Human blood–brain barrier tissue model to characterize lipid nanoparticle delivery and transport mechanisms

Abstract

Crossing the blood-brain barrier (BBB) to deliver lipid nanoparticles (LNPs) for central nervous system (CNS) therapies remains a major challenge. Here, we present a compartmentalized, human-derived 3D BBB tissue model incorporating five cell types, integrating a BBB with a protein composite scaffold system, to support parenchymal elements within a single integrated brain tissue system. This tissue model recapitulates key structural and functional features of the native human BBB, including enhanced tight junction formation, neuronal maturation, polarized endothelial morphology, low permeability, and a more homeostatic microenvironment. The silk-collagen composite provided physiologically relevant extracellular matrix stiffness that supported long-term culture stability and parenchymal development. Translational utility for evaluating CNS-penetrating LNPs was demonstrated by showing that LNP transport efficiency and parenchyma penetration capability in vitro correlated with in vivo brain delivery following systemic administration in mice. Furthermore, this in vitro tissue model enables mechanistic investigation of LNP transport via receptor modulation using siRNA knockdown and pharmacological inhibition, revealing scavenger receptor class B type I (SR-B1) and insulin receptor (INSR) as key mediators of receptor-dependent transcytosis. By enabling the integrated assessment of permeability, transport mechanisms, and toxicity within a single human-relevant in vitro tissue platform, this model serves as a tool to bridge the translational gap between LNP design, in vitro screening, and in vivo validation, to support the optimization of CNS drug delivery. Among the tested formulations, LNP1 exhibited superior BBB penetration, neuronal transfection, and low toxicity, highlighting its potential as a promising lead candidate for CNS mRNA therapeutics for neurological diseases.

1. Introduction

The blood-brain barrier (BBB) is a highly selective, tightly regulated interface that protects the central nervous system (CNS) by restricting the entry of over 98% of small molecules and nearly all macromolecules from the systemic circulation into the brain parenchyma [1]. Although this physiological barrier is essential for maintaining neural homeostasis, it poses a major challenge for CNS drug delivery and therapeutic development for neurological diseases, such as Alzheimer’s, Parkinson’s, and amyotrophic lateral sclerosis (ALS) [2]. The functional integrity of the BBB is mainly contributed by a complex multicellular structure known as the neurovascular unit (NVU) [3], which comprises endothelial cells, pericytes, astrocytes, neurons, and microglia [2]. These cells form a tightly regulated microenvironment through cell-cell and cell-matrix interactions that control transport, immune response, and homeostasis [4]. In addition to this cellular complexity, the extracellular matrix (ECM) of the brain parenchyma plays a crucial role in supporting the parenchymal structure and the blood-brain barrier (BBB) function [5,6]. Extracellular matrix components such as collagen and laminin promote neurite extension and cell adhesion and also influence the BBB tight junction integrity and gene expression profiles in endothelial cells [5].

Despite extensive research efforts, the efficient delivery of therapeutic modalities to target and cross the BBB remains a challenge. A wide range of CNS delivery strategies have been explored, including antibody-based delivery [7],cell-penetrating peptides (CPP) and peptide-conjugates [8], exosomes and extracellular vesicles [9], viral vectors [10], small molecule prodrugs, and polymeric nanoparticles [11]. These platforms aim to enhance transport across the BBB via diverse mechanisms, such as receptor-dependent transcytosis, membrane permeabilization, or cellular uptake [12,13]. Delivery strategies vary in invasiveness, from non-invasive systemic administration to more direct methods such as intrathecal cerebrospinal fluid (CSF) injection, intracerebroventricular delivery, or intranasal delivery [1,14]. These strategies can be used individually or in combination to improve CNS drug distribution and therapeutic efficacy. Lipid nanoparticles (LNPs) have emerged as a leading platform for CNS drug delivery, particularly due to their modular design and capacity for systemic delivery [11,15]. Brain-penetrating LNPs typically cross the BBB through receptor-dependent transcytosis pathways, engaging endothelial transport receptors-such as transferrin receptors (TfR), insulin receptors (INSR), scavenger receptor class B type 1 (SR-B1), and low-density lipoprotein receptor-related protein (LRP), among others [16–18]. However, evaluating BBB transport efficiency and the mechanistic behavior of such LNPs with in vitro BBB models or in vivo models remains a challenge. Although animal models provide physiological relevance, they are more challenging to use for evaluating the LNP performance and studying the transport mechanism simultaneously due to the complex physiological features combined with difficulties in accurate tracking, as well as the species translational gaps with humans [16]. On the other hand, current in vitro models usually fall short of replicating the comprehensive physiological and functional complexities of the human BBB [19,20]. 2D monolayer or conventional 2D transwell systems lack multicellular spatial architecture and ECM signaling [21,22]. 3D organoids and spheroids offer tissue-level complexity but lack a defined vessel-parenchyma compartmentalized structure, limiting quantitative permeability analysis [23–25]. Microfluidic BBB-on-chip systems simulate shear stress and allow real-time TEER monitoring via impedance spectroscopy. However, standardization across platforms remains limited, and many systems lack comprehensive NVU complexity, have low throughput, and rely on materials like Polydimethylsiloxane (PDMS) that may absorb small molecules [25,26].

To address these limitations, we developed a compartmentalized 3D human BBB tissue model that incorporates all cellular components of the NVU and mimics the native BBB and brain parenchyma structure and functions (Fig. 1a). This architecture combined a transwell-based BBB barrier with a 3D silk-collagen scaffold, previously developed in our lab [27–30], supporting neural and glial populations. The system reconstructs both the vascular and parenchymal compartments, providing a physiologically relevant platform for directional transport, barrier integrity measurement, brain parenchyma penetration evaluation, and receptor-level mechanism research, to support the accurate quantitative assessment of LNP library performance. Using this platform, we identified LNP1, a formulation containing cholesterol and polyethylene glycol (PEG)-lipid components, as a top-performing candidate of those screened. LNP1 showed dose- and time-dependent transfection in the 3D parenchyma and functional delivery to neurons in vivo, with a strong correlation between in vitro and in vivo outcomes. Furthermore, siRNA knockdown of SR-B1 and INSR in the endothelial layer significantly reduced LNP1 transport, confirming receptor-dependent facilitation of transport. These results highlight both the predictive accuracy of the tissue model and its utility for mechanistic evaluation. Importantly, the high brain penetration efficacy, consistent superior brain distribution, and minimal cytotoxicity results also highlight LNP1 as a potential candidate for future CNS delivery platforms.

Fig. 1. 3D human blood-brain barrier (BBB) tissue model recapitulating vascular and neuronal features of the neurovascular unit.

(a) Schematic illustration of the native human BBB structure and the corresponding cellular composition incorporated into the 2D and 3D in vitro BBB model. (b) Schematic depiction of the preparation of the 3D triculture brain scaffold consisting of human-induced neural stem cells (hiNSCs), primary human astrocytes, and HMC3 human microglial cells. (c) schematic description of our 3D human BBB model at the cellular resolution, highlighting the spatial arrangement of the neurovascular unit. (d-e) RT-qPCR analysis of endothelial junctional markers CLDN5 (d) and PECAM1 (e) in 2D vs 3D model inserts, with or without 600 nM CDK4/6 inhibitor (PD 0332991) treatment over time. (f) Barrier permeability measurement using apparent permeability coefficients (P_app) calculated from 20 kDa FITC-dextran diffusion across 2D and 3D BBB inserts, with or without PD treatment at day 7. (g) Transendothelial electrical resistance (TEER) measured between the apical and basolateral compartments with 2D and 3D BBB models on day 4 and day 7. (h) Immunofluorescence images of tight junction protein expression in the 2D IM-HBMEC-only model (DAPI, blue; PECAM1, green; CLDN5, red). (i) Immunofluorescence images of tight junction protein expression in the 2D triculture BBB model (DAPI, blue; PECAM1, green; CLDN5, red). (j) Tight junction staining in the 3D BBB model reveals mature cobblestone-like endothelial morphology (DAPI, blue; PECAM1, green; CLDN5, red) (scale bar = 50 μm). (k) Quantification of the endothelial cell perimeters across 2D and 3D models. (l-n) RT-qPCR analysis of MAP2, NeuN, and TUJ1 expression in 2D vs. 3D triculture brain parenchyma, with or without PD treatment at different time points. (o) Immunostaining of the 2D triculture brain compartment shows expression of astrocytic and neuronal markers (DAPI, blue; GFAP, green; TUJ1, red) (scale bar = 50 μm). (p) Immunostaining of the 3D triculture scaffold demonstrates a dense, mature neuronal network and a higher polarized astrocyte population compared to the 2D model. (DAPI, blue; GFAP, green; TUJ1, red) (scale bar = 50 μm). (q) Quantification of MMP-1 and Angiopoietin-2 cytokine release from 2D and 3D triculture models using RayBiotech cytokine array, with signal intensity measured by ImageJ. (Data represent mean ± s.d. of at least n = 3 for each condition, *p < 0.05, **p < 0.01, ***p < 0.005, and ****p < 0.001. Statistical analyses were performed using the one-way ANOVA and Tukey’s multiple comparison test).

Here, we demonstrated that this physiologically relevant, human-derived, five-cell 3D BBB model can serve as a robust and accurate tissue platform to evaluate LNP transcytosis efficiency across the BBB, assess downstream brain parenchymal penetration, and investigate transport mechanisms through receptor interventions. Ultimately, this tissue model aims to bridge the translational gap between in vitro screening and in vivo performance, accelerating the development of effective CNS-targeting therapeutics.

2. Results

2.1. Development of a compartmentalized 3D human BBB model recapitulating neurovascular architecture and function

To reconstruct the human BBB, a physiologically relevant architecture is required for the robust evaluation of LNP permeability, transport mechanisms, and post-barrier brain parenchyma distribution. To address this need, we developed a 3D in vitro BBB tissue model composed of five human-derived cell types, which aims to recapitulate the sophisticated physiological structural, and microenvironmental features of the human neurovascular unit (NVU). Using commercially available transwell inserts, human-immortalized brain microvasculature endothelial cells (IM-HBMECs) were seeded on the apical side, while primary human astrocytes and pericytes were seeded on the underside of the insert to mimic the vascular interface. The basolateral compartment incorporated human-induced neural stem cells (hiNSCs), primary human astrocytes, and a human microglia cell line (HMC3) in a 2D monolayer culture or in a 3D brain-like triculture scaffold [30], designed to replicate the brain parenchyma and the immune regulatory environment (Fig. 1a–c). By incorporating both vascular and parenchyma elements, this five-cell type model supports detailed investigations into drug transport, distribution mechanisms, and potential adaptations associated with aging or neurodegenerative-related BBB status.

To assess barrier integrity, gene expression in the apical insert population was evaluated over time. The 3D BBB model requires a defined maturation period to establish stable barrier integrity and neurovascular organization; therefore, molecular and functional analyses were performed at Day 4 and Day 7, corresponding to fully established barrier conditions. Day 4 was also selected to align with subsequent nanoparticle transport studies, ensuring consistency between barrier characterization and functional evaluation. In the 2D BBB model, a time-dependent decline in claudin 5 (CLDN5) expression was observed between days 2 and 7 post-seeding. In contrast, the 3D brain tissue model exhibited a significant 3-fold upregulation in CLDN5 mRNA expression by day 7 relative to the 2D model (Fig. 1d), suggesting enhanced tight junction formation under 3D conditions. Functionally, the 3D BBB model showed at least a 3-fold lower permeability to 20 kDa FITC-dextran (Fig. 1f) and significantly higher transendothelial electrical resistance (TEER) measurements by day 7 compared to the 2D model (Fig. 1g). Continued TEER increase in the 3D model from day 4 to day 7 further supported barrier maturation. These findings highlight that 3D scaffold integration promotes tight junction gene expression and barrier function, establishing a more robust BBB phenotype.

Previous studies have shown that cyclin-dependent kinase 4/6 (CDK4/6) inhibition can induce endothelial cell quiescence [23,31,32], we hypothesized that CDK4/6 inhibition might reinforce tight junctions in our model. Endothelial cell quiescence is associated with reduced proliferation, loss of activated tip-stalk cell phenotype, and reinforcement of apical-basal polarity, and these effects support the formation of mature blood-brain barrier networks [32]. To test this hypothesis, the apical insert was treated with 600 nM CDK4/6 inhibitor (PD 0332991) in the apical compartments of the transwell model. After 7 days of the treatment, PECAM1 gene expression was significantly upregulated in both the 2D and 3D models (Fig. 1e), indicating the development of endothelial tight junctions. This change corresponded with the significant reduction in the permeability coefficient in the 3D model (Fig. 1f), indicating that CDK4/6 inhibition effectively promoted tight junction maturation via endothelial quiescence.

Maturation of the neuronal network was also significantly improved in the 3D brain scaffold compared to the 2D system. RT-qPCR results revealed a significant increase in mRNA levels of mature neuronal markers by day 7, where expression of the Microtubule-Associated Protein 2 (MAP2), Neuronal Nuclei (NeuN), and βIII-tubulin (TUJ1) increased 5-fold, 15-fold, and 2-fold, respectively, compared to the traditional 2D triculture brain parenchyma system (Fig. 1l–n). These transcriptomic changes led to a higher proportion of structurally mature neurons. This differential response to CDK4/6 inhibition likely reflects the presence of extracellular matrix cues and multicellular architecture in the 3D system, which support context-dependent neuronal maturation following cell cycle exit. In contrast, the 2D system lacks these structural and paracrine interactions, limiting the translation of cell cycle arrest into enhanced maturation marker expression. In parallel, cytokine analysis revealed that the 3D triculture model secreted significantly lower levels of barrier-disruptive cytokines, angiopoietin-2 and MMP-1, reduced by 50% and 90%, respectively, compared to the 2D model (Fig. 1q). This suggests a more homeostatic and barrier-supportive microenvironment in the 3D system.

Immunofluorescence staining further supported the enhanced functionality of the 3D BBB model system. Strong TUJ1 expression and glial fibrillary acidic protein (GFAP) enrichment in the 3D scaffold (Fig. 1p) indicated neuronal polarization, axon growth, stabilization of the microtubule network of neurons, and astrocytic support in the triculture system [16]. In parallel, the astrocytes in the 3D model also exhibited reduced cell body size and process length compared to the 2D model, indicating astrocyte polarization and functional maturation (Fig. 1o and p). Consistent with these findings, morphological changes associated with barrier organization across the endothelial monolayer, 2D BBB, and 3D BBB model inserts were also observed by immunostaining (Fig. 1h–j). The results indicated that the existence of 3D brain scaffolds generated a mature and cobblestone-like morphology change in the endothelial inserts, including well-defined and continuous dense PECAM1 expression and significantly increased cell perimeter (Fig. 1k), indicating a strong and mature blood-brain barrier phenotype. We hypothesize that the presence of the mature neurons, astrocytes, and microglia embedded in the 3D collagen ECM with the protein scaffolds provided a more human physiologically relevant homeostasis microenvironment to stimulate the parenchyma-barrier crosstalk, further driving the BBB and NVU polarization and differentiation into cobblestone-like phenotypes than the 2D triculture system.

Altogether, the 3D BBB tissue model successfully reproduced key features of the human BBB, including enhanced tight junction gene expression, robust mature neuronal network maturation, cobblestone-like endothelial cell morphology, reduced barrier-disruptive cytokines, and decreased barrier permeability. These improved features provided a foundation for accurately evaluating LNP permeability and transport across the BBB and enabled mechanistic evaluation of CNS drug delivery, post-barrier brain parenchyma distribution, and disease modeling in a controlled, physiologically relevant in vitro platform.

2.2. In vitro screening of lipid nanoparticle formulations reveals superior CNS delivery by LNP1

Having validated the cellular organization, barrier functionality, and morphology features of the 3D BBB model, we next evaluated its translational utility in predicting the transport and distribution of LNPs across the BBB and into the parenchyma. Our primary goal was to determine if the tissue model could provide accurate in vitro prediction and quantification of LNP transport before conducting in vivo animal studies, thereby reducing animal usage and improving early-stage CNS drug candidate selection. To assess the model’s prediction capability, we evaluated a library of LNPs encapsulating enhanced green fluorescence protein (EGFP) mRNA and compared their delivery performance across different versions of the BBB tissue models by studying increasing physiological complexity: a conventional 2D transwell BBB model, and our 3D triculture BBB tissue model, using a commonly used 2D HEK293T monolayer as a transfection control. By capturing LNP distribution and barrier-transport ability, this approach enabled a systematic comparison of screening capability between standard in vitro assays and our physiologically relevant 3D tissue model.

In previous LNP library screenings conducted in our lab with the 2D HEK239T cell model, due to the lack of structural and functional complexity of the BBB, we failed to accurately predict CNS delivery capability, showing minimal differences in transfection efficiency across the LNP formulations (Fig. S2e). To improve resolution, we tested a subset of LNPs (LNP1, LNP6, and LNP9, Fig. 2a, Fig. S1) using a traditional 2D transwell system. In this system, IM-HBMECs were seeded on the apical insert, and hiNSCs were seeded in the basolateral chamber. LNPs were applied to the apical side, and EGFP transfection efficiency was assessed by both flow cytometry and fluorescence imaging of the apical IM-HBMECs and basolateral hiNSCs. Both LNP1 and LNP9 demonstrated equal levels of EGFP expression in both endothelial cells and hiNSCs. The EGFP intensity was measured by flow cytometry and fluorescence imaging of the IM-HBMECs and hiNSCs (Fig. S2a–d). These results suggested comparable delivery efficiency between LNP1 and LNP9 in the 2D transwell model.

Fig. 2. Evaluation of LNP BBB Penetration and Brain Parenchymal Distribution Capability.

(a) Chemical structure of LNP1, LNP6 and LNP9. (b) Illustration of 3D BBB in vitro model for LNP BBB penetration assessment. (c) Quantification of normalized EGFP intensity per cell across PBS, LNP1, LNP6, and LNP9 treated conditions in the parenchyma scaffold, based on ImageJ analysis. (d-g) Confocal images of the 3D triculture brain scaffold 24 h after apical chamber administration of 40uL of LNP1 (e), LNP6 (f), LNP9 (g) carrying EGFP mRNA, and PBS control (d). EGFP expression and TUJ1 immunostaining signal are detected. (DAPI, blue; EGFP, green; TUJ1, red). (h-j) Time-course evaluation of 20uL LNP1-EGFP-mRNA transfection at 6h (h), 24h (i), and 48h (j), showing progressive increase in EGFP signal. (DAPI, blue; EGFP, green; TUJ1, red). (k) Quantification of time-dependent EGFP expression in the parenchyma scaffold after treatment with 20ul of LNP1-EGFP-mRNA for 6-, 24-, and 48 h using ImageJ. (l-n) Dose-dependent transfection visualized by EGFP intensity in brain parenchyma 24 h after treatment with 10uL (l), 20uL (m), or 40uL (n) LNP1-EGFP-mRNA. (DAPI, blue; EGFP, green; TUJ1, red). (o) Quantification of dose-response in the parenchyma scaffold following 10uL, 20uL, and 40uL of LNP1-EGFP-mRNA for 24 h, normalized EGFP intensity per cell count using ImageJ. (p). Absorbance values for Lactate dehydrogenase (LDH) assay performed on apical compartment media to assess cytotoxicity 24 h post-treatment with LNP1, LNP6, LNP 9, or lysis buffer (positive control). (Data represent mean ± s.d. of at least n = 3 for each condition, *p < 0.05, **p < 0.01, ***p < 0.005, and ****p < 0.001. Statistical analyses were performed using the one-way ANOVA and Tukey’s multiple comparison test).

However, when the same LNPs were assessed in our 3D BBB tissue model, significant differences emerged (Fig. 2c–g). LNP1 demonstrated robust EGFP expression throughout the parenchyma scaffold, indicating efficient BBB crossing and penetration capability (Fig. 2e). In contrast, LNP9 demonstrated only approximately 15% of the EGFP intensity observed with LNP1 (Fig. 2g), while LNP6 exhibited no detectable EGFP signal (Fig. 2f), consistent with its expected lack of transfection efficiency in neural tissues. These findings emphasize the ability of the 3D BBB model to distinguish functional differences in LNP performance that are not captured in the simpler 2D models (Fig. S2a–d).

To better characterize the predictive capability of this 3D BBB tissue model, we conducted time-course and dose-response analysis using LNP1. Confocal microscopy revealed a clear dose-dependent increase in EGFP expression within the brain parenchyma at 24 h post-treatment, with higher fluorescence intensity observed in the 40uL group compared to 10uL and 20uL (Fig. 2l–o). The time course analysis showed that no EGFP signal was detectable at 6 h post-treatment. However, fluorescence signal appeared at 24 h and peaked at 48 h, suggesting continuous LNP penetration and mRNA expression within the 3D matrix over time (Fig. 2h–k). The quantification data revealed that LNP1 achieved a 6-fold higher transfection efficiency compared to LNP9 at 24 h (Fig. 2c), confirming superior delivery capability. Importantly, the 3D BBB tissue model responded to both dose and time variables, providing a robust platform to screen a range of LNP formulations. The positive correlation between LNP dose and EGFP intensity, as well as the time-dependent increase in signal, demonstrated the model’s sensitivity for high-resolution delivery research (Fig. 2k–o).

Notably, this BBB tissue model also enabled the evaluation of LNP safety. As a measure of cell damage, lactate dehydrogenase (LDH) activity in the apical chamber media was quantified 24 h post-LNP treatment to measure the activity of cytoplasmic enzymes released by damaged cells. The results indicated that LNP1 induced minimal cytotoxicity, with LDH levels corresponding to approximately 7% of maximal cytotoxicity (lysis buffer-treated cells). In contrast, LNP6 triggered a 13-fold increase in LDH release relative to LNP1, and LNP9 induced a 7-fold increase, indicating higher cytotoxicity (Fig. 2p). These data support the safety profile of LNP1 and further verify its selection for further mechanistic characterization and in vivo validation.

Taken together, these results highlight that our 3D BBB tissue model offers a significant advantage over conventional 2D models for evaluating CNS drug delivery. Unlike HEK293T cell monolayers and 2D transwell, the compartmentalized structure allowed independent quantification of LNP delivery across both barrier and neural parenchyma simultaneously, which provides more accurate predictions about LNP BBB transport efficiency, brain parenchyma penetration, dose-/time-responsiveness, and safety assessment.

In addition to reducing the need for early-stage animal usage and addressing ethical issues, this BBB tissue model enables rapid analysis of LNP formulations in a human-relevant tissue context. By recapitulating the main physiological barriers, cellular compositions, and tissue architecture, it bridges the gap between conventional 2D in vitro screening and in vivo drug development research. Among the candidates tested, LNP1 consistently exhibited the highest EGFP expression, strongest parenchyma penetration capability, and lowest cytotoxicity, identifying it as the most promising CNS delivery vehicle candidate within the test library assessed herein. Based on its performance in our 3D BBB model, LNP1 was selected as the lead formulation for subsequent mechanistic research, in vivo validation, and future therapeutic development.

2.3. SR-B1 and INSR dependent LNP1 transcytosis across the blood-brain barrier

Following the identification of LNP1 as a lead formulation with superior BBB transcytosis and brain parenchyma penetration in our 3D BBB tissue model (Fig. 2), we next investigated the underlying mechanism that enables the superior transport performance relative to other LNPs in the library. Considering the highly selective nature of BBB and the limited permeability of most LNPs, we hypothesized that LNP1 utilizes receptor-dependent transcytosis pathways to facilitate its efficient transport across the endothelial barrier (Fig. 3a). To investigate this hypothesis, we compiled a panel of candidate receptors previously reported to mediate transcytosis of LNPs across the BBB, including transferrin receptor (TfR), scavenger receptor class B type I (SRB1), Low-Density Lipoprotein receptor (LDLR), and insulin receptor (INSR) [13, 33]. These receptors were selected based on their relevance to the lipid composition of LNP1, LNP6, and LNP9, which consist primarily of ionizable lipid, cholesterol, DOPC, and DMG-PEG in varying ratios [34] (Fig. 2a).

Fig. 3. Investigating the Transport Mechanisms of LNPs Across the BBB.

(a) Illustration of intravenous LNP administration proposed BBB-crossing mechanism. (b-d) Receptor-blocking experiments to evaluate LNP transport mechanisms. The apical BBB insert was pre-treated with antibodies or small molecules targeting key receptors, transferrin receptor (TfR), scavenger receptor class B type 1 (SRB1), low-density lipoprotein receptor (LDLR), and insulin receptor (INSR), to inhibit receptor-mediated transcytosis. Following receptor blockade, the BBB model was treated with LNP1 (b), LNP6 (c), or LNP9 (d), and flow cytometry was used to assess changes in LNP penetration and distribution in the basolateral (brain parenchyma) compartment. (e) RT-qPCR analysis of SRB1, INSR, and PECAM1 expression following siRNA-mediated knockdown of SRB1. (f) Chemiluminescent protein quantification of GAPDH, INSR, and PECAM1 post-siRNA-SRB1 treatment. (g) RT-qPCR analysis of SRB1, INSR, and PECAM1 expression following siRNA-mediated knockdown of INSR. (h) Chemiluminescent protein quantification of GAPDH, INSR, and PECAM1 after siRNA-INSR treatment. (i) Flow cytometry analysis of LNP1 transfection efficiency in the BBB insert following SRB1 and INSR receptor blocking with siRNA treatment or small molecule and antibody treatment. (j) Flow cytometry analysis of LNP1 transfection efficiency in the basolateral brain parenchyma compartment after SRB1 and INSR receptor blockade with siRNA treatment or small molecule and antibody treatment. (k) Immunofluorescence staining and confocal imaging of SRB1 protein expression on the surface of BBB insert (DAPI, blue; SRB1, green). (l) SRB1 protein expression on BBB insert after siRNA-SRB1 treatment, visualized by immunostaining and confocal imaging (DAPI, blue; SRB1, green). (m) Immunofluorescence staining and confocal imaging of INSR protein expression on the surface of BBB insert (DAPI, blue; INSR, green). (n) INSR protein expression on BBB insert after siRNA-INSR treatment, visualized by immunostaining and confocal imaging (DAPI, blue; INSR, green). (Data represent mean ± s.d. of at least n = 3 for each condition, *p < 0.05, **p < 0.01, ***p < 0.005, and ****p < 0.001. Statistical analyses were performed using the one-way ANOVA and Tukey’s multiple comparison test).

To investigate the receptor involvement, we used monoclonal antibodies to selectively block TfR, LDLR, and INSR, respectively, and employed the small molecule inhibitor BLT-1 to inhibit SRB1 [16]. Following receptor inhibition, LNP1, LNP6, and LNP9 encapsulating EGFP mRNA were administered to the apical compartment of the 2D five-cell BBB system for 24 h, and EGFP fluorescence in the basolateral chamber was quantified by flow cytometry to assess transfection efficiency. Among the receptors tested, blocking either SRB1 or INSR led to a significant reduction of the EGFP signal for both LNP1 and LNP9 treatment (Fig. 3b–d), indicating the involvement of both receptors in the transcytosis of these two LNPs. As expected, LNP6 showed minimal EGFP expression and was not significantly affected by receptor blocking (Fig. 3c), consistent with its poor delivery profile.

To further confirm the specificity of receptor-dependent transcytosis, we conducted siRNA treatment to knockdown SRB1 and INSR expression in the endothelial layer. Knockdowns were performed for 48 h before LNP treatment, and chemiluminescence electrophoresis analysis confirmed a significant reduction in SRB1 and INSR protein levels, without disturbing the tight junction marker PECAM1 expression (Fig. 3e–h). RT-qPCR showed reduced SRB1 mRNA expression, but INSR knockdown led to a paradoxical upregulation of INSR mRNA despite effective suppression of the protein (Fig. 3g and h). This suggests a potential compensatory feedback mechanism, as siRNA-mediated silencing is transient, with mRNA levels capable of rebounding while protein levels remain suppressed due to slower turnover [35]. Immunofluorescence staining further validated the reduction in SRB1 and INSR expression in siRNA-treated transwell inserts compared to nontreated controls (Fig. 3k–n). To determine the functional consequences of receptor knockdown and affected LNP1 transcytosis, we quantified the EGFP signal in both apical and basolateral compartments following treatment with LNP1-EGFP-mRNA. Blocking INSR via antibody or siRNA reduced endothelial uptake by ~30% and decreased basolateral delivery by 50–70% (Fig. 3j). Similarly, SRB1 knockdown by siRNA reduced basolateral transfection by approximately 50%, without affecting apical uptake. BLT-1 treatment further suppressed parenchymal delivery, reducing the parenchymal EGFP signal to ~20% of control (Fig. 3j) and BBB insert signal to ~40% (Fig. 3i), underscoring the role of SRB1 in mediating efficient transcytosis.

These results indicated that both SRB1 and INSR are critical for LNP1 uptake and transcytosis and play a particular role in mediating parenchyma delivery. The different effects of chemical inhibition versus siRNA knockdown may reflect compensatory regulation of alternative receptor transport pathways or receptor crosstalk. Notably, the reduction of EGFP signal in the basolateral population suggests that LNP1 transcytosis impacts not only endothelial uptake but also functional delivery to downstream neural cells. This level of resolution is key to adequate screening, where both facets of delivery are considered within the BBB tissue model.

In summary, our data revealed that LNP1 utilized SRB1 and INSR receptor-dependent transcytosis for efficient BBB transcytosis in the in vitro model. These findings provide mechanistic insights into the superior delivery performance of LNP1 relative to other formulations in the library and provide a potential direction for further receptor-dependent LNP optimization.

2.4. In vivo validation of LNP1 brain delivery and distribution

After demonstrating the superior BBB penetration and brain parenchyma delivery of LNP1 in our 3D in vitro tissue model (Fig. 2) and identifying SRB1 and INSR as key mediators of its receptor-dependent transcytosis mechanism (Fig. 3), we next evaluated whether these in vitro findings translated into in vivo performance. Specifically, we aimed to validate the predictive power of our 3D BBB in vitro tissue model by evaluating the organ-level distribution and brain penetration efficiency of LNP1 in mouse models.

To assess the in vitro-in vivo correlation, we compared the in vivo delivery performance of LNPs that were previously examined in the 3D BBB model. In vitro, LNP1 demonstrated the highest EGFP expression in the brain parenchyma compartment, while LNP9 showed approximately 15% of LNP1’s delivery capacity, and LNP6 showed a minimum signal. These differences provided a benchmark for evaluating translational accuracy. We intravenously administered BALB/c mice with single or double dose of LNP1 or LNP9 containing firefly luciferase (Fluc) mRNA, and luciferase expression was monitored by in vivo imaging (IVIS) and quantified by RT–qPCR (Fig. 4a). Robust bioluminescence was detected in the head region and perfused brain of LNP1-treated mice, whereas LNP9-treated animals exhibited signals that were ~25-fold lower, approaching PBS control levels (Fig. 4d and e). LNP1 has also shown markedly higher luminescence signal intensity in peripheral organs compared with LNP9 (Fig. 4f). These findings demonstrated that LNP1 showed strong distribution in vivo, consistent with the findings from our 3D in vitro tissue model.

Fig. 4. In Vivo Biodistribution and Neuronal Transfection by LNP1-mRNA Formulations.

(a) Schematic overview of intravenous LNP administration and tissue processing workflow for LNP distribution assessment. (b) Schematic illustrating CRE recombinase mRNA-mediated activation of tdTomato expression in Ai14 reporter mice. (c) RT-qPCR quantification of Fluc mRNA level in perfused brain following single or double intravenous injections of LNP1-Fluc-mRNA and LNP9-FlucmRNA, demonstrating a 121-fold increase in brain accumulation with LNP1 double dosing. (d) IVIS bioluminescence imaging of the head region and perfused brains from BALB/c mice 3 h after double intravenous injection of PBS, LNP1-Fluc-mRNA, and LNP9-Fluc-mRNA, demonstrating a stronger brain signal in the LNP1-treated group. (e) Normalized luminescence intensity of the brains of mice intravenously injected with PBS, LNP1-Fluc-mRNA, and LNP9-Fluc-mRNA. The intensity was normalized to the LNP9 group. (f) Luminescence quantifications of the organs from mice treated with PBS, LNP1-Fluc-mRNA, and LNP9-Fluc-mRNA. (g) Confocal microscopy of perfused brain sections (hippocampus and pons) from Ai14 transgenic mice 7 days post-treatment with double intravenous injection of LNP1-CRE-mRNA or PBS control. Strong tdTomato fluorescence was observed in NeuN + neurons, confirming BBB penetration and neuronal transfection in vivo. (Data represent mean ± s.d. of at least n = 3 for each condition, *p < 0.05, **p < 0.01, ***p < 0.005, and ****p < 0.001. Statistical analyses were performed using the one-way ANOVA and Tukey’s multiple comparison test).

To further quantify tissue-specific accumulation and assess dose-response, we administered either a single or double IV doses of LNP1-Fluc-mRNA and LNP9-Fluc-mRNA and analyzed mRNA expression across major organs using RT-qPCR. Luciferase mRNA expression in the LNP1-dosed brain was 121-fold higher in double-dosed animals compared to single-dosed animals (Fig. 4c), while the expression in other peripheral organs, including the liver, spleen, kidney, and lung, was approximately less than 10-fold higher (Figs. S3a,c,e,g). As expected, LNP9 showed only low and limited increases in brain accumulation upon repeat dosing. These findings suggest that the initial dose may serve to prime the BBB, enhancing endothelial cell engagement and facilitating more efficient LNP1 brain penetration during the second administration. This finding highlights a potentially novel LNP delivery strategy: priming the BBB with an initial dose to improve subsequent CNS distribution. The observed dose-dependent enhancement further highlights the relevance of LNP1 delivery and its potential for robust CNS penetration under optimized dosing conditions.

To further investigate cellular localization and functional transgene delivery to neurons, we performed an in vivo study with Ai14 transgenic mice (Fig. 4a and b). The mice were intravenously administered with single or double LNP1-CRE-mRNA or LNP9-CRE-mRNA, and tdTomato expression was evaluated using IVIS seven days post-injection. LNP1-treated Ai14 mice exhibited tdTomato signal in the brain (Fig. 4g), while LNP9-treated animals displayed a weaker signal in both CNS and peripheral distribution (Fig. S3k). Both LNP1 and LNP9 showed strong tdTomato signal in the liver, while LNP1 showed stronger distribution in brain, spleen, lung, and kidney compared to LNP9.

In fixed brain sections from LNP1-treated mice, we observed tdTomato expression in the choroid plexus, suggesting LNP delivery to the blood-CSF barrier, and we observed tdTomato signal with vascular-like morphology in the hippocampus, pons, and midbrain (Fig. S3i–j), indicating potential transport and distribution routes before BBB transcytosis. Importantly, confocal imaging of perfused brain sections from LNP1-treated Ai14 mice revealed colocalization of tdTomato signal with the NeuN + nuclei in both the hippocampus and pons regions (Fig. 4g). These data confirmed that LNP1 not only reaches the brain but also successfully delivers mRNA to neurons, enabling functional gene expression. This is a critical distinction, as many delivery systems fail to achieve post-BBB neuronal targeting.

Taken together, these in vivo findings supported the predictive validity of our 3D in vitro BBB tissue model. The consistency between in vitro and in vivo data, including both organ-level distribution and cell-type resolution, demonstrated that our BBB tissue model captures critical aspects of LNP transport and can reliably predict the CNS delivery potential. Moreover, the ability of LNP1 to cross the BBB, transfect neurons, and show dose-response distribution demonstrates its potential as a lead formulation for therapeutic development. These results indicate the translational relevance of physiologically accurate in vitro 3D BBB tissue models and establish a strong foundation for their use in preclinical LNP screening and optimization for CNS-targeted drug delivery.

3. Discussion

Robust in vitro models of the human BBB are critical for preclinical drug testing, bridging the in vitro–in vivo correlation gap, and guiding the development of CNS-targeted therapeutics. Previous silk-based models developed by our group have progressively advanced the integration of NVU components within a physiologically relevant matrix. Earlier work incorporated neurons, astrocytes, and microglia into 3D silk scaffolds to explore neural tissue functionality [27–30,36,37], while more recent efforts co-cultured endothelial cells with astrocytes and pericytes to generate vessel-like structures and assess NVU mechanics in brain inflammation contexts [38]. However, these systems lacked a compartmentalized architecture and were not optimized for evaluating functional barrier properties or directional molecular transport.

To address these limitations, we developed a compartmentalized, human-derived 3D BBB–NVU tissue model that represents a significant advancement in the design of physiologically relevant platforms for CNS drug delivery. By integrating key neurovascular unit components within a silk-collagen scaffold and resolving both the apical endothelial barrier and basolateral brain parenchyma, this model overcomes critical limitations of existing in vitro systems. Unlike organoids and spheroids [22,25], which often lack spatial separation between the vascular and the parenchymal compartments, our model enables compartment-specific permeability measurement and localized evaluation of therapeutic delivery. This spatial resolution is essential for capturing directional transport and evaluating parenchymal penetration, which are not accessible in spheroid-based systems. The model also addresses the scalability and reproducibility challenges associated with iPSC-derived systems. The use of immortalized human brain microvascular endothelial cells (IM-HBMECs) [16,20,39], human-induced neural stem cells (hiNSCs) [36–38,40], primary astrocytes, pericytes, and a microglia cell line allows for consistent tissue generation and long-term culture stability. In contrast to microfluidic BBB chips [21,26], which often require specialized equipment and suffer from limited throughput, our transwell-compatible format supports high-content screening and batch production of mature neural tissues. This compatibility with standard formats enhances accessibility and positions the model for translational applications.

Barrier integrity was confirmed through elevated TEER values, continuous tight junction morphology, and reduced FITC-dextran permeability. Importantly, the 3D tissue model also supported a more homeostatic environment, with significantly lower levels of MMP-1 and Ang-2 secretion than the 2D system. In addition, the silk-collagen composite scaffold supported long-term culture for up to 8 months and promoted an extended neurite network within the mature neural compartment, leading to sustained neuronal functional gene expression, enabling extended studies without repeated seeding. The mechanical properties of the silk-collagen composite scaffold (Young’s modulus ~0.4 kPa) fall within the range of native brain tissue (Young’s modulus 0.1–1 kPa) and support long-term neuronal health and BBB maturation [30,38]. These findings highlight the ability of the model to support BBB maturation and function. While previously reported transwell-based systems [41–43] demonstrated basic barrier formation, they plateaued at low TEER values and exhibited weak junctional staining, limiting their predictive power. Similarly, recent 3D bioprinted BBB constructs [44] and hydrogel-based 3D NVU platforms [45] have achieved structural complexity but lack therapeutic validation or in vivo correlation. Our model fills this gap by combining physiological fidelity with functional readouts that extend beyond permeability to capture downstream delivery and transfection. Importantly, this platform represents a baseline human-derived neurovascular model under homeostatic conditions and does not yet incorporate disease-associated BBB remodeling. Alterations in BBB integrity and transporter expression under pathological states such as Alzheimer’s disease, neuroinflammation, or Parkinson’s disease may influence nanoparticle transport and should be considered in future model extensions.

The predictive capacity of the model was demonstrated through the screening of lipid nanoparticle (LNP) formulations. LNP1, an unmodified formulation, showed robust transfection in the parenchymal compartment and strong in vivo correlation following systemic administration in mice. Compared to Han et al.’s high-throughput BBB platform [42], where the lead LNP achieved ~4 × 10⁴ p/s/cm²/sr luminescence in the mouse brain, our LNP1 produced signals in the range of 2.02 × 10⁵ to 5.33 × 10⁵ p/s/cm²/sr, which is nearly an order of magnitude higher. Moreover, in relation to other recently reported BBB-crossing LNP formulations [46] that yielded brain luminescence within the 10⁵–10⁶ p/s/cm²/sr range, our LNP1 performed comparably, underscoring both its translational potential and the predictive power of our model. Notably, when double-dosed, LNP1-Fluc-mRNA yielded a 17-fold increase in luciferase mRNA distribution in the perfused brain and a 25-fold increase in normalized luminescence intensity compared to the double-dosed LNP9 formulation, further demonstrating its superior brain-penetrant capacity. Importantly, LNP1 achieved this performance without surface modifications, highlighting the intrinsic delivery capability of the formulation and the sensitivity of our model to detect meaningful differences in CNS penetration.

Mechanistic studies further revealed that LNP1 utilizes multiple receptor-dependent transcytosis pathways, with SR-B1 and INSR identified as key facilitators. Pharmacological inhibition and siRNA knockdown of either receptor significantly reduced LNP1 uptake in both endothelial and parenchymal compartments. These findings support a model of composition-dependent, receptor-facilitated transport rather than direct ligand-mediated receptor targeting, suggesting that lipid composition may enable interaction with endogenous endothelial transport pathways. This dual-pathway transport mechanism underscores the model’s ability to dissect complex delivery routes and supports the rational design of next-generation nanomedicines. While a previous microfluidic BBB model observed a decreased LNP fluorescence signal in the vascular channel upon SR-B1 inhibition, it did not observe a reduction in the perivascular channel or quantify the cellular uptake of LNPs in parenchymal tissue [16]. In contrast, our study enabled the quantification of both endothelial and downstream parenchymal delivery, providing a more complete view of the transport process. This distinction is critical for evaluating therapeutic modalities such as LNP-mRNA, where parenchymal transfection is the ultimate determinant of efficacy.

While our tissue model offers significant advantages, there are several limitations that may be addressed in future studies. The TEER values of our IM-HBMECs layers are relatively low compared to iPSC-derived models but are consistent with previously reported ranges for immortalized endothelial cells. Nevertheless, the inclusion of all NVU cellular components and a 3D brain tissue scaffold significantly enhanced tight junction integrity, leading to improved TEER and reduced permeability over time. In line with previous findings [47], FITC-dextran permeability was inversely correlated with TEER in immortalized endothelial models at lower resistance levels (10–50 Ω cm²). However, in iPSC-derived BBB systems, increasing TEER did not correspond to reduced permeability, suggesting that extremely high TEER may not be essential for evaluating macromolecular transport, such as LNP delivery across the BBB. In addition, the current in vitro configuration does not fully capture the dynamic formation of protein corona that occurs upon systemic exposure in vivo, which can influence nanoparticle biodistribution and cellular interactions. This limitation is partially addressed by complementary in vivo studies, and future work will incorporate serum-conditioned nanoparticles and perfusion-based systems to better recapitulate physiological exposure conditions. The absence of physiological shear stress remains a common limitation in transwell-based BBB systems. A recent study [48] demonstrated that applying shear stress using an Annular Shaking Dish (AS) increased the TEER from 20 to 50 Ω cm² and upregulated tight junction protein expression. However, this approach requires specialized equipment and setup. While incorporating shear may improve barrier function in our model system, further evaluation is needed to determine whether the possible benefits justify the added complexity of the shear system. Another practical consideration arises in the mechanistic studies. The dense cellular network of the 3D scaffold makes viable single-cell dissociation technically challenging for flow cytometry. To overcome this, we employed a 2D BBB model, preserving cellular composition while enabling robust apical and parenchymal readouts. Future directions include developing optimized dissociation protocols, applying RT-qPCR or fluorescent labeling (e.g., DiR), and leveraging imaging-based analyses to strengthen mechanistic interrogation directly in 3D. Scalability and high-throughput adaptation are critical for translational drug screening. While our current model uses a standard 24-well format, it is adaptable to a 96-well format for higher throughput applications [42]. In addition, our 3D brain tissue scaffold manufacturing process allows for the generation of approximately 300 brain tissue scaffolds per batch. Importantly, seeded scaffolds maintain a mature phenotype and functional stability for at least 8 months, enabling long-term experiments and reducing variability from repeated seeding [28,49]. Finally, LNP1 remains unmodified in this study. Optimization of LNP1 represents one of the future directions to enhance CNS delivery. Incorporating surface modifications, such as targeting ligands, peptides, or antibodies [50], may improve delivery specificity and broaden the therapeutic potential, especially when combined with the double-dosing strategy that has shown promising results in our mouse model. Building on the identification of SRB1 and INSR as the facilitators of LNP transcytosis in vitro, future studies may explore in vivo validation of these pathways by dosing mice with receptor-targeted, antibody-conjugated LNP1 or by using a receptor knockdown or knockout animal model to evaluate barrier crossing capability and parenchyma distribution. We further recognize that direct in vivo receptor perturbation represents an important next step to validate receptor-dependent transport mechanisms in a physiologically complex environment, and approaches such as pharmacological inhibition or conditional endothelial targeting may provide deeper mechanistic insight. These efforts would further elucidate LNP transport mechanisms and support the rational design of next-generation CNS targeting nanomedicines.

In conclusion, our 3D BBB–NVU model offers a distinct advantage over existing platforms by integrating structural complexity with functional relevance, enabling directional transport studies, parenchymal delivery quantification, and long-term tissue stability. In this study, LNP1 emerged as a promising lead candidate for CNS-targeted mRNA therapeutics, achieving robust brain transfection without modification and demonstrating strong in vitro–in vivo correlation. Together, the model and LNP formulation establish a versatile and predictive foundation for translational research in CNS drug delivery, disease modeling, and high-throughput screening.

4. Materials and methods

4.1. Transwell insert triculture construction

Cell culture and 3D brain-like triculture tissue model fabrication are described in detail in the SI Appendix. To ensure a consistent description of the co-culture and integration process, we defined Day 0 as the day when IM-HBMECs were seeded into the transwell insert. The days prior to endothelial seeding are referred to as Day −1, Day −2, and Day −3, during which supporting glial and pericyte cell types were seeded sequentially on the basolateral side. Day 1 refers to the day when the transwell insert were integrated with either a 2D or a 3D brain tissue scaffold in the basolateral chamber to assemble the complete BBB model (Fig. S4a). On Day −3, 6.5 mm Transwell^® inserts with 0.4 μm Pore Polycarbonate Membrane (Cat# 3413, Corning) were pre-coated to facilitate cell adhesion. The apical membrane surface was coated with 25 μg/mL fibronectin. The underside of the insert and the corresponding well surface were coated with 2 μg/cm² poly-L-lysine for 1 h in a humidified incubator at 37 °C with 5% CO₂ and subsequently rinsed once with sterile PBS to remove the excessive residues. Transwell inserts were inverted and 60 μL of primary human astrocyte suspension (6.25 × 10⁴ cells in astrocyte medium) were carefully pipetted onto the basolateral side of each insert. The inverted plate was returned to the incubator for at least 2 h to allow cell attachment. Following the incubation, inserts were returned to the upright position, and the astrocyte medium was added to both basolateral and apical chambers. Plates were incubated overnight at 37 °C with 5% CO₂ to support initial astrocyte growth. On Day −2, the transwell plate was removed from the incubator and the astrocyte medium was gently aspirated from both chambers. Transwell inserts were then inverted and 60 μL of human brain vascular pericyte suspension (1.25 × 10⁴ cells in pericyte medium) were carefully pipetted onto the basolateral side of each insert. This seeding density corresponded to an astrocyte-to-pericyte ratio of 5:1. The inverted plate was returned to the incubator for at least 2 h to allow cell attachment. Following incubation, inserts were returned to the upright position, and 600 μL of a 5:1 volumetric mixture of astrocyte and pericyte medium was added to both basolateral and apical chambers. Plates were incubated overnight at 37 °C with 5% CO₂ to support astrocyte-pericyte co-culture establishment. On Day 0, once astrocyte and pericyte co-culture reached approximately 90% confluency, the mixed medium in the apical chamber was gently removed. 100 μL of IM-HBMECs suspension (5 × 10⁴ cells in endothelial medium) was carefully seeded into the apical chamber of each insert. Inserts were then returned to the incubator for at least 5 h to allow the cell attachment to the membrane. Following attachment, the apical chamber was topped up with endothelial medium to a final volume of 200 μL, and the plates were incubated overnight at 37 °C with 5% CO₂ to establish the tri-culture BBB insert. On Day 1, the established BBB insert was integrated with either a 2D triculture brain parenchyma layer or a 3D triculture brain tissue scaffold to assemble the 2D or 3D BBB model, respectively (Fig. S4a).

4.2. 2D and 3D BBB model assembly and CDK4/6 inhibitor treatment

BBB models were constructed using either a 2D triculture layer or a 3D brain tissue scaffold placed in the basolateral chamber of a 6.5 mm Transwell^® system. On Day −3, the basolateral chamber wells were coated with 2 μg/cm² poly-L-lysine for 1 h in a humidified incubator at 37 °C with 5% CO₂, followed by a single sterile PBS rinsing to remove the excess reagents. For the 2D model, a triculture cell suspension composed of hiNSCs, human astrocytes, and HMC3 at a ratio of 20:5:1 (total 7 × 10⁴ cells in 1,000 μL volume) was seeded directly into each basolateral well. For the 3D model, previously prepared triculture silk scaffolds were placed into each pre-coated basolateral well. In both setups, cultures were maintained in the triculture medium, consisting of Neurobasal medium supplemented with 1% Antibiotic-Antimycotic, 1% GlutaMAX, 2% B-27, and 1% Astrocyte growth supplement. The medium was refreshed every other day. On Day 1, the previously established triculture BBB insert was integrated to each well to complete the BBB model. The basolateral chamber medium was switched to a mixed formulation (astrocyte:pericyte:triculture medium at a ratio of 5:1:6) and 600 μL of the mixture was added to each well every other day to maintain dual-compartment culture conditions. Beginning on Day 2, CDK4/6 inhibitor treatment was initiated in the designated treatment group. PD 0332991 isethionate (Cat# 4786, Tocris Bioscience^™) was freshly prepared in endothelial medium at a final concentration of 600 nM and added to the BBB insert apical chamber (200 μL per insert). Untreated control groups received endothelial medium only. Medium containing PD 0332991 or vehicle was refreshed on Day 3 and Day 5 during routine medium changes. On Day 4 and Day7, endpoint assays were performed, including transendothelial electrical resistance (TEER) measurement, FITC-dextran permeability assays, and sample collection for RT-qPCR and immunofluorescence staining, as described in the following sections.

4.3. RNA isolation and sample preparation

Total RNA was extracted using RNAdvance Cell V2 Kit (Cat# A47942, Beckman Coulter) according to the manufacturer’s instructions, with minor optimization for both 2D monolayer, 3D scaffold and mouse tissue samples. For 2D culture, culture medium was aspirated, and cells were rinsed once with PBS. A total of 200 μL of freshly prepared lysis buffer containing Proteinase K (50 mg/mL) was added directly to each well. In vitro cell Lysates were then incubated for 20 min at room temperature to complete the cell disruption and protein digestion. For 3D triculture scaffold samples, 200 μL lysis buffer was added directly onto the scaffold, followed by mechanical homogenization using a 5 mm stainless steel bead in 2010 Geno/grinder (Cat# S6006, Spex, Sample-Prep) at 1000 rpm for two 30-s cycles to ensure efficient disruption. Lysates were then incubated at 37 °C for 60 min to complete the cell disruption and protein digestion. For mouse tissue samples, RNA was isolated using RNAdvance Tissue Total RNA Isolation Kit (Cat# A32646, Beckman Coulter). Tissues were lysed by 200 μL of freshly prepared lysis buffer containing Proteinase K (50 mg/mL) and then homogenized using a 5 mm bead at 1000 rpm for five 30-s cycles to ensure efficient tissue disruption. Lysates were then incubated for 60 min at 37 °C to complete the cell disruption and protein digestion. Phenol: chloroform: isoamyl alcohol mixture (Cat# 77617, Sigma-Aldrich) was used for purification to remove excess lipids and debris from the tissue sample, before magnetic bead binding. All lysates were then transferred to a 96-well plate (Cat# AB0932, Thermo Fisher Scientific) compatible with the magnetic bead separation. Total RNA was bound to paramagnetic beads by adding 200 μL of pre-mixed Bind-BBC and 100% isopropanol, followed by gentle mixing and incubation at room temperature for 5 min. The RNA-bound beads were washed subsequently with wash buffer and 70% ethanol. To remove the genomic DNA, samples were treated with DNase I solution (Cat# AM2224, Ambion) for 30 min at 37 °C. RNA was rebound to the beads and washed again. Magnetic beads with RNA were air-dried for 10 min. Finally, RNA was eluted in 40 μL nuclease-free water and transferred to a 96-well plate. RNA concentration, integrity, and purity were assessed using a NanoDrop 8000 spectrophotometer. Samples were stored at −80 °C until further use.

4.4. Real-time quantitative PCR (RT-qPCR)

Gene expression was quantified by RT-qPCR analysis. A total of 13 ng of RNA per sample was reverse transcribed into complementary DNA (cDNA) using SuperScript^™ IV VILO^™ Master Mix kit (Cat# 11756500, Invitrogen^™) on a Proflex 96-well PCR system (Cat# 4484075, Thermo Fisher Scientific). To analyze the gene expression, quantitative PCR was performed using TaqMan^™ Gene Expression Assays (Table S1) on a QuantStudio^™ 7 Pro Real-Time PCR System (Applied Biosystems) following the manufacturer’s protocol. Gene expression levels were normalized against the housekeeping gene GAPDH. All reactions were performed in technical triplicates. Relative gene expression was calculated using the ΔΔCt method below:

$$\mathit{\text{Relative}}\mathit{\text{Gene}}\mathit{\text{Expression}}=2^{{-\Delta \Delta C}_t}$$

Where:

$${\Delta C}_t=C_{t,\mathit{\text{target}}\mathit{\text{gene}}}-C_{t,\mathit{\text{GAPDH}}}$$

$$\Delta \Delta \text{Ct}={\Delta C}_{t,\mathit{\text{sample}}}-{\Delta C}_{t,\mathit{\text{GAPDH}}}$$

4.5. Immunofluorescence staining

Cell cultured in 2D or 3D triculture scaffolds were fixed with 4% paraformaldehyde (PFA, Cat# 15713S, Electron Microscopy Sciences) + 4% sucrose (Cat# S5–500, Fisher Scientific) for 1 h at room temperature. Following fixation, samples were permeabilized with 0.02% Triton X-100 (Cat# X100–100 ML, Sigma-Aldrich) in DPBS for 10 min. Non-specific binding was blocked by incubating samples in blocking buffer containing 4% normal goat serum (Cat# 50062Z, ThermoFisher Scientific) in 0.02% PBST for 1 h at room temperature. Samples then were incubated overnight at 4 °C with the primary antibodies diluted in blocking buffer: PECAM1 (1:1000, Abcam, ab9498) for IM-HBMECs, GFAP (1:500, Invitrogen, 13–0300) for human astrocyte, TUJ1 (1:500, Abcam, ab78078) for hiNSCs, NeuN (1:100, Cell Signaling, 24307) for post-mitotic neurons of the central nervous systems, SRB1 (1:1000, Invitrogen, PA1–16788) for scavenger receptor class B type I, INSR (1:10, Invitrogen, MA5–13778) for insulin receptor. The next day, samples were washed four times with DPBS for 10 min on an orbital shaker at room temperature. Fluorophore-conjugated secondary antibodies, diluted by blocking solution, were applied and incubated for 1.5 h at room temperature in the dark on an orbital shaker. Following secondary staining, samples were again washed four times with DPBS for 10 min each: Goat anti-Mouse IgG (H + L) Highly Cross-Adsorbed Secondary Antibody, Alexa Fluor^™ Plus 488 (1:500, Invitrogen, A32723), Goat anti-Rat IgG (H + L) Highly Cross-Adsorbed Secondary Antibody, Alexa Fluor^™ Plus 488 (1:500, Invitrogen, A48262), Goat anti-Mouse IgG (H + L) Highly Cross-Adsorbed Secondary Antibody, Alexa Fluor^™ Plus 647 (1:500, Invitrogen, A32728), Goat anti-Rabbit IgG (H + L) Cross-Adsorbed Secondary Antibody, Alexa Fluor^™ 488 ((1:500, Invitrogen, A-11008). Nuclei were counterstained with Hoechst 33342 (2’-[4-ethoxyphenyl]-5-[4-methyl-1-piperazinyl]-2,5′-bi-1H-benzimidazole) (Cat# R37605, Invitrogen), prepared at a dilution of 2 drops/mL in DI water. A total of 250 μL of the diluted stain was added to each sample and incubated 20min at room temperature in the dark prior to imaging. Images were acquired on a Leica TCS SP8 confocal microscope.

4.6. Image analysis

Confocal Imaging was performed using filter set for Hoechst (Ex/Em: 350/470 nm), GFP/FITC (Ex/Em: 488/514 nm), Texas Red (Ex/Em: 540/605 nm), and Alexa Fluor 647 (Ex/Em: 640/680 nm). All images within the same experiment were acquired using the same laser power and gain setting to ensure consistency. Maximum intensity projection images were generated using ImageJ. For quantification of EGFP transfection efficiency, z stacks images were captured at 1 μm intervals along the z-axis. Images were split into individual channels to isolate EGFP (green) and Hoechst (blue) signals. EGFP fluorescence intensity was quantified by measuring the integrated density in the green channel. To determine the cell number, the Hoechst channel was processed by applying an intensity threshold followed by binarization and watershed segmentation to separate adjacent cells. The number of Nuclei was quantified using the “Analyze Particles” function. EGFP signal was normalized to the number of Hoechst-positive nuclei to calculate the mean EGFP intensity per cell.

4.7. BBB model morphology quantification methods

All image analyses were performed using the Deep Learning module of AIVIA software (version 15.0, Leica Microsystems). Distinct analysis pipelines were optimized for different imaging contexts, including 2D IM-HBMECs, 2D BBB and 3D BBB groups, to ensure accurate segmentation of endothelial cells. A custom deep learning model was trained to generate binary masks delineating the cells of interest. Each detection was manually reviewed and fine-tuned to verify segmentation accuracy and ensure complete coverage of the targeted cellular regions.

Quantitative features such as segmented cell area and perimeter were extracted for each detected object. In cases where segmentation performance was suboptimal, classifier parameters were iteratively refined to maintain consistency and accuracy across the dataset.

All quantitative data were analyzed and visualized using GraphPad Prism (version 10.5). Statistical analyses were conducted using one-way ANOVA, followed by appropriate post hoc tests, as specified in the figure legends. Significance thresholds and exact n values are provided for each comparison.

4.8. Permeability measurement

On Day 4 and Day 7 after the 2D and 3D BBB model was integrated. 600 μL of fresh endothelial medium was added to a new 24 well plate, BBB inserts were transferred to the plate, 150 μL of the culture medium containing 1 mg/mL of 20 kDa FITC-dextran (Cat# FD20S, Sigma-Aldrich) was introduced into the apical chamber of the insert and incubated in dark in a humidified incubator at 37 °C with 5% CO₂. Sixty minutes later, stop permeation by removing the insert, the medium in the lower channel was collected and thoroughly mixed by pipetting. Then a 100 μL of medium from the receiver tray was collected and transferred to wells of the black 96-well opaque plate for fluorescence measurement. The FITC-dextran concentration was determined by FITC fluorescence intensity using a cytation5 plate reader (excitation/emission ~ 485nm/535 nm). The apparent permeability coefficient (P_app) was calculated according to the equation below:

$$P_{\mathit{\text{app}}}=\frac{V_l\times C_l}{A\times C_u\times t}$$

Where V_l (mL) indicates the volume 0.6 and C_l (ug/mL) indicates the concentration of the FITC-dextran in the basolateral chamber. A indicates the surface area of the transwell insert (0.33 cm²), C_u (ug/mL) is the FITC-dextran concentration on the apical insert, and t (s) represents the diffusion time of FITC-dextran (3600s). All samples were read in technical triplicate.

4.9. Transendothelial electrical resistance (TEER) measurement

On Days 4 and 7, Transendothelial Electrical Resistance (TEER) was measured using an STX4 electrode (Cat# EVM-EL-03–03-01, World Precision Instruments) and connected to EVOM [2] meter (WPI). The instrument applies a 12.5Hz alternative current (AC) square-wave voltage signal to assess barrier integrity across the transwell models. Prior to measurement, the electrode was sterilized by soaking in 70% ethanol for 15 min, followed by equilibration in pre-warmed cell culture medium. The EVOM [2] meter was set to OHMS mode, and the electrode was carefully inserted vertically into each transwell without disrupting the cell layer. After 10 s of stabilization, the resistance value was recorded. Each well was measured in triplicate, and the average resistance value was calculated. Final TEER values (Ω·cm²) were calculated using the following equation:

$$\mathit{\text{TEER}}=(R_{\mathit{\text{sample}}}-R_{\mathit{\text{blank}}})\times A$$

Where Rsample is the raw resistance value of each insert (Ω), Rblank is the resistance of an empty insert with medium only, and A is the surface area of the insert (0.33 cm²).

4.10. Cytokine profiling

Medium was collected from the basolateral chamber of 2D and 3D models 4 days post media change. Angiogenic cytokines were analyzed using the Human Angiogenesis Antibody Array C1 (Cat# AAH-ANG-1, RayBiotech) according to the manufacturer’s instructions. Chemiluminescent signals were imaged on X-ray film, and cytokine spot intensities were quantified using ImageJ software after background subtraction. Cytokine levels in the 2D condition were set to 100%, and relative expression in the 3D condition was calculated as a percentage of the 2D values.

4.11. LNP formulation and characterization

The lipidoids were synthesized as per our previous reports [34]. Briefly, the bioreducible lipids were synthesized by heating an amine and acrylate in Teflon-lined glass screw-top vials for 48 h, and the crude products were purified using flash chromatography. LNPs were prepared using a NanoAssemblr microfluidic system (Precision Nanosystems). Briefly, lipidoids, cholesterol (Sigma), phospholipids (DOPC, Avanti Polar Lipids), and methoxypolyethylene glycol (DMG-PEG2000) (Avanti Polar Lipids) were all dissolved in 100% ethanol at 10 mg/mL, and mixed at weight ratios of 16:4:2:1. Firefly luciferase (Fluc) mRNA (Cat# L-7202, TriLink Biotechnology) or Cre recombinase (Cre) mRNA (Cat# L-7211, TriLink Biotechnology) was encapsulated into LNPs for in vivo delivery, while enhanced green fluorescent protein (EGFP) mRNA (Cat# L-7201, TriLink Biotechnology) was encapsulated for in vitro studies. The hydrodynamic diameter and polydispersity index (PDI) of the LNPs were characterized by dynamic light scattering (DLS) using a Zetasizer Nano ZS (Malvern Instruments).

4.12. In vitro BBB LNP distribution with fluorescence imaging and flow cytometric analysis

On Day 4 following integration of the 2D or 3D BBB model, culture medium containing LNP-EGFP-mRNA was introduced to the apical chamber of each transwell insert and incubated at 37 °C for 24 h. After incubation, the insert and the 2D brain parenchyma samples were imaged using a KEYENCE fluorescence microscope to visualize the EGFP expression intensity. After imaging, cells were disassociated using 0.25% Trypsin-EDTA and neutralized with their respective culture medium. Single-cell suspension was analyzed using an Attune NxT Flow Cytometer (BD Biosciences) using a 488 nm laser and FITC detection filter (typically ~530/30 nm) to detect EGFP fluorescence. Forward scatter (FSC) and side scatter (SSC) voltage values were optimized and consistently used across the same experiment. The EGFP-positive gate was set based on PBS-treated control samples, ensuring a clear separation of positive and negative populations. At least 10,000 cells per sample were analyzed. Flow cytometry data were analyzed using FlowJo-v10. EGFP expression was reported as the mean fluorescence intensity (MFI) of the EGFP-positive population, normalized to the MFI of the untreated control group.

Following treatment with different LNP formulations, 3D triculture scaffolds in the BBB model were analyzed by confocal imaging using a Leica TCS SP8 microscope. The immunostaining, image acquisition, and analysis methods were described earlier.

4.13. LDH cytotoxicity assay

Cytotoxicity was measured by lactate dehydrogenase assay (LDH) (Cat# TOX7, Sigma-Aldrich) following the manufacturer’s protocol. Cell culture medium was collected 24 h after the LNP treatment. LDH released into the medium from the damaged cells was quantified by comparing with the lysis buffer-treated positive controls. LDH assay mixture was added to each sample and incubated for 20 min at room temperature in the dark. Absorbance was measured at 490 nm with background subtraction at 690 nm using a Cytation5 microplate reader (BioTEK). The percentage of cytotoxicity was calculated using the following equation:

$$\mathit{\text{Cytotoxicity}}(\%)=\frac{({\mathit{\text{LDH}}}_{\mathit{\text{sample}}}-{\mathit{\text{LDH}}}_{\mathit{\text{background}}})}{({\mathit{\text{LDH}}}_{p-\mathit{\text{ctrl}}}-{\mathit{\text{LDH}}}_{\mathit{\text{background}}})}\times 100$$

All measurements were performed in technical triplicates. Values were normalized against the maximum LDH release and reported as mean ± s.d.

4.14. Receptor blocking assay

To evaluate the role of receptor-dependent transcytosis in LNP transport across the BBB insert, both small molecule inhibitor and monoclonal antibody blocking strategies were employed to block the candidate receptors involved in the LNP RMT process. For scavenger receptor class B type I (SR-B1) inhibition, 200 μM block lipid transport-1 (BLT-1, Cat# SML0059–5 MG, Sigma-Aldrich) was added to the apical chamber and incubated at 37 °C for 2 h prior to the LNP incubation. For antibody-mediated receptor blocking, 25ug/mL of the following antibodies were used: Anti-Insulin Receptor beta antibody (Cat# ab227831, Abcam), Anti-Transferrin Receptor antibody (Cat# ab84036, Abcam), and Anti-ALK-1 antibody(Cat# ab68703, Abcam). Each antibody was added to the apical chamber and incubated at 37 °C for 2 h to block the corresponding receptors, insulin receptor, Transferrin receptor, and low-density lipoprotein receptor, respectively. Following blocking treatment, the inserts were gently rinsed twice with the fresh medium to remove the unbound inhibitors or antibodies. LNP-EGFP-mRNAs were then introduced into the apical chamber and introduced as described for downstream transfection or transport analyses.

4.15. siRNA-mediated receptor knockdown

At Day 4 post-BBB model integration, transwell inserts were transferred to a new plate for siRNA-mediated gene knockdown. siRNA (Table S2) targeting SCARB1 (siRNA-SRB1, Cat# 4390825, Assay ID s2650, Thermo Fisher Scientific), and INSR (siRNA-INSR, Cat# AM51331, Assay ID 29, Thermo Fisher Scientific) were used to silence target genes expression in the endothelial layer on the BBB insert. Cells were transfected with a total volume of 200uL containing 25 pmol of siRNA per insert with Lipofectamine^™ RNAiMAX Transfection Reagent (Cat# 13778150, Thermo Fisher Scientific) in serum-reduced OptiMEM^™ I Reduced Serum Medium (Cat# 31985062, Gibco), following the manufacturer’s protocol. siRNA and lipofectamine were diluted separately in the Opti-MEM, combined at a 1:1 vol ratio, and incubated for 5min at room temperature to form siRNA-lipid complexes. The siRNA-lipid complexes were added dropwise to the apical chamber of each insert containing fresh Opti-MEM. Cells were incubated for 24 h at 37 °C to allow the transfection. At 24 h post-transfection, the medium containing the siRNA-lipid complexes was aspirated, and both the apical and basolateral chambers were replenished with the complete endothelial growth medium. Cells were maintained for an additional 24 h to allow recovery. At 48 h post-transfection, LNP-EGFP-mRNAs were then introduced into the apical chamber and incubated as described in the LNP transfection and transport assay protocols for downstream fluorescence imaging, and flow cytometry. A subset of the siRNA-treated samples were harvested for mRNA and protein expression analysis to validate gene knockdown efficiency.

4.16. Chemiluminescence electrophoresis protein detection

Protein expression of INSR, SR-B1, and PECAM1 after siRNA-mediated knockdown was quantified using the Jess^™ Simple Western System (ProteinSimple) according to the manufacturer’s protocol. Cells were harvested 48 h post-transfection and lysed in ice-cold RIPA buffer (Cat# 89900, Thermo Fisher Scientific) supplemented with protease inhibitor cocktail (Cat# 78429, Thermo Fisher Scientific). The lysate was incubated on ice for 30 min, followed by centrifugation at 14,000×g for 15 min at 4 °C to remove debris. The supernatants were collected and the total protein concentration was measured by Pierce^™ BCA Protein Assay Kit (Cat# 23225, Thermo Fisher Scientific). Protein samples were diluted to 400ug/mL in 0.1x Sample buffer (Cat# SMW002, ProteinSimple) and mixed with Fluorescent 5X Master Mix, containing DTT, SDS, and molecular weight standards. The mixture was denatured at 95 °C for 5 min prior to loading. The assay was run using the 12–230 kDa Jess separation module (Cat# SMW002, ProteinSimple). A volume of 5uL of each sample, along with primary antibodies, and secondary antibodies was loaded into the designated wells of the capillary cartridge. The capillary cartridge was then loaded into the Jess instrument, and the chemiluminescence-based immunoassay was run using Compass for Simple Western software. The following primary antibodies were used: anti-PECAM1 (1:10, Abcam, ab9498) anti-SRB1 (1:10, Invitrogen, PA1–16788), anti-INSR (1:10, Abcam, ab227831), and anti-GAPDH (1:10, Cat# ab181602, Abcam). HRP-conjugated secondary antibodies from ProteinSimple were used for chemiluminescent detection: Anti-Rabbit Secondary HRP Antibody (Cat# 042–206, ProteinSimple), and anti-Mouse Secondary HRP Antibody (Cat# 042–205, ProteinSimple). Protein band signals were detected and quantified using Compass software.

4.17. In vivo LNP administration via the intravenous route

BALB/c mice (n = 3 for each group) were administered LNP1-FlucmRNA and LNP9-Fluc-mRNA via intravenous tail vein injection. Two dosing regimens were evaluated: a single intravenous injection or a double IV injection spaced 4 h apart. Three hours following the final injection, mice received an intraperitoneal injection of 100 μl of D-luciferin substrate at the concentration of 15 mg/mL and were incubated for 10 min prior to imaging. The whole body and head bioluminescence signals were acquired using the In vivo Imaging System (PerkinElmer). Following imaging, the mice were then anesthetized and transcardially perfused with 10 mL PBS to remove circulating blood and unbound nanoparticles. Brains and peripheral organs were immediately reimaged using the IVIS system to quantify post-perfusion brain signal intensity.

From a separate cohort of mice receiving the same treatment, organs (including the brain, heart, spleen, kidney, lung, and liver) were harvested and snap-frozen in dry ice. Tissues were stored at −80°Cfor downstream Fluc mRNA quantification by RT-qPCR, as described previously.

In a parallel experiment, Ai14 reporter mice (n = 3 for each group) were intravenously injected with LNP1-CRE-mRNA and LNP9-CRE-mRNA using the same IV administration protocol and dosing regimens. One-week post-injection, mice were transcardially perfused with 10 mL PBS followed by 4% paraformaldehyde (Cat# 15713S, Electron Microscopy Sciences) to fix the tissue for histological analysis. Major organs (including the brain, heart, spleen, kidney, lung, and liver) were collected and imaged using the IVIS system to assess tdTomato expression. Brains were post-fixed in 4% paraformaldehyde at 4 °C overnight and then incubated in a 30% sucrose (Cat# S5–500, Fisher Scientific) solution in PBS at 4 °C for up to 3 days before cryosectioning.

4.18. Immunofluorescence staining for brain tissue sections

The brain was immersed in optimal cutting temperature (O.C.T.) compound and cryosectioned at a thickness of 30 μm using a Leica CM1950 cryostat (Leica Biosystems). Coronal brain sections were mounted onto Superfrost Plus microscope slides and stored at −20 °C until further processing. For immunostaining, the slides were equilibrated to room temperature and permeabilized in 0.2% Triton X-100 in PBS for 1 h. To block non-specific antibody binding, sections were incubated in blocking buffer containing 4% normal goat serum in 0.02% PBST for 1 h at room temperature. Samples were then incubated overnight at 4 °C with the primary antibody against NeuN (1:100, Cell Signalling, 24307) diluted in blocking buffer. The next day, slides were washed four times with DPBS for 10 min on an orbital shaker at room temperature. Fluorophore-conjugated secondary antibody, Goat anti-Rabbit IgG (H + L) Cross-Adsorbed Secondary Antibody, Alexa Fluor^™ 488 ((1:500, Invitrogen, A-11008), diluted in blocking solution, was applied for 1.5 h at room temperature in dark on an orbital shaker. Following incubation, samples were washed four times with DPBS for 10 min each. Nuclei were counterstained with Hoechst 33342, prepared at a dilution of 2 drops/mL in DI water. A total of 250 μL of the diluted stain was added to each sample and incubated 20min at room temperature in the dark. Fluorescence images were acquired using Leica TCS SP8 confocal microscope. Regions of interest were selected based on tdTomato and NeuN signal expression. Image analysis and colocalization were performed using ImageJ software.

Appendix A. Supplementary data

Supplementary data to this article can be found online at https://doi.org/10.1016/j.biomaterials.2026.124202.

Data availability

Data will be made available on request.