Breathing lung-on-chip: a versatile tool for assessing respiratory toxicity across multiple therapeutic modalities

Abstract

Inhalation administration of therapeutics is a crucial method for treatment of respiratory diseases, offering direct access to the target organ. However, the progression of candidate drugs is frequently impacted by clinical dose level limitations due to lung histopathological findings or functional effects identified in in vivo studies. Addressing these safety concerns is crucial in advancing compounds with the right safety profile. To that end, there is a need for predictive in vitro model systems to evaluate lung toxicities, including inflammatory responses across various modalities. This study aimed to assess the predictive capability of the AlveoliX Lung-on-Chip (^AXLung-on-Chip) model in determining respiratory toxicity of eight inhaled substances of varying modalities. Experiments using a two-dimensional (2D) culture were conducted to assess cellular responses, optimize dose settings and study design. Differentiation between compounds with lower and higher inflammatory potential was not possible in the 2D model. In contrast however, the response following treatment in the ^AXLung-on-Chip model was more pronounced, and the use of multiple endpoints enabled differentiation based on their inflammatory potential. Our study also indicated a potential increased sensitivity in cytokine response following treatment when mechanical stretch was incorporated in the ^AXLung-on-Chip. Comparison to in vivo toxicology studies demonstrated that the ^AXLung-on-Chip model predicted drug-induced inflammatory responses, capturing a spectrum of lung pathologies from mild toxicity to severe inflammatory damage, and illustrates the potential of the ^AXLung-on-Chip to identify inhaled compound toxicity across various modalities.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00204-025-04269-9.

Introduction

Inhalation is an important therapeutic delivery route for treating respiratory diseases and has the benefit of achieving direct access to the target organ. However, the progression of candidate drugs is often impacted by clinical dose level limitations, which arise from lung histopathological findings or functional effects observed in safety in vivo studies. Tools to identify, understand and mitigate safety concerns early during drug discovery are of utmost importance for the progression of compounds with the right safety profile (Easter et al. 2009; Hornberg et al. 2014).

Traditionally, Good Laboratory Practice (GLP) in vivo studies have been considered the gold standard in toxicology. However, the low throughput, high cost, translational limitations and the growing emphasis on the 3Rs (Replacement, Reduction, Refinement) are driving the field to develop new strategies on safety assessment, including more advanced in vitro models. With the advancement of new chemical modalities beyond traditional small molecules (Czechtizky et al. 2022), an expanding range of therapeutics and delivery methods, such as nucleic acid-based therapeutics, lipid nanoparticles (LNPs) (Chow et al. 2021; Xu et al. 2021), peptides, antibody fragments, and antibody–drug conjugates (Blanco et al. 2022; Czechtizky et al. 2022; Verma et al. 2023; Wang et al. 2023), are currently employed, which necessitate evaluation in new assays developed for safety predictions (Blanco et al. 2022; Chow et al. 2021; Czechtizky et al. 2022; Verma et al. 2023; Wang et al. 2023; Xu et al. 2021). For example, phosphorothioate containing Antisense Oligonucleotides (ASOs) are therapeutics that can modulate cellular RNA levels with high selectivity but have been linked to cytotoxicity and innate immune responses in preclinical studies (Frazier 2015; Pollak et al. 2022). No inhaled ASO has yet been approved, possibly reflecting the challenges of turning a promising drug modality into an approved medicine. Similar challenges exist for inhaled biotherapeutics, which can induce perivascular and peribronchiolar mononuclear inflammatory cell infiltrates, and increased macrophages, in toxicology studies in non-human primates (NHPs) and rodents (Hall et al. 2021). Indeed, even more established modalities such as small molecules are relatively often associated with adverse responses in the respiratory tract that may limit further development (Cook et al. 2014). This highlights the importance of investigating potential toxicity as early as possible in predictive in vitro models during drug development.

Whilst simple 2D models offer multiple advantages in terms of throughput, their simplicity can present limitations. For example, we have previously developed a high-content 2D imaging assay capable of identifying small molecules with the potential to cause lung irritation, however, this assay lacks an immune component, model complexity and does not have an inflammatory readout (Fitzpatrick et al. 2024). Respiratory epithelial monocultures on porous membranes cultured at air-liquid interface (ALI) are a more complex model commonly used for studying respiratory biological- and safety-related questions in vitro. Previously we have shown that using Transepithelial electrical resistance (TEER) measurement in a bronchial epithelial ALI model treated with small molecules was a highly predictive method to identify in vivo lung irritant compounds. However, this model required high concentrations to distinguish between lung toxic and non-toxic compounds, was only validated for small molecules, and lacked an immune component and biomechanical stimuli (Balogh Sivars et al. 2018). Hence, it is of key importance to continue developing lung in vitro models capable of recapitulating a local inflammatory response for identifying and predicting potential lung inflammatory responses arising from a range of modalities. Significant advancements have been made in the development of organ-specific in vitro models containing physiologically relevant components, known as Microphysiological systems (MPS). MPS models, which offer a greater complexity than standard two-dimensional (2D) models, are increasingly used in pre-clinical drug assessment (Pognan et al. 2023) due to their improved ability to recapitulate the in vivo cellular environment by incorporating factors like biomechanical stimulation, substrate stiffness and biomaterials (Guenat et al. 2020; Nizamoglu et al. 2023). In particular, it has been shown that more complex models have the capability to predict inflammatory responses due to improved cellular complexity, a three-dimensional (3D) structure and relevant biomechanical forces, and thus could help bridge the gap between in vitro culture and animal models (Jensen and Teng 2020).

The aim of the current study was to investigate the ability of the AlveoliX Lung-on-Chip model (^AXLung-on-Chip) to predict respiratory toxicity of inhaled compounds. The ^AXLung-on-Chip model represents a highly advanced barrier model via its complex 3D composition and its unique breathing motion. This human lung model incorporates a primary-derived alveolar epithelial cell line ^AXiAEC (AlveoliX), lung microvascular cells and immune cells (Sengupta et al. 2023, 2022; Stucki et al. 2018). The ^AXiAEC cells, previously extensively characterized, exhibit characteristics of both type 1 and type 2 pneumocytes, highlighted by expression of markers such as HTI-56, SFTPC and ABCA3 (Dobbs et al. 1999; Sun et al. 2021). Under dynamic conditions, cells undergo cyclic stretching simulating the in vivo respiratory movement. This cyclic stretch enhances the expression of F-actin filaments and alveolar epithelial markers at the gene level, including MUC1, AQP5, CAV1, SFTPC, thereby highlighting the relevance of biomechanical stimulation (Sengupta et al. 2023). Moreover, cyclic strain induces alignment in vascular endothelial cells (Matsumoto et al. 2007), has been shown to activate signaling pathways leading to increased surfactant production (Diem et al. 2020) and activates metalloproteinases involved in barrier regulation, cellular proliferation and inflammation (Grannemann et al. 2023). This study assessed the effect of treatments at various concentrations under both static and dynamic conditions. Given that the physical barrier is the first line of defense within the immune system, maintained by junctional proteins in the epithelial cells, the assessment of barrier integrity (measured by TEER) was included. Additionally, monitoring of pro-inflammatory cytokines, damage-associated molecular patterns (DAMPs) released during tissue injury, and cytotoxicity measurements in response to treatment were included as crucial endpoints (Burgoyne et al. 2021; Gopallawa et al. 2023). Assessment of these endpoints was used to inform on the toxic profile of inhaled compounds.

To determine whether the ^AXLung-on-Chip in vitro model can accurately predict in vivo lung toxicity outcomes, we assembled a focused validation set of substances with well-characterized in vivo toxicity profiles. This approach allowed us to directly compare in vitro responses with established in vivo results, thereby evaluating the model’s translational relevance. Eight substances with varying in vivo lung toxicity profiles representing four different drug modalities (Table 1) were tested in the ^AXLung-on-Chip model. The validation set included two RNase H dependent ASO “gapmers” targeting different transcripts (ASO1 and ASO2), an LNP with encapsulated mRNA coding for green fluorescent protein (LNP-eGFP), three small molecules with different targets (Tiotropium, AZ1, AZ2), and Cadmium chloride (CdCl₂), an inorganic salt and environmental pollutant. Previous internal in vivo toxicity studies have shown lung inflammatory responses of varying histopathologic severity following administration of AZ1, ASO1, ASO2, LNP-eGFP, and AZ2. AZ1 induced high severity lung inflammation and increased alveolar macrophages in rats after 7 days of inhalation dosing, causing widespread lung parenchymal damage. ASO1 and ASO2 dosed once intratracheally (IT) in rats resulted in lung inflammatory cell infiltrates and increased alveolar macrophages after 72 h. Although the inflammatory cell infiltrates were of high severity, no parenchymal damage was observed. Lung histopathology following LNP-eGFP administration in rats was limited to low severity inflammatory cell infiltration 24 h after a single IT dose. In 3-month inhalation studies in rat and cynomolgus monkey, AZ2 induced a low severity inflammatory macrophage response in the lung, characterized by aggregations of alveolar macrophages associated with secondary mononuclear inflammatory cell infiltrates. CdCl₂ was included in the validation set as a positive control as it is a known toxic air pollutant linked to obstructive lung disease and lung cancer (Genchi et al. 2020; Wang et al. 2024). Bacterial lipopolysaccharide (LPS) was also used as a positive control for proinflammatory induction (Albano et al. 2022; Forti et al. 2010). Tiotropium, a potent clinically approved drug widely utilized in the treatment of chronic obstructive pulmonary disease (COPD) served as a negative control compound (Oba et al. 2008), and has also been shown to be non-irritant in previous in vitro studies (Balogh Sivars et al. 2018; Fitzpatrick et al. 2024). To ensure consistency with clinically relevant dosing paradigms and the design of the in vivo studies, LNP-eGFP and ASOs were administered only once in vitro, mirroring their single administration in the corresponding in vivo experiments. In contrast, AZ1, AZ2 and Tiotropium were dosed repeatedly in vitro, reflecting the repeated dosing regimen used in their in vivo studies. This distinction aligns with clinical practice, where small molecules such as AZ1, AZ2 and Tiotropium are more frequently dosed whilst modalities like LNPs and ASOs are administered far less frequently, due to their long-acting effects and extended half-lives (Bäckström et al. 2024; Mehta et al. 2023; Moumné et al. 2022).

Table 1.

Substance assay validation set

Test substance	Modality/type	Preclinical/clinical findings*
Tiotropium	Small molecule; acetylcholine receptor antagonist	Negative control. No evident lung histopathology findings reported in subchronic and chronic toxicology studies (FDA 2004). Marketed product for COPD (Oba et al. 2008)
AZ1	Small molecule kinase inhibitor	Lung inflammation and increased alveolar macrophages observed at high severity in rat 7-day inhalation study. Associated lung parenchymal damage characterized by air space consolidation, alveolar fibrin deposits and alveolar type 2 cell hyperplasia
AZ2	Small molecule kinase inhibitor	Alveolar macrophage aggregates and mononuclear inflammatory cell infiltrates present at low (minimal) severity in rat and cynomolgus monkey 3-month inhalation studies. Histopathology findings consistent with lung clearance in response to repeated deposition of inhaled drug particulates (Nikula et al. 2014)
ASO1	Antisense oligonucleotide	Lung inflammatory cell infiltrates (high severity) and increased alveolar macrophages (low severity) observed in rat 3-day single dose IT study
ASO2	Antisense oligonucleotide	Lung inflammatory cell infiltrates (high severity) and increased alveolar macrophages (low severity) observed in rat 3-day single dose IT study
LNP-eGFP	Lipid nanoparticle encapsulating eGFP encoding mRNA	Lung inflammatory cell infiltrates (low severity) observed in several rat 24-h single IT dose studies
CdCl2	Toxic inorganic salt	Positive control. Drives lung inflammation and injury in rodents (Larson-Casey et al. 2020), and induces oxidative stress, cytotoxicity and barrier dysfunction in vitro (Albano et al. 2022; Forti et al. 2010). Obstructive lung disease and lung cancer are reported in exposed humans (Genchi et al. 2020; Wang et al. 2024)
LPS	Outer membrane component of gram-negative bacteria	Positive control. Potent microbial activator of innate immune system. Induces lung inflammation and injury in rodents and humans (Rosadini & Kagan 2017; Tsikis et al. 2022)

Table summarizing the test substances used in this study, their corresponding modality and associated preclinical findings

^*Preclinical histopathology findings listed for AZ1, AZ2, ASO1, ASO2 and LNP-eGFP were observed in AstraZeneca toxicity studies (unpublished data)

Methods

^AXLung-on-Chip model

The ^AXLung-on-Chip model, based on the ^AXBarrier-on-Chip technology (AlveoliX) (Sengupta et al. 2023, 2022), comprises of the microfluidic plate (AX12) (Fig. 1a), two electro-pneumatic devices (^AXActuator, ^AXExchanger), and a connecting unit (^AXDock; Fig. 1b). The AX12 is designed in a 96-well plate format, consisting of two modular chips, each with six individual units and a support plate with pneumatic valves. Each unit features an inlet, a cell compartment, and an outlet, interconnected on the basal side by microfluidic channels and pneumatically controlled valves (Fig. 1a). The apical cell compartment and basal fluidic chamber are separated by a 3.5 µm thin elastic porous membrane (3 µm pores, 8 × 10⁵ pores/cm²) where cells are cultured. The ^AXExchanger allows for medium exchange, TEER measurement, and other handling procedures. To initiate actuation, the AX12 is placed in the ^AXDock within an incubator, followed by initiation of 3D cyclic stretch using the ^AXActuator, which employs negative pressure curves to deflect the porous membrane.

Fig. 1.

The ^AXLung-on-Chip model enables robust, immune competent alveolar barrier model with dynamic breathing motion. a Schematic representation of the AX12 chip design, with top and side views. b Overview of the ^AXBarrier-on-Chip system used for operating the AX12 chip; image provided by AlveoliX. c Illustration of the immune competent alveolar barrier model incorporating breathing motion on the AX12 porous membrane (grey). Alveolar epithelial cells (green) are cultured on the apical side of the membrane, and lung microvascular endothelial cells (red) are cultured on the basolateral side. Immune cells, PBMCs (purple), are cultured on the epithelial side of the membrane. d TEER progression over time in static and dynamic alveolar barrier models under untreated conditions (static: n = 9–113, dynamic n = 14–184, mean ± standard error of the mean). e Representative immunofluorescence image of ZO-1 tight junction protein (green), the endothelial marker PECAM-1 (red) and Hoechst nuclear staining (blue) in the alveolar co-culture barrier model. f 3D reconstruction of immunofluorescence image of alveolar co-culture model in AX12 chip, ZO-1 (green), PECAM-1 (red), Hoechst (blue)

Cell culture

^AXiAECs and primary human lung microvascular endothelial cells (hLMVEC) were provided by AlveoliX (Sengupta et al. 2022). ^AXiAECs were cultured in AX Alveolar Epithelial Medium (^AXAEM, AlveoliX) and used at passages 30 or 31, while hLMVEC were cultured in AX Lung Microvascular Medium (^AXL.MVM, AlveoliX) and used at passage 7 or 8. For 2D and chip experiments, peripheral blood mononuclear cells (PBMCs, AlveoliX) were added to epithelial cells along with the compound treatments. In the 2D experiments, ^AXiAECs were seeded in 96-well plates (3596, Corning) at a density of 62.5 × 10³ cells/cm² and cultured to confluency in ^AXAEM until treatment. In the ^AXLung-on-Chip model, hLMVECs were first seeded on the basolateral side of the AX12 membrane and allowed to adhere for 2 h. The AX12 plates were then closed and filled with AX Epithelial/Endothelial Alveolar Barrier Medium (^AXE2-ABM, AlveoliX) using the initial filling function on the ^AXExchanger. ^AXiAECs were then seeded on the apical side of the membrane in ^AXE2-ABM. Cells were maintained at 37 °C, 5% CO₂, and the medium was replaced every 2–3 days. MLE-12 pulmonary mouse tumor cells (ATCC) were seeded at 5 × 10³ cells/well in a 96-well plate (655946, Greiner) using DMEM/F-12 medium (11039-21, Gibco) supplemented with 10% FBS (A5256701, Gibco). Hepa 1–6 mouse hepatoma cells (ATCC) were seeded at 10 × 10³ cells/well (655946, Greiner) in DMEM + GlutaMAX (21885-025, Gibco) medium supplemented with 10% FBS. MLE-12 and Hepa 1–6 cell experiments were conducted between passages 8 and 16.

LNP formulation

The eGFP-containing LNPs were prepared as described previously (Philipp et al. 2023), using MC3 as ionizable lipid and eGFP mRNA at a 10:1 w/w ratio of lipid to mRNA (MC3/Chol/DSPC/DMPE-PEG at 50/38.5/10/1.5 mol percent, with 3.08 mol of MC3 per mole mRNA). The DLin-MC3-DMA was synthesized by AstraZeneca, cholesterol was purchased from Sigma-Aldrich, DSPC from Avanti Polar Lipids, DMPE-PEG2000 from NOF Corporation and eGFP mRNA from TriLink Biotechnologies. For characterization of the prepared LNPs, the z-averaged particle diameter was measured using a ZetaSizer Nano (Malvern Instruments Inc.), while the final mRNA concentration and encapsulation efficiency were assessed using the Quant-it RiboGreen Assay Kit (ThermoFisher Scientific).

Cell treatments

All test substances except for LPS (E-Coli 026:B26, Sigma) and CdCl₂ (202908-10G, Sigma Aldrich) were synthesized and provided by AstraZeneca. Human cells were treated on the apical side with respective compound concentrations diluted in cell culture medium, upon confluency in the 2D experiments and upon barrier formation in the ^AXLung-on-Chip experiments. In 2D experiments, LNP-eGFP was administered at 0.0625, 0.25, 1 and 4 µg/ml; ASO1, ASO2, Tiotropium and AZ1 at 0.4, 2, 10 and 50 µM; AZ2 and CdCl₂ at 0.2, 1, 5 and 25 µM, and a vehicle of 0.5% DMSO. In ^AXLung-on-Chip experiments LNP-eGFP was administered at 0.1 and 1 µg/ml; ASOs at 6 and 54 µM; AZ1 and AZ2 at 6 and 60 µM; Tiotropium at 0.6, 6 and 60 µM. CdCl₂ were used as a positive control at 200 µM for repeated administration (RA) and 400 µM for single administration (SA) regimen, to ensure a strong toxic effect (Cao et al. 2015; Forti et al. 2010). LPS was consistently administered at 100 ng/ml according to previous experience (Sengupta et al. 2022) and literature (Honda and Inagawa 2023). Vehicle composition varied by compound and individual experiment. LNP-eGFP used 0.5% PBS; ASOs used 5.4% water (static), 0.6% PBS (dynamic); LPS used 5.4% water (static) and 0.5%, 0.6%, 10% PBS or 5.4% water (dynamic); CdCl₂ used 0.4%, 0.5% DMSO (static) and 0.4, 0.8, or 1.8% DMSO (dynamic); RA experiments for AZ1, AZ2, Tiotropium, CdCl₂ and LPS used 0.9% DMSO. For SA, cells were dosed at 0 h and cultured up to 72 h (day 3). In the RA protocol, dosing started at 0 h and was repeated daily up to day 7, for a total of 8 doses. During RA, half of the apical medium was replaced with freshly prepared compound solution at the same concentration. Control cells were treated with the vehicle corresponding to each test compound. For ASO treatment of mouse cells, Hepa1-6 and MLE-12 were incubated at 37 °C, 5% CO₂ for 24 h before treatment with ASO1 or ASO2, respectively, or 0.5% PBS as vehicle in culture media for 24 h.

TEER measurement in the ^AXLung-on-Chip

TEER was measured over time in the ^AXLung-on-Chip using a 96-well plate electrode (STX100MC96, World Precision Instruments) and an Epithelial Volt/Ohm Meter (EVOM3, World Precision Instruments). Background TEER (Ω) was obtained from wells without cells. Background-subtracted TEER values were multiplied by the surface area of the well for final TEER values in Ω x cm². For each well, TEER values (Ω x cm²) at a given timepoint were first normalized against 0 h timepoint and further divided by the average of the corresponding vehicle group and expressed as percent reduction.

Cytokine analysis

In the 2D experiment medium was sampled on days 1 and 3 following SA treatments, and on days 1, 3, 8 for RA. In the chip experiment medium was collected from the apical compartment on day 3. Cytokine quantification (G-CSF, IL-10, IL-1β, IL-6, IL-8, IP-10, MCP-1, MIP-1α, MIP-1β, TNF-α) was conducted using EMD Millipore´s MILLIPLEX® MAP Human Cytokine magnetic bead kit (HCYTOMAG-60 K, Merck KGaA, Darmstadt, Germany). Samples and standards were diluted two to three fold in assay buffer, and analyzed following manufacturer’s instructions, using Bio Rad Luminex 200® for fluorescent quantification. Fold changes for individual cytokines, as represented in cytokine heat maps, were calculated by dividing each measurement by the average level of the corresponding vehicle.

Lactate dehydrogenase (LDH) analysis

LDH activity was assessed using the LDH-Glo assay kit (J2380, Promega). Medium was sampled on days 1 and 3 for SA treatments, and on days 1, 3, 8 for RA in 2D experiments, and at day 3 in the chip experiments (apical compartment). Complete cell membrane lysis was achieved by treating cells with 1% Triton-X 100 (X100, Sigma), which served as maximum cytotoxicity signal. Following manufacturer’s instructions, samples were diluted 1:10 in LDH buffer, and LDH was further quantified. Results were reported as percentage of maximum cytotoxicity.

Immunofluorescence imaging

Cells cultured on AX12 plates were fixed with Image-iT Fixative Solution (R37814, ThermoFisher Scientific) and stored in PBS. The chips were disassembled and stained following the manufacturer’s instructions. Briefly, the membranes were detached using the ^AXDisassembly tool (AlveoliX), blocked with 2% BSA (A9418, Sigma), and stained with the following antibody dilutions in 2% BSA: ZO1-AF488 (1:100 dilution, 339,188, Invitrogen), PECAM1-AF555 (1:100 dilution, 61255s, Cell Signaling), Hoechst nuclear staining (1:2000 dilution, 62249 ThermoFisher Scientific). The membranes were mounted on glass coverslips (631–1339, VWR) and imaged with Nikon Eclipse Ti-E Spinning Disk Confocal microscope. Images were processed using ImageJ software.

Gene expression analysis

Epithelial cells and PBMCs were lysed separately in 150 µl RLT Buffer (79216, Qiagen) and RNA was extracted in Qiacube extractor (Qiagen) using RNeasy Mini (74104, Qiagen) and Micro kit (74004, Qiagen) respectively, with DNAse incubation. RNA was quantified for cDNA synthesis using the High-Capacity cDNA Reverse transcription kit (4,368814, Applied Biosystems). For MLE-12 and Hepa 1–6 cells, three technical replicates per condition were pooled in RLT Buffer (79216, Qiagen) and RNA and cDNA were obtained as described above. Target gene assessment was performed using human specific TaqMan gene expression assays FAM dye (TSLP Hs00263639_m1, IL33 Hs04931857_m1, Thermofisher) and 2 × TaqMan Gene Expression FAST Master Mix (4444557, Applied Biosystems) on a QuantStudio Real-Time PCR System (Applied Biosystems). CT values were extracted with Quantstudio Real-time PCR software (Applied Biosystems), and data analysis involved normalization to housekeeping genes (GAPDH Mm99999915_g1 for Hepa 1–6, HPRT Mm00446968_m1, for MLE-12, and HPRT1 Hs99999909_m1 for ^AXiAECs and PBMCs, Thermofisher), log transformation, and visualized as fold change to control (2^−ΔΔCT).

Total protein quantification

Total protein content in each LNP-eGFP extraction was determined using Pierce bicinchoninic acid (BCA) Protein Assay Kit (23225, Thermo Scientific) according to the microplate protocol. Samples were diluted in 1X Cell Extraction Buffer PTR and absorbance was measured at 562 nm using the CLARIOstar plate reader (BMG LABTECH). The total amount of protein per sample was determined from the internal assay 4-parameter standard curve.

Quantification of cellular GFP levels

Quantification of GFP in cell lysates was performed using GFP ELISA kit from Abcam (ab171581) according to manufactures protocol specified for adherent cell lysates. Epithelial and PBMC cells were lysed separately in 100 µl 1X Cell Extraction Buffer PTR from GFP ELISA kit (ab171581, Abcam). Absorbance was measured at 450 nm using a CLARIOstar plate reader (BMG LABTECH). Normalization was performed based on volume and protein content per well, with technical duplicates analyzed for each sample.

AZ2 nonclinical toxicology studies and AZ quantification in lung tissue

Groups of rats (Han Wistar CRL: WI(Han), 7–10 weeks old, 10/sex/group) were given AZ2 by inhalation (nose-only, 20 min/day) at estimated achieved doses of 0 (air control), 0 (placebo control), 85.1, 184 or 603 µg/kg/day for 13 weeks. For the achieved dose of 1.48 mg/kg/day groups of 7–9 weeks rats (Han Wistar CRL: WI(Han), 10/sex/group) were given AZ2 by inhalation (nose-only, 2 h/day) for 13 weeks. Groups of cynomolgus monkeys (2 years 2 months-2 years 3 months old, 3/sex/group) were given AZ2 by inhalation (10 min/day) at achieved doses of 0 (placebo control), 26.4, 56.4 or 188 µg/kg/day for 13 weeks.

Quantification of AZ2 in lung tissue samples

Two sub-samples of cynomolgus monkey lung tissue were taken from the accessory lobe 24 h after last inhalation, and AZ2 concentrations were determined using qualified methods. Tissues were homogenized with deionized water at a 1:3 (w/v) ratio or higher, followed by liquid–liquid extraction by liquid chromatography with mass spectrometric detection (LC–MS/MS). The lower limit of quantification was 1 nmol/kg and the upper limit of quantification 1000 nmol/kg. The mean tissue concentration of the two aliquots was calculated, applying the dilution factor for homogenization volume and assuming a density of 1 kg/L for conversion from nmol/kg to µmol/L.

Histopathological assessment of lung from AZ2-dosed animals

For both rat and NHP toxicology studies, lungs and trachea were removed intact at necropsy and fixed using a combination of intratracheal instillation of 10% neutral buffered formalin followed by immersion in this fixative for 48 h. Representative samples of fixed lung tissue were collected from all lobes, embedded in paraffin, sectioned, and stained with haematoxylin and eosin (H&E). Light microscopic examination of the H&E-stained lung sections was performed by a toxicologic pathologist.

Data and statistical analysis

All data are presented as mean ± standard deviation (SD) unless otherwise specified. The sample size (n) refers to individual replicates. Statistical analyses were performed in GraphPad Prism (version 10.1.2). For comparisons of treatment groups to the corresponding vehicle control, ordinary one-way ANOVA was used with either Šídák’s or Dunnett’s multiple comparisons test, as appropriate. When the assumption of normality was violated, the non-parametric Kruskal–Wallis test with Dunn’s multiple comparisons was applied. When variances were unequal, the Brown–Forsythe and Welch ANOVA were used. For comparisons between two groups, unpaired t-tests were performed. For analyses with repeated measures over time, repeated-measures two-way ANOVA or a mixed-effects model was used, depending on data structure and missingness. Statistical significance thresholds were defined as follows: *p < 0.05, **p < 0.01, ***p < 0.001. The specific tests used for each analysis are reported in figure legends. Comprehensive results for all statistical tests, including 95% confidence intervals, mean differences, and p-values, are provided in Supplementary Document 1 and 2.

Results

The ^AXLung-on-Chip model enables robust immune competent co-culture with dynamic breathing motion

The alveolar ^AXLung-on-Chip tri-culture barrier model, employed in this study, was composed of ^AXiAEC co-cultured with hLMVEC on opposite sides of a porous membrane. PBMCs were applied to the apical side of the alveolar co-culture to ensure close physical proximity to the epithelial cells and enhance the response to treatments administered apically (Fig. 1c). The co-culture develops a tight and functional barrier (> 1000 Ω x cm²) over approximately 10 days in both static and dynamic settings (Fig. 1d). Immunofluorescence staining conducted during the plateau phase of the TEER measurements showed strong tight junction staining in the epithelial layer, indicated by Zonula Occludens 1 (ZO-1), and typical cell–cell junctions in the microvascular layer, marked by platelet endothelial cell adhesion molecule 1 (PECAM-1; Fig. 1e and f).

Assessment of cellular responses and experimental design in 2D epithelial-PBMC co-culture

To evaluate the sensitivity of ^AXiAEC co-cultured with PBMCs regarding cytokine secretion and cytotoxicity and to inform the design of the upcoming ^AXLung-on-Chip experiments, ^AXiAEC cells were cultured in 2D 96-well plates and treated with test compounds across a wide dose range to identify optimal concentrations and timepoints. Reflecting the appropriate dosing frequency for each modality used, three compounds were dosed once (SA), while four compounds were added repeatedly (RA) over eight days (Fig. 2a and b). LPS treatment increased the cytokine secretion above control levels for most cytokines on days 1 and 3 in both SA and RA (Fig. 2c and d), confirming the co-culture’s ability to secrete cytokines. Overall, cytokine responses to SA-compounds were comparable between days 1 and 3, with slight increases observed on day 3 (Fig. 2c and d). Both ASO1 and ASO2 treatments resulted in dose-dependent cytokine increases on days 1 and 3 (Fig. 2c and e), with the strongest response at the highest concentrations. Notably, ASO2 treatment also induced cytokine production at lower concentrations. In contrast, LNP-eGFP treatment elicited cytokine reductions or minimal increases across all concentrations. Cytotoxicity remained low across treatments and timepoints (Fig. 2e and Supplementary Fig. 1a). Following RA, Tiotropium unexpectedly demonstrated a dose-dependent cytokine increase at the two highest concentrations, peaking on day 3 (Fig. 2c and d). The highest dose of AZ2 caused a transient minor cytokine elevation on day 1. Moreover, CdCl₂ at 25 µM elevated several cytokine levels, with the strongest response on day 3 (Fig. 2d). Most RA-test substances showed the highest cytokine response on day 3 (Fig. 2c, d and Supplementary Fig. 1b) with minimal cytotoxicity (Fig. 2e). However, by day 8, LDH levels increased across all groups, including the vehicle control, indicating reduced cell viability (Supplementary Fig. 1c and Supplementary Fig. 1d). These findings highlight day 3 as the optimal timepoint for cytokine measurements in future ^AXLung-on-Chip experiments.

Fig. 2.

Initial cellular response profiling in 2D Epithelial–PBMC co-culture across timepoints. Illustration of experimental workflow indicating compound addition and sampling timepoints for a SA and b RA experiments in 2D. c, d Mean cytokine fold change compared to vehicle following SA or RA at day 1 (c) or day 3 (d) (n = 2). e Cytotoxicity as percentage of maximum LDH release at day 3 following SA or RA (n = 1–2). In heatmaps, increasing purple intensity indicates greater upregulation, light blue indicates downregulation, and white indicates no change compared to control. In bar graphs data are presented as mean ± SD

3D ^AXLung-on-Chip model demonstrates capability to identify inflammatory compounds with varying modalities after single administration

To study the safety profile of the tested substances, evaluate the models’ sensitivity and assess the influence of biomechanical stimuli on cellular responses, we tested the response of the ^AXLung-on-Chip model following SA treatment with compounds from the validation-set (Fig. 3a), reflecting their in vivo dosing frequency (Table 1) as well as previous in vitro experience, including results from the 2D experiment.

Fig. 3.

The 3D ^AXLung-on-Chip model identifies inflammatory compounds following single administration. a Illustration of experimental workflow, indicating compound addition and sampling timepoints. b Mean cytokine fold change compared to vehicle in the ^AXLung-on-Chip model (day 3 apical sampling) under static (i) (n = 3–5) or dynamic (ii) (n = 3–15) conditions. Statistical analysis was performed using ordinary one-way ANOVA with Šídák’s multiple comparisons test. c) Mean cytokine fold change following CdCl₂ treatment under static (blue) and dynamic (orange) conditions (n = 5 for static, n = 15 for dynamic; Unpaired t-test comparing CdCl₂ responses in static versus dynamic condition). d Cytotoxicity as percentage of maximum LDH release at day 3 following SA (blue) and dynamic condition (apical sampling) (n = 2–3; ordinary one-way ANOVA with Šídák’s multiple comparisons test). e Percentage change in TEER normalized to changes in vehicle control over 3 days post exposure under static (blue) and dynamic (orange) condition (n = 5–15 CdCl₂, n = 3 ASOs, n = 1–3 LNP; repeated measures two-way ANOVA or a mixed effect model was performed for statistical analysis following treatment versus vehicle). In heatmaps, increasing purple intensity indicates greater upregulation, light blue indicates downregulation, and white indicates no change compared to vehicle control. In bar graphs and curve graphs data are presented as mean ± SD

As expected, based on their inflammatory potential, LPS treatment induced a strong cytokine response, present in both static and dynamic conditions, whereas CdCl₂ treatment triggered a less pronounced increase (Fig. 3b). ASO1 and ASO2 exhibited an inflammatory signature, with the strongest response at 54 µM. Under dynamic conditions, ASO1 treatment induced a stronger cytokine response, with several cytokines reaching levels comparable to those triggered by LPS. The response to ASO2 was in general less pronounced compared to ASO1. No inflammatory response was observed for LNP-eGFP. To assess robustness and further explore differences between static and dynamic conditions, CdCl₂ treatment was repeated with a larger number of replicates with results indicating a significantly stronger treatment effect on G-CSF and IL-6 in dynamic conditions over static (Fig. 3c). Under both conditions, LDH results revealed an overall non-cytotoxic effect, apart from significant increases after CdCl₂ treatment (Fig. 3d). Moreover, the barrier integrity decreased significantly following SA of CdCl₂ in static condition (Fig. 3e, Supplementary Fig. 2a i, Supplementary Fig. 2b i). In contrast, TEER increased significantly following ASO2 treatment under static condition.

Compatibility of ^AXLung-on-Chip with oligonucleotide and lipid nanoparticle modalities

Gene expression analysis was performed to confirm ASO uptake by evaluating target knockdown efficiency, as well as to investigate potential changes in DAMP levels. Since the test ASOs are mouse-specific in design, their activity was validated in mouse Hepa 1–6 and MLE cells, where dose-dependent significant target knockdown was observed (Fig. 4a). As expected, ASO treatment did not yield detectable target knockdown in the human ^AXiAEC cells cultured on-chip, indeed the gene targeted by ASO1 was undetectable in this model (Fig. 4b), suggesting that observed toxicity responses are influenced by other factors than the target gene knock-down. ASO toxicities are often based on other factors like hybridization-dependent off-target effects and hybridization-independent but sequence-dependent toxicity, which is highly dependent on design, chemistry, and nucleotide sequence (Andersson 2022; Goyenvalle et al. 2023). Given the induction of cytokines in the absence of increased LDH levels, additional cellular stress markers were evaluated by measuring changes in alarmins expression. While ASO2 exposure caused only minor changes in alarmins gene expression, ASO1 treatment resulted in increased TSLP levels in both static and dynamic settings and elevated IL-33 more markedly under static setting (Fig. 4c). These findings suggest a potential cellular damage response following ASO1 treatment, consistent with the strong inflammatory response and partial barrier impairment observed previously (Fig. 3). Additionally, GFP protein quantification confirmed cellular uptake of LNPs and translation of delivered mRNA, with comparable GFP levels in static and dynamic conditions (Fig. 4d). As expected, the highest dose resulted in the highest cellular GFP levels, which were greater in epithelial cells compared to PBMCs (Fig. 4d).

Fig. 4.

Evaluation of ASO target knockdown and eGFP delivery by lipid nanoparticles. a Relative gene expression of ASO1 and ASO2 target genes normalized to housekeeping genes, in treated versus vehicle-treated mouse cells; i) Hepa 1–6 ii) MLE (n = 3; Ordinary one-way ANOVA with Dunnett’s multiple comparison test versus vehicle). b Relative gene expression of ASO2 target gene normalized to housekeeping genes in treated versus vehicle-treated ^AXiAEC epithelial cells and PBMCs from the ^AXLung-on-Chip experiment on day 3, under static (blue) and dynamic (orange) conditions (n = 3). c Relative gene expression of TSLP and IL-33 normalized to housekeeping genes in ASO1- and ASO2-treated versus vehicle-treated ^AXiAEC epithelial cells from the ^AXLung-on-Chip experiment (day 3), under static (blue) and dynamic (orange) conditions (n = 3; Ordinary one-way ANOVA with Dunnett’s multiple comparison test versus vehicle). d GFP levels (as determined by ELISA) in ^AXiAEC epithelial cells and PBMCs from the ^AXLung-on-Chip experiment (day 3), following treatment with the indicated LNP-eGFP concentrations under static (blue) and dynamic (orange) conditions (n = 1–3). All data are presented as mean ± SD

3D ^AXLung-on-Chip model enables identification of inflammatory substances after repeated compound administration

Following the SA experiment, the ^AXLung-on-Chip model was employed to evaluate the safety profiles of compounds intended for RA (Fig. 5a), reflecting their in vivo dosing frequency and prior in vitro data. As expected, LPS treatment induced a robust increase in cytokine secretion under both static and dynamic conditions, with fold changes ranging from 3.5 to 327 compared to the vehicle control (Fig. 5b), without causing cytotoxicity (Fig. 5c). The ^AXLung-on-Chip model accurately reproduced the non-inflammatory, non-cytotoxic profile of Tiotropium, as expected (Fig. 5b and c). No inflammatory response was observed following AZ1 treatment, although this compound increased LDH levels relative to the vehicle control (Fig. 5b and c). AZ2 induced the most pronounced cytokine response, showing a dose-dependent significant increase for several cytokines, in both static and dynamic settings, and also increased LDH levels compared to vehicle control (Fig. 5b and c). Similar to results in SA regimen, CdCl₂ treatment induced an inflammatory response for a limited set of cytokines under static condition, while biomechanical stimulation elicited a more pronounced response (Fig. 5b). Not surprisingly, repeated CdCl₂ treatment also elevated LDH levels in both static and dynamic conditions (Fig. 5c). A notable difference in TEER was observed between the negative control Tiotropium and treatment with AZ1, AZ2, and CdCl₂ (Fig. 5d). Although no statistically significant differences were observed across the treatment period, AZ1, AZ2 and CdCl₂ treatments showed a trend of decreasing barrier integrity, while Tiotropium maintained TEER levels comparable to vehicle levels. Notably, on day 1, AZ1 at 60 µM and AZ2 at both concentrations significantly reduced TEER under static condition (Supplementary Fig. 3 a i).

Fig. 5.

3D ^AXLung-on-Chip model identifies inflammatory toxic compounds following repeated administration. a Illustration of experimental workflow for RA ^AXLung-on-Chip experiment. b Mean cytokine fold change compared to vehicle in ^AXLung-on-Chip model (day 3, apical sampling) under i static and ii dynamic condition (n = 3). Statistical analysis was performed using ordinary one-way ANOVA with Dunnett’s multiple comparisons test. c Cytotoxicity as percent of maximum LDH release, day 3 following RA under static (blue) and dynamic (orange) condition (apical sampling) (n = 3; ordinary one-way ANOVA with Dunnett’s multiple comparisons test). d Percentage change in TEER normalized to changes in vehicle control over 8 days of RA under static (blue) and dynamic (orange) condition (n = 3). A repeated measures two-way ANOVA or a mixed effect model was performed for statistical analysis following treatment versus vehicle. In heatmaps, increasing purple intensity indicates greater upregulation, light blue indicates downregulation, and white indicates no change compared to control. In bar and curve graphs data are presented as mean ± SD

AZ2 exemplifies translation of chip findings to corresponding in vivo study observations

To understand differences between the in vitro and in vivo responses to test compounds, we compared the ^AXLung-on-Chip data following AZ2 treatment under dynamic settings with lung histopathology from 3-month dry powder inhalation studies in rat and cynomolgus monkey. Lung histopathological assessment in cynomolgus monkeys showed a low severity inflammatory macrophage response at 0.188 mg/kg/day, in both males and females (Fig. 6a), characterized by aggregations of alveolar macrophages associated with secondary mononuclear inflammatory cell infiltrates (Fig. 6a iii). Although the infiltrates were mainly restricted to the perivascular interstitium, scattered mononuclear inflammatory cells were also visible within adjacent alveoli. The AZ2-related lung findings were minimal in severity, multifocally distributed throughout the parenchyma, were not associated with histopathologic evidence of air-blood barrier damage (e.g. alveolar oedema, hemorrhage, or ATI cell necrosis) and no disruption of lung parenchymal architecture was observed. No AZ2-related microscopic findings were present in the lung at the lower dose of 0.0564 mg/kg/day (Fig. 6a ii). Similarly, in the 3-month rat inhalation study, a low severity inflammatory macrophage response occurred in lung from males and females at 1.48 mg/kg/day (Fig. 6b iii), with no AZ2-related lung findings at the lower dose of 0.603 mg/kg/day (Fig. 6b ii). AZ2 quantification in homogenized lung tissue showed that the lower dose in cynomolgus monkey of 0.0564 mg/kg/day corresponded to a lung tissue concentration ranging between 3.3 nM-4.3 µM, while the histopathological lung toxic dose 0.188 mg/kg/day corresponded to a 5–40 µM tissue concentration (Fig. 6c). Interestingly, AZ2 treatment in vitro led to increased cytokine levels at both 6 and 60 µM, with the most pronounced response observed at the higher concentration (Fig. 6d). Specifically, G-CSF, IL-1β, IL-6, IL-8, MIP-1α, MIP-1β and TNF-α were significantly upregulated at 60 µM. The higher concentration also induced increased cytotoxicity and showed a trend toward reduced barrier integrity, whereas the lower dose of 6 µM was associated with a safer profile (Fig. 6e and Fig. 6f). Taken together, the in vitro responses to AZ2 in the human ^AXLung-on-Chip model occurred at concentrations comparable to those encountered in vivo, underscoring the model’s utility for translational toxicology studies.

Fig. 6.

AZ2 in vitro ^AXLung-on-Chip results correlate with in vivo findings. a, b Lung histology after AZ2 administration in a male rat and b male cynomolgus monkey, i vehicle group showing normal microarchitecture with empty air spaces, ii mid dose-group (0.603 mg/kg/day in rat, 0.0564 mg/kg/day in cynomolgus monkey) showing normal lung histology, iii high dose-group (1.48 mg/kg/day in rat, 0.188 mg/kg/day in cynomolgus monkey) showing histologically abnormal lung with macrophage aggregates within alveoli (arrows) and mononuclear inflammatory cells surrounding a blood vessel (asterisks). Bar = 100 µm. c Total AZ2 lung tissue concentration in cynomolgus monkey from mid dose and high dose group (as determined by LC–MS/MS analysis). d Cytokine fold change versus vehicle following AZ2 treatment at 6 and 60 µM in the ^AXLung-on-Chip experiment in dynamic condition (apical sampling) (n = 3; ordinary one-way ANOVA and Dunnett’s multiple comparison test). e Percent of maximum cytotoxicity (LDH release) following AZ2 treatment in the ^AXLung-on-Chip experiment in dynamic condition (apical sampling) (n = 3 ordinary one-way ANOVA and Dunnett’s multiple comparison test). f TEER changes in percentage compared to control following AZ2 treatment in the ^AXLung-on-Chip experiment in dynamic condition (n = 3; repeated measures two-way ANOVA). Data are presented as mean ± SD

Discussion

To enable the development of safe and efficacious novel drugs to patients, it is essential to identify and mitigate potential safety concerns as early as possible during drug discovery. To address the need for predictive in vitro tools to evaluate adverse effects in the lung toxicities of inhaled candidate drugs, we assessed the ^AXLung-on-Chip model for determining respiratory toxicity induced by a range of inhaled modalities. This model incorporates important variables, including a 3D tri-culture environment with immune cells, and the possibility to include biomechanical stimuli resembling breathing motion. Evaluation of cellular responses using eight inhaled substances of varying modalities and toxicity levels demonstrated that the ^AXLung-on-Chip model is a competent tool for enhancing the understanding of inhaled compound toxicity.

Given the well-known limitations of 2D cell culture in replicating the complex tissue dynamics found in vivo, and the generally superior performance of 3D models in this respect (Guenat et al. 2020; Kimura et al. 2025), in this study we undertook a comparative analysis of 2D versus 3D culture systems to evaluate the relative capability in predicting inhaled compound toxicity (Table 2). Initial experiments performed using the 2D model provided valuable insights for study design and dose settings. Despite observing an evident cytokine response to stimuli in the 2D model for most test compounds, it was challenging to distinguish between compounds with lesser or greater inflammatory risk. Indeed, the negative control Tiotropium exhibited a strong positive signal at the highest concentration under 2D, but not under 3D conditions, despite absence of reported signs of local toxicity in in vivo studies. This aligns with previous observations that cells are more sensitive to drug treatment when cultured in 2D due to the lack of normal morphology and polarization (Jensen and Teng 2020). This is further underscored when one considers the therapeutic dosage and potency of this compound. Without considering kinetics, accumulation, or distribution factors, a conservative estimate of clinical Tiotropium, would estimate lung exposure to be more than 50 times lower than the highest in vitro concentration used in our study. This is based on assumptions of a daily dosing of 18 µg Tiotropium, a maximum deposition of 20% (Brand et al. 2007), and an average volume of reported human lung lining fluid (Fröhlich et al. 2016). The 3D model, despite higher dosing concentrations, accurately reported Tiotropium as negative, consistent with its clinical safety profile. In contrast, compounds with inflammatory potential (AZ1, AZ2) showed a limited increase in cytokines in the 2D model. Specifically, AZ2 treatment had a weak response in 2D, while it was correctly identified as inflammatory in the more complex ^AXLung-on-Chip 3D model (Table 2). Although AZ1 stimulus did not induce cytokines in either model, likely due to its intended anti-inflammatory pharmacological effect, it caused TEER reduction and cytotoxicity increase in the ^AXLung-on-Chip model, where the ability to measure TEER enabled identification of its toxicity. The differences in responses between models highlight the robustness of the 3D environment, likely attributed to its complex microenvironment enriched with extracellular matrix (ECM) molecules crucial for regulating cell functions (Jensen and Teng 2020). The ^AXLung-on-Chip model enhances cell–cell and ECM interactions, which are essential for maintaining alveolar homeostasis (Bissonnette et al. 2020) and barrier function (Blume et al. 2017; Vitucci et al. 2024; Wang et al. 2020), as well as improves innate immune signaling. Overall, the 3D ^AXLung-on-Chip model more accurately classified the compounds according to expectations based on in vivo data (Table 2).

Table 2.

Summary of in vivo findings compared to in vitro responses to the validation substances

Test substance	Toxicity findings in vivo	Toxicity-response in 2D	Toxicity-response in static AXLung-on-Chip		Toxicity-response in dynamic AXLung-on-Chip
		aFold increase cytokines	aFold increase cytokines	bTEER change %	aFold increase cytokines	bTEER change  %
AZ1	Yes	No	No	− 42	No	− 35
AZ2	Yes	2.9	45	− 77	33	− 61
Tiotropium	No	22.8	1.3	No	1.3	No
ASO1	Yes	5.7	14.8	No	45	− 40
ASO2	Yes	8.9	3.5	No	2	No
LNP-eGFP	Yes	No	No	No	No	No
CdCl2	Yes	7.6	6.4 SA 1.3 RA	− 36 SA − 52 RA	5.2 RA 5.2 SA	− 32 SA − 42 RA

Overview of in vitro responses from 2D, static and dynamic ^AXLung-on-Chip models for all compounds in the validation set

^aAverage mean of fold increases across the 10 cytokines in the panel

^bMaximum TEER reduction across the experiment reported in table

Having established the need for a 3D environment, we examined cellular responses under both static and dynamic conditions. Our findings revealed increased release of G-CSF and IL-6 in dynamic over static conditions following SA of CdCl₂. These cytokines are associated with processes in immune cell recruitment (Bengalli et al. 2017; Shan et al. 2024), a common histopathology finding in inhalation toxicology studies. This effect was observed also after RA of CdCl₂ and SA of ASO1 which showed significantly elevated cytokine levels compared to vehicle in dynamic but not static conditions, although with less statistical power. Supporting this, a recent RNA sequencing study demonstrated that mechanical strain in a lung-on-chip model elevated cytokine levels, activated host defense pathways, and suppressed cell cycle and cell proliferation processes (Bai et al. 2022). Moreover, previous studies emphasize the importance of stimulating breathing motion for improved sensitivity of lung-on-chip models following treatment of nanoparticles and IL-2 (Huh et al. 2012; Sengupta et al. 2023). These observations underscore the potential value of incorporating biomechanical stimuli for increased cytokine sensitivity following treatments, or when including endpoints involving processes such as immune cell recruitment. Complementing this approach, the ^AXLung-on-Chip features a thin 3.5 µm membrane enabling epithelial-endothelial crosstalk and immune cell infiltration. This membrane closely resembles the in vivo alveolar barrier (Gehr et al. 1978) and is to our knowledge the thinnest among the commercially available membrane-based models, such as the commonly used polyester or polycarbonate Transwell® inserts, which are about 10 µm (Zakharova et al. 2021).

Establishing translatability and understanding the exposure versus effect relationship to in vivo outcomes is crucial in developing predictive in vitro toxicology models. To this end we ran comparable in vitro-in vivo studies with the known lung toxicant small molecule AZ2. Our study demonstrates promising comparability, particularly in the ability to recapitulate dose-dependent responses. Considering doses, the low in vitro concentration (6 µM) was comparable to the measured AZ2 lung tissue concentration in the low-dose group (0.0564 mg/kg/day), while the highest in vitro concentration (60 µM) was approximately twice the highest measured in vivo lung concentration in the high-dose group (0.188 mg/kg/day). In vitro, only the high dose of AZ2 exhibited a strong inflammatory response and changes in barrier function and cytotoxicity. In contrast, the low dose did not show these effects, which aligns with similar observations made in vivo. Indeed, a clear dose-dependent effect was evident in vivo showing normal lung histology in the low dose group and histologically abnormal lungs with alveolar macrophage aggregates and mononuclear inflammatory cell infiltrates at low severity in the high dose group. These observations illustrate the ^AXLung-on-Chip model’s ability to identify safety profiles in a dose-dependent manner and highlighting the importance of using predicted clinical doses and human lung concentrations when putting a potential response into context in the risk assessment of the compound.

The compound validation set used in this study included substances with known lung toxicity profiles established by histological assessment in preclinical safety studies (AZ1, AZ2, ASO1, ASO2, LNP-eGFP), and compounds whose lung effects in vivo are well documented in the literature (Tiotropium, CdCl₂, LPS) (Table 1). When comparing to in vivo findings, the aim was to assess which in vitro endpoints could be informative as to the compound’s safety profile. Compound-specific comparisons revealed alignment between in vitro and in vivo findings for most test items (Table 2). However, some differences were observed and the severity in response was difficult to correlate. The negative control Tiotropium and positive controls CdCl₂ and LPS were successfully identified in the ^AXLung-on-Chip as non-toxic and toxic, respectively. The ^AXLung-on-Chip model also accurately identified the ASOs as proinflammatory without affecting barrier function, correlating well with the lung inflammatory cell infiltrates and increased alveolar macrophages observed in rats. However, whilst ASO2 showed a less pronounced cytokine and alarmin response in the ^AXLung-on-Chip versus ASO1, this difference was not mirrored in the severity of the lung histopathology observed in the single dose 3-day rat studies. Despite this, to the authors’ knowledge, this represents the first published in vitro lung model demonstrating the potential to recapitulate ASO-induced innate lung inflammation in vivo. Another minor discrepancy was observed with AZ1 treatment, which in vivo induced high severity inflammation, alveolar macrophage aggregates, and lung parenchymal damage in rats, whilst in vitro exposure caused moderate barrier reduction, increased LDH levels but without a corresponding cytokine elevation. These results align with the compound’s tissue-damaging effect; however, the absence of cytokine responses could represent a limitation in the model, species-specific differences, or be a result of in vitro/in vivo differences linked to the compounds pharmacological anti-inflammatory effect. Additionally, despite evidence of strong correlation between in vitro and in vivo responses, AZ2 treatment induced a pronounced cytokine and barrier response in the ^AXLung-on-Chip model, while in vivo observations of inflammatory cell responses in rats and cynomolgus monkeys were of low severity with no associated tissue damage. Treatment with LNP-eGFP also showed a discrepancy between in vivo and in vitro results. No toxicity in any conditions tested in vitro was observed, despite evidence of GFP expression. This result was unexpected since previous lung-delivered LNP studies (Friis et al. 2023) have reported mild inflammatory responses (Friis et al. 2023). However, whilst inflammatory responses have been previously identified, there exists a high degree of variation between responses (Hassett et al. 2019; Lemdani et al. 2024; Omo-Lamai et al. 2025). Additionally, other in vitro study results have also not shown a clear inflammatory profile (Forster Iii et al. 2022; Omo-Lamai et al. 2025; Tahtinen et al. 2022), except when applying significantly higher concentrations (Nguyen et al. 2025). Therefore, it is challenging to interpret the data from our model and to determine whether our lack of response is a true false negative, reflects low sensitivity in our in vitro model, species differences, or limitations in the ability of in vitro studies to accurately replicate in vivo outcomes from LNP-studies. Further studies would be required to establish utility for this model in safety assessments for LNP-based modalities. Additionally, whilst this data demonstrates that the ^AXLung-on-Chip model detects drug-induced inflammatory responses and identifies compounds linked to mild to severe in vivo lung toxicity, the severity observed in vivo did not always correlate with in vitro severity. Based on this small dataset, further studies are required to evaluate the model’s capacity for compound ranking.

Whilst the model showed marked responses to known inflammatory agents the study was limited in its number of biological replicates, which may reduce statistical power and increase uncertainty around effect estimates. While efforts were made to maximize rigor within these constraints, future studies with larger cohorts will be important to validate and extend these results. It is also noteworthy that pharmacodynamic, pharmacokinetic, and species differences were not considered when comparing in vitro and in vivo findings. Factors influencing concentrations, such as media composition and the presence of undissolved compounds in in vivo lung tissue could also impact interpretation of these results. Furthermore, the direct application of molecules in liquid form does not replicate aspects of particle deposition and dissolution, which could influence responses related to concentration, time gradients and inflammatory responses. Conversely, however, the usage of liquid administration in in vitro models has the advantage of allowing more precise control of applied concentrations, indeed it should be noted that IT administration of solutions in liquid form is often carried out in vivo in order to simplify study parameters. Within this investigation, IT instillation was employed for ASOs and LNP-eGFP animal studies (Table 1), which allows for a more direct in vitro to in vivo comparison. Altogether, further refinement and validation are needed to address these differences and to better evaluate the model’s predictive power and its capability to accurately grade inflammatory potential of test compounds.

Our study illustrates the potential of the ^AXLung-on-Chip as a valuable tool for predicting inhaled compound toxicity across various modalities. This study demonstrates the advantages of 3D cell culture over 2D which enables better differentiation between non-inflammatory and inflammatory compounds and highlights the potential benefits of including biomechanical stimuli. Moreover, the study suggests that a combination of different models and read-outs is favorable for comprehensive and reliable safety profiling. The comparison of in vitro data to in vivo findings underscores the translatable nature of the ^AXLung-on-Chip model, paving the way for improved preclinical assessment of inhaled therapeutics.

Supplementary Information

Below is the link to the electronic supplementary material.

Data availability

The data underlying this study are proprietary and cannot be shared publicly due to company confidentiality restrictions.

Declarations

Conflict of interest

The authors declare the following competing interests: G.R., L.F. and N.H. are employees at AlveoliX and N.H. holds equity in AlveoliX AG. The remaining authors (L.J., J.C., J.L., P.A., C.B., J..H, A.O. and P.F.) are employees and/or shareholders of AstraZeneca.

Ethical approval

Human cell material provided by AlveoliX was anonymized and used with patient’s consent and ethical approval from Kantonale Ethikkommission Bern. Rat and cynomolgus monkey studies from where data is presented, were performed at Charles River Edinburgh in the UK and thus conducted under the UK Home Office Act (project license number PBAD559F8) by the issue of license under the animals (Scientific Procedures) Act 1986. The regulations conform to EU Directive 2010/63/EU and achieve the standard of care required by the US Department of Health and Human Services’ Guide for the Care and Use of Laboratory Animals.