Human peripheral blood was collected from healthy volunteers after receiving written, informed consent. This study was performed following the principles outlined in the Declaration of Helsinki. The blood donation protocol and use of blood for this study were approved by the institutional ethics committee of the Jena University Hospital (permission number 2207–01/08).

Cell culture and incubation procedures were performed at 37°C and 5% CO₂ in a cell culture incubator, unless stated otherwise. They were sub-cultured (when reached 80% of confluence) using trypsin-EDTA solution (Capricorn Scientific, Ebsdorfergrund, Germany). Cells used in this study were tested negative for mycoplasma using the Venor®GeM Classic kit (Minerva Biolabs, Berlin, Germany).

Primary Human Umbilical Vein Endothelial Cells were obtained from Promocell (single donor, C-12200) seeded at a density of 2 × 10⁴ cells/cm² in ECGM MV (Promocell, Heidelberg, Germany) with supplements and 1% penicillin/streptomycin (Gibco/Thermo Fisher Scientific). Medium exchange was performed every 3 days until seeding into biochips.

Caco-2 cells (acCELLerate GmbH, Hamburg, Germany) were seeded at a density of 2 × 10⁴ cells/cm² and were cultured in DMEM with 4.5 g/L glucose (Lonza, Cologne, Germany) supplemented with 10% fetal calf serum (FCS, Capricorn Scientific), 1X MEM non-essential amino acids (Capricorn Scientific), 1 mM sodium pyruvate solution (Capricorn Scientific), 5 mg/mL holo-transferrin (Merck, Darmstadt, Germany) and 1X AAS. Medium was exchanged every three to 4 days until further seeding into biochips.

Human peripheral blood was collected in EDTA K3E tubes (Sarstedt, Nürmbrecht, Germany) and Serum Gel CAT tubes (Sarstedt) from healthy volunteers. Serum was obtained from each donor by centrifugation at 2,000 g for 10 min at room temperature (RT) and was stored at −20°C until use as supplement in cell culture medium.

Peripheral blood mononuclear cells (PBMCs) were isolated from EDTA blood by density gradient centrifugation as previously described (Kaden et al., 2023a). using lymphocyte separation medium (Capricorn Scientific). PBMCs were seeded into 6-well plates at a density of 1 × 10⁶ cells/cm² in X-VIVO 15 medium (Lonza) supplemented with 10% human autologous serum and 1X AAS. After 1 h of incubation, PBMCs were washed twice with X-VIVO 15 medium without supplements to remove non-adherent cells. Subsequently, adherent monocytes were differentiated to monocyte-derived macrophages (MDMs) in X-VIVO 15 medium supplemented with 10% human autologous serum, 10 ng/mL macrophage colony-stimulating factor (M-CSF, Peprotech, Hamburg, Germany), 10 ng/mL granulocyte-macrophage colony-stimulating factor (GM-CSF, Peprotech) and 1X AAS for 5 days before seeding into biochips.

Undifferentiated HepaRG cells were obtained from Biopredic International (Rennes, France) and were cultured as described before (Gripon et al., 2002). Cells were seeded at a density of 2.7 × 10⁴ cells/cm² in culture flasks and were cultured in William’s Medium E (Gibco/Thermo Fisher Scientific, Darmstadt, Germany) containing 10% FCS, 2 mM L-glutamine (Gibco/Thermo Fisher Scientific), 5 μg/mL insulin (Merck), 50 µM hydrocortisone-hemisuccinate (Biomol, Hamburg, Germany) and 1X AAS. The cells were cultured for 2 weeks with medium exchange every three to 4 days before HepaRG differentiation was initiated. Differentiation was induced by adding 2% dimethyl sulfoxide (DMSO, Merck) to the culture medium for at least 2 weeks as previously described (Gripon et al., 2002), (Cerec et al., 2007). Only fully differentiated HepaRG cells between passage 15-20 were used for biochip experiments.

Expandable human upcyte liver sinusoidal endothelial cells (LSECs) were purchased from Biotrend Chemikalien GmbH (Cologne, Germany). The cells were freshly thawed and seeded at a density of 1.3 × 10⁴ cells/cm² on collagