New Approach Methodologies (NAMs) for Preclinical and Translational Evaluation of mRNA–Lipid Nanoparticle (LNP) Therapeutics

Abstract

Messenger RNA–lipid nanoparticle (mRNA-LNP) therapeutics have emerged as a versatile drug modality, enabling in vivo protein expression for vaccines, cancer immunotherapy, and the treatment of genetic and metabolic diseases. Although mRNA–LNP platforms achieved rapid clinical success during the COVID-19 pandemic, most candidates fail to progress beyond preclinical development, largely due to the limited capacity of conventional animal models to predict human-relevant efficacy, safety, inflammation, biodistribution, and population heterogeneity. New approach methodologies (NAMs), encompassing advanced in vitro and ex vivo human systems and in silico computational models, offer a promising strategy to address these translational gaps while reducing reliance on animal testing. In this review, we evaluate commonly used animal models for mRNA–LNP development and highlight key areas where animal findings have shown poor concordance with human clinical outcomes. We then provide a comprehensive overview of emerging NAM technologies, including in vitro and ex vivo-based platforms such as two-dimensional and three-dimensional cell culture systems and microphysiological platforms; as well as in silico tools, including physiologically based pharmacokinetic (PBPK), quantitative systems pharmacology (QSP), and artificial intelligence (AI) models for formulation design and delivery optimization. Collectively, this review highlights how systematic adoption of NAMs can improve human predictivity, accelerate development timelines, and support more efficient, ethical, and translational robust mRNA–LNP drug development.

Graphical Abstract

1. Introduction

Lipid nanoparticle (LNP)–encapsulated messenger RNA (mRNA) has emerged as a versatile therapeutic modality for in vivo protein expression. [1] A typical synthetic mRNA consists of a 5’ cap, untranslated regions, an open reading frame, and a poly(A) tail, with extensive optimization strategies employed to enhance stability and modulate innate immune activation. [2] These efforts include chemical nucleoside modifications, advanced purification methods, and rational sequence design. [1] LNPs, composed of ionizable lipids, phospholipids, cholesterol, and polyethylene glycol (PEG)-lipids, protect mRNA from degradation and enable intracellular delivery, while also influencing biodistribution, immune recognition, and transfection efficiency. Together, mRNA design and LNP formulation define the safety, efficacy, and translational performance of this new modality. [3–5]

Key advantages of mRNA–LNP platforms include fast modular design, rapid manufacturing, and scalable production. The COVID-19 pandemic provided a real-world demonstration of these capabilities, with first-generation SARS-CoV-2 vaccines progressing from sequence identification to phase I clinical testing within months. [6] Nowadays mRNA–LNP technologies have rapidly transitioned from experimental tools to clinically viable drug platforms with growing regulatory and commercial momentum. Beyond COVID-19, mRNA–LNP therapeutics are being actively explored for a wide range of infectious diseases and oncology applications, with multiple candidates showing promising preclinical and early clinical outcomes. [2, 7] In cancer immunotherapy, development has increasingly shifted from dendritic cell–based ex vivo approaches toward LNP-mediated in vivo mRNA delivery, offering improved scalability and manufacturing simplicity. [8] In parallel, targeted LNP strategies have enabled in vivo cell engineering, including selective delivery of chimeric antigen receptor (CAR) mRNA to T cell subsets, highlighting the potential of mRNA–LNPs to expand access to cellular immunotherapies. [9]

Many mRNA-LNP therapeutics are currently under development because of their unique advantages. A review identified 84 ongoing mRNA therapeutic trials, including 46 lipid-carrier-based mRNA therapeutic trials as of August 2023. [10] However, the broader translation of mRNA-LNP therapeutics remains challenging and has shown relatively low success rates. For example, a 2025 ClinicalTrials.gov-based analysis reported 238 global mRNA clinical programs, but only 34% had advanced beyond Phase I.[11] This low rate of successful clinical translation shows the substantial challenges associated with advancing mRNA–LNP technologies from preclinical animal models to human applications.

Two representative failure cases further illustrate this issue.

• MRT5005, developed by Translate Bio, is a biosynthetic, codon-optimized mRNA encoding the cystic fibrosis transmembrane conductance regulator (CFTR), delivered via LNP–formulated aerosol. In vitro studies and in vivo evaluations in non–cystic fibrosis rodent models and nonhuman primates (NHP) demonstrated CFTR protein expression after MRT5005, suggesting potential therapeutic benefit for cystic fibrosis [12]. However, in the Phase I/II RESTORE-CF study (NCT03375047), MRT5005 failed to demonstrate clinical efficacy, as no improvement was observed in forced expiratory volume in 1 second (FEV₁), the primary efficacy endpoint, in patients with two severe class I and/or II CFTR mutations. [13]

• mRNA-2416, developed by Moderna, is an LNP-encapsulated mRNA encoding human OX40 ligand (OX40L) for cancer immunotherapy. Preclinical studies showed durable tumor regression with OX40 agonism, particularly when combined with programmed cell death protein 1 (PD-1) blockade and demonstrated antitumor activity using mRNA-based OX40L expression strategies. [14, 15] Despite these promising preclinical results, the first-in-human Phase I/II clinical trial of mRNA-2416 (NCT03323398) did not meet its predefined primary efficacy endpoint of tumor objective response rate in an exploratory cohort of patients with advanced solid tumors or lymphoma. [16]

Together, these examples highlight the persistent difficulty of translating robust preclinical efficacy of nanomedicine therapeutics into meaningful clinical outcomes. Although these compounds exhibited promising activity in animal models, barriers between preclinical systems and human biology limited their clinical translation. [17]

There is increasing consensus that conventional animal models have limited ability to recapitulate human disease mechanisms and to predict human safety and efficacy of interventions. More than 90% of drug candidates that demonstrate acceptable safety and efficacy in animals ultimately fail in clinical development, most commonly due to insufficient efficacy or unanticipated human toxicities. [18] These failures reflect fundamental biological and mechanistic differences between animal models and human pathophysiology, as animal systems cannot adequately capture the heterogeneity and complexity of human disease. [19] Beyond scientific limitations, reliance on animal testing also raises ethical, economic, and logistical concerns that can slow therapeutic innovation and development. [20]

In response, New Approach Methodologies (NAMs) have emerged as a regulatory-relevant framework to improve human predictivity while reducing animal use. NAMs broadly refer to innovative testing strategies, models, and analytical tools that can improve the relevance, efficiency, or human predictivity of nonclinical evaluation, and support replacement, reduction, or refinement (3Rs) of conventional animal testing. In the context of drug development, NAMs encompass a broad range of in vitro, in silico, ex vivo, and other innovative approaches that are compliant with the principles of 3Rs and are applied in a fit-for-purpose manner to support regulatory decision-making. [21] Regulatory agencies have increasingly endorsed the use of NAMs. The FDA Modernization Act 2.0 formally permits scientifically justified alternatives to animal testing, while both the FDA and European Medicines Agency have advanced guidance and pilot initiatives supporting risk-based, human-relevant approaches to preclinical evaluation. [18, 22–24] Together, these developments provide a clear policy foundation for the expanded adoption of NAMs in medicines development. According to these regulatory guidance, NAMs are defined not by the novelty of individual technologies, but by their intentional integration into safety and efficacy assessment to enable more direct evaluation of human-relevant biological mechanisms. [25, 26]

NAMs offer a promising solution to enable more human-relevant assessment of mRNA-LNP therapeutics, while substantially reducing development time and cost and aligning with the principles of 3Rs animal use. Given the growing volume of preclinical nanomedicine data available for model training and validation, NAMs are well positioned to play an increasingly central role in future mRNA-LNP development. In this review, we examine the key translational gaps between animal studies and human outcomes for mRNA-LNPs, discuss the opportunities and challenges associated with applying NAMs to this emerging modality, and outline future research directions needed to advance their effective implementation.

2. Current Landscape of Animal Models for mRNA-LNP Therapeutics and Their Limitations

Current preclinical drug development relies heavily on animal models, which can be broadly categorized into healthy animals, models with spontaneous (naturally occurring) disease, genetically modified animals, and induced disease models generated through chemical or surgical interventions. [27] In this review, we focus specifically on animal models commonly used for the preclinical development of mRNA-LNP therapeutics and discuss their respective advantages and limitations. Given the extensive use of mRNA-LNP platforms in cancer research, Table 1 summarizes representative oncology animal models used in mRNA-LNP studies. Animal models for other disease indications are discussed separately in the text where relevant.

Table 1.

Comparisons of animal models used in mRNA-LNP therapeutics development in oncology.

Model Type	Functional Immune System	Human Tumor Representation	Key Advantages	Key limitations
Syngeneic Mouse Tumor	Yes (murine)	No (murine tumor)	Enables evaluation of immune activation, tumor-immune interaction, and in combination with checkpoint blockade	Fully murine tumor microenvironment; species-specific antigens and immune responses limit human translatability
Human cancer cell line–derived xenograft (CDX)	No (immunodeficient host)	Yes (human cell line)	Easy to establish, reproducible, human tumor cell lines	Lacks functional immune system; limited tumor heterogeneity; poor assessment of immunotherapy mechanisms
Patient-derived xenograft (PDX)	No (immunodeficient host)	Yes (patient tumor tissue)	Preserves patient tumor architecture and heterogeneity	Absence of human immune components; stromal replacement by murine microenvironment over time
Humanized mouse model	Partial (human immune reconstitution)	Yes (human tumor)	Allows study of human immune–tumor interactions; relevant for immuno-oncology	Incomplete immune reconstitution; graft-versus-host disease (GVHD); high variability and cost
Genetically Engineered Mouse Models	Yes (murine)	Endogenous murine tumor	Intact immune system and tumor co-evolution; physiological tumor initiation	Non-human antigens and immune context; species-specific nanoparticle biodistribution
Non-Human Primate (NHP)	Yes (primate)	Typical none (non-tumor studies)	Closest immune physiology to humans; useful for safety and biodistribution	High cost, ethical constraints, limited tumor modeling, small group sizes

2.1. Syngeneic Mouse Tumor Models

In the development of anti-tumor mRNA–LNP therapeutics, syngeneic mouse tumor models are widely used to evaluate antitumor efficacy and immune activation. These models involve implantation of mouse-derived tumor cell lines into genetically matched, immunocompetent mice, enabling assessment of vaccine-induced immune responses in the context of an intact immune system. [28] Syngeneic models have successfully predicted the clinical activity of immune checkpoint inhibitors such as anti–PD-1/PD-L1 and anti–CTLA-4 antibodies but have failed to accurately forecast the performance of other immunostimulatory agents, including agonistic anti-CD137 antibodies and STING agonists. This divergence likely reflects fundamental differences in immune contexture between syngeneic tumors and human cancers: checkpoint inhibitors act by releasing inhibitory constraints on partially functional T cells, whereas agonistic strategies require the presence of immunologically competent effector cells—conditions that are often met in syngeneic models but are uncommon in advanced human tumors characterized by chronic immune exhaustion and adaptive resistance. [28]

Syngeneic mouse models have been extensively applied to evaluate mRNA-LNP cancer vaccines, particularly in combination with immune checkpoint blockade, across multiple tumor types. [29–31] These models allow robust assessment of immune activation, including antigen-specific CD8⁺ T-cell responses, immune cell infiltration, and cytokine production. [32, 33] However, their translational relevance is limited by the short duration of tumor development and the absence of prolonged tumor–immune co-evolution. Implanted tumors often elicit strong de novo immune responses, whereas human cancers typically arise over years and exist in immunosuppressive microenvironments with dysfunctional or exhausted tumor-reactive T cells. As a result, syngeneic models tend to overestimate the magnitude and durability of vaccine-induced antitumor immunity observed in clinical settings. [28]

2.2. Human Cancer Cell Line–Derived Xenograft (CDX) Models

Human cancer CDX models involve implantation of established human cancer cell lines into immunodeficient mice. These models are highly reproducible, cost-effective, and enable rapid tumor growth and straightforward measurement, making them well suited for early proof-of-concept studies. [34, 35] CDX models have been used to evaluate mRNA-LNP therapeutics, including combination strategies with bispecific antibodies where binding to human antigens is critical for efficacy. [36, 37] They are also valuable for assessing mRNA-LNP biodistribution and functional mRNA expression within tumors. By incorporating human tumor cells, CDX models enable evaluation of tumor-intrinsic responses to mRNA-LNP delivery. [38] However, the lack of a functional immune system limits their ability to assess immune-mediated mechanisms, making them unsuitable for evaluating cancer vaccine immunogenicity.

2.3. Patient-Derived Xenograft (PDX) Models (Non-Humanized)

PDX models are generated by implanting patient tumor tissues into immunocompromised mice and are widely regarded as more faithful representations of human cancers than conventional xenografts, as they preserve native tumor architecture and intratumoral heterogeneity. [39] Consequently, PDX murine models have been widely employed to evaluate the therapeutic efficacy and delivery performance of mRNA-LNP formulations. For example, Gu et al. utilized a colorectal cancer PDX model to demonstrate that LNP-encapsulated tumor necrosis factor–related apoptosis-inducing ligand (TRAIL) mRNA effectively induced tumor cell apoptosis in vivo. [40] Similarly, Wang et al. validated that their mRNA-LNP system restored HER2 and TP53 expression and resensitized previously unresponsive tumors to HER2-targeted antibody therapy in a HER2-low patient-derived PDX model. [41] To date, PDX applications in mRNA–LNP development have primarily focused on assessing tumor delivery efficiency and functional transgene expression in solid tumors, leveraging their superior preservation of patient-specific tumor phenotypes. [42, 43] However, the lack of a functional immune system, species-specific differences in lipid metabolism, and limited scalability restrict the utility of PDX models for evaluating immunogenicity and for early-stage formulation optimization of mRNA-LNP therapeutics.

2.4. Humanized Mouse Models

Humanized mouse models are immunodeficient mice engrafted with functional human immune components, enabling in vivo evaluation of human-specific immune responses, immunogenicity, and immunotherapy efficacy that cannot be captured by conventional CDX or non-humanized PDX models. [44] These models are therefore particularly valuable for mRNA-LNP vaccine development, allowing assessment of human innate and adaptive immune responses. [45] In a recent study, Fick et al. established a humanized mouse model by engrafting human peripheral blood mononuclear cells (PBMCs) into mice to assess the efficacy of LNP-encapsulated human IFNα2 mRNA in suppressing human tumor metastasis. [46] Wang et al. developed a humanized B cell knock-in mouse model to investigate whether mRNA-LNP–encoded germline-targeting HIV Env immunogens could simultaneously prime, boost, and drive affinity maturation of competing broadly neutralizing antibody precursor lineages, thereby providing preclinical proof of concept for germline-targeting vaccination strategies. [47] However, humanized mouse models are limited by high cost, experimental variability, and incomplete or immature immune reconstitution. In addition, mouse physiology continues to dominate LNP biodistribution, restricting the direct translatability of mRNA–LNP delivery profiles to humans. [48, 49]

2.5. Genetically Engineered Mouse Models (GEMMs)

GEMMs harbor targeted genetic modifications that drive spontaneous disease development in an immunocompetent host, enabling physiologically relevant in vivo studies of gene function, disease progression, and therapeutic response. [50] GEMMs have been used to evaluate mRNA-LNP therapeutics in genetic disease settings, including inducible deletion of the ornithine transcarbamylase (OTC) gene to assess OTC mRNA-LNP therapy in a model of OTCD, as well as Ugt1 knockout mice for evaluating LNP-encapsulated UGT1A1 mRNA in Crigler–Najjar syndrome. [51, 52] Beyond disease modeling, GEMMs enable mechanistic interrogation of mRNA-LNP delivery and biodistribution using genetic reporters and pathway-specific knockouts. Sago et.al established a high-throughput in vivo system using genetically engineered reporter and SpCas9 knock-in mouse models to address whether distinct LNP formulations could overcome liver-dominant tropism and achieve efficient, functional mRNA and CRISPR component delivery to nonhepatic targets, particularly endothelial cells. [53] Paunovska et al. investigated the contribution of the low-density lipoprotein receptor (LDLR)–apolipoprotein E (ApoE)–mediated pathway to hepatocyte uptake of mRNA-LNP, using Ldlr⁻/⁻ and ApoE⁻/⁻ mouse models to delineate this delivery mechanism. [54] Overall, GEMMs provide a powerful platform for studying mRNA–LNP activity in immunocompetent settings. However, high cost, long development timelines, and reliance on mouse-specific genetics and physiology limit their scalability and predictive value for human mRNA expression and nanoparticle biodistribution. [55]

2.6. Nonhuman Primates (NHP) Models

NHP models are typically employed at late preclinical or investigational new drug (IND)-enabling stages, after proof-of-concept has been established in rodents, to support translational relevance and regulatory decision-making. For mRNA-LNP therapeutics, NHPs more closely recapitulate human immune architecture, lipid metabolism, liver physiology, and biodistribution than mouse models, enabling more predictive assessment of pharmacokinetics (PK), dose–exposure relationships, durability of mRNA expression, and innate immune responses. [56–58] As a result, NHPs are particularly valuable for evaluating systemic safety, immunogenicity, and clinical dosing strategies that cannot be reliably inferred from murine studies. However, their use is limited by high cost, long timelines, low throughput, and ethical and logistical constraints, restricting NHP studies to select candidates and precluding their application in early-stage formulation screening or mechanistic discovery. [59, 60]

2.7. Non-Oncology Disease Models

Given the broad therapeutic potential of mRNA–LNPs beyond oncology, disease-specific murine models have been widely used to evaluate candidate therapies. Because many current systemically administered mRNA-LNPs predominantly accumulate in the liver, liver diseases—particularly metabolic dysfunction–associated steatohepatitis (MASH)—have been a major focus. For example, delivery of hepatocyte growth factor (HGF), epidermal growth factor (EGF), or HNF4A mRNA via LNPs improved liver injury, fibrosis, and cirrhosis in multiple mouse models, including diet-induced and CCl₄-induced disease settings. [61, 62] These disease-specific models enable evaluation of mRNA–LNP efficacy, target engagement, and disease-modifying mechanisms within relevant pathological microenvironments that cannot be captured in healthy animals. However, many rely on simplified or singular induction methods and fail to reflect the multifactorial nature of human disease, resulting in species-specific biology, variable disease severity, and limited translational predictability for clinical outcomes.

2.8. Summary

Preclinical development of mRNA-LNP therapeutics commonly relies on multiple animal models to capture different aspects of delivery, expression, immunogenicity, efficacy, and safety. Rodent models remain valuable for early formulation screening, biodistribution assessment, dose optimization, and proof-of-concept efficacy studies, while NHP studies can provide additional translational insight because of their closer physiological, immune, and anatomical similarity to humans. Together, these models have contributed substantially to the development of mRNA–LNP vaccines and therapeutics by enabling integrated evaluation of systemic exposure, tissue distribution, innate and adaptive immune responses, and tolerability. However, no single animal model fully recapitulates human biology, and multi-model strategies can be costly, resource-intensive, variable, and ethically challenging. These considerations highlight the need to use animal models judiciously and to complement them with more predictive human-relevant approaches. In the following section, we examine key challenges in mRNA–LNP clinical development and major barriers to translational predictability.

3. Translational Gaps of Animal Models in mRNA–LNP Development

Here, we discussed the potential reasons behind this translational gap in mRNA-LNP context from inflammation, delivery efficiency and biodistribution, as well as population heterogeneity aspects. (Figure 1)

Figure 1. Key biological and translational gaps between humans and mice that may affect mRNA-LNP pharmacology and safety interpretation.

Differences in inflammatory responses, anti-PEG antibody interactions, liver sinusoidal structure, and population heterogeneity may affect mRNA-LNP biodistribution, clearance, efficacy, and safety. These factors highlight key translational gaps when extrapolating mouse data to human outcomes. PK/PD, pharmacokinetics/pharmacodynamics; PEG, polyethylene glycol.

3.1. Exaggerated Inflammatory Responses in Humans

In humans, mRNA-LNPs are frequently recognized as danger signals that trigger strong innate immune activation. [6, 63] Human innate immune systems interpret mRNA-LNP exposure as a combination of membrane perturbation and aberrant cytosolic RNA, leading to rapid activation of type I interferon (IFN)–dominated programs. [64, 65] Following endosomal uptake, mRNA-LNPs robustly activate Toll-like receptor (TLR)7 and TLR8 in human immune cells, resulting in pronounced cytokine and interferon responses. [66, 67]

In contrast, mouse TLR8 differs from human TLR8 in ligand recognition because of structural divergence in the ectodomain. Murine TLR8 does not respond to the same single-stranded RNA (ssRNA) ligands or TLR8-sleective synthetic agonists that activate human TLR8, leading to a systematic underestimation of innate immune activation and inflammatory toxicity in mice. [68, 69] Consequently, mice appear more resistant to mRNA-induced inflammation, tolerating high levels of IFN-α and producing elevated levels of IL-1 receptor antagonist, which attenuates IL-1β–mediated inflammatory responses. [67] Clinically, excessive type I IFN responses accelerate immune-mediated clearance of mRNA-LNPs, reduce intracellular mRNA translation, and ultimately diminish therapeutic protein expression. These inflammation-driven clearance mechanisms are therefore a critical contributor to the failure of murine models to predict human efficacy and tolerability. [6, 63]

Recent work by Hellgren et al. demonstrated that human innate immune responses to mRNA COVID-19 vaccines closely resemble those in NHPs, both phenotypically and mechanistically, although the magnitude of inflammation is generally lower in humans. [58] These findings highlight that NHPs better model human innate immune responses than rodents, and that the reliance on mouse models alone may underestimate reactogenicity and safety risks in preclinical mRNA vaccine development.

3.2. Species-Specific Differences in Delivery Efficiency and Biodistribution

Delivery efficiency represents a major translational gap for mRNA-LNP therapeutics, arising from both human-specific immune recognition and fundamental anatomical differences between species. A key contributor is PEG, which is incorporated into LNPs to enhance colloidal stability but is weakly immunogenic. Due to widespread environmental exposure from pharmaceuticals, cosmetics, and food packaging, a substantial proportion of humans harbor pre-existing anti-PEG antibodies; approximately 72% of individuals have detectable anti-PEG antibodies prior to mRNA vaccination. [70] Upon administration, PEGylated mRNA-LNPs may interact with circulating anti-PEG antibodies, triggering complement activation and immune-mediated clearance before effective cellular transfection occurs. These anti-PEG antibodies can be boosted by LNPs and may influence PEGylated nanoparticle–immune cell interactions and mRNA-LNP vaccine reactogenicity. [71] This process may contribute to variability in delivery efficiency, antigen expression, and inflammatory responses across individuals. In contrast, most laboratory animals lack pre-existing anti-PEG immunity, leading to systematic overestimation of mRNA-LNP stability and delivery efficiency in preclinical models.

Beyond immune-mediated clearance, long-underappreciated structural differences in vascular and lymphatic biology further limit cross-species translation. [72] In mice, liver sinusoidal endothelial cells exhibit large, high-density fenestrae that facilitate nanoparticle extravasation and preferential hepatocyte transfection. [72] In humans, however, sinusoidal fenestrae are smaller and less dense and are further reduced under common pathological conditions such as fatty liver disease or fibrosis due to capillarization. [73, 74] As a result, mRNA-LNPs in humans are more likely to remain intravascular and be taken up by immune cells rather than target parenchymal cells, shifting tissue and cellular distribution and amplifying innate immune sensing, which are the effects that are poorly captured in murine models. [75]

These species-dependent differences have been systematically characterized by Hatit et al., who evaluated 89 LNP formulations across mouse models bearing murinized, humanized, or primatized livers. [49] LNP delivery profiles in primatized livers showed substantially stronger correlation with human outcomes than those observed in murine livers, highlighting the limited predictive value of rodent models for human mRNA-LNP performance. Transcriptomic analyses further revealed species-specific differences in mRNA translation efficiency, endocytic pathways, and intracellular trafficking. [56] Additional contributors include differences in circulatory path length and systemic exposure time, which influence protein corona formation and nanoparticle aggregation prior to tissue uptake. [56] However, cross-species comparisons in siRNA studies are often confounded by differences in target gene sequences and siRNA payload design, making these effects more difficult to quantify directly

Together, species-specific heterogeneity poses major challenges for translating mRNA–LNP delivery and expression from animals to humans, as formulations that perform robustly in rodents often fail to maintain equivalent efficacy or safety in higher species. Importantly, the dominant mechanisms of hepatic LNP uptake-adsorption of endogenous ApoE followed by LDLR mediated endocytosis are generally considered conserved across mice, NHPs, and humans. [54] This suggests that translational failure arises primarily from upstream PK, anatomical, and immune-mediated differences rather than from fundamental disparities in receptor engagement.

3.3. Human Population Heterogeneity

Population heterogeneity represents a major gap in translational drug development and contributes to variability in both safety and efficacy observed in clinical trials and real-world clinical practice. Preclinical animal models frequently fail to recapitulate the pathophysiologic heterogeneity of human disease as well as the pharmacologic and pharmacogenomic diversity that shapes individual drug responses, thereby underestimating the breadth and magnitude of inter-individual variability in clinical outcomes, including rare but clinically significant adverse events and differential therapeutic responses. [19] For example, sex and age have been identified as significant modifiers of responses to mRNA COVID-19 vaccines, indicating that the influence of host-specific factors on mRNA-LNP induced immune responses that are difficult to anticipate from preclinical studies alone. [76, 77] Collectively, the limited representation of genetic, age-related, and sex-dependent diversity in standard preclinical models, together with human-specific distributions of anti-carrier antibodies and variability in complement sensitivity, contributes to translational gaps in both safety and efficacy prediction, as rare but high-impact immune-mediated outcomes may only emerge upon human exposure.

4. NAM Technologies for mRNA–LNP Development: Current Capabilities and Limitations

4.1. In Vitro Experimental Systems

A variety of in vitro experimental systems have been developed to study the delivery efficiency, cellular responses, and functional expression of mRNA-LNPs. The major categories of in vitro experimental systems used for mRNA-LNP evaluation and their strengths are summarized in Figure 2.

Figure 2. Human-relevant in vitro platforms for evaluating mRNA-LNP delivery and translational performance.

Model systems progress from two-dimensional monocultures to more physiologically relevant co-culture, three-dimensional, and microphysiological platforms. Greater biological complexity can improve assessment of tissue penetration, immune interactions, efficacy, toxicity, and organ-level responses, while introducing tradeoffs in throughput, reproducibility, and standardization.

4.1.1. Two-Dimensional (2D) Cell-Based Monoculture Systems

2D cell-based systems, in which immortalized cell lines, primary human cells, or iPSC-derived cells are cultured as monolayers are commonly used to evaluate drug behavior in vitro. In mRNA-LNP research, 2D cell-based monoculture systems are typically used for high-throughput screening of LNP formulations, assessment of cellular uptake and intracellular trafficking, quantification of functional mRNA expression, and evaluation of cytotoxicity and innate immune activation. [78] While highly scalable and cost-effective, these systems lack tissue architecture and physiological complexity, limiting their ability to predict in vivo biodistribution and therapeutic efficacy.

Immortalized cell lines originate from cells harboring chromosomal abnormalities or mutations, or can be intentionally produced using human telomerase reverse transcriptase (hTERT), and are widely used in research due to their ability to proliferate indefinitely and provide reproducible experimental results. [79, 80] Various immortalized cell lines are available and differ by phenotypes and origin for mRNA-LNP research. Patel et.al reported cell-based assay to evaluate potency of mRNA-LNP vaccines using 2D systems with HepG2, Vero, HeLa, Hep-2, A549, and Caco-2 cell lines. [81] All these cell lines are epithelial or epithelial-like adherent cells, with EC50 of mRNA-LNP functional expression between 10–100 ng, and adding ApoE can significantly increase their total potency. With reported potency curves, these cell lines can serve as a good reference of mRNA-LNP in vitro potency and cytotoxicity. Macrophage cell lines such as RAW 264.7 and THP and dendritic cell line DC2.4 have been used to evaluate mRNA-LNP trafficking or pro-inflammatory cytokines for vaccine development. [78] Despite their flexibility and ease of access, a major limitation of immortalized cell lines is that they do not fully represent cells encountered in vivo, which can reduce the translational relevance of experimental findings. [78] Moreover, these cells do not undergo normal cellular aging and are inherently selected for high proliferative capacity, a phenotype often associated with altered endocytic and intracellular trafficking mechanisms. [82]

Primary human cells are isolated directly from human tissues and relatively preserve the physiological characteristics and functions of their original in vivo state, but they exhibit a limited lifespan in culture. [83] They have been widely used to evaluate mRNA-LNP delivery in vitro to characterize LNP transfection efficiency and mRNA expression. [49, 84, 85] Compared with the HepG2 cell line, primary human hepatocytes more faithfully recapitulate native human liver physiology—such as metabolic enzyme activity, transporter expression, and lipid accumulation—making them more predictive for assessing mRNA–LNP safety, efficacy, and translational relevance. However, their broader application is constrained by high cost, donor-to-donor variability, limited availability, short-term viability in culture, and low scalability. In addition, the culture and propagation of these cells in 2D alters transcriptional profiles and can reduce physiologic relevance. [78, 86]

Building upon conventional 2D cell line models and primary human cells, induced pluripotent stem cell (iPSC)–based platforms have emerged as a versatile and human-relevant in vitro approach for preclinical research. iPSCs can be expanded indefinitely and differentiated into a wide range of disease- and tissue-specific human cell types, enabling reproducible modeling of human biology with greater physiological fidelity than immortalized cell lines. [87] The iPSC-derived 2D cultures offer key advantages as NAMs, including human genetic relevance, and scalability. [88] In the context of mRNA-LNP development, iPSC-derived cells are increasingly used to assess cellular uptake, functional mRNA expression, cytotoxicity, and innate immune responses in otherwise difficult-to-transfect cell types, such as cardiomyocytes, neurons, and endothelial cells. [89–91] The iPSC-derived hepatocytes were also used to evaluate mRNA-LNP uptake as a function of LNP size and time postdosing. [78] These features position iPSC-based systems as a powerful complementary platform for screening modified LNP formulations for cardiac targeting, and neurotoxicity prediction, while their simplified architecture still necessitates integration with more complex models to support translational assessment.

4.1.2. Co-Culture Systems

To provide a more physiologically relevant in vitro model than traditional 2D monocultures, co-culture systems have been introduced and are generally classified as direct or indirect (Transwell) models. In direct co-culture systems, different cell types are grown together in the same environment, allowing direct physical contact and cell–cell interactions. For example, Svensson et al. employed a co-culture of antigen-loaded liver non-parenchymal cells and OVA-specific CD8⁺/CD4⁺ T cells to investigate how adjuvant plus mRNA-LNP strategies modulate non-parenchymal cell immune phenotypes and their capacity for antigen presentation and T-cell activation, as measured by T-cell proliferation. [92] Zhu et al. established a direct co-culture system consisting of C2C12 cells (a mouse myoblast cell line) and bone marrow–derived dendritic cells to compare mRNA-LNP expression and transfection efficiency in immune versus non-immune cells. [93] Such direct co-culture systems hold great potential for mRNA-LNP cancer vaccine research, as they enable the simultaneous incorporation of tumor cells and immune cells to directly assess antigen presentation, immune activation, and tumor responses in a controlled in vitro setting. However, direct cell–cell contact complicates the disentanglement of cell-specific contributions and the analysis of LNP distribution among different cell types. Moreover, the requirement for a shared culture medium can introduce bias by differentially affecting the growth, viability, and phenotype of each cell type, thereby confounding interpretation of intercellular interactions and treatment responses. [94]

In contrast, indirect co-culture systems use a physical barrier to separate cell populations while allowing the exchange of soluble factors, facilitating clearer attribution of cell-specific responses and mechanistic investigation of mRNA-LNP transport and signaling. [95] Using this approach, Han et al. established a Transwell-based co-culture of endothelial and neuronal cells to screen mRNA-LNPs capable of crossing the blood–brain barrier and transfecting neuronal cells. [96] Likewise, Cao et al. applied a Transwell co-culture system with Müller cells in the upper chamber and retinal pigment epithelial ARPE-19 cells in the lower chamber to demonstrate that mRNA-LNPs can undergo transcellular transport (transcytosis) through Müller cells and subsequently be re-uptaken by ARPE-19 cells. [97] Despite their improved physiological relevance, co-culture systems remain limited by simplified cellular composition and artificial culture conditions that cannot fully recapitulate the complexity, spatial organization, and dynamic signaling of in vivo tissues. In addition, the mechanical properties of the indirect co-culture system substrate represent an important and often overlooked confounder: properties such as substrate stiffness are potent regulators of cell internalization and endocytosis. [98, 99] This is particularly relevant for Transwell-based co-cultures, as the polymer membranes are typically superphysiologically stiff and need to be coated with extracellular matrix components (e.g., collagen, fibronectin, Matrigel, or other basement-membrane mimetics), which may shift mechanotransduction and uptake behavior away from in vivo states. [100]

4.1.3. Advanced Three-Dimensional (3D) Models

Advanced 3D in vitro models more closely recapitulate native tissue architecture, cell–cell and cell–matrix interactions, and physiological gradients than conventional 2D cultures, resulting in more physiologically relevant cellular behaviors and improved predictive power for drug delivery, efficacy, toxicity, and translational relevance. [101, 102] Here, we discuss the application of spheroid- and organoid-based 3D in vitro models in the development of mRNA-LNP therapeutics.

Spheroids are 3D aggregates of cells grown in vitro that self-organize into spherical structures, mimicking key aspects of tissue spatial organization and partially restoring native tissue morphology and function. [103] Tumor spheroid–based 3D models have been widely used to better recapitulate the structural, cellular, and diffusion barriers present in vivo, thereby enabling more physiologically relevant evaluation of mRNA-LNP penetration, intratumoral distribution, cellular uptake, and transgene expression compared with conventional 2D cultures. For example, Kim et al. used L929 fibroblast spheroids to assess mRNA-LNP penetration and transgene expression within dense 3D cell aggregates [104]; Chen et al. generated Raji lymphoma spheroids to evaluate mRNA-LNP–engineered CAR-T cell infiltration and antitumor activity in a 3D tumor context [105]; Zhang et al. employed glioblastoma stem cell spheroids to investigate mRNA-LNP delivery efficiency, gene expression or editing outcomes, and associated phenotypic responses in a clinically relevant 3D brain tumor model [91]; Taylor et al. used 3D spheroids derived from ovarian cancer cell lines to demonstrate that spheroid models reveal culture-dependent differences in mRNA-LNP cargo expression that are not predicted by 2D monolayers, underscoring the advantage of spheroids for evaluating LNP delivery efficacy, penetration, and transgene expression in tumor-relevant settings. [106] Despite their improved physiological relevance over 2D cultures, spheroids lack key features of native tissues, such as vasculature, immune components, and dynamic fluid flow, which limits their ability to fully recapitulate in vivo tumor heterogeneity. Variability in spheroid size, cellular composition, and nutrient or oxygen gradients can reduce experimental reproducibility and complicate data interpretation. [107] Another limitation of spheroid models is their limited ability to recapitulate the tumor extracellular matrix: in many solid tumors, desmoplasia and extracellular matrix crosslinking strongly shape the transport environment and penetration barriers, features that are typically absent or simplified in spheroids. [108, 109]

