NGRA in Cosmetic Safety: Moving From the Frameworks and Case Studies to Wide Adoption

Abstract

The cosmetic safety evaluation has undergone a paradigm shift driven by animal testing bans now implemented across countries representing one-third of the global population and nearly half of worldwide cosmetics sales. The legislative mandates have accelerated the development and adoption of Next Generation Risk Assessment (NGRA), an innovative framework that proposes making of context-specific safety decisions using exposure-driven, mechanism-based approaches based on non-animal test methods. Significant progress has been achieved in replacing animal tests for specific endpoints including skin sensitization, irritation, and phototoxicity through validated New Approach Methodologies (NAMs). However, major gaps remain for one-for-one replacement of systemic toxicity, reproductive and developmental toxicity, and carcinogenicity animal tests. Recent advances demonstrate that NGRA can successfully address many of these complex endpoints through integrated approaches combining physiologically-based kinetic (PBK) modeling, high-throughput transcriptomics and in vitro phenotyping to derive Bioactivity-Exposure Ratios (BERs) for risk characterization. Numerous case studies across diverse chemical types, including coumarin, benzophenone-4, caffeine, and benzyl salicylate, have demonstrated NGRA's utility for making protective safety decisions without animal data. These examples highlight critical considerations that need to be addressed in NGRA, including metabolite assessment, dose metric selection, and quantitative uncertainty characterization. While like-for-like replacement of all animal tests remains unachievable, NGRA provides a scientifically robust pathway for protective assessments where an absence of systemic effects is the primary goal, particularly for cosmetic ingredients. The challenge now lies in harmonizing methodologies, building regulatory confidence in these approaches, and evolving institutional structures to enable routine NGRA implementation across global regulatory regimes.

1. Animal Ban in Cosmetics – Brief History

The efforts to end animal testing for cosmetics and personal care products combine topics of ethics, scientific innovation, and consumer-driven calls for change. Depending on local regulatory definitions, this category includes products for oral care, hygiene and cleansing, sun protection, skin care and deodorants as well as color cosmetics. Because of the widespread use of these products, safety testing for ingredients used in these products historically involved procedures on animals. Starting in the 1980s, animal welfare organizations launched public awareness campaigns that caused a major shift in consumer sentiment against the idea of animals being used for this purpose. Concurrently, scientific advancements offered new possibilities for ensuring safety of some consumer products without the need to use animals. In vitro methods using human cells, computer modeling, and tests on donated human tissue promise to be not only more ethical but perhaps even more relevant to human biology. In response to the citizen-driven petitions and promise of scientific advances, the European Parliament legislated the change in the practice of testing cosmetic ingredients’ safety. In 2004, it banned testing finished cosmetic products on animals within its borders. The scope expanded in 2009 to include ingredient testing. The final, and most crucial, step came in 2013 with a full marketing ban that prohibited the sale of cosmetic products or ingredients that had been tested on animals anywhere in the world, effectively forcing global brands to adopt animal-free methods if they wanted access to the European Union (EU) market.

This landmark EU legislation created a domino effect, inspiring nations around the world (Figure 1) to follow suit and cementing a new global standard for “ethical” cosmetics. Nearly one-third of the global population now lives in countries where animal testing of cosmetic products and ingredients is banned or restricted; India alone accounts for 17.7% of world population. The EU and UK combined represent another 6.3% of global population. Brazil and Mexico add considerable Latin American representation at 4.3% combined. Together, there are 18 countries that have legislatively banned or restricted the use of animals in testing safety of cosmetics, and/or marketing/import of animal-tested cosmetics – they span five continents demonstrating truly global momentum for this animal welfare standard. The United States does not have a nation-wide ban on animal testing for cosmetics; however, the landscape is evolving through a combination of state-level (California, Nevada, Illinois, Virginia, Maryland, Maine, Hawaii, New Jersey, New York, Oregon and Louisiana) legislative acts and recent federal-level efforts to encourage the greater use of alternative methods [1]. China's recent policy changes (2021–2023) are also important as efforts are made to eliminate mandatory animal testing for most imported ordinary cosmetics even though animal testing is still required for “special use” cosmetics (hair dyes, sunscreens, etc.). Importantly, these countries account for nearly half of worldwide cosmetics sales (approximately $235 billion), because of higher per-capita spending in these markets on beauty products [2]. For example, the EU remains the largest single market at 18.5% of global sales. Brazil is the third-largest national cosmetics market globally, representing 6.8%. South Korea alone is estimated to represent 2.7% of the global market for cosmetics, well above other countries in terms of per-capita spending.

Figure 1.

World map showing countries with bans and restrictions on animal use in safety testing of cosmetic products and ingredients. See color legend for explanation. Information is current as of May 2026.

2. Regulatory Requirements for Cosmetic Ingredients

The safety assessment process for cosmetic ingredients has undergone a dramatic transformation over the past two decades, shifting from largely animal-dependent testing protocols to non-animal assays, many of which are now recognised for defined regulatory purposes through OECD Test Guidelines, validated test methods, or accepted Integrated Approaches to Testing and Assessment (IATA) (Table 1). Historically, regulatory frameworks required extensive animal testing batteries covering multiple toxicological endpoints, from acute toxicity and skin irritation to reproductive effects and carcinogenicity [3]. The EU ban on animal testing for cosmetics provided a considerable impetus for speeding up the development and validation of non-animal New Approach Methodologies (NAMs) [4]. Today, many traditional animal tests have been successfully replaced by validated non-animal tests, while remaining gap areas are being addressed through weight-of-evidence approaches combining read-across, in silico predictions, and exposure assessments. Table 1 provides a comprehensive comparison of traditional animal-based testing methods and their modern non-animal alternatives, organized by safety endpoint, including OECD Test Guideline (TG) numbers and current validation status. This overview illustrates both the progress achieved in developing NAMs and the remaining scientific challenges that continue to drive innovation in cosmetic safety science. For example, several required tests have been fully or partially replaced – skin irritation/corrosion, eye irritation (partial), skin sensitization (IATA), phototoxicity, dermal absorption.

Table 1.

Cosmetic Ingredient Safety Evaluation: Animal vs. Non-Animal Tests.

Safety Endpoint	Traditional Animal Test Method(s)	OECD TG #	Non-Animal Test Methods	OECD TG #	Validation Status1
ACUTE TOXICITY
Oral	Oral LD50 in rats/mice	420, 423, 425	In silico QSAR models; Cytotoxicity assays (3T3); Read-across		Partial replacement; QSAR accepted for classification
Dermal	Dermal LD50 in rats/rabbits	402	In silico QSAR models; In vitro skin absorption + cytotoxicity; Read-across	TG 428 (absorption)	Partial replacement
Inhalation	Inhalation LC50 in rats	403, 433, 436	In silico models; In vitro lung cell cultures; Read-across		Limited alternatives; waived for non-volatiles
SKIN EFFECTS
Skin Irritation/ Corrosion	Draize rabbit skin test	404	Reconstructed human epidermis: EpiSkin™, EpiDerm™, SkinEthic™	439 (corrosion), 431 (irritation)	Validated method
Skin Corrosion	Rabbit skin application	404	Transcutaneous electrical resistance (TER); Rat skin TER; Corrositex®	430, 435, 435	Validated method
Skin Sensitization	Guinea Pig Maximization Test; Buehler test; Local Lymph Node Assay	406, 406, 429, 442A, 442B	Direct Peptide Reactivity Assay; KeratinoSens™; h-CLAT; U-SENS™; LuSens; IL-8 Luc; SENS-IS; GARDskin; Integrated Approaches (IATA)	442C, 442D, 442E, 442F, 497, 442H	Validated as integrated approaches; Defined Approaches accepted (2021+)
Phototoxicity	Mouse or guinea pig + UV exposure		3T3 Neutral Red Uptake Phototoxicity Test; Reconstructed human skin phototoxicity	432	Validated method
Photoallergenicity	Mouse or guinea pig + UV		No validated alternative yet; SENS-IS under development; Combined phototoxicity + sensitization data		Gap area; case-by-case assessment
EYE EFFECTS
Eye Irritation/ Corrosion	Draize rabbit eye test	405	Bovine Corneal Opacity & Permeability; Isolated Chicken Eye; RhCE: EpiOcular™, SkinEthic™; Fluorescein Leakage; Short Time Exposure	437, 438, 492, 460, 491	Validated for specific categories; tiered testing strategies required
Serious Eye Damage	Rabbit eye test	405	BCOP; RhCE models; Integration with pH, structure	437, 492	Validated for severe irritants
REPEATED DOSE TOXICITY
Subacute (28-day)	Rats	407, 410
Subchronic (90-day)	Rats	408, 409, 411	In vitro organ models (liver, kidney); Transcriptomics; Read-across; QSAR; Threshold of Toxicological Concern (TTC) Multi-organ Tissue Chips; Systems toxicology; PBPK modeling		Gap area; case-by-case assessment and integrated approaches emerging
Chronic (>1 year)	Rats	452, 453	Rarely required for cosmetics		Not typically needed
GENETIC TOXICITY
Bacterial Mutagenicity	None (in vitro)	471			Already non-animal
In Vitro Chromosomal Aberration	None (uses cells)	473			Already non-animal
In Vitro Micronucleus	None (uses cells)	487			Already non-animal
In Vivo Genotoxicity (follow-up)	Rodent micronucleus; Rodent chromosome aberration; Comet assay	474, 475, 489	In silico models; Enhanced in vitro batteries (including 3D models); Read-across		Mode of action based approach for in vitro positives
Gene Mutation (mammalian)	Mouse lymphoma assay (in vitro)	476, 490			Already non-animal
ABSORPTION & TOXICOKINETICS
Dermal Absorption	Rat/human skin (ex vivo)	427, 428	In vitro skin absorption: human skin, RhS; Franz diffusion cells; QSAR	428	Validated method
Systemic ADME	Rats (in vivo)	417	In vitro metabolism (hepatocytes, microsomes); PBPK modeling; In silico ADME prediction		Scientifically valid methods
Dermal Penetration	Animal or human skin	428	Human skin in vitro; RhS models; Computational models	428	Validated method
REPRODUCTIVE & DEVELOPMENTAL TOXICITY 2
Developmental Toxicity	Rats (oral, prenatal); Rabbits (oral, prenatal)	414	Embryonic Stem Cell Test; Reprotracker®; devTOX quickPredict™; Transcriptomics; Receptor-based assays; steroidogenesis assays; Read-across, Threshold of Toxicological Concern (TTC)	455–458 Endocrine activity assays (see below)	Validated methods for embryotoxicity and some endocrine activity; therefore, there is only partial coverage across the full reproductive cycle
Reproductive Toxicity (Fertility)	Rats, multi-generation	415, 416, 421, 422, 443
Reproductive Screening	Rats (combined repeat/repro)	421, 422
ENDOCRINE DISRUPTION
Estrogenic Activity	Uterotrophic assay (rats)	440	Stably Transfected Transcriptional Activation Assay: ER binding, ER trans-activation; Human cell reporter assays	455, 457	Validated method
Androgenic Activity	Hershberger assay (rats)	441	AR binding and transactivation assays	458	Validated method
Thyroid Disruption	Rat thyroid assays	407, 408	In vitro thyroid peroxidase inhibition; Receptor binding assays		Partial coverage
Steroidogenesis	Rat studies	Various	H295R Steroidogenesis Assay	456	Validated method
CARCINOGENICITY
Long-term Carcinogenicity	Rats/mice (2-year bioassay)	451, 453	Genotoxicity battery; Cell transformation assays; QSAR models; Read-across; Weight of evidence		Rarely required for cosmetics; predicted from other data
SPECIAL ENDPOINTS
Nanomaterial Characterization	Various animal studies		Physicochemical characterization; In vitro cellular uptake/toxicity; Barrier model studies (intestinal, skin)		Case-by-case assessment; guidance documents available
Immunotoxicity	Rodent immune function tests	407 (extend.)	Human cell-based immune assays; PBMC; Cytokine release assays		Limited alternatives
Neurotoxicity	Rats (behavioral, histology)	418, 419, 424, 426	Human neuronal cell cultures; Electrophysiology; Organoids (brain)		Emerging area

¹Validation status indicates whether methods are formally recognized for defined uses, for example through OECD Test Guidelines, validated test methods, or regulatory integrated approaches. “Partial”, “case-by-case”, or “gap area” indicates that non-animal methods can contribute to assessment but do not yet provide complete endpoint coverage or universal regulatory acceptance.

²Note that because the micromass test requires primary rat embryo cells and the whole embryo culture and Zebrafish assays involve dosing during the embryonic stage, their status as “non-animal” tests is contentious and as such they are not listed in this table.

Also, many of the previously and currently required tests were already not using live animals – in vitro genotoxicity tests (Ames, chromosomal aberration, micronucleus using cell cultures, etc.). Endpoints that are now partially replaced by NAMs include acute toxicity and endocrine disruption screening. The major gaps that remain are acceptable methods for repeated dose systemic toxicity, reproductive and developmental toxicity, and carcinogenicity.

3. Next Generation Risk Assessment (NGRA): A Leading Framework for Cosmetic Safety and Beyond

NGRA is a framework to evaluation of cosmetic ingredient safety testing that represents a paradigm shift – it advocates moving beyond traditional hazard-based testing toward exposure-driven, mechanism-based risk assessment using exclusively non-animal tests [5]. Proposed originally by US EPA in response to the large-scale testing of industrial and environmental chemicals [6], NGRA implementation advanced most prominently in cosmetic safety where alternative methods must be developed for the industry to continue innovation. Embraced as a collaborative approach by leading cosmetics companies, regulatory agencies, and academic researchers, NGRA is focused on the use of exposure-led and tiered approaches that integrate multiple streams of evidence. These include exposure assessment, in vitro assays using human-relevant cell systems and tissues, in silico computational modeling, high-throughput screening, and read-across from existing data, and together can provide a protective basis for safety decision making [7]. Deriving a point of departure (PoD) from a set of assays with broad biological coverage means that although this approach is grounded in the Adverse Outcome Pathway (AOP) concept [8], which describes the sequence of key events from initial molecular interactions through cellular and tissue responses to potential adverse effects, a defined AOP may not always be needed to make a safety decision. In addition, data on human cell-based models and cells from exposed humans has been used for the past decade in cancer hazard evaluations [9]. By combining sophisticated exposure modeling (to determine realistic human exposure levels) with in vitro toxicity testing at relevant concentrations, NGRA enables safety decisions based on margin of exposure calculations without generating new animal data [10]. Major companies and trade associations have advocated for NGRA framework as an approach to assure consumer safety under the ban on animal testing of cosmetics [11] and to meet consumer demand for such products. NGRA represents not just a replacement for animal testing, but a more scientifically advanced, human-relevant approach to cosmetic safety assessment that many experts believe will eventually become the global standard across all industries, extending far beyond cosmetics to pharmaceuticals, chemicals, and food additives [6,12,13].

4. Moving NGRA From a Framework Stage to Wider Acceptance

Development and regulatory adoption of the non-animal tests for skin and eye irritation and skin sensitization are a clear success in terms of implementing NGRA for those specific endpoints [14–16]. Additional recent methods (Table 2) and case study (Table 3) publications have showed how regulatory requirements for systemic and organ-specific toxicity endpoints may be addressed by NGRA and NAMs.

Table 2.

NGRA Methods for Systemic Toxicity Endpoints. Methods listed represent NAM-based evidence streams used within integrated, exposure-led NGRA workflows, rather than stand-alone replacements for traditional animal studies.

Systemic Endpoints Addressed	NGRA Methods & Tools	Ref.
General systemic toxicity across multiple organs	• Cell stress panel (multiple cell types) • Pharmacological profiling • High-throughput transcriptomics (multiple cell types) • Dose-response • PBK models for internal exposure • Bioactivity-Exposure Ratio (BER) calculations • Bayesian uncertainty modeling	[18]
	• Cell stress panel (multiple cell types) • Pharmacological profiling • High-throughput transcriptomics (multiple cell types) • Dose-response • PBK models for internal exposure • Bioactivity-Exposure Ratio (BER) calculations • Bayesian uncertainty modeling	[17]
	• IVIVE and dose-metrics consideration for in vitro studies • Biologically effective dose assessment • Considerations of chemical volatility, stability, hydrophobicity, binding, solubility analysis	[19]
	• Sensitivity analysis for BER calculations • IVIVE parameter and PBK tool analysis	[20]
	• Data integration and decision analysis approach • Structured documentation of data, decisions & uncertainties • Alternative Safety Profiling Algorithm (ASPA)-assist software	[21]
Liver toxicity	• Primary human hepatocytes and HepaRG cells • Toxicogenomics data • Gene co-expression network analysis • Dose-response (transcriptomics PODs)	[22]
Hormonal disruption (estrogenic)	• Three in vitro estrogenic assays • Dose-response (BMCL05 and EC50 values) • Dietary Comparator Ratio approach	[23]
Hormonal disruption (androgenic)	• AR-CALUX assay • Exposure Activity Ratios (EARs) • Dietary Comparator Ratio approach	[24]
Lung/respiratory toxicity	• Human respiratory tract models • Computational in vitro NAMs • Adverse Outcome Pathway framework	[25]
Developmental and Reproductive Toxicity (DART)	• Cell stress panel (multiple cell types) • Pharmacological profiling • High-throughput transcriptomics (multiple cell types) • Dose-response • PBK models for internal exposure • ReproTracker and devTOX quickPredict™ pluripotent stem cell assays • Bioactivity-Exposure Ratio (BER) calculations	[26]
Population variability in toxicodynamics	• Lymphoblast Cell Lines from 100+ humans of broad ancestry • Dose-response in cytotoxic effects of tested compounds • Bayesian dose-response modeling to derive PODs in each tested individual and then a chemical-specific toxicodynamic variability factor 5% (TDVF05)	[27]

Table 3.

NGRA Case Studies for systemic toxicity endpoints that demonstrate fit-for-purpose applications to specific chemicals, endpoints, and exposure scenarios. Confidence is based on the overall weight of evidence, including exposure relevance, biological coverage, metabolite coverage, dose-metric selection, and uncertainty characterisation.

Chemical/Ingredient	Endpoint(s) Addressed	Main Findings	Ref.
Coumarin, Caffeine, Sulforaphane	• Exposure assessment (dermal exposure)	• Developed framework for dermally applied products using PBK models constructed exclusively from NAM-derived ADME properties • Conservative Cmax estimates refined using sensitivity analysis • Demonstrated that framework increases confidence in PBK predictions	[28]
Caffeine	• General systemic toxicity	• Demonstrated 10-step read-across framework using NAMs and PBK modeling • Read-across based on structural similarity and mode of action • Derived Margin of Internal Exposure of 27-fold – sufficiently conservative for human health	[38]
Genistein and Daidzein	• General systemic toxicity	• Developed and validated human PBPK models using bottom-up approach • Used in vitro ADME data with dermal modules including first-pass metabolism • Demonstrated PBPK model utility for NGRA of cosmetic ingredients	[29]
Benzophenone-4 (BP-4)	• General systemic toxicity	• Performed NGRA using PBK models and in vitro data without human data • Concluded that consumer exposure to BP-4 is low; biological responses only at much higher exposure	[30,31]
Octocrylene	• General systemic toxicity	• Applied NGRA framework to explore risks for vulnerable groups • Combined existing risk assessment data with bioactivity data and identified potential risk drivers • Highlighted need for more data to assess vulnerable populations	[32]
Avobenzone	• General systemic toxicity	• Performed NGRA using PBK models and in vitro data without human data • Compared in vitro PoD with plasma concentrations to demonstrate exposure is low risk	[33]
Hippophae rhamnoides (sea buckthorn) fruit extract	• General systemic toxicity	• Estimated safe usage levels using three non-animal approaches: TTC, PoD, and history of safe consumption • Concluded that TTC was most conservative	[39]
Coumarin	• General systemic toxicity • Genotoxicity	• Applied NGRA without using animal or human data • Integrated PBK for internal exposure estimation • Used battery of NAMs (phenotypic & transcriptomics) • Demonstrated safety with margins >100	[34]
Phenoxyethanol	• General systemic toxicity	• Applied NGRA without using animal or human data • Integrated PBK for internal exposure estimation • Used battery of NAMs (published data, pharmacological profiling, transcriptomics & cell stress) • Low margin of internal exposure for major stable metabolite a key uncertainty in the assessment	[36]
Benzyl Salicylate	• General systemic toxicity • DART	• Performed ab initio NGRA using skin absorption, hepatocyte metabolism, PBPK modeling • Used battery of NAMs (phenotypic & transcriptomics) • Identified that metabolite was toxicological driver • Derived margins of internal exposure (x16-338) • Both traditional approach & NGRA showed safe use	[37]
Sodium Isethionate	• General systemic toxicity • DART	• Integrated occupational exposure modeling and PBK modeling with chemical-specific ADME data • Used battery of NAMs to calculate PODs • Predicted plasma Cmax and derived BERs	[35]
Troglitazone	• Liver toxicity	• Used mitochondria toxicity data from multiple liver cell models • Developed human population PBPK model • Explored IVIVE uncertainties (cell type/dose metric)	[40]
HC Yellow No. 13	• Liver toxicity	• NGRA focused on liver steatosis • In vitro data using human liver cells • Predicted internal liver exposure • Calculated BER to demonstrate safety	[41]
Propylparaben	• Reproductive toxicity	• Developed a read-across framework using NAMs • Focused on filling reproductive toxicity data gap • Integrated structural similarity, in silico predictions, toxicogenomics, in vitro data, and PBK modeling • Focused on chemical and biological similarity	[42]
Flutamide (and metabolite OH-flutamide)	• Hormonal disruption (androgenic)	• Focused on parent vs active metabolite bioactivity • Used in vitro AR-CALUX assay and PBK modeling • Showed that metabolite testing is important	[43]
Doxorubicin	• Cardiotoxicity	• Evaluated impact of dose metrics (Cmax vs AUC) • Integrated PBK modeling with human cell-based assays • Developed population-based PBK model validated against pharmacokinetic data • Concluded AUC is better indicator than Cmax	[44]

Table 2 summarizes recent advances in NGRA that have produced a suite of innovative methods and tools for evaluating systemic toxicity without relying on traditional animal testing. To address general systemic toxicity, integrated approaches that combine cell stress panels, pharmacology screening panels, high-throughput transcriptomics across multiple cell types, and dose-response modeling have been developed [17,18]. These studies proposed a three-step workflow that (i) establishes points of departure (PoDs) from NAMs assay-derived phenotypic and transcriptomic data using approaches that incorporate Bayesian uncertainty modeling, (ii) derives estimates of internal exposure levels by using physiologically-based kinetic (PBK) models, and then (iii) calculates the Bioactivity-Exposure Ratio (BER) that is used to inform risk characterization. Supporting methodologies to this workflow have addressed critical technical considerations such as in vitro to in vivo extrapolation (IVIVE), assessment of biologically effective doses, and chemical-specific properties including volatility, stability, and solubility [19]. Sensitivity analyses for BER calculations and PBK parameters were also tested to establish robustness of the overall approach [20]. In addition, structured data integration frameworks and decision-support software like the Alternative Safety Profiling Algorithm (ASPA) have been developed to help organize complex datasets and document uncertainties [21].

For organ-specific and other specialized endpoints, targeted NGRA methods have emerged. For liver toxicity assessment, primary human hepatocytes and HepaRG cells combined with toxicogenomics and gene co-expression network analysis have been proposed [22]. Hormonal disruption can be evaluated through validated in vitro assays for both estrogenic [23] and androgenic [24] activity, and utilizing Dietary Comparator Ratios and Exposure Activity Ratios to contextualize potential risks. Respiratory toxicity assessment can leverage human respiratory tract models integrated with computational NAMs within an AOP framework [25]. For developmental and reproductive toxicity (DART), specialized assays using pluripotent stem cells (e.g., ReproTracker^® and devTOX quickPredict^™) can be used in conjunction with the general transcriptomic, pharmacology profiling and cell stress approaches [26]. Finally, to address inter-individual susceptibility, lymphoblast cell lines from over 100 individuals of diverse ancestry enable characterization of population variability in toxicodynamics through Bayesian modeling that derives chemical-specific toxicodynamic variability factors [27]. Collectively, these methods represent a paradigm shift toward mechanistically-grounded, human-relevant safety assessment that quantitatively addresses uncertainty and population variability.

In addition to the methodological work and frameworks, many case studies demonstrate the practical application of NGRA approaches across diverse chemicals and endpoints, showcasing both the versatility and evolving maturity of these methods (Table 3). Several studies have focused on establishing foundational NGRA frameworks for cosmetic ingredients applied dermally. For example, in cases of coumarin, caffeine, and sulforaphane PBK models were constructed exclusively from NAM-derived ADME properties, demonstrating that conservative internal exposure estimates could be refined through sensitivity analysis [28]. This dermal exposure framework was further validated through subsequent studies on genistein and daidzein, which incorporated first-pass metabolism in bottom-up human PBPK models [29]. Additional examples include sunscreen ingredients including benzophenone-4 [30,31], octocrylene [32], and avobenzone [33]. These studies concluded that consumer exposures to tested compunds from cosmetics applied on a skin were sufficiently low and also pose minimal biological risk.

A large number of case studies listed in Table 3 focused on systemic toxicity endpoints. Coumarin is a proof-of-concept case for fully animal-free NGRA addressing both general systemic toxicity and genotoxicity, achieving safety margins exceeding 100-fold through integration of PBK modeling with phenotypic and transcriptomic assays [34,35]. Phenoxyethanol case study revealed the importance of considering metabolites, because a major stable metabolite introduced key uncertainties despite otherwise acceptable margins [36]. Similarly, benzyl salicylate case study identified a metabolite as the potential driver of hazardous effects, yielding margins of internal exposure ranging from 16 to 338-fold – information that aligned with conclusions from traditional animal-based testing approaches [37]. Caffeine case study included a 10-step read-across analysis based on similarity in structure and mode of action, yielding a 27-fold margin of internal exposure that was deemed adequately protective [38]. The case study of sodium isethionate integrated occupational exposure modeling with PBK approaches and NAM-derived points of departure to calculate BER and it was concluded that adequate safety exists for workers exposed to sodium isethionate under current levels of factory-specific risk management. [35]. Sea buckthorn fruit extract safety evaluation compared three non-animal approaches: threshold of toxicological concern (TTC), NAMs-based PoDs, and history of safe consumption, finding TTC to be most conservative approach to risk management [39].

In addition, several case studies listed in Table 3 focused on organ-specific and mechanism-based effects. For liver effects, troglitazone case study used mitochondrial toxicity data from multiple liver NAMs assays and a population PBPK model to determine how IVIVE uncertainties relate to cell type and dose metrics [40]. A case study of HC Yellow No. 13 focused specifically on liver steatosis in studies that used human liver cells and derived BERs for risk evaluations [41]. Reproductive toxicity case studies are exemplified by propylparaben read-across, which integrated structural similarity, in silico predictions, toxicogenomics, in vitro data, and PBK modeling to fill data gaps [42]. Hormonal disruption case study of flutamide emphasized the critical importance of evaluating active metabolites, because the OH-flutamide metabolite showed substantially different androgenic activity than the parent compound [43]. Finally, doxorubicin cardiotoxicity case study incorporated population-based PBK modeling validated against clinical pharmacokinetic data, demonstrating that area-under-the-curve (AUC) serves as a better dose metric than maximum concentration (C_max) for this endpoint [44]. Collectively, these case studies illustrate that NGRA approaches can successfully characterize safety across diverse chemical structures and toxicological endpoints. They also highlight recurring themes including the importance of metabolite consideration, dose metric selection, and quantitative uncertainty characterization in building confidence for regulatory decision-making.

5. Expert Opinion

Despite major progress in development and validation of in vitro and in silico NAMs for many hazard traits (Table 1), it is our opinion that the commonly stated objective that NAMs directly replace all animal tests is neither achievable nor desirable. Direct replacements for systemic, reproductive and developmental toxicology, and carcinogenicity studies are not available, and there is increasing consensus that one-for-one replacement, or even integrated approaches that aim to replicate the results of particular guideline animal tests, is neither necessary to address many safety questions on consumer-use chemicals nor practical. There is also increasing recognition that, while OECD guideline animal studies have historically supported protective safety decision-making, they may not be the tools of first choice in a chemical safety testing paradigm designed from first principles using today’s technologies.

Instead, NGRA is a regulatory science workflow that begins with characterization of human exposures and then proceeds to tiered, hypothesis-driven hazard evaluation that makes use of mature tools such as TTC, read-across, PBK modelling, and in vitro bioactivity assays to inform risk characterization that can be used to make safety decisions that are reliably protective of systemic health effects. The utility of NGRA has been illustrated with numerous case studies shown in Table 3, with the clear provison that a protective safety assessment conclusion can only be determined with high confidence where the intended use case for the chemical being assessed does not require eliciting systemic effects. Thus, for consumer or occupational safety assessments for cosmetics (as well as certain other use cases), NAM-based decision-making is achievable today. To gain acceptance of protective, exposure- and bioactivity-based assessments, our focus needs to be on building confidence in the lower-tier approaches (i.e., use of weight of evidence, use of protective in vitro PoDs alongside PBK modelling) through harmonization, including but not limited to consensus on how to consistently and reproducibly utilize in vitro phenotypic and transcriptomics assays that adequately cover biological targets of concern and from which one can derive bioactivity-based PoDs for comparison to internal exposure predictions.

Where safety decisions cannot be made using these lower-tier approaches, either due to the conservatism of a bioactivity-based assessment or because systemic hazard identification is required, more work is needed to better distinguish between potentially innocuous bioactivity measured in vitro and an in vivo adverse health effect. Nonetheless, it would be a mistake to posit that no useful safety decisions can be made until we have every pathway quantitatively mapped and validated. Ultimately, the context of the safety decision (problem formulation) guides the data needs of the safety assessment, and as Table 2 shows, tools exist today that can be applied to address a great many decision contexts, particularly when the goal is protection against any adverse effects, rather than identifying specific adverse outcomes.

The context of the decision that needs to be made should guide the construct needed to ensure any NAMs used are fit-for-purpose to answer safety questions, rather than replicate animal tests. If our strategy for adoption of science-informed decision-making does not involve replacement of animal tests, then our validation strategy should also move away from only benchmarking with animal tests. Existing animal data remain useful as historical comparators, particularly for assessing whether NAM-based approaches generate protective toxicity values in the absence of human data. However, animal studies are not perfect reference standards because of species differences, strain variability, and potential limitations in human relevance. NAMs can provide human-relevant and mechanistic information, but may also have limitations in biological coverage, metabolism and chronicity. Flexible validation frameworks that consider the circumstances in which the data are generated and the protection goals for the specific decision context need to be considered [45].

Beyond validation strategies, the practical adoption of approaches such as NGRA requires systemic change [46]. Change takes time, but it is clearly occurring, albeit in a stepwise and case-by-case manner across different areas:

• Evolving Regulatory Landscape: Regulatory agencies are seeing more and more submissions that include new data streams and incorporate NGRA workflows. Adaptations to standard information requirements under EU REACH (Registration, Evaluation, Authorisation, and Restriction of Chemicals)have opened a window of opportunity. Even though the early success of NAMs in regulatory submissions has been modest [47,48], better understanding of what gaps can be addressed by NAMs and further experience with these new data will inevitably result in greater success. Even in the absence of clear regulatory guidance, for the first time, dossiers for cosmetic ingredients that incorporate NAM data are being submitted by industry to regulatory authorities.

• Continued Technological & Scientific Advances: The pace of NAM development shows no sign of slowing. For instance, many organizations are now looking to artificial intelligence-enabled workflows to help synthesize, structure, and interpret data in a transparent and consistent manner. However, the developers of new approaches need to better understand the contexts of use for their technologies [49], a process that is become more achievable because of the greater visibility of discussions of both NGRA and NAMs at recent major scientific meetings such as EuroTox, Society of Toxicology and World Congress on Alternatives and Animal Use in the Life Sciences.

• Cross-Discipinary Education and Workforce: Interdisciplinary and regulatory-science focused training is becoming more popular in both masters and doctoral programs. Some companies that make chemicals for a variety of uses are beginning to change their approach to staffing by hiring computational modellers, bioinformaticians, cell biologists, and chemists alongside risk assessors and exposure scientists to form the multidisciplinary teams needed to implement NGRA.

Cumulatively, these developments demonstrate that the momentum that has been building since the publication of the “Toxicity Testing in the 21st Century – A Vision and a Strategy” report [50] is being translated into real-world applications, a trend to which the ban on the use of animals for cosmetic safety evaluation has clearly contributed. The future value of NGRA lies not in like-for-like replacement of animal tests, but in enabling more human-relevant, exposure-led, mechanistically informed, and transparent safety decisions. The question going forward is not whether NGRA for cosmetics is scientifically possible, but how our regulatory and institutional structures will evolve in parallel with these scientific advances to enable its routine use.

Highlights.

• Animal testing bans now cover one-third of global population and half of cosmetics sales

• Validated alternatives exist for skin sensitization, irritation, and phototoxicity endpoints

• Few validated alternative methods exist for systemic toxicity assessments

• NGRA integrates PBK modeling, transcriptomics, and NAMs for protective safety decisions

• NGRA has been shown to be generally protective across many chemicals without animal data