A human-airway-on-a-chip for the rapid identification of candidate antiviral therapeutics and prophylactics
A microfluidic bronchial airway chip lined with differentiated human airway epithelium and pulmonary endothelium reproduces strain-dependent influenza virulence, cytokine output and neutrophil recruitment, and when drugs are delivered at clinically achievable blood concentrations under flow it separates candidates that work in cell lines from those that also work in hamsters challenged with SARS-CoV-2.
- Senior authors
- Donald E. Ingber
- Correspondence
- Donald E. Ingber
Research areas & themes
Citation
Si L, Bai H, Rodas M, Cao W, Oh CY, Jiang A, Moller R, Hoagland D, Oishi K, Horiuchi S, Uhl S, Blanco-Melo D, Albrecht RA, Liu WC, Jordan T, Nilsson-Payant BE, Golynker I, Frere J, Logue J, Haupt R, McGrath M, Weston S, Zhang T, Plebani R, Soong M, Nurani A, Kim SM, Zhu DY, Benam KH, Goyal G, Gilpin SE, Prantil-Baun R, Gygi SP, Powers RK, Carlson KE, Frieman M, tenOever BR, Ingber DE. A human-airway-on-a-chip for the rapid identification of candidate antiviral therapeutics and prophylactics. Nature Biomedical Engineering. 2021. Volume 5, issue 8, pages 815-829.
DOI 10.1038/s41551-021-00718-9. PMID 33941899. PMCID PMC8387338.
One-sentence contribution
A microfluidic bronchial airway chip lined with differentiated human airway epithelium and pulmonary endothelium reproduces strain-dependent influenza virulence, cytokine output and neutrophil recruitment, and when drugs are delivered at clinically achievable blood concentrations under flow it separates candidates that work in cell lines from those that also work in hamsters challenged with SARS-CoV-2.
Executive summary
Repurposing approved drugs is one of the fastest routes to a therapeutic during a pandemic, but the cell line assays that generate candidates do not reproduce airway tissue architecture, interferon competence, immune cell recruitment or the time-varying drug exposures that follow oral dosing. This work applies a two-channel microfluidic chip in which primary human bronchial basal stem cells differentiate at an air liquid interface above primary pulmonary endothelium under continuous perfusion. Chips supported replication of six influenza A strains with the relative differences seen clinically, recruited perfused human neutrophils that transmigrated and cleared infected cells, and produced cytokine profiles that scaled with strain virulence rather than with viral load. Oseltamivir acid delivered through the vascular channel reproduced both the efficacy and the two-day treatment window seen in patients, and the protease inhibitor nafamostat delivered to the airway channel extended that window to four days. The group then pivoted to SARS-CoV-2, first using spike-pseudotyped particles under biosafety level 2 conditions. Eight drugs active against pseudoparticle entry in a liver cell line were retested on chips at their reported human maximum plasma concentrations under flow, and only amodiaquine, toremifene and clomiphene remained active, while hydroxychloroquine, chloroquine and arbidol did not. Amodiaquine subsequently reduced native SARS-CoV-2 in hamsters given prophylactically, therapeutically and in a transmission setting, whereas hydroxychloroquine did not.
Scientific context
The paper states its gap in terms of the mismatch between the assays used to nominate repurposing candidates and the human tissue those candidates must act in. Cell lines frequently carry defects in interferon responses, because the cell cycle arrest that interferon imposes selects against competence during continuous passage. Neither cell lines nor conventionally cultured primary airway cells form the mucociliary pseudostratified epithelium of the living airway. Explanted human respiratory tissue avoids that problem but is scarce and short lived. Organoids give a more functional epithelium but do not permit an air liquid interface, epithelial to endothelial cross-talk, mucociliary clearance or recruitment of circulating immune cells. Across all of these, and in Transwell co-cultures that do support an air liquid interface, drugs are applied statically, so the exposure profile that a patient experiences after oral dosing is not represented. The authors place their work explicitly against the confusion over hydroxychloroquine and chloroquine in early 2020, where cell line activity was not borne out clinically. They note the programme was funded by DARPA and the National Institutes of Health two years before the pandemic in anticipation of biothreat challenges, and that their COVID-19 work began on 13 January 2020, one day after the viral genome sequence was released.
Central question
Can a perfused human airway chip that reproduces airway tissue structure, innate immune responses and clinically relevant drug exposure predict which approved drugs will work against respiratory viruses better than static cell line assays do.
Experimental strategy
Validation precedes application throughout, and the logic is to establish clinical mimicry with a virus and a drug whose human behaviour is already known before asking the model anything new. The chip places differentiated airway epithelium at an air liquid interface on one face of a porous membrane and primary pulmonary microvascular endothelium under flow on the other, with pore size chosen to permit immune cell transmigration. Fidelity was assessed at several levels, cell type composition and junctional structure, protease and receptor expression against a commonly used cell line, replication kinetics of six influenza strains spanning three subtypes and both clinical isolates and laboratory strains, barrier permeability, cytokine output from the vascular effluent, and recruitment and transmigration of perfused primary human neutrophils, with donor variability tested across five epithelial donors. Drugs are then delivered by the route that matches their clinical use, oseltamivir acid through the vascular channel to mimic blood levels after oral dosing and hepatic conversion, protease inhibitors into the airway channel to mimic inhaled delivery, and always at concentrations set by published human maximum plasma concentration values rather than by what is convenient in vitro. Treatment window is tested by adding drug at times spanning twenty-four hours before to ninety-six hours after infection. For SARS-CoV-2 the work proceeds in stages defined by containment, spike-pseudotyped luciferase particles at biosafety level 2 for entry inhibition with vesicular stomatitis virus glycoprotein particles as a specificity and toxicity control, then native virus in cell lines, then hamsters in higher containment for prophylaxis, therapy and animal to animal transmission, with pharmacokinetics measured in hamsters to set dosing and with oral and subcutaneous routes compared. Proteomics was used to ask why one antimalarial behaved differently from its close relatives.
Key findings
- Differentiated airway epithelium on chip formed a pseudostratified mucociliary layer with basal and ciliated cells in proportions resembling the human airway, continuous tight junctions and an endothelium with adherens junctions, and expressed TMPRSS2, TMPRSS4, TMPRSS11D and TMPRSS11E at higher levels than MDCK cells, with ACE2 messenger RNA and protein rising on differentiation (Figures 1b to 1f, Supplementary Figure 1).
- Influenza infection introduced into the air channel infected the epithelium, disrupted tight junctions, caused cilia loss and increased barrier permeability, and disrupted endothelial adherens junctions without detectably infecting the endothelium, which the authors align with vascular leakage observed in patients (Figures 1c and 1g, Supplementary Figure 2a). Undifferentiated basal epithelium was much less susceptible.
- Across six strains, H1N1 and H3N2 propagated by three to four logs over twenty-four to forty-eight hours while H5N1 grew more slowly, and the two H3N2 strains replicated roughly tenfold better than the H1N1 strains and caused more barrier disruption and cilia loss (Figure 2a, Figure 1g, Supplementary Figure 2b). Five epithelial donors gave similar infectivity (Supplementary Figure 2c).
- Perfused primary human neutrophils adhered to the activated endothelium within minutes, transmigrated into the epithelium over hours, targeted nucleoprotein-positive cells and cleared virus over one to two days, with more recruitment for H3N2 than H1N1 and reduced titres for both when neutrophils were present (Figures 2b to 2d, Supplementary Figure 3a). Neutrophils also increased cytokine production and barrier damage (Supplementary Figures 3b and 3c), and the authors note that neutrophils can protect or harm depending on context.
- Three clinical isolates of differing virulence produced cytokine and chemokine levels that ranked with clinical severity, H5N1 highest despite the lowest replication (Figure 2e). This dissociation between viral load and inflammatory output is the clearest demonstration that the chip reports something beyond replication.
- Oseltamivir acid at 1 micromolar in the vascular channel reduced progeny titre, preserved barrier function and tight junctions, and lowered cytokine output (Figures 3a to 3d), an effect the authors compare with the approximately one log reduction reported in a randomised trial.
- Nafamostat and Trasylol delivered to the airway channel reduced H1N1 and H3N2 titres, preserved barrier and junction integrity, lowered cytokines, and blocked cleavage of hemagglutinin into HA1 and HA2 (Figure 3e, Supplementary Figures 4a to 4e).
- In a time-course experiment, oseltamivir worked only when added within forty-eight hours of infection, matching its clinical recommendation, and combining it with nafamostat extended effective treatment to ninety-six hours (Figure 3f). The authors describe the combined effect at later times as synergistic. The observation is a doubling of the effective window in this model.
- Eight approved drugs inhibited spike-pseudotyped particle entry in Huh-7 cells in a dose-dependent manner without detectable toxicity (Figure 4a). The authors note that Huh-7 cells express little ACE2, lack TMPRSS2 and derive from liver.