Organoid is a 3D in vitro model derived from stem cells or primary tissues that differentiates and self-organizes to recapitulate key aspects of the architecture, cellular diversity, and function of the corresponding organ in vivo. [110] Compared with traditional 2D cell culture, organoid culture systems have the unique advantage of conserving parental gene expression and mutation characteristics, as well as long-term maintenance of the function and biological characteristics of the parental cells in vitro. [110] A key advantage of organoids is their multilineage cellular diversity: unlike spheroids, which typically contain a limited subset of cell types, organoids develop dynamic, heterogeneous populations that more closely recapitulate the cellular complexity of native tissues. [111, 112] Sela et.al established human iPSC-derived cortical brain organoids to show that acetylcholine-conjugated LNP can effectively penetrate through this model and mediate mRNA functional expression throughout the complex, 3D neural structures, indicating successful delivery in a 3D tissue-like context. [113] This provides direct evidence that organoids can be used as advanced in vitro models to assess the performance of mRNA-LNP systems in a setting that better mimics human tissue architecture than 2D cultures. However, a fundamental limitation of organoids is the lack of functional vascularization and perfusion, which constrains tissue architecture and efficient nutrient/oxygen delivery. Consequently, as organoids (including cortical brain organoids) grow beyond the oxygen diffusion limit, they often develop hypoxia and necrotic cores. [114–116] Emerging vascularized organoid are increasingly overcoming these limitations by integrating endothelial cell networks, lumenized vascular structures, and controlled microfluidic perfusion. For example, recent vascularized organoid-on-chip platforms have established functional perfusable endothelial networks around spheroids, mesenchymal and pancreatic islet spheroids, and blood vessel organoids, enhancing organoid growth, maturation, and function. [117, 118] These advances are expected to improve the physiological relevance of organoid-based NAMs and may provide more predictive platforms for evaluating mRNA–LNP transport, tissue penetration, local expression, and safety in human-relevant tissue environments. [119]

4.1.4. Microphysiological Systems and Organ-on-Chip

Microphysiological systems are a broad class of advanced in vitro models that include organoids, spheroids, and organoid-on-chip platforms. Among these, organ-on-chip platforms culture living cells within microfluidic devices to mimic the native fluidic environment of organs while reproducing key features of cellular architecture and tissue function. By providing precise control over liquid flow and biochemical gradients, these systems offer a highly tunable and dynamic environment for modeling organ function and therapeutic responses. [120] The organ-on-chip platforms offer key advantages for mRNA-LNP development by recapitulating human-relevant tissue architecture, fluid flow, and multicellular interactions, enabling more predictive assessment of LNP delivery, biodistribution, efficacy, and toxicity than static 2D or simple 3D models. Jeger-Madiot et.al applied a lymphoid organ-chip model based on a microfluidic chip seeded with human PBMC to assess the memory B-cell response to COVID mRNA vaccine. This organ-on-chip system can be used to study B lymphocytes proliferation and antibody production after prime and boost administration of mRNA-LNPs. [121] Recently, Zhai et.al developed a human lymphoid follicle chip that integrates a biomimetic intramuscular vaccination module to mimic muscle injection, lymphatic drainage, and adaptive immune activation. The team demonstrated the possibility of predicting the efficacy and immunogenicity of intramuscular mRNA-LNP vaccines in a fully human in vitro model. [122] Neiman et.al developed a phenotypic cardiac microphysiological system constructed from a human induced pluripotent stem cell cardiomyocytes Cre-reporter line to screen the mRNA-LNP formulations with optimal transfection efficiency in cardiomyocytes. [123] However, the broader application of organ-on-chip system is limited by technical complexity, high cost, low throughput, and challenges in standardization, which can hinder large-scale formulation screening and routine adoption in early-stage development. [120]

Although microphysiological systems have traditionally been limited by relatively low throughput, this constraint is increasingly being addressed through advances in device miniaturization, plate-based formats, integrated sensors, automation, and parallelized culture systems. For example, recent platforms have incorporated 96-device plate formats, programmable fluidic control, real-time sensing, and compatibility with high-throughput imaging and screening workflows. [124–126] These developments suggest that while throughput remains an important practical limitation for many microphysiological systems implementations, scalability is an active area of innovation and may improve the feasibility of microphysiological systems-based screening in mRNA-LNP development and translational research.

4.2. Ex Vivo Assays

Ex vivo assays refer to experimental systems in which viable cells, tissues, organs, or biological fluids are isolated from a living organism and evaluated outside the body under controlled laboratory conditions. Unlike the in vitro systems discussed above, ex vivo models retain more of the native biological context of the source tissue, including donor-specific characteristics, multicellular composition, tissue architecture, immune components, and endogenous extracellular matrix. In this section, we focus primarily on human ex vivo assays, as these systems are particularly valuable for capturing human-specific responses to mRNA–LNPs, including donor-dependent variability, innate immune activation, cytokine release, complement activation, and tissue-relevant safety signals.

4.2.1. Human Whole Blood Assays

Human whole blood assays are ex vivo platforms in which freshly collected anticoagulated human blood is exposed directly to mRNA–LNP formulations, preserving circulating immune cells, platelets, plasma proteins, complement factors, and soluble mediators. Compared with simplified in vitro cell culture systems, these assays are particularly useful for capturing early human blood–nanoparticle interactions, including cytokine and chemokine release, immune cell activation, complement activation, hemocompatibility, and donor-dependent variability. [127] Published mRNA–LNP studies have used human whole blood assays to evaluate formulation-dependent inflammatory responses, interferon pathway activation in disease-relevant donor samples, and complement-related safety signals. [128, 129] Despite their human relevance, these assays are limited by short-term viability, donor-to-donor variability, sensitivity to blood handling and anticoagulant selection.

4.2.2. Human PBMC

Human PBMC assays use isolated peripheral blood mononuclear cells, mainly lymphocytes, monocytes, NK cells, and dendritic cells, to evaluate immune responses to mRNA–LNPs. Compared with whole blood assays, PBMC assays provide a more standardized and experimentally controlled platform that is compatible with cryopreserved donor samples, batch testing, and focused assessment of monocyte- and lymphocyte-mediated cytokine release. [130] In mRNA–LNP development, PBMC-based studies have been used to assess cytokine responses to various LNP formulations, including IL-1β, IL-6, TNF-α, IFN-γ, and IL-8; and to identify monocyte-dependent innate immune activation or formulation/adjuvant-driven inflammatory differences. [131–133] However, PBMC assays may underestimate blood–nanoparticle interactions, complement activation, hemocompatibility, and plasma protein corona effects, and results can also be influenced by donor variability, cell isolation, and cryopreservation processes.

4.2.3. Human Plasma/Serum Study

Human plasma or serum studies are acellular ex vivo assays in which mRNA–LNP formulations are incubated with human plasma or serum to evaluate interactions with soluble blood components. Unlike PBMC or whole blood assays, these systems isolate plasma-mediated mechanisms, such as protein corona formation, opsonization, complement activation, aggregation, and changes in colloidal stability, without confounding effects from cellular uptake or cytokine production by blood cells. [134] These assays are particularly useful for understanding how human biofluids reshape LNP surface properties and influence downstream delivery, immunogenicity, and safety. For example, recent mRNA–LNP studies have shown that human serum exposure can induce complement activation markers such as C5a, sC5b-9, and Bb. [131] However, plasma/serum studies cannot capture cell-mediated immune activation, tissue uptake, or organ-specific toxicity, and therefore should be interpreted as complementary to whole blood and PBMC assays.

4.2.4. Tissue Explant Models

Tissue explant models are ex vivo systems in which freshly isolated human tissues or precision-cut tissue sections are maintained outside the body while preserving native tissue architecture, extracellular matrix, resident immune/stromal cells, and organ-specific cellular organization. [135] Tissue explants are well suited for evaluating tissue-level mRNA-LNP uptake, transfection efficiency, local mRNA expression, inflammatory responses, and tissue toxicity in a physiologically relevant microenvironment. In RNA-LNP research, human skin explants have been used to optimize self-amplifying RNA-LNP formulations, human ocular tissue explants have been used to assess retinal mRNA-LNP delivery and inflammation, and ex vivo human placenta perfusion models have been used to evaluate LNP transfer across the maternal–fetal interface. [136–138] Precision-cut liver slices can be used to assess liver-targeted mRNA-LNP expression, biodistribution, and toxicity, while precision-cut lung slices are useful for evaluating tissue exposure and safety in the context of inhaled LNP formulations. [139, 140] However, these models are limited by tissue availability, donor variability, short-term viability, reduced scalability, and lack of systemic PK or inter-organ interactions.

4.3. In Silico and Computational Modeling Approaches

In silico and computational modeling approaches are central components of NAMs, using mathematical and data-driven models to simulate biological processes, disease progression, and therapeutic responses. These tools complement in vitro and in vivo studies by integrating molecular, cellular, PK, and pharmacodynamic (PD) data to generate mechanistically informed predictions. By quantitatively linking molecular interactions to phenotypic outcomes, in silico models support hypothesis generation and translation across biological scales and species. They also enable rapid evaluation of dosing, efficacy, and toxicity scenarios, reducing reliance on animal studies and improving experimental efficiency. In this section, we review key computational NAMs applied to mRNA–LNP development and discuss practical considerations for their implementation.

4.3.1. Physiologically Based Pharmacokinetic (PBPK) Models

Mathematical models are widely used to characterize and predict the PK of drugs in animals and humans by providing a quantitative framework to describe drug exposure, response, and time course across dosing regimens. [141] PBPK models are a mechanistic class of PK models that explicitly incorporate physiological characteristics of the target population to predict drug behavior. These models are constructed using systems of differential equations parameterized with physiological and biochemical variables, enabling a quantitative description of drug absorption, distribution, metabolism, and excretion (ADME). [142] A number of commercial PBPK software platforms are now available with integrated modeling frameworks that substantially reduce coding barriers, including Symcyp PBPK Simulator (Certara, https://www.certara.com/software/simcyp-pbpk/), GastroPlus (SimulationsPlus, Simulations Plus | Modeling & Simulation Software | Consulting for Discovery Through Commercialization ), PKSIM (Open Systems Pharmacology, https://www.open-systems-pharmacology.org/ ), and MATLAB SimBiology (MathWorks, https://www.mathworks.com/products/simbiology.html ).

PBPK models have become increasingly valuable for mRNA-LNP development by linking nanoparticle properties and mRNA kinetics to tissue distribution, cellular uptake, and intracellular expression across species. These models support prediction of dose–exposure–response relationships, evaluation of formulation and route-of-administration effects, and translation from preclinical studies to humans to inform first-in-human dosing strategies. [143] Parhiz et.al developed a PBPK model for systemically administered luciferase mRNA–LNP in mice to characterize the PK/PD relationship and capture the time delay between mRNA delivery and pharmacological response. [144] In this model, the body is represented by major organ compartments connected through the blood circulation, reflecting normal mouse physiology.

Blood flows from peripheral tissues through the lungs before being distributed to organs such as the heart, liver, kidneys, and spleen. Additional compartments were included to represent gastrointestinal organs and remaining tissues, with blood from the spleen and gastrointestinal tract passing through the liver prior to returning to the systemic circulation. [144] Within each tissue compartment, the model describes LNP transport from blood into tissues, cellular uptake, and intracellular degradation. LNP degradation releases mRNA into the cytosol, where it is translated into luciferase and subsequently eliminated; all intracellular processes are represented using first-order kinetics. Model fitting was performed using radiolabeled LNP PK data and time-course luciferase expression data measured in blood, lung, heart, kidney, spleen, and liver from mouse studies. Physiological parameters such as tissue volumes, blood flow rates, and organ connectivity were obtained from experimentally measured data for a 25-g mouse and from established databases in commercial PBPK software and literatures. [145, 146] Drug-specific parameters related to mRNA–LNP properties, such as tissue-specific uptake rates and luciferase expression kinetics, were either estimated from in vivo data or informed by in vitro experiments. A mechanistic understanding of mRNA–LNP ADME processes, together with well-defined drug-specific parameters, is critical for successful PBPK model development. [142] Following model development, validation is essential to demonstrate predictive performance. In this case, the same PBPK framework was used to simulate biodistribution and mRNA expression for a different targeted LNP by updating only target-related parameters, successfully capturing expression profiles for the new formulation. [144] This highlights an important application of PBPK modeling in reducing animal use: predict PK profiles in vivo for new LNP designs. If model validation is unsuccessful and simulations fail to adequately capture the experimental data, this may indicate that important biological mechanisms influencing drug PK have not been incorporated into the model. In such cases, PBPK modeling can be used as a hypothesis-generating tool to identify potential sources of discrepancy between model predictions and observations and to guide the design of future experiments. [142, 147]

PBPK modelers have the flexibility to include ADME processes that are believed to have a meaningful impact on drug PK, while simplifying or omitting others depending on the modeling objective. For example, the ApoE–LDLR–mediated LNP uptake pathway was not explicitly modeled in the PBPK framework developed by Parhiz et.al [144] but was incorporated in the PBPK model developed by Wang et.al [148]. In Wang et al.’s study, the primary goal was to characterize mRNA–LNP transport and translation across species; therefore, species- and organ-specific heterogeneity in LDLR expression was incorporated into the model as a key determinant of ADME. Both the Parhiz and Wang models are well suited to address their respective clinical questions, despite differing in how the ApoE–LDLR interaction is represented. As George Box famously stated, “All models are wrong, but some are useful.” Accordingly, PBPK models developed by different groups may differ in structure and assumptions, reflecting the specific context, purpose, and clinical questions they are designed to address.

PBPK models can serve as foundational frameworks to reduce animal experiments by enabling in silico prediction of PK for new LNP formulations using formulation-specific parameters measured in vitro. This will allow virtual evaluation of LNP formulation or targeting changes, helping prioritize LNP designs before in vivo testing. By combining drug-specific inputs with physiological parameters, PBPK models can also extrapolate mRNA–LNP biodistribution and expression across species. This capability reduces reliance on animal studies for first-in-human dose selection and supports early efficacy prediction. Regulatory agencies such as FDA and EMA formally recognize the use of PBPK modeling in drug development and regulatory submissions. [149] Notably, there are currently limited PBPK case examples specific to mRNA–LNP therapeutics; however, PBPK models developed for other nanomedicine platforms and siRNA-based drugs provide valuable and relevant references for model development and application. [150, 151]

4.3.2. Quantitative Systems Pharmacology (QSP) Models

Quantitative systems pharmacology (QSP) models are mechanistic models that integrate drug exposure with biological pathways and disease mechanisms to predict pharmacological effects. Unlike PBPK models, which primarily describe drug ADME, QSP models focus on linking drug concentrations to target engagement, downstream biology, efficacy, and safety outcomes. [152] As a result, QSP models typically involve large numbers of equations and parameters, making parameter estimation and model calibration challenging. [153] Commonly used software platforms for QSP model development include MATLAB, R, Julia, and specialized systems biology tools such as OmniPath (network reconstruction from the literature), CellNOpt (model fit to experimental data), MaBoSS (model analysis), and Cytoscape (visualization), which generally require substantial programming and modeling expertise. [154] More recently, AI-guided approaches for QSP model construction, parameterization, and model reduction are being actively explored to lower technical barriers and accelerate model development. [155]

Dasti et al. developed a quantitative systems pharmacology (QSP) model for an mRNA–LNP COVID-19 vaccine to characterize immune responses following vaccination in humans and to optimize dosing schedules. [156] The model integrates two biological scales: a tissue scale and a molecular scale. At the tissue scale, compartments representing the injection site, draining lymph nodes, and blood describe the uptake of LNP-encapsulated mRNA by antigen-presenting cells, and subsequent activation of T- and B-cell–mediated immune responses. At the molecular level, the model captures mRNA release from LNPs, translation into antigenic protein, and antigen expression on the cell surface. Model parameters were obtained from published literature or estimated using clinical data from human COVID-19 mRNA vaccine studies. The model predicts immune cell dynamics at the injection site and lymph nodes, as well as vaccine-induced anti-RBD IgG antibody responses. Model validation was performed using independent clinical data describing anti-RBD IgG concentrations following administration of different doses of the BNT162b2 mRNA vaccine that were not used during model calibration. [156] Overall, this QSP model demonstrates predictive capability across dose levels and provides a quantitative tool to support optimization of mRNA–LNP vaccine dose and dosing regimens to improve efficacy while ensuring safety.

One important application of QSP models is to generate virtual animals or virtual patient populations by incorporating biological and parameter variability, enabling simulation of heterogeneous responses to treatment and supporting evaluation of efficacy and safety across diverse populations. [157] Miyazawa et.al develop a QSP modeling framework to describe the mRNA-LNP enzyme replacement therapy for Crigler-Najjar syndrome. [158] The model characterized the systematically administered mRNA-LNP entering liver, distinct kinetics in hepatocytes and Kupffer cells, as well as recycling pathways that impact target protein expressions. Once successfully validated, the model can be used to simulate virtual animals by introducing reasonable variability in the parameters. These virtual cohort simulations have the potential to recapitulate population heterogeneity in mRNA-LNP therapeutics efficacy, and therefore reduce animal use in future study design.

Another application of QSP model is first-in-human (FIH) dose selection. Traditional empirical FIH dose selection methods rely largely on animal toxicology endpoints and simple exposure scaling, which often fail to capture the complex intracellular delivery, translation, and delayed pharmacodynamic effects of mRNA–LNP therapeutics. [159] As a result, QSP models are increasingly needed to support FIH dose selection by mechanistically linking preclinical drug exposure to target engagement and downstream biological responses, enabling quantitative prediction of safe and efficacious dose ranges in humans. [143] For example, Apgar et al. developed a QSP model to integrate preclinical data for a modified UGT1A1 mRNA–LNP therapy for Crigler–Najjar syndrome; by accounting for species differences in target protein production and clearance, the model enabled rational selection of an appropriate FIH dose. [160]

Quantitative systems toxicology (QST) models extend QSP frameworks to mechanistically describe the biological pathways underlying drug-induced toxicity and immune activation. QST models can integrate LNP biodistribution, innate immune sensing, cytokine signaling, and downstream tissue responses to predict mRNA-LNP therapeutics immunogenicity and toxicity, such as excessive inflammatory activation or organ-specific adverse effects. Although currently there are no published, classical QST models specifically for mRNA-LNP to date, lessons can be learned from small molecules and biologics. [161, 162] One notable example is the DILIsym QST model, which mechanistically predicts drug-induced liver injury (DILI) by integrating in vitro mechanistic data with drug exposure and biological pathways of liver toxicity across species. DILIsym successfully predicted the human hepatotoxicity of the migraine drugs telcagepant and MK-3207, which was not identified by routine preclinical testing and ultimately led to termination of their clinical development. In contrast, the model predicted that the next-generation compound ubrogepant would have a favorable liver safety profile, a prediction that was later confirmed in phase 3 clinical trials and supported its FDA approval without liver safety warnings. [163]

In summary, QSP and QST models have strong potential to reduce or replace animal use in mRNA–LNP drug development by mechanistically predicting efficacy and toxicity across species. Their ability to integrate virtual population heterogeneity and variability enables more informed decision-making for dose selection and drug approval. Key challenges include high model complexity, data requirements for parameterization, and uncertainty in poorly characterized immune and toxicity pathways. Careful definition of model scope, transparent assumptions, and integration of high-quality in vitro and clinical data validation are critical for successful application as NAMs. [164]

4.3.3. AI and Machine Learning Tools

Artificial intelligence (AI) and machine learning models learn patterns from complex datasets to make accurate predictions and support decision-making. Compared with traditional statistical models, AI/ML approaches can capture nonlinear and high-dimensional structure–activity relationships that are difficult to define a priori in multi-component LNP systems, but they require sufficiently large, well-annotated, and experimentally consistent datasets for training and validation.. [165] In a traditional machine learning workflow, the first step is to collect a dataset that includes features (input variables describing the system) and outcome of interest (the target to be predicted). [166] For mRNA–LNP therapeutic development, model input features may include formulation descriptors, physicochemical properties, in vitro or ex vivo assay readouts, process parameters, and, when available, preclinical biodistribution or expression data. The predicted outcomes may include organ- or cell-specific mRNA expression, immunogenicity or inflammatory potential, tissue biodistribution, critical quality attributes, and therapeutic efficacy or response. These outcomes can be binary, categorical, or continuous, depending on the scientific question. There is no single fixed number of samples needed for all machine learning models, but empirical studies often find that hundreds to thousands of samples are needed for robust performance. [167] The full dataset is then split into training and testing sets, commonly using an 80:20 or 70:30 ratio, to allow independent model evaluation. The machine learning model is trained using only the training dataset. Unlike classical statistical regression, where equations and parameters are explicitly defined, machine learning models often contain hundreds or thousands of hyperparameters that are learned automatically from the data. Commonly used algorithms in this context include tree-based models such as random forests and gradient-boosting methods. [168] After training, the same model (with fixed hyperparameters) is applied to the test dataset to evaluate how well it generalizes to unseen data. Model performance is typically assessed using metrics such as the AUC–ROC (Area Under the Receiver Operating Characteristic curve), which measures the model’s ability to distinguish between outcome classes across all classification thresholds. An AUC–ROC value in test dataset of 0.5 indicates no predictive ability, while values closer to 1.0 indicate excellent discrimination; in practice, an AUC–ROC greater than 0.8 is often considered indicative of a strong and useful predictive model. [169]

AI and machine learning approaches are increasingly used to design and optimize ionizable lipids and multi-component LNP formulations by learning complex relationships between chemical structure, formulation composition, and mRNA delivery performance. For example, the AGILE platform applies deep learning to identify cell-specific lipid preferences, enabling tailored LNP design for optimal delivery to different cell types. [170] Trained through large-scale self-supervised pretraining followed by fine-tuning with experimental transfection data, AGILE rapidly narrows vast virtual lipid libraries to a small set of high-performing candidates, substantially reducing experimental screening and animal use. The same team also developed a machine learning model using LNPs composed from 584 different ionizable lipids to predict their transfection efficiency in HeLa cells. [171] The model was then expanded to discover ionized lipid chemistry and structure that can improve LNP target delivery in lung and spleen in vivo (animal models). This AI-powered lipid discovery platform is anticipated to provide a useful tool facilitating the future development of delivery systems for RNA therapeutics, with less reliance on animal experiments.

Similarly, Duan et.al launched a machine learning framework named LipidAI, where the model can predict expression of mRNA with different ionizable lipids structures using 475 LNP samples in mouse models. [172] Wang et al. developed an AI-driven machine learning platform trained on nearly 20 million virtual ionizable lipid structures to predict apparent pKa and mRNA delivery efficiency, enabling rational design and rapid identification of high-performance LNP lipids with reduced experimental screening. [173] These approaches demonstrate that data-driven models can link LNP chemistry and in vitro performance to in vivo biodistribution and mRNA expression, enabling model-informed formulation selection, reducing reliance on animal studies, and accelerating mRNA–LNP therapeutic development. However, how these AI-guided LNP design can be translated into human still remain to be explored.

Other examples included using machine learning frameworks to bridge in vitro properties—such as physicochemical characteristics and protein corona fingerprints—with in vivo biodistribution and mRNA expression, enabling more predictive, model-informed formulation selection with reduced animal use. Hanafy et al. developed PRELIVE, a predictive framework that links LNP composition and in vitro protein-corona profiles to organ-level biodistribution, demonstrating the feasibility of predicting in vivo behavior from in vitro measurements. [174] Similarly, deep learning–based lipid optimization platforms trained on large experimental datasets enable virtual screening of millions of ionizable lipids and identification of high-performing formulations for targeted mRNA delivery. [175]

In parallel, machine learning models are increasingly applied to optimize manufacturing processes and critical quality attributes, including particle size, polydispersity, and encapsulation efficiency. Duffrène et.al developed a data-driven framework using microfluidic-generated mRNA–LNP datasets to predict nanoparticle size and encapsulation efficiency from lipid composition and process parameters, enabling efficient formulation optimization with limited experimental input. [176] Similarly, Maharjan et al. applied machine learning to optimize microfluidic process conditions and lipid mixing ratios across 24 mRNA–LNP formulations, improving prediction of critical quality attributes and overall manufacturing efficiency. [177]

4.3.4. Integrated and Emerging Tools

Emerging integrative frameworks that bring together different types of computational models, such as PBPK and QSP, may further support mRNA-LNP therapeutic development. One example is the digital twin, which can be understood as a virtual version of a product, process, or biological system. Unlike a static model, a digital twin can be updated with new experimental, manufacturing, or clinical data and used to simulate how the system may behave under different conditions. [178] Although digital twins have not yet been widely applied to mRNA-LNP therapeutics, their use is expected to grow as richer datasets and more mature multiscale models become available. We anticipate that digital twins could be used in mRNA-LNP development to simulate behavior under alternative LNP formulation, dosing, biological, or manufacturing scenarios before committing to costly animal experiments or clinical studies.

The emergence of AI co-scientists represents another promising future opportunity. An AI co-scientist is a large language model-based AI system designed to assist human researchers by generating, refining, and validating scientific hypotheses, thereby augmenting human creativity and accelerating scientific discovery. [179] In the context of mRNA-LNP development, such systems could integrate chemical knowledge, formulation expertise, and in vitro and in vivo data to autonomously propose and prioritize new LNP formulations optimized for targeted organ delivery, enhanced expression, and improved safety. Another important opportunity is the use of AI-assisted knowledge extraction and natural language processing could then identify patterns across studies to infer potential drug–drug interactions (DDIs) for mRNA-LNPs, such as interactions driven by shared metabolic pathways, immune activation, or lipid excipient effects. These models could support early DDI risk assessment and guide safer clinical development without extensive empirical testing. [180–182]

4.3.5. Validation for Computational Models

Validation is essential for ensuring that the in silico tools or computational models provide reliable and biologically meaningful predictions rather than only fitting the data used for model development. Internal validation strategies, such as bootstrapping or cross-validation, can help assess model stability and reduce overfitting. External validation using independent experimental datasets is particularly important for evaluating generalizability and translational relevance. Iterative comparison between predicted outputs and experimental observations can then be used to refine model assumptions, update parameters, and improve predictive performance. This good practice is also aligned with broader principles of transparency and rigor in model development, helping ensure that computational predictions are reproducible, interpretable, and suitable for informing drug development.

4.4. Context of Use: Matching NAMs to Specific mRNA–LNP Development Decisions

A central limitation in the current implementation of NAMs for mRNA–LNP development is that different technologies are often discussed as broadly interchangeable “alternatives” to animal models. In practice, however, the value of a NAM depends strongly on its context of use: the specific drug-development question, biological mechanism, decision point, route of administration, therapeutic indication, and acceptable level of uncertainty. [183] Therefore, NAMs should be selected and interpreted according to the question they are intended to answer, rather than ranked generically by technological complexity. (Figure 3)

Figure 3. Context of Use diagram for NAMs in mRNA-LNP therapeutics.

Physicochemical, in vitro, and ex vivo data can be integrated into in silico models to address distinct translational questions. PBPK models are well suited for biodistribution and exposure prediction, QSP models support efficacy, safety, immunogenicity, and mechanism-based interpretation, and AI/ML approaches enable formulation exploration and data-driven candidate optimization. PBPK, physiologically-based pharmacokinetics; QSP, quantitative systems pharmacology; AI/ML, artificial intelligence and machine learning.

For early LNP formulation screening, the primary goal is to prioritize LNP compositions with acceptable physicochemical properties, cellular uptake, mRNA expression, endosomal escape, and preliminary cytotoxicity. High-throughput 2D assays and AI/ML models are well suited for this purpose because they can evaluate large formulation spaces before animal testing. Immortalized cell lines provide scalable first-pass screening, while primary human cells and iPSC-derived cells are more appropriate when human target-cell biology is central to the product. [78, 81] However, these systems should be used for formulation triage, not for predicting whole-body biodistribution or clinical efficacy.

For delivery and biodistribution, NAM selection should depend on the route of administration and target tissue. Intravenous liver-directed mRNA–LNPs require models that capture serum protein interactions, ApoE/LDLR-mediated uptake, hepatic cell tropism, endothelial barriers, and inflammatory clearance;[4, 54] relevant NAMs include primary human hepatocytes, liver co-cultures, liver spheroids, liver-on-chip systems, and PBPK models. By contrast, intramuscular vaccines require models of injection-site uptake, lymphatic drainage, antigen-presenting-cell activation, and lymph-node responses, whereas inhaled products require airway epithelial, mucus-barrier, macrophage, and mucociliary-clearance components. [121, 184, 185] Therefore, NAMs developed for intravenous liver delivery should not be assumed to generalizable to other routes.

For immunogenicity and reactogenicity, NAMs should focus on human innate immune sensing, cytokine release, complement activation, anti-PEG responses, and excessive inflammation. [66] Human PBMC assays, dendritic-cell/macrophage co-cultures, cytokine-release assays, complement assays, and lymphoid organ-chips are more relevant than standard mouse models. [186] For vaccines, these systems can assess antigen presentation, B-cell activation, antibody production, and T-cell priming, but should be used to compare immune mechanisms and risk rather than predict long-term protection. [6]

For predicting oncology therapeutic response, NAM selection should depend on the mechanism of action. Cancer vaccines require models that capture antigen expression, antigen presentation, T-cell priming, immune-cell infiltration, and tumor-mediated immune suppression. [187] Tumor spheroids and organoids are useful for assessing LNP penetration and transgene expression in 3D tumor architecture, but immune-competent tumor organoids, tumor–immune cell co-cultures, and tumor microphysiological systems are better suited for evaluating immunotherapy mechanisms. [188, 189]

For clinical pharmacology questions, such as first-in-human dose selection, NAMs should enable quantitative translation. PBPK models can link formulation properties, tissue exposure, cellular uptake, and mRNA expression across species, while QSP models can connect mRNA expression to downstream pharmacology, biomarkers, and therapeutic outcomes. [158, 160] Rather than defining a single dose, these models should estimate a plausible human dose range, quantify uncertainty, and identify assumptions that can be updated with early clinical data.

For population heterogeneity, ex vivo and in silico NAMs can help evaluate variability that is poorly captured in standardized animal models. Donor-diverse primary cell panels, human serum panels, PBMC assays, and virtual patient populations generated by PBPK/QSP models can estimate response variability and identify subgroups at higher risk of reduced delivery, excessive reactogenicity, or altered protein expression. [190]

4.5. Summary

With the rapid advancement of human-relevant in vitro cell-based systems, ex vivo assays, and in silico predictive models, the development of mRNA–LNP therapeutics can increasingly rely on NAMs to assess efficacy and safety. Importantly, NAM selection should be guided by the specific context of use, including the scientific question being addressed and the type, quality, and relevance of available data. Given the novelty of the mRNA–LNP modality and the limited number of approved human products, animal models currently remain necessary as interim or confirmatory tools. However, the systematic integration of in vitro, ex vivo and in silico approaches, supported by accumulating modality-specific data, holds strong potential to progressively reduce animal use and enhance the human predictivity of preclinical development.

5. Future Opportunities and Development Directions

5.1. Critical Use Cases for Applying NAMs in mRNA-LNP Development

For mRNA-LNP therapeutics, the application of NAMs should be guided by safety and efficacy questions that are mechanism-specific and insufficiently predicted by traditional animal models. Key assessment domains include acute toxicity, chronic toxicity and organ injury, immune responses, and efficacy. [18] Acute toxicity assessments of both LNP-associated cytotoxicity and dose-dependent inflammatory effects are often species-specific and inadequately captured in rodent models. [191] Similarly, chronic toxicity and organ injury, particularly those arising from repeated dosing and hepatic or splenic accumulation, remain difficult to evaluate in animals due to interspecies differences in lipid metabolism, tissue distribution, and clearance pathways. [192] Immune responses represent another urgent area for NAM development, as animal immune systems have limited translatability for predicting innate immune activation, cytokine release, complement activation, and the formation of anti-LNP or anti-PEG antibodies in humans. [57, 58] Finally, efficacy evaluation of mRNA–LNP products should emphasize human-relevant functional readouts that directly link target protein expression to biological activity using NAM-based systems, rather than relying solely on expression levels or disease animal models with limited predictive value. Collectively, prioritizing NAM strategies for toxicity, immune response, and efficacy assessments has the potential to enhance human relevance, reduce reliance on animal testing, and support more efficient regulatory decision-making in mRNA–LNP drug development.

5.2. Decision Frameworks for Reducing or Replacing Animal Testing

Decision frameworks for reducing or replacing animal testing should be explicitly driven by context of use, with NAM selection tied to clearly defined regulatory questions such as acute toxicity risk, immune activation potential, or functional efficacy. [25] Risk-based decision trees should be applied to determine whether NAM data are sufficient to replace, augment, or de-risk animal studies based on exposure duration, dosing regimen, route of administration, and known modality-specific liabilities. [193] To support decision-making, results from complementary NAM assays should be integrated using structured confidence-scoring approaches that evaluate biological relevance, technical robustness, and concordance across endpoints, rather than relying on single-assay outcomes. [194] Validation of NAMs within this framework should be fit-for-purpose, focusing on demonstrated utility for specific regulatory decisions—such as first-in-human dose selection or immunotoxicity risk assessment—rather than requiring full equivalence to animal models. [25] This enables consistent, transparent, and scientifically justified reductions in animal testing while maintaining regulatory confidence in human safety assessments.

Despite their growing utility, current NAM platforms still have important limitations for mRNA-LNP evaluation. In particular, organ tropism and systemic biodistribution remain challenging to fully recapitulate using in vitro or ex vivo systems, as these outcomes are shaped by complex interactions among circulation dynamics, vascular barriers, immune clearance, tissue-specific uptake, and whole-body physiology. Therefore, while NAMs can provide valuable mechanistic and screening-level insights, immediate replacement of animal testing remains challenging for endpoints requiring integrated systemic assessment. Further development and validation of NAM platforms will be needed to support their broader application.

5.3. Data Integration and Open Science Opportunities

The development of in silico tools for predicting the safety and efficacy of mRNA–LNP products requires access to large, high-quality datasets. Existing preclinical and clinical data generated from mRNA–LNP programs should be systematically leveraged to train and refine these models. For example, Zenhausern et al. evaluated a panel of mRNA–LNP formulations with diverse lipid chemistries following intravenous administration in NHP, generating a rich dataset that can serve as a valuable resource for investigating systemic physiological responses to LNP treatment in a translationally relevant species. [195] To enable broader application of such data, there is a need for the development of open-access, standardized databases for mRNA–LNP formulations, biological responses, and outcomes, which would substantially accelerate Ai–based model development and regulatory adoption.

5.4. Training, Communication, and Cultural Change

Successful implementation of NAMs requires targeted cross-disciplinary education that equips regulators, industry scientists, and academic researchers with a shared understanding of emerging technologies, their limitations, and their appropriate context of use. [25] Training programs should integrate expertise across toxicology, immunology, biomedical engineering, pharmacology, computational modeling, and regulatory science to ensure consistent interpretation of NAM-derived data. Alignment among regulators, industry, and academia should be strengthened through early and continuous communication, including precompetitive data sharing and joint pilot studies.

6. Conclusions

In summary, this review highlights the urgent need for NAMs in the development of mRNA–LNP therapeutics, given the well-documented translational gaps of animal models in predicting human efficacy, safety, inflammation, delivery and biodistribution, and population heterogeneity. We discuss how human-relevant in vitro and ex vivo experimental systems and in silico modeling tools can address these limitations by enabling mechanism-informed, quantitative, and context-specific assessments that are more directly applicable to human biology. Particular emphasis is placed on the development of in silico models, including the importance of fit-for-purpose and model validation to ensure that NAMs are applied appropriately to specific scientific questions. Looking ahead, we identify key use cases for NAMs in mRNA–LNP development and emphasize the need for coordinated data sharing, cross-sector communication, and cultural change across academia, industry, and regulatory agencies. Together, these efforts are essential for integrating NAMs responsibly, improving translational predictability, and reducing reliance on animal testing.

• Defines animal-model limitation and translational gaps in mRNA-LNP drug development

• Provides a comprehensive summary of mRNA-LNP NAMs

• Outlines workflows for in vitro, ex vivo, and in silico NAMs for mRNA-LNP

• Maps NAMs to key decisions in mRNA-LNP drug development