gene was defined using the biochemical category, the direction of deregulation, and available information upon upstream pathways and transcriptional regulators. This revealed that GCs induce a DNA damage response, characterized by upregulation of target genes of the transcription factor p53.

In comparison, NGCs affect various pathways, such as NR activation, cell proliferation, and oxidative stress responses. These pathways may all contribute to the NGC MOA if they support survival or proliferation of mutated cells or induce mutations themselves via generation of reactive species. Thus, gene expression profiles representing certain pathways or functions may contribute to characterization of a compound’s carcinogenic potential after short time treatment of rodents. Since different NGCs are associated with different MoAs, distinct gene expression signatures representing major MoAs are needed. This lends itself to the AOP concept with signatures representing well defined MIEs or KEs as way forward. The publication of extensive transcriptomics data sets representing liver gene expression profiles after short-term treatment of rats with GCs, NGCs and other chemicals then made engagement into development of such signatures possible. This was taken up by the Carcinogenomics Project (https://hesiglobal.org/emerging-systems-toxicology-for-assessment-of-risk-committee/). The major objectives of this project are (1) determination of gene expression signatures representing major MIE/KEs for rat carcinogens on (a) a general mechanistic level, associated with (b) signature induction thresholds based on known carcinogenic doses, and (2) identification of cancer driver gene mutations in few mutated cells only within a pool of cells without DNA sequence mutations, before visible tumour appearance, employing highly sensitive sequencing methods with very low error rates, also called error-corrected next-generation sequencing (ecNGS). ⁶

With respect to objective 1, the different carcinogens used for generating the available short-term rat liver transcriptomics data (TG-GATEs, DrugMatrix, IMI Marcar), were assigned to the different MIEs, gene expression data of the corresponding samples were compiled and quality controlled and analysed by several experts using their usual workflows. This yielded several overlapping gene expression signatures per MIE (a list of genes associated with a certain expression pattern). Studies with rats with either p53 or any of the NRs knocked out and treated with three reference compounds each are being used to confirm and adapt signatures derived from WT rats.

Investigations with respect to objective (2) “cancer driver mutation detection” are ongoing in collaboration with the HESI Genetic Toxicology Technical Committee (GTTC).^18,23,27

With respect to detection and characterization of carcinogens using in vitro models, which is the goal for NAM approaches, more research appears necessary. It is to be noted that for GC, well-established in vitro models are available. For subtopics like characterization of compounds which are AMES-neg but potentially false-positive in chromosome damage assays, HESI has developed an in vitro transcriptomic biomarker to predict the probability that an agent is DNA Damage-Inducing (DDI) or non-DDI. ¹⁶ This approach is currently undergoing biomarker qualification with the Food and Drug Administration (FDA).

Concerning NGCs, simple in vitro models will not suffice to cover the various MoAs employed by NGCs to induce cancer, as exemplified by the comparison of 3 GCs and 3 NGCs applied to both a 2D primary rat hepatocyte model and to rats in vivo. ⁸ On the basis of the AOP concept, insights gained from investigations of in vivo studies employing omics methods in addition to traditional parameters, like the development of NR-MIE gene expression signatures mentioned above, could be used to develop in vitro models for specific single NGC MoAs.

Advanced Cellular Models for Carcinogenicity Assessment

Dr Kelly Evans (AstraZeneca) focused on the challenges associated with traditional preclinical drug safety models, which often do not accurately represent human physiology. Many current in vitro preclinical safety models use either 2D-cell cultures or cancer cell lines, neither of which accurately represent the patient population due to their simplified nature or cancer background. Therefore, there is often a translational disconnect between safety findings found in a dish and those observed in animal studies or in the clinic.

In the context of complex, multi-stage processes such as carcinogenicity, there is a clear need for more human-relevant, advanced cell models that can better recapitulate the complex structure and mechanisms of human tissues.

To this end, advanced human cell models, including organoids and organ-on-a-chip systems, are currently under development to help overcome these issues. Due to their 3D arrangement and increased complexity through inclusion of multiple cell types, they provide the opportunity to assess carcinogenicity within a physiologically relevant, human-appropriate system. Such systems have the advantage of being generally long-lived, allowing assessment of repeat dosing or carcinogenicity that may only be detected over an extended period.^5,7,13

Kelly Evans used the complexity of evaluating the safety of cell-based therapies to present new models and methodologies. Multicellular organoids could hold promise in providing material for organ transplantation, because of their unique potential for tissue repair. A model of bone marrow representing the possibility to test the immune response was presented. In addition, as data from human trials continue to expand, there is also scope for a significant impact of computational and machine learning models to make predictions of key parameters involved in immune adverse events. ³

Identifying Potential Carcinogenicity Risks through Innovative Genetic Toxicology Methods and Target Safety Assessments

Dr Joanne Elloway highlighted the pivotal role of target safety assessments, along with advancements in genetic toxicology and sequencing in predicting carcinogenicity during drug development.

Target safety reviews for carcinogenicity prediction focus on evaluating compounds for potential cancer risks early in development by reviewing the biology of the target, expression profiles, pathway roles (eg, cell cycle and DNA repair), and human genetics. These reviews assess biological pathways and molecular targets to identify potential oncogenic pathways and genetic markers, using published and data repositories.

Error-corrected next-generation sequencing is rapidly emerging as a valuable, highly sensitive, and accurate method for detecting and characterizing mutations in any cell type, tissue or organism from which DNA can be isolated. EcNGS is under extensive investigation for the use in genetic toxicology and has been applied to experiments assessing mutagenicity and carcinogenicity to quantify drug-/chemical-induced mutations and mutational spectra associated with cancer risk. It has potential applications in genotoxicity assessment as a new readout for traditional models, for mutagenesis studies in 3D organotypic cultures, and for detecting off-target effects of gene editing tools. ²⁰

These methodologies may be conducted without animal testing, thereby facilitating early identification of carcinogenic risks.