- On chips pretreated for twenty-four hours at published human maximum plasma concentrations and maintained under flow, only amodiaquine, toremifene and clomiphene reduced entry, by 59.1, 51.1 and 28.1 percent, while hydroxychloroquine, chloroquine, arbidol, verapamil and amiodarone did not (Figure 4b, Table 1). The authors point out that the three that failed on chip also failed in clinical trials. The active amodiaquine metabolite desethylamodiaquine gave about 60 percent inhibition at a clinically relevant concentration (Supplementary Figure 8).
- Quantitative mass spectrometry showed that amodiaquine perturbed the airway epithelial proteome more broadly and differently than chloroquine or hydroxychloroquine, with the most affected proteins related to cilium regulation and lysosomal function (Figures 5a to 5c). The authors offer this as something that may be responsible for the greater effect on entry, which is a proposal rather than a mechanism established here.
- Against native SARS-CoV-2, amodiaquine and desethylamodiaquine inhibited Vero E6 infection with half-maximal inhibitory concentrations of 7.5 and 9.9 micromolar (Supplementary Figure 9), and amodiaquine at 10 micromolar reduced viral load by about three orders of magnitude in ACE2-expressing A549 cells (Figure 6a).
- In hamsters, subcutaneous amodiaquine begun one day before intranasal challenge reduced lung subgenomic nucleocapsid RNA by roughly 70 percent at day three with reduced nucleocapsid staining in lung sections (Figures 6b and 6c). In a co-caging transmission model where all vehicle animals became infected within two days, amodiaquine reduced nucleocapsid RNA by 90 percent and titres by more than one log (Figure 6d).
- Oral amodiaquine at 75 milligrams per kilogram gave pharmacokinetics comparable to subcutaneous dosing with higher tissue than plasma levels, prevented infection to a similar degree, and in the same model hydroxychloroquine at a dose reported to give clinically relevant lung exposure had no significant effect (Figure 6e, Supplementary Figure 11). RNA sequencing of treated infected hamsters showed downregulation of inflammatory gene sets including TNF alpha and NF kappa B signalling, IL-6 JAK STAT3 and interferon gamma (Supplementary Figure 12).
- Treatment begun one day after infection gave about 70 percent inhibition at day three and no detectable nucleocapsid transcript in lung at day seven (Figure 6f).
Mechanistic model
The study is a model validation and drug discovery paper rather than a mechanistic one, and it does not establish how amodiaquine acts against SARS-CoV-2. For nafamostat the mechanism is supported directly, since the protease inhibitors blocked cleavage of hemagglutinin precursor into its subunits by TMPRSS11D and TMPRSS2, which is required for entry, and the chip expresses those proteases at levels above the cell lines conventionally used.
For amodiaquine the paper establishes an effect on entry of spike-pseudotyped particles and on infection by native virus in cells and animals, but not the molecular target. The proteomic comparison with chloroquine and hydroxychloroquine shows divergent host proteome effects concentrated on cilium regulation and lysosomal proteins, and the authors write that this may be responsible for the greater effect on entry. That is a proposal, and since the pseudoparticle system reports only entry the activity against native virus could involve additional steps not tested here.
The claim that carries the most weight, that the chip predicts clinical outcome better than static cell line assays, is supported by concordance rather than by a mechanism. Drugs that failed on chip also failed in trials, drugs that succeeded on chip succeeded in hamsters. The authors are explicit that a caveat applies, namely that drug absorption into the device material and protein binding were not quantified in this study, so the concentration actually reaching the cells is not known. They also note that full pharmacokinetic profiles rather than fixed maximum concentrations could be recapitulated in future work.
Conceptual or technical advance
The work demonstrates that an airway chip can reproduce several features of human influenza that had not been available together in one in vitro system, notably strain-dependent virulence ranking, the dissociation between viral load and inflammatory output seen with H5N1, endothelial disruption without endothelial infection, and the full sequence of neutrophil adhesion, transmigration and clearance under flow. It then shows that the model discriminates among drugs in a way that static assays do not, which is the practical contribution. The combination finding, that a serine protease inhibitor doubles the window during which a neuraminidase inhibitor remains useful, is directly actionable because most patients present late. On the coronavirus side, the study establishes a staged discovery route that begins under biosafety level 2 with pseudotyped particles, which lowered the barrier to entry for laboratories without high containment during the early pandemic, and it nominated amodiaquine on evidence that subsequently supported the initiation of Phase 2 trials in Africa by Medicines for Malaria Venture with Wits University and by the ANTICOV programme, a development the authors report.
Relationship to the broader research program
The tenOever laboratory's contribution, as stated in the author contributions, was to develop the hamster COVID-19 infection model and to test drug efficacy against native SARS-CoV-2 in vivo, together with colleagues at Mount Sinai. The study was conceived by Si, Bai and Ingber at the Wyss Institute, who also developed the discovery pipeline, with the Frieman group at Maryland testing amodiaquine and its metabolite against native virus. The connection to the tenOever programme runs through the golden hamster model of SARS-CoV-2, which that laboratory developed and applied across its own pandemic-era work, and through the transcriptional readout of the host inflammatory response, which is the laboratory's characteristic way of scoring infection. Reading this paper alongside the laboratory's own SARS-CoV-2 studies shows the same animal model and the same host response framing serving as a shared platform for collaborations with groups bringing different front ends, which is a category 3 synthesis available from the corpus rather than a claim made here.
Related publications
- Si and colleagues, 2020 preprint, predecessor to this article, from the same collaboration. The bioRxiv report of the drug repurposing findings, published on 13 April 2020, which the authors state contributed in part to the initiation of amodiaquine clinical trials.
- Benam and colleagues, cited as reference 17, methodological foundation, from the Ingber laboratory. The earlier human airway chip, from which this device differs in membrane chemistry, material properties and pore size, the last change permitting immune cell transmigration.
- Thacker and colleagues, 2020, and Zhang and colleagues, 2020, companion, from other laboratories. Contemporaneous organ chip studies of SARS-CoV-2 infection that the authors distinguish from this work on the grounds that they used alveolar epithelium or cell lines rather than airway epithelium and did not address drug repurposing.
Limitations and boundaries
The SARS-CoV-2 chip experiments used spike-pseudotyped particles, which report entry only and do not replicate, so the chip data speak to prophylaxis against initial infection rather than to therapy, and the authors say that integrating these chips into higher containment laboratories will be necessary to study native virus on chip. The authors state that they did not quantify drug absorption into the device or protein binding, so nominal and delivered concentrations may differ, and that fixed maximum plasma concentrations were used rather than full pharmacokinetic profiles. Amodiaquine carries a black-box warning in the United States for rare agranulocytosis and liver damage with high or prolonged dosing, which the authors raise alongside their suggestion that a short course could be considered, and no safety or efficacy claim in humans follows from the animal data presented. Hamster group sizes were small, between three and eight animals per condition. Endpoints in animals were subgenomic nucleocapsid RNA, plaque titre and histology rather than survival or clinical scoring. The influenza work uses one chip format, a limited number of strains and healthy donor cells, with no comorbidity, age or prior immunity represented, and the model lacks adaptive immunity, alveolar tissue and systemic circulation. The neutrophil experiments use cells from blood donors perfused at physiological concentration for a short period and do not model sustained recruitment. The proteomic explanation for why amodiaquine differs from its relatives is a correlation with pathway annotation, not a demonstrated mechanism. Finally, the concordance argument, that chip results match clinical trial outcomes, rests on a small number of drugs chosen for testing on the basis of prior reports.
Audience summaries
25 words
A perfused chip of human airway tissue reproduced influenza severity differences and immune cell recruitment, and picked out amodiaquine, which protected hamsters from SARS-CoV-2 where hydroxychloroquine failed.
75 words
Drug candidates nominated in cell lines often fail in people. A microfluidic chip lined with differentiated human airway epithelium over pulmonary endothelium reproduced strain differences in influenza severity, cytokine output and neutrophil recruitment, and matched the clinical effect and two-day treatment window of oseltamivir, which a protease inhibitor doubled. Tested at human blood concentrations under flow, hydroxychloroquine and chloroquine failed against pseudotyped SARS-CoV-2 while amodiaquine worked, and amodiaquine went on to protect hamsters from infection.
150 words
The assays used to nominate repurposing candidates use cell lines with impaired interferon responses, no airway architecture, no circulating immune cells and static drug exposure. This study used a two-channel chip with primary human bronchial epithelium differentiated at an air liquid interface above perfused pulmonary endothelium. It reproduced replication differences among six influenza strains, greater barrier damage from H3N2 than H1N1, cytokine output that tracked clinical severity rather than viral load for H5N1, and neutrophil adhesion, transmigration and virus clearance. Oseltamivir acid delivered at blood-like concentrations reproduced its clinical efficacy and its two-day window, which the protease inhibitor nafamostat extended to four days by blocking hemagglutinin cleavage. Eight drugs active against spike-pseudotyped particles in a liver cell line were retested on chip at human maximum plasma concentrations, and only amodiaquine, toremifene and clomiphene remained active. Amodiaquine reduced native SARS-CoV-2 in hamsters prophylactically, therapeutically and in transmission, while hydroxychloroquine did not.
Documented publication relationships
No explicit publication relationship was recorded in the reviewed graph.