Extrapolation steps in quantitative in vitro to in vivo extrapolation to advance in vitro-based chemical risk assessment for human health

Abstract

The transition towards Next Generation Risk Assessment (NGRA) aims at increasing human relevance and efficiency of chemical risk assessment, by using human-based in vitro and in silico methods. Quantitative in vitro to in vivo Extrapolation (QIVIVE) is central to this approach and enables the derivation of health-based guidance values (HBGV) from in vitro data. However, consensus on the required extrapolation steps within QIVIVE is missing. To find out which extrapolation steps are necessary when executing a QIVIVE analysis, this review identified and analyzed potential key extrapolation steps across 212 research articles and regulatory documents. Six key extrapolation steps were identified: 1. Extrapolations based on choice of in vitro cell model 2. Extrapolation from nominal concentration to target concentration in vitro, 3. In vitro target concentration to in vivo target compartment concentration (dose metric alignment), 4. In vivo target compartment concentration to external dose, 5. Effect extrapolation, and 6. Extrapolation of exposure duration. Furthermore, uncertainties and variations within the steps should be considered. Although extrapolation steps 2 to 4 are increasingly standardized, challenges are more prominent regarding the execution of in vitro model choice, effect and exposure duration extrapolation. To advance the field of NGRA and gain trust in QIVIVE, further QIVIVE case studies and in vitro studies are needed that report results systematically on the different extrapolation steps.

Graphical abstract

Highlights

• •Six key extrapolations steps for QIVIVE in chemical risk assessment were identified.
• •Dose extrapolation steps are increasingly standardized across QIVIVE frameworks.
• •Effect and exposure duration extrapolation remain the least harmonized steps.
• •Systematic reporting of in vitro studies is essential to advance NGRA.

1. Introduction

The risk assessment of chemicals aims to identify exposure levels of chemicals that are safe for human health. However, traditional risk assessment relies heavily on animal testing, which involves high costs and raises ethical concerns. These limitations have increased the demand for more human-relevant approaches that can translate in vitro findings into human-relevant predictions. Developments in the understanding of the toxicodynamics and toxicokinetics of chemicals have paved the way for a risk assessment based on in vitro experiments with human cell models. Integration of this concept into Next Generation Risk Assessment (NGRA) aims to improve human relevance and efficiency of testing by moving away from animal testing and switching to in vitro and in silico methods. For a definition of NGRA, see Kasteel (Kasteel, 2021). This article focuses on the integration of in vitro methods into NGRA.

Quantitative in vitro to in vivo extrapolation (QIVIVE) is an essential part of NGRA, as QIVIVE is seen as the combination of extrapolation steps that are needed to translate results from in vitro methods to quantitative in vivo predictions. While multiple definitions of QIVIVE exist, this article follows one of the definitions given by Chang, et al. (Chang et al., 2022): The extrapolation of an in vitro concentration and bioactivity measurement to an in vivo dose and toxicological endpoint, by a combination of toxicokinetic and toxicodynamic modeling.

One of the main aims in the risk assessment of chemicals is to derive a health-based guidance value (HBGV), i.e. the exposure or dose below which no adverse effects or unacceptable risks are expected to occur in humans. A HBGV is historically derived from animal studies (E. S. Committee et al., 2021), however, the ambition of NGRA is to rely more on in vitro studies using human cells, for the reasons mentioned above. For this, it is important to assess whether a HBGV can be derived from a combination of in vitro results from human cells, in silico results and, when needed, results from a reduced amount of higher-tier animal tests. A combination of extrapolation steps is needed to derive a HBGV from these data. QIVIVE can be used for these extrapolations.

Whilst the extrapolation of the in vitro concentration leading to a certain effect to an in vivo dose is often seen as the main step in QIVIVE (OECD, 2025; Yoon et al., 2015), several frameworks describe the need for other extrapolation steps (Magurany et al., 2023; Zwartsen et al., 2025). Although a variety of studies have applied QIVIVE (Fabian et al., 2019; Louisse et al., 2017; Moreau et al., 2024; Strikwold et al., 2013; Wetmore et al., 2012; Wetmore, 2015), the extrapolation steps that are considered necessary for QIVIVE differ. This shows that consensus has not yet been reached on the extrapolation steps that are required in QIVIVE.

On the other hand, when deriving a HBGV based on animal data, consensus does exist about the required extrapolation steps. Institutes responsible for performing risk assessments at a global and European level (i.e., the World Health Organization and the European Food Safety Authorization) (E. S. Committee, 2012; WHO, 2020), describe the process of deriving a HBGV based from animal experiments as follows:

• 1.Identification of the critical endpoint for the chemical
• 2.Determination of the Point of departure (PoD)
• 3.Application of assessment factors (AFs) or otherwise account for differences between the experimental setup and population, and the target (human) population and exposure scenario, e.g. covering interspecies, intraspecies and exposure duration differences, or route-to-route extrapolation

In contrast to animal-based risk assessment, such a concrete list of extrapolation steps does not exist for in vitro-based risk assessment. A consensus on the required extrapolation steps to obtain a HBGV from in vitro data would help harmonize developments in the NGRA field and increase regulatory potential of in vitro-based risk assessment.

To address this knowledge gap, this review aims to identify the extrapolation steps required in QIVIVE within chemical risk assessment. Therefore, articles and frameworks were collected that perform, discuss or suggest one or more QIVIVE extrapolation steps. These identified extrapolation steps are critically discussed, and from this, the key steps that are required in QIVIVE will be distinguished. Each identified extrapolation step is discussed in detail in view of its importance in the context of deriving a HBGV based on in vitro data. Current challenges regarding the extrapolation steps and suggestions for further improvements will be provided. Finally, the identified QIVIVE extrapolation steps, based on the obtained insights, are discussed and put forward.

This work contributes to the development of a harmonized QIVIVE approach that is applicable for deriving a HBGV from human-based in vitro models. Identifying all needed extrapolation steps will support in streamlining QIVIVE applications, and the reporting and assessment thereof.

2. Methods

To identify extrapolation steps that are part of QIVIVE, a search in Embase, Scopus and PubMed was performed on the 21st of January 2026. After deleting duplicates, the search query resulted in a total of 1654 articles. The search terms used in this study are provided in appendix table A1.

To facilitate the literature search, ASReview was used in this study (van de Schoot et al., 2021). ASReview requires an annotated subset of the identified articles and uses an active learning algorithm to make predictions on the relevance of the complete set of articles, based on the title and abstract. This way, not all articles need to be studied in order to collect the most relevant articles. ASReview presents articles to the user, always showing the most relevantly predicted article, the user then marks this article as relevant or irrelevant. This information is used by ASReview to adjust the sorting of the articles when needed. At some point, relevant papers run out and stop appearing. A stopping point needs to be selected to quit the reviewing process.

In this study, a random subset of 431 articles (approximately 26% of 1654 articles) was first taken from the complete set of articles and used as the initial annotated dataset for ASReview. This subset contained 18.3% relevant articles.

Then, the stopping point was determined in a statistical way, via a method that has been shown to provide high recall, or true positive rate (Callaghan and Muller-Hansen, 2020). Assuming that the complete set of articles has the same fraction of relevant and irrelevant articles, and that 95% of articles can be identified using this method, the stopping point was set at 287 relevant articles found (95% of 18.3% of 1654). An overview of the article selection can be seen in appendix fig. 1.

First, the subset of 431 articles was manually labeled by two independent reviewers as relevant or irrelevant based on the title and abstract. The inclusion criteria for marking a paper as relevant were: 1) the paper describes or executes at least one extrapolation step that the authors of the articles considered to be part of QIVIVE, and 2) the paper is written in the context of chemical risk assessment.

The exclusion criteria for excluding articles were: 1) the paper focuses on (nano) particles, including metal containing (nano)particles, as the risk assessment of particles is different than that of other chemicals and may require different extrapolation steps. 2) the paper only extrapolates parameters from in vitro assays to in vivo parameters to be used in a kinetic model. These papers were left out because they do not aid in identifying the necessary extrapolation steps in QIVIVE. And 3) the paper only studies a certain technique that can be used as a QIVIVE step but does not discuss the place or usage of the technique in a risk assessment context. These articles were left out because they do not directly lead to a better understanding of QIVIVE as a concept.

After manually labeling 431 articles as relevant or irrelevant based on title and abstract, this data was used by ASReview to sort the remaining articles based on relevance. Articles presented by ASReview were then marked as relevant or irrelevant. When the pre-determined stopping point was reached, i.e., when 287 relevant articles were found, the full text of the 287 articles were screened.

Several articles (n = 76) were excluded after full-text screening based on the exclusion criteria, leading to a total of 211 relevant articles. This list of 211 articles is available as supplementary file 1. Each of the articles was then assessed and annotated for the specific extrapolation steps executed or described, starting from an initial set of expected extrapolation steps. Any newly encountered steps were added to the list iteratively. These annotations were used to collect information about the different extrapolation steps in QIVIVE. Using these annotations, the included articles were selected based on their direct relevance on the identified extrapolation steps, leading 70 included articles.

Additional records (n = 35) were included when they were identified as key literature by other articles or were frequently cited. These articles were added to the dataset, to obtain a more complete overview of the literature, and to not be bound by the search strategy alone. Documents written by regulatory authorities such as OECD and EFSA were also added under this category, as these were not uncovered in the search strategy. The addition of these records lead to the 105 included articles. A flow diagram of the literature search can be seen in Fig. 1.

Fig. 1.

PRISMA flow diagram of articles selection. Based on the template given by Page, et al. (Page et al., 2021).

3. Results

In addition to the articles that described a relevant QIVIVE extrapolation step, seven frameworks for executing QIVIVE in chemical risk assessment were identified in the literature. These frameworks aim to provide a complete overview of all QIVIVE extrapolation steps, although a framework does not necessarily cover all steps. Six potential QIVIVE extrapolation steps were identified from the reviewed articles and frameworks. Although terminology varied across sources, the extrapolation steps described in these articles and frameworks could all be grouped under the following categories (not necessarily in this order):

• 1.Extrapolations needed based on choice of in vitro cell model
• 2.Extrapolation from nominal concentration in vitro to target concentration in vitro
• 3.Extrapolation from in vitro target concentration to in vivo target compartment concentration (dose metric alignment)
• 4.Extrapolation from target concentration in vivo to external dose
• 5.Extrapolation from in vitro effect to in vivo effect
• 6.Extrapolation from acute to (sub-) chronic exposure

Additionally, next to extrapolation steps, an often discussed and essential point considered when executing QIVIVE is the analysis of uncertainties and variability. A graphical overview of the identified extrapolation steps is given in Fig. 2, starting from in vitro concentration-response data ending with in vivo dose-response data that can be analyzed further for the derivation of a HBGV. In practice however, if the goal is to derive a HBGV, it may not be strictly necessary to derive the entire in vivo dose-response curve. Instead, one could use an in vitro-based point of departure (PoD). Nevertheless, because it is difficult to judge how a given magnitude of an in vitro response translates to a relevant in vivo effect size, the estimation of an in vivo dose–response curve is considered to be an ambitious but important step in the QIVIVE workflow.

Fig. 2.

overview of identified QIVIVE extrapolation steps. The extrapolation from in vitro concentration-response curve to in vivo dose-response curve is shown. The 6 identified extrapolation steps are numbered.

Firstly, an extrapolation should be made based on in vitro assay choice (Fig. 2, extrapolation 1). Cells from different sources can have different characteristics, leading to quantitatively different results across different toxicity assays. For example, primary cells may show different characteristics than cells from immortalized cell lines. As this introduces a quantitative difference, an extrapolation can be made to account for this.

Extrapolation steps 2 to 4 cover the dose extrapolation, from nominal in vitro concentration to external dose in vivo (Fig. 2, extrapolation 2 to 4): First, the nominal in vitro concentration is extrapolated to the concentration in a relevant (sub)compartment in the in vitro setting, e.g., a free concentration in the medium or an intracellular concentration. Secondly, the concentration in the relevant in vitro (sub)compartment is then linked to a tissue or fluid compartment in a kinetic model, often a Physiologically Based Kinetic (PBK) model, representing a tissue or fluid in a human. However, as in vitro and in vivo concentrations change over time, dose metrics are typically used to summarize the concentration-time curves into single representative concentration values (e.g. maximum concentration, C_max), which allows for the extrapolation of in vitro to in vivo concentrations. This choice of dose metric can cause a quantitative difference in predicted result. This dose metric selection subsequently facilitates the third extrapolation step, which consists of extrapolating the in vivo target concentration to an external dose by reverse dosimetry.

Effect extrapolation should be considered in the context of key events within Adverse Outcome Pathways (AOPs). An AOP describes a molecular initiating event (MIE) as an initial stressor to a cell, followed by a key event (KE) as a response to this MIE, the KE is then often followed by multiple KEs, ultimately leading to an adverse outcome (AO). AOPs are created according to the following principles: 1. AOPs are not stressor specific, 2. AOPs are modular, 3. AOPs are a pragmatic unit of development, not meant to capture every possible biological interaction and 4. Modular AOPs can be assembled into AOP networks (S. f. A. o. AOPs, 2026; OECD, 2017a).

These AOPs can inform on effect extrapolation. And in this way, the extrapolation from key event to adverse outcome is often studied (Fig. 2, extrapolation 5b). Additionally, it is suggested that there could be differences in required dose that results in the same magnitude of effect in a KE between in vitro and in vivo, caused by a lack of feedback mechanisms or homeostasis in vitro (Zwartsen et al., 2025; Zhang et al., 2018). However, the extrapolation from an in vitro key event to an in vivo key event is not studied in the dataset (Fig. 2, extrapolation 5a), highlighting a potential knowledge gap.

Lastly, exposure duration extrapolation should be considered (Fig. 2, extrapolation 6). Differences in exposure time between in vitro and in vivo can have effects on the predicted toxic doses and the type of effect observed. Exposure duration in vitro is often short, hampering modeling of chronic and sub-chronic exposures. This makes toxicities driven by long-term exposures or by accumulation more difficult to study in vitro. When these effects need to be studied, an extrapolation needs to be added to consider the difference in exposure duration.

Additionally, variability in target population can be considered for a more translatable risk assessment. In the identified literature, this is usually done by applying an assessment factor to account for intraspecies differences or by including intraspecies differences in physiology and toxicokinetics in the reverse dosimetry. Next to the variation, the approach used to account for the uncertainty introduced throughout the various extrapolations is often discussed in the identified literature. This is not an extrapolation step in itself, but an important analysis to execute along with the other steps. This approach differs for each extrapolation step, as the magnitude of the uncertainty introduced also differs between steps. While this topic will not be discussed in detail, it is added to the overview of QIVIVE, to accentuate its importance (Fig. 2).

The identified frameworks commonly addressed the extrapolation steps that are part of the dose extrapolation (steps 2–4) (Table 1). The other steps were only considered by some researchers. This shows that agreement on the extrapolation steps needed for QIVIVE is still lacking, and that work can still be directed towards harmonizing the extrapolation steps. By describing the current state and discussing the relevance of each of these steps, we aim to uncover the required extrapolation steps for QIVIVE in NGRA. The information available in literature on each step will be discussed in detail.

Table 1.

Overview of published QIVIVE frameworks and extrapolation steps described in them.

Reference / source	1. Choice of in vitro cell model1	2. Nominal concentration to target concentration in vitro	3. Dose metric alignment	4. Target concentration in vivo to external dose	5. In vitro effect to in vivo effect	6. Extrapolation from acute to (sub-) chronic exposure	Uncertainty and variation analysis
EFSA (2025) (E. P. o. P. P. Products et al., 2025)	x	x	x	x		x	x
Magurany et al. (2023) (Magurany et al., 2023)	x	x	x	x		Partly	x
OECD (2025) (OECD, 2025)	x	x	x	x			Partly
Wambaugh et al. (2025) (Wambaugh et al., 2025)	x	x	x	x			x
Yoon et al. (2015) (Yoon et al., 2015)		x		x
Zhang et al. (2018) (Zhang et al., 2018)	x	x	x	x	x		Partly
Zwartsen et al. (2025) (Zwartsen et al., 2025)	x	X	x	x	x	x	x

¹The need for careful selection of an assay is often discussed. The fact that the selection influences the results quantitatively, and could be seen as an extrapolation step, is not.

3.1. Step 1: Extrapolations based on choice of in vitro cell model

3.1.1. What is known?

The choice of in vitro cell model can influence the ultimate outcome of the hazard assessment. Different cell models may represent specific parts and functionalities of in vivo biology to a varying degree. This impacts the biological relevance of the in vitro outcome, which can in turn lead to different toxicity measures for the same compound. (Hondebrink et al., 2017; LeCluyse et al., 2012; Tukker et al., 2020). Cell model selection thus introduces variability and inherent uncertainty in the hazard characterization, which must be accounted for in NGRA.

It has been shown that the use of different cell models in vitro leads to variations in the measured potency of a chemical. For example, altered potency for a chemical has been measured between primary, cell-line derived and iPSC-derived neuronal cells (Hondebrink et al., 2017; Tukker et al., 2020; Tukker et al., 2018). This difference has been shown to come from the ratio of excitatory and inhibitory neurons, the seeding density, and the maturity of the cell model.

Liver cell models differ in their capacity to measure the potency of a chemical (LeCluyse et al., 2012). 2D cultures may fail to maintain phenotypic characteristics over time and can miss certain toxicity endpoints. More advanced models such as 3D cultures and co-cultures, on the other hand, better replicate the in vivo biology and increase complexity of the model. However, this increased complexity comes at the cost of lower throughput, decreased reproducibility, and increased technical challenges.

3.1.2. Limitations and suggestions

According to Astashkina, et al. (Astashkina et al., 2012), intercellular interactions are necessary for a cell to maintain normal processes. The microenvironment needed for a cell does not only require cell-cell interactions, but also cell-matrix interactions, tissue architecture, immune signaling, oxygen gradients, and mechanical forces. These processes are all disrupted in a traditional 2D cell culture. Without these interactions, such in vitro systems may lead to less realistic toxicity measures. To better capture these interactions, test strategies in which simpler and more advanced in vitro models are integrated are being worked on (LeCluyse et al., 2012; Sung, 2022; Aravindakshan et al., 2025).

Because cell models have varying characteristics, they could be complementary to each other in a hazard assessment (Hondebrink et al., 2017). For example, Liu, et al. (Liu et al., 2020) use two cell lines to assess the toxicity of anthraquinones on the liver. Effective concentrations were collected from HepG2/C3A cells and HuH-7 cells, in order to increase the chance of capturing the toxicity of the chemicals. This way, the difference between cell models can be used to obtain a more robust toxicity assessment. Studies that include multiple cell models of a tissue are also useful for future decision making when choosing a cell model (Sakolish et al., 2025; Lujan et al., 2019). To make the information obtained by these studies reusable, it is important to thoroughly characterize the cells used in an in vitro assay (Tukker et al., 2018; Lujan et al., 2019; OECD, 2017b).

To help characterize in vitro cell models, a unified template following OECD guidance can be used to simplify the documentation of in vitro methods (Krebs et al., 2019). Machine learning-based tools that automate the documentation of in vitro methods according to these templates can improve the standardization of documentation (Houweling et al., 2025). Standardized documentation is relevant for QIVIVE because it enables better comparison of effect concentrations across studies and supports the selection of appropriate assessment factors when combining data from different cell models.

It is currently difficult to quantify the uncertainty introduced by cell model choice, which complicates the extrapolation and combination of findings across studies. In vivo studies handle the uncertainty of model choice by using fixed assessment factors, without explicitly quantifying the uncertainty. However, in vitro approaches have the potential to characterize this uncertainty more explicitly, for example through systematic comparison studies. If this uncertainty would be quantified, this would present an advantage of in vitro-based risk assessment over traditional in vivo approaches.

3.2. Step 2: Nominal concentration to target concentration in vitro

3.2.1. What is known?

It is commonly accepted that the nominal concentration in an in vitro experiment may not represent the concentration that induces an adverse effect. (Kasteel, 2021; Zwartsen et al., 2025; Armitage et al., 2021a; Bloch et al., 2022; Broeders et al., 2013; Fischer et al., 2017; Fisher et al., 2019; Gulden et al., 2001; Henneberger et al., 2020; Henneberger et al., 2019; Kisitu et al., 2019; Kramer, 2010; Lin et al., 2025; Ouattara et al., 2011; Proenca et al., 2021; Proenca et al., 2019; Zwartsen et al., 2026). In an in vitro set-up, the added substance of interest may become unavailable to elicit a (toxicological) response e.g., because it binds to plastic, glass, or medium lipids and proteins, or is evaporated, or is transformed by abiotic degradation or cellular metabolism (Fischer et al., 2017; Kramer, 2010). Conversely, chemicals that accumulate in cells can result in intracellular concentrations that exceed the nominal concentration. Both processes mean that using the nominal concentration as a proxy for the effective concentration can lead to either an over- or underestimation of the actual effective concentration.

For example, Henneberger, et al. (Henneberger et al., 2021) showed that when using the nominal concentration as opposed to the free plasma concentration, the hazard of a chemical can be overestimated by a factor of 60. On the other hand, when aiming to estimate intracellular concentrations, the use of nominal concentrations may lead to underestimations for substances that accumulate in cells (Gulden et al., 2001). Similar conclusions are shown by Broeders, et al. (Broeders et al., 2013) and Kasteel (Kasteel, 2021).

The ideal approach in in vitro-based risk assessment is to use the effective concentration within a relevant in vitro compartment that actually causes an effect (Kisitu et al., 2019; Kramer, 2010; Armitage et al., 2014; Groothuis et al., 2015). This is known as the Biologically Effective Dose (BED). If this concentration cannot be measured, the concentration needs to be used that approximates this effective concentration most closely.

The free concentration in the medium is generally considered more biologically relevant than the nominal concentration, as it represents the concentration available to interact with the molecular target, and it is therefore closer to the BED (Kasteel, 2021; Zwartsen et al., 2025; Armitage et al., 2021a; Broeders et al., 2013; Fischer et al., 2017; Fisher et al., 2019; Gulden et al., 2001; Henneberger et al., 2020; Henneberger et al., 2019; Kisitu et al., 2019; Kramer, 2010; Lin et al., 2025; Ouattara et al., 2011; Proenca et al., 2021; Proenca et al., 2019; Zwartsen et al., 2026). However, when the mode of action (MOA) is initiated intracellularly, or when a chemical accumulates highly, the intracellular concentration is an even better proxy for the BED (Kasteel, 2021; Zwartsen et al., 2025; Broeders et al., 2013; Fischer et al., 2017; Fisher et al., 2019; Henneberger et al., 2020; Proenca et al., 2021; Proenca et al., 2019; Zwartsen et al., 2026). When the free intracellular concentration can not be distinguished, total intracellular concentration can be used instead. Alternatively, the cell-associated concentration (concentration in and on cells) can be measured, but this can overestimate intracellular exposure for lipophilic chemicals, it is nonetheless appropriate when the cell membrane itself is the chemical's target receptor.

To give guidance for selecting the most appropriate in vitro target concentration, i.e. the BED, Groothuis, et al. (Groothuis et al., 2015) developed a flowchart, which has been adopted by the OECD in their guidance document on good in vitro method practices (OECD, 2018). The choice of appropriate in vitro target concentration is based on the MOA of the studied effect and on the feasibility of the measurement of the target concentrations. In some situations, when the chemical is very hydrophilic, non-volatile and does not bind to serum proteins, the nominal concentration will be close to the free concentration and can be used instead, this saves efforts and resources.

After selecting the in vitro target concentration, the next step is to either measure this concentration directly in vitro, or to predict the concentration using an in silico model. Directly measuring a concentration is seen as the more reliable option, with in silico modeling as an alternative if measuring is not feasible (Kasteel, 2021; Armitage et al., 2021a; Bloch et al., 2022; Broeders et al., 2013; Gulden et al., 2001; Henneberger et al., 2020; Kramer, 2010; Proenca et al., 2021). An overview of published in silico models for predicting in vitro target concentrations and their applicability domains is provided by Proenca, et al. (Proenca et al., 2021).

3.2.2. Limitations and suggestions

Although various studies have improved the extrapolation of nominal in vitro concentrations to target concentrations, several limitations remain. One of these considerations is the applicability domain of in silico models which predict in vitro kinetics of a chemical over compartments. Especially the distribution of organic acids and ionizable chemicals often deviate from the predictions of equilibrium predictions (Armitage et al., 2021a; Henneberger et al., 2020; Henneberger et al., 2019; Proenca et al., 2021; Proenca et al., 2019).

It is suggested to compare multiple in silico models that predict in vitro target concentrations Lin, et al. (Lin et al., 2025). For these models, a sensitivity analysis can even be used to ensure the critical parameters are described well.

There are no validated kinetic models for predicting the distribution of a chemical over in vitro compartments yet and guidance about the validation of these models does not exist yet. This would be an important step forward for their usage in risk assessment. For the current usage of these kinetic models, the applicability domain is important to consider, as well as executing the sensitivity analysis and refining the most sensitive parameters. Each use-case needs a careful selection of a model based on chemical properties of the chemical and duration and type of the experiment. When possible, the in vitro target concentration should be measured experimentally instead. Currently, this is often not done and is highly recommended for future in vitro research.

A further consideration when selecting the in vitro target concentration is that chemical transformation during the assay is not only a loss process affecting the parent compound. Abiotic processes such as hydrolysis and autoxidation, as well as cellular metabolism, generate transformation products, which may be toxicologically active themselves (Huchthausen et al., 2022). When these products drive a substantial part of the response, attributing the effect solely to the parent concentration can lead to incorrect potency assessments and errors in the complete QIVIVE. Therefore, it is important to characterize the stability of a chemical over the course of an assay.

3.3. Step 3: In vitro target concentration to in vivo target compartment concentration

3.3.1. What is known?

The next step in the dose extrapolation is the extrapolation of an in vitro target concentration to an in vivo target compartment concentration. In this step, two choices need to be made: the choice of target compartment in vivo and the dose metric to be used.

The OECD guidance document on QIVIVE for developmental neurotoxicity (OECD, 2025) suggests that the in vitro target concentration must closely correspond to the in vivo target compartment concentration (or internal dose metric) at the target tissue. The in vivo target concentration of a chemical should consider the biological matrix (tissue or plasma concentration) and the duration and amount of exposure.

These pairings of in vitro and in vivo target concentrations should be considered on a case-by-case basis. An overview of in vitro metrics and their preferred pairings to in vivo compartments as found in the literature are given in Table 2.

Table 2.

In vitro target concentrations and their preferred in vivo compartment pairings.

In vitro concentration	Preferred in vivo compartment pairing	Note	Source
Free Intracellular concentration	Free concentration in tissue fraction of tissue compartment	Unbound chemical within the cell, seen as closest to the biologically effective dose (BED)	(Groothuis et al., 2015; Dimitrijevic et al., 2022)
Total intracellular concentration	Concentration in tissue fraction of tissue compartment	Total chemical within the cell	(Groothuis et al., 2015; Dimitrijevic et al., 2022)
Cell-associated concentration	Concentration in tissue	Total chemical within and bound to the cell. Extracellular membrane and cell membrane is included	(Proenca et al., 2021)
Free concentration in medium	Free concentration in plasma	Shown to have high concordance	(Dimitrijevic et al., 2022; Fischer et al., 2019; Honda et al., 2019)
Total concentration in medium (free + bound)	Total plasma concentration	Seen as sub-optimal, can over-estimate risk, not recommended	(Groothuis et al., 2015; Dimitrijevic et al., 2022)
Nominal concentration	Steady-state plasma concentration	Seen as sub-optimal, can over-estimate risk, not recommended	(Fischer et al., 2019)

The selection of compartment pairings also depends on the structure of the PBK model. Most PBK models do not consider tissue, intercellular fluid and/or blood fractions of tissue as separate compartments, due to a lack of information to parametrize them. In such cases, simpler in vivo compartments such as the total tissue concentration have to be used instead, which is a limitation that should be taken into account when interpreting QIVIVE results.

As the target concentration in vivo is not constant over time, comparing in vitro and in vivo concentrations requires the use of dose metrics. These are mathematical functions that can be applied to concentration-time profiles and produce a single output value. When the same dose metric is used in vitro and in vivo, these values can be compared.

The choice of appropriate dose metric depends on the MOA, the type of exposure, the critical effect, and whether it is reversible (OECD, 2025; Zwartsen et al., 2025; WHO, 2020; Groothuis et al., 2015; OECD, 2018; EPA, 2006a; Li et al., 2021; Lin et al., 2024; Mielke et al., 2019). For acute exposure, the maximum concentration (Cmax) is often appropriate. For chronic or repeated exposure, steady-state concentrations may be used as a basis. However, even under steady-state conditions, concentrations may still fluctuate in a reproducible pattern, e.g. due to repeated dosing. Consequently, a steady-state concentration may refer either to the peak steady-state concentration or to an average steady-state concentration. This distinction should be considered when using steady-state concentrations. When cumulative exposure is relevant for the critical effect, the Area Under the Curve (AUC) is recommended. Lastly, when the effect depends on average exposure over time or when exposure is irregular, the Time Weighted Average (TWA) can be used (Zwartsen et al., 2025; Groothuis et al., 2015; Lin et al., 2024; Ramanarayanan et al., 2022). These dose metrics are graphically summarized in Fig. 3.

Fig. 3.

Overview of dose metrics shown in concentration time curves after a. single and b. continuous exposure. Maximum concentration (Cmax) in red in the single exposure graph, steady state concentration (Css) in red in the continuous exposure graph. The area under the curve (AUC) is shown in blue and the time weighted average (TWA) in brown. The concentration metric on the y-axis could be free concentration, cellular concentration etc. (for interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

3.3.2. Limitations and suggestions

A dose metric is always a simplification of the full concentration-time curve, reducing it to a single value and losing some information with each dose metric. Because of this, it can be informative to evaluate multiple dose metrics and their inherent simplifications. If the preferred dose metric for the risk assessment is unclear, multiple dose metrics can be used and compared to each other (Zwartsen et al., 2025; Honda et al., 2019; Li et al., 2021; Lin et al., 2024; Ramanarayanan et al., 2022; Yoon et al., 2012). If no reasoning for the selection of a certain dose metric can be used, the most sensitive one could be selected as a worst-case scenario assumption.

For future research, it is recommended to standardize dose metric and target concentration selection for QIVIVE using guidance and decision frameworks, like the one presented by Groothuis, et al. (Groothuis et al., 2015). It needs to explicitly address compartment pairings and dose metric selections for different classes of chemicals, MOAs, endpoints and exposure scenarios. This would improve the reproducibility and comparability of QIVIVE-based risk assessments.

3.4. Step 4: Target concentration in vivo to external dose

3.4.1. What is known?

In QIVIVE, the in vivo target compartment concentration identified in step 3 is extrapolated to an external dose, using reverse dosimetry. This external dose can then be used as the point of departure for deriving a HBGV. Reverse dosimetry is commonly executed by using kinetic models that simulate the absorption, distribution, metabolism and excretion (ADME) of a chemical in the human or animal body (Louisse et al., 2017).

One type of these models is compartmental kinetic models. These can be used for screening and prioritization of chemicals (Paini et al., 2019; Punt et al., 2022; EPA, 2006b) and for the estimation of systemic exposure (OECD, 2021). However, these models perform adequately only when exposure is chronic, and biotransformation is slow and cannot predict concentrations at a specific target tissue (Armitage et al., 2021b). Because reverse dosimetry for QIVIVE requires predicting the external dose corresponding to a concentration at a target tissue, simple compartmental models are generally not suited. PBK models are the preferred choice of kinetic model, as these multi-compartment models mimic the physiology of the studied organism and provide a mechanistic approach (OECD, 2021; Xie et al., 2023).

When selecting or creating a PBK model for reverse dosimetry, the minimal compartments needed are: the target organ showing the critical effect, the compartment where the external dose enters the body (i.e. the compartment(s) influencing absorption rate and fraction) and the sites of metabolism and excretion for the chemical. In PBK models these compartments are usually linked by a blood compartment (Punt et al., 2022; Xie et al., 2023). Additionally, the model should try to include all other organs or tissues of physiological and biochemical relevance to the biokinetics of the chemical in question, such as the blood-brain barrier or the placental barrier, however, data for these compartments are currently scarce (Louisse et al., 2017; EPA, 2006b).

Both generic and chemical-specific PBK models can be used for reverse dosimetry in QIVIVE. Generic models can be used for a range of different chemicals or chemical groups, whereas chemical-specific models are developed for a single chemical and produce more robust predictions due to their more detailed ADME characterization. Both types of models can in theory be parameterized using only in vitro and in silico data, which is important for their use in NGRA (Punt et al., 2022; Najjar et al., 2022). Generic models are especially efficient when assessing a large number of chemicals and should be used for early tiers of risk assessment, whereas specific models are more appropriate when unique toxicokinetic behavior of the chemical is known (Louisse et al., 2017; Paini et al., 2019; OECD, 2021; Najjar et al., 2022).

3.4.2. Limitations and suggestions

When performing reverse dosimetry, an external exposure scenario has to be defined for mapping of the internal concentration to an external dose. The outcome of the reverse dosimetry depends heavily on the chosen exposure scenario and target population within the PBK model, and these should be chosen carefully (Wetmore et al., 2014; Tan et al., 2007; Ring et al., 2017; McNally and Loizou, 2015).

Kinetic models should ideally be validated against in vivo data, however, such data are often missing. The OECD guidance document on validation of PBK models (OECD, 2021) identifies five criteria of which only one requires in vivo data, meaning that models can be largely validated in a full in vitro and in silico context, which is important for NGRA.

To increase regulatory trust in kinetic models for reverse dosimetry, good modeling practices (GMP) should be followed (Najjar et al., 2022). This includes stepwise explanation of the model creation and parameter derivation. OECD and EFSA have both published reporting templates for PBK models (OECD, 2021; Aagaard et al., 2014). Following these templates improves the interpretability and reusability of models across studies.

3.5. Step 5: In vitro effect to in vivo effect

3.5.1. What is known?

In QIVIVE, not only the dose, but also the effect needs to be extrapolated from in vitro to in vivo. The preceding steps ensure that the in vitro and in vivo concentrations are expressed in a comparable metric, but even when the concentrations are matched, effects observed in vitro cannot be assumed to occur at the same concentration in vivo.

Cell stress in vitro, for example, does not necessarily mean organ damage in vivo. Since a variety of physiological and biological processes in whole organisms are not included in in vitro assays, this extrapolation step is best split into two steps: extrapolation of the effect from in vitro to in vivo (Fig. 2, extrapolation step 5a), and extrapolation from molecular-level effects to an adverse outcome in vivo (Fig. 2, extrapolation step 5b). This extrapolation step works in both directions, in vitro effects could both under- and overestimate in vivo effects. However, the literature shows that this extrapolation is usually performed in one overarching step (Hartung, 2018; Leist et al., 2017).

To support this last part of the extrapolation of an effect (5b), AOPs (as explained earlier in section 3) can be used as a scaffolding to link the in vitro effect to the in vivo effect. It is essential to provide evidence that the effects studied in vitro and in vivo can be mapped to each other, to make sure a risk assessment is undertaken for the correct adverse outcome. In this way AOPs play a role in the hazard assessment of a chemical (OECD, 2017a; Perkins et al., 2019; Tebby et al., 2022; Ter Braak et al., 2024). At the moment of writing, the AOP-wiki contains 520 AOPs (S. f. A. o. AOPs, 2026), and more are still being added, meaning that considering an AOP when doing a risk assessment of a chemical is increasingly feasible.

Efforts are currently being made to develop quantitative Adverse Outcome Pathways (qAOPs) (Conolly et al., 2017). qAOPs aim to quantify the relation between MIEs, KEs and adverse outcomes by creating and quantifying Key Event Relationships (KERs). qAOPs can be developed based on existing AOPs, but their structure may differ (Perkins et al., 2019). By quantifying the activation of each KE by the previous one, qAOPs can be used to simulate effect-effect relationships along the entire pathway. This is largely done with in vitro data, as studying a single KE in vivo is difficult. This allows the creation of dose-effect relationships for an adverse outcome based on in vitro data, which can then be used for risk characterization.

The development of qAOPs is inherently more complex than qualitative AOPs, as the quantification of each KER is an added source of complexity. Several qAOP case studies have demonstrated this approach in practice, showing that predictions made using in vitro-derived effect concentrations can be extrapolated to human equivalents doses that are in agreement with known in vivo toxicity data (Lin and Lin, 2020; Jarzina et al., 2022).

3.5.2. Limitations and suggestions

The step from KE to adverse outcome is captured by the AOP. It should be noted however that an impact on an early KE does not necessarily result in an adverse outcome. Feedback loops and repair mechanisms in vivo may prevent progression along the AOP. Because fewer biological steps separate late KE's from the adverse outcome, they are generally more predictive of a specific AO. While late KEs are not often studied, because of the higher difficulty of measurement, it would be interesting to study the link between in vitro and in vivo effects along the entire AOP. qAOPs are especially useful for this problem, as they quantify the dependency on the previous KE, however, the lack of in vivo data is a remaining problem.

Additionally, AOPs can be used as a mechanistic starting point for integrated testing strategies (S. f. A. o. AOPs, 2026; Tebby et al., 2022; Ter Braak et al., 2024) and for grouping chemicals based on shared toxicity pathways (Silva et al., 2015; Bil et al., 2025), allowing for more efficient risk assessment of chemical groups (OECD, 2017a).

Even with qAOPs, it remains unclear what the effect is of supporting biological systems like homeostasis or the immune system on the difference between in vitro effects and in vivo effects. These could cause the effect to be more or less pronounced in vivo. A way to address this is not commonly discussed in the literature and requires more attention.

One potential method of examining this uncertainty is the creation of response-response relationships between in vitro and in vivo. This is only possible for chemicals where in vivo data exist or can be obtained. So, this could most likely only be used in case-studies, but would be informative about the differences between in vitro and in vivo effects and could serve as a basis for read-across to other chemicals. This would provide a starting point for understanding the assumptions underlying effect extrapolation.

3.6. Step 6: Extrapolation from acute to (sub-)chronic exposures

3.6.1. What is known?

The duration of exposure to a chemical in vivo and in vitro is often different. In vitro toxicity assays tend to be shorter (Macko et al., 2021), because of the limited cell viability in vitro, whereas exposure to a chemical in real life can be as long as an entire lifetime. This difference in exposure duration should be considered when executing a QIVIVE analysis. Acute effects are typically defined as effects occurring following a single dose within 24 h (E. S. Committee et al., 2021), but this definition is dependent on the regulatory context and endpoint. For these effects, the difference in exposure duration between in vitro and in vivo is relatively small, as a single short-term in vitro exposure can reasonably represent an acute in vivo exposure scenario. By selecting a relevant dose metric for the exposure and toxic effect, the dose over the short exposure duration can be captured, allowing acute effects to be studied in vitro (Groothuis et al., 2015; Li et al., 2021). However, when looking at (sub-)chronic toxicities that extend beyond 24 h, effects become more difficult to capture in vitro, and the discrepancy between in vitro and in vivo exposure duration can theoretically span multiple orders of magnitude.

Efforts are being made to increase the duration of in vitro toxicity assays (Macko et al., 2021). A longer assay could be more informative about longer or repeated exposure to a chemical in vivo. For example, in a 28-day in vitro assay, the EC₅₀ for viability of HepaRG cells after exposure to valproic acid or bosentan decreased respectively 33-fold and 21-fold when compared with day one toxicity (Klein et al., 2016). This illustrates that toxicity can develop over time in vitro and may not be apparent in short-term assays. In this example, the EC₅₀ values were corrected by nominal concentration over time. However, when interpreting repeated-dosing results, it is important to measure not only the nominal concentration, but also the effective concentration, such as intercellular concentration, over time.

When using repeated dosing, lipophilic compounds may accumulate intracellularly while the nominal concentration seems stable. This accumulation can occur faster in vitro than in vivo, as the chemical is directly added to the cells without the distribution and elimination processes present in vivo (Bloch et al., 2022; Kramer et al., 2015). As a result, observed effects may appear to be caused by prolonged exposure, but are actually caused by intracellular accumulation in the assay. Selecting an appropriate dose metric, such as the intracellular AUC, can partly account for this. Therefore, a longer assay does not automatically capture chronic toxicity. Chronic toxicities should therefore be understood through phenotypes and mechanisms and not be linked to exposure duration alone (Donato et al., 2022). The effective concentrations should be measured over time to make sure that the toxic effect can be attributed to the relevant chronic exposure.

3.6.2. Limitations and suggestions

The considerations described in the previous section also have implications for the kinetics in vitro. In long-term and repeated-dose assays, the in vitro kinetics (step 2) become time-dependent. Distribution, degradation, evaporation and cellular accumulation all change over the course of the assay, and repeated dosing or medium refreshment changes the concentration-time profile (Kramer et al., 2015). Characterization of these kinetics is thus important, and may be done with in silico models. However, steady-state in silico models that estimate in vitro distributions may misrepresent exposure in long-term assays. Instead, for such long-term assays, dynamic in vitro kinetic models are better suited, as they are built specifically for predicting in vitro distributions over time (Bloch et al., 2022; Nunes et al., 2023). Selecting the correct in silico model or measuring a relevant concentration over time, is an important consideration in the extrapolation of acute to chronic exposures.

A way to link (sub)acute in vitro findings to chronic exposures in vivo is to use kinetic modeling (Zwartsen et al., 2025; Nunes et al., 2023). A kinetic model can simulate the relevant concentration after repeated or chronic exposure, allowing the concentration-time profile to be compared between the in vitro experiment and the in vivo situation using an appropriate dose metric. However, this approach assumes that short-term exposure in vitro and long-term exposure in vivo lead to the same effects and same magnitude of the effect. This assumption cannot always be made.

In traditional risk assessment using animal studies, effects from shorter-term studies are extrapolated to long-term risk using time-scaling approaches, ranging from assessment factors (AFs) to linear time scaling or scaling relative to expected lifespan (ECETOC, 2003). Importantly, subacute in vivo studies are generally not used for deriving chronic HBGVs, a (sub)chronic study would be conducted instead. The same limitation thus applies to both in vitro and in vivo approaches: short-term data alone are generally insufficient for predicting chronic effects (Malkiewicz et al., 2009).

Instead, an approach based on a mechanistic understanding of chronic toxicity seems more suitable, as in vitro experiments can currently not be maintained for months or years. When a mechanistic understanding of chronic toxicity is available, it can be used to design assays that capture the relevant toxicity mechanisms. Only when chronic effects are represented in an in vitro assay, conclusions can be made about the long-term effects of chemicals in vitro. When this understanding is still lacking, AFs can be applied as a conservative solution. However, an appropriate AF for in vitro acute to chronic toxicity is still unknown. With growing mechanistic understanding, these AFs could potentially be refined and reduced.

4. Discussion

This review identified six extrapolation steps required to perform QIVIVE for in vitro-based risk assessment of chemicals and examined the current state of knowledge, limitations and suggestions for each step. Of the included publications in this review, 46 publication focused on performing a QIVIVE-based risk assessment of chemical substances. Interestingly, while the reverse dosimetry (step 4) has been described in 40 of the 46 publications (87%), the other steps have been described less frequently. For example, the extrapolation from nominal to target vitro concentration (step 2) has been reported in 12 of the 46 publications (26%), and the extrapolation from in vitro target concentration to in vivo target compartment concentration (step 3) was described in 3 publications (7%). Similar numbers were found for the extrapolation of in vitro effects to in vivo effects (step 5) (15 publications; 33%), the extrapolation from acute to (sub-) chronic exposures (step 6) (4 publications; 9%) and for the extrapolation based on choice of in vitro assay (step 1) (3 publications; 7%). Although the dose extrapolation steps are increasingly standardized in QIVIVE frameworks, the extrapolation of effect, cell model choice and exposure duration are mentioned less in guidance from leading authorities. This lack of alignment is further illustrated by differing views on the execution of certain extrapolation steps. In particular, dose metric selection and compartmental pairing still lack consensus, and the execution of effect and exposure time extrapolation is also unclear. However, qAOPs can be used to mechanistically link in vitro effects to in vivo adverse outcomes, and to quantify the extrapolation of effects. The development of (q)AOPs is still ongoing and will hopefully lead to more standardized execution of effect extrapolation.

So far, qAOPs have been proposed as a potential solution for effect extrapolation. However, they may also serve as an overarching framework connecting and informing several QIVIVE steps (Perkins et al., 2019; Paini et al., 2022). Because qAOPs specify which MIEs and KEs are mechanistically relevant to an adverse outcome, they can provide biological reasoning for the choice of an in vitro cell model (step 1), ensuring in vitro systems that capture relevant KEs or MIEs are chosen (Jarzina et al., 2022). Similarly, the mechanistic basis of qAOPs can inform dose metric and target compartment selection (step 3). The tissue or cell type in which a KE occurs points to the relevant target compartment, while the temporal aspect of the quantified KERs can inform which dose metric is most suited. Lastly, as discussed before, qAOPs can quantitatively inform effect extrapolation (step 5). In this way, qAOPs are not only a tool for effect extrapolation, but a potential scaffolding for aligning multiple extrapolation steps. In combination with correct in vitro concentration measurements and reverse dosimetry using PBK models, qAOPs can form a mechanistically based framework for risk assessment of chemicals (Perkins et al., 2019; Paini et al., 2022). It is, however, important to note that the field of qAOPs is quickly developing and not yet sufficiently mature for regulatory applications. Fully quantified AOPs remain scarce, and deriving KERs is still costly. Nevertheless, as the number of qAOPs increase, their integration across multiple QIVIVE steps is a promising approach towards a more mechanistically based NGRA.

The uncertainty introduced in the different extrapolation steps within QIVIVE is an often-discussed topic in the examined literature. While this is not an extrapolation step, and was thus not discussed in the previous sections, it is an important aspect that should be considered when executing a QIVIVE analysis.

Traditional chemical risk assessment generally relies on default AFs, to address differences between the experimental setup and population and the target (human) population, including interindividual variability, and the uncertainties we have in the magnitude of these differences (Martin et al., 2013). However, in QIVIVE-based risk assessment, no standardized method exists for managing differences and uncertainties. Creating a standardized method is hindered by the large variety of in vitro and in silico methods, and the lack of validation of them. Additionally, in vitro methods vary in how well they capture human biology. This variation in biological relevance is currently difficult to quantify. Nevertheless, creating a standardized approach for uncertainty management is important for QIVIVE, and different approaches are currently being developed.

One approach mirrors the traditional approach by applying AFs, such as a factor of 20 for in vitro to in vivo differences and 5 for inter-individual human variability (Combes, 2005). However, without solid scientific evidence, this suggestion has been criticized for obscuring rather than clarifying uncertainty (Martin et al., 2013; Bosgra et al., 2012). Other advancements focus on chemical-specific adjustment factors (CSAFs), which use toxicokinetic data to refine traditional AFs and better account for known variability (Wetmore et al., 2014; Dourson et al., 2022; Kasteel and Westerink, 2021; Lu et al., 2025; Shi et al., 2022). These CSAFs replace parts of the standard AF values when the variability or uncertainty is quantified, leading to mechanistic-based handling of uncertainties. This improves on standard AFs that are used in traditional risk assessment, by making the uncertainties and variations more interpretable.

Furthermore, a way to refine these CSAFs is by quantifying inter-individual differences. This can be done by population-based PBK models that incorporate distributions of human physiological parameters and enzyme abundances, including differences between ethnic groups (Wetmore et al., 2014; Shi et al., 2022; Darwich et al., 2010), and use Monte-Carlo simulations to enable more targeted risk assessments (McNally and Loizou, 2015; Bosgra et al., 2012). In a similar way, uncertainties in parameters can be added to the model and studied as well. Conservative estimates can be made using this method, but exact guidance for the application of this method is still lacking.

Despite these advances, significant challenges remain in quantifying uncertainties across all extrapolation steps in QIVIVE analyses, particularly regarding effect and exposure time extrapolations. While PBK models and in vitro kinetic research have improved the quantification of population variability and dose extrapolation uncertainties, gaps persist in linking in vitro effects to in vivo outcomes and in extrapolating findings to chronic exposure scenarios. Methods such as propagating uncertainties in dose extrapolation through Monte-Carlo offer alternatives to default AFs (Zwartsen et al., 2025; McNally and Loizou, 2015; Martin et al., 2013; Wambaugh et al., 2019), but when uncertainties cannot be quantified, AFs and adequate reporting on the approach taken remain necessary. Therefore, continued research should focus on characterizing these uncertainties. For example, by developing qAOPs and refining models for chronic exposure, as a higher mechanistic knowledge can lead to lower uncertainties. Ultimately, comprehensive quantification of uncertainties is essential for robust and reliable in vitro-based chemical risk assessments.

Currently, the application of in vitro models in regulatory frameworks is low and mostly limited to hazard identification and mechanistic understanding. To stimulate broader application towards risk assessment, a clear QIVIVE approach is needed, including standardized methods to deal with uncertainty. This work aims to contribute to this approach. Building confidence in QIVIVE will require case studies that apply all extrapolation steps to chemicals where in vivo dose-response data are available for comparison with QIVIVE-derived dose-response data. This in vivo comparison is not a prerequisite for QIVIVE to be valid, but such comparisons provide empirical evidence and build regulatory trust. As a next step, once sufficient case studies have been done, patterns across chemicals and endpoints should be examined, to identify where QIVIVE is most reliable. This could form a starting point for defining conditions under which QIVIVE can be executed without an in vivo comparison, most likely for well-defined chemical groups or well-established AOPs. At that point, QIVIVE could turn into a regulatory tool. How long this process will take is unclear, but a harmonized approach can help in the development.

Additionally, in vitro toxicity studies should be designed and reported in a way that enables their use in QIVIVE analyses. This requires considering extrapolation steps during experimental design. For example, measuring in vitro kinetics alongside the toxicity data so that the extrapolation from nominal concentration to target concentration can be executed. Reporting detailed information about the experimental setup and a detailed characterization of the cell system, including dosing, medium composition and considerations for choosing an in vitro method, supports the execution of multiple extrapolation steps. This is not common practice yet, but will be crucial for advancing NGRA.

Regulatory institutes and academic research can both benefit by advancing the harmonization of QIVIVE approaches. A harmonized QIVIVE framework could lead to more reliable and translatable health-based guidance values. Contributing to this harmonization includes developing guidance for dose metric and compartment pairing, as well as guidance on the implementation of chemical-specific assessment factors for the different steps. Transparent and standardized documentation of in vitro and in silico models, including their applicability domain, is important for reusability and harmonization across NGRA applications.

Several limitations of the literature search strategy should be acknowledged. The use of machine learning-assisted screening with ASReview has an inherent limitation of not all articles being shown to the user. This makes review of the literature more efficient, but potentially less complete. Additionally, the exclusion criteria used can lead to relevant articles being unnoticed. Finally, the field of in vitro-based risk assessment is rapidly developing, meaning that recent frameworks or relevant articles may not have been identified when published after the search date.

A last consideration is the chemical applicability domain of the identified extrapolation steps. Currently, articles on (nano-)particles and nanomaterials were excluded, for their differences in toxicodynamics and –kinetics from chemicals. (Nano-)particles and nanomaterials require distinct approaches because of their dissolution and particle-specific uptake (Dekkers et al., 2016). Extending and validating QIVIVE for (nano-)particles is an important direction for future research.

We believe that future research to advance and practically implement QIVIVE should be directed towards the following gaps:

• •Development of qAOPs and response-response relationships to support effect extrapolation.
• •Methods for extrapolating exposure duration from in vitro to chronic in vivo scenarios
• •Quantification of uncertainties introduced across all extrapolation steps
• •Development of knowledge on the kinetics of a chemical in vitro, either by measuring or modeling.
• •Standardized documentation of in vitro and in silico models.

Together, these developments will ultimately lead to a more reliable and advanced in vitro-based risk assessment.

5. Conclusion

This review identifies the critical extrapolation steps of QIVIVE for the NGRA of chemicals using human cell-based in vitro models. We identified six extrapolation steps. While a shared foundation exists across recent QIVIVE frameworks, alignment on how to execute certain steps remains incomplete. Significant progress has been made in standardizing dose extrapolation, but considerable uncertainties remain in effect extrapolations, in the quantification of differences introduced by cell model choice, and in the extrapolation of exposure duration. The development of mechanistically-based frameworks such as (q)AOPs will be crucial for improving QIVIVE as part of NGRA. Addressing these challenges is essential for achieving regulatory acceptance of in vitro-based risk assessment, eventually enabling the derivation of HBGVs from in vitro and in silico data, without the need for animal testing.

CRediT authorship contribution statement

M.J.M. Benschop: Conceptualization, Methodology, Investigation, Resources, Writing – original draft, Writing – review & editing, Visualization. J. Minnema: Conceptualization, Methodology, Resources, Writing – review & editing. E.E.J. Kasteel: Writing – review & editing. A.S. Kienhuis: Conceptualization, Methodology, Writing – review & editing, Supervision. A.G. Oomen: Conceptualization, Methodology, Writing – review & editing, Supervision. J. Vriend: Writing – review & editing, Supervision. R.H.S. Westerink: Conceptualization, Methodology, Writing – review & editing, Supervision. A. Zwartsen: Writing – review & editing. B.G.H. Bokkers: Conceptualization, Methodology, Resources, Writing – review & editing, Supervision.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Appendix

Table A1.

Databases searched on the 21st of January 2026, number of results and search terms used.

Database	Number of results	Search string
Embase	1288	(‘in vitro in vivo extrapolation’/exp. OR ‘quantitative in vitro to in vivo extrapolation’/exp. OR ‘qivive’:ti,ab,kw OR ‘ivive’:ti,ab,kw OR ((‘in vitro in vivo’ NEAR/3 ‘extrapolat*’):ti,ab,kw) OR ((‘in vitro to in vivo’ NEAR/3 ‘extrapolat*’):ti,ab,kw) OR ((‘reverse’ NEAR/3 ‘dosimetr*’):ti,ab,kw) OR ((‘in vitro’ NEAR/3 ‘based’ NEAR/3 ‘risk assessment’):ti,ab,kw) OR (((‘uncertainty’ NEAR/3 ‘factor’):ti,ab,kw) AND ‘nam’:ti,ab,kw) OR (((‘assessment’ NEAR/3 ‘factor’):ti,ab,kw) AND ‘nam’:ti,ab,kw) OR ((‘in vitro’ NEAR/3 ‘based’ NEAR/3 ‘hazard assessment’):ti,ab,kw) OR ((‘in vitro’ NEAR/3 ‘based’ NEAR/3 ‘safety assessment’):ti,ab,kw)) AND ([article]/lim OR [article in press]/lim OR [review]/lim) AND [english]/lim
Scopus	1525	TITLE-ABS-KEY (“qivive”) OR TITLE-ABS-KEY (“ivive”) OR TITLE-ABS-KEY (“in-vitro-in-vivo” W/3 “extrapolat*”) OR TITLE-ABS-KEY (“in-vitro to in-vivo” W/3 “extrapolat*”) OR TITLE-ABS-KEY (“reverse” W/3 “dosimetr*”) OR TITLE-ABS-KEY(“in vitro” W/2 based W/2 “risk assessment”) OR (TITLE-ABS-KEY(“uncertainty” W/2 “factor”) AND TITLE-ABS-KEY(“NAM”)) OR (TITLE-ABS-KEY(“assessment” W/2 “factor”) AND TITLE-ABS-KEY(“NAM”)) OR TITLE-ABS-KEY(“in vitro” W/2 based W/2 “hazard assessment”) OR TITLE-ABS-KEY(“in vitro” W/2 based W/2 “safety assessment”) AND LANGUAGE(‘English’)
PubMed	1182	“qivive”[Title/Abstract] OR “ivive”[Title/Abstract] OR “in vitro in vivo extrapolat*”[Title/Abstract] OR “reverse dosimetry”[Title/Abstract:∼3] OR “in vitro based risk assessment”[Title/Abstract:∼10] OR “uncertainty factor NAM”[Title/Abstract:∼10] OR “Assessment factor NAM”[Title/Abstract:∼10] OR “in vitro based hazard assessment”[Title/Abstract:∼10] OR “in vitro based safety assessment”[Title/Abstract:∼10]

Appendix B. Supplementary data

Full overview of the 211 relevant articles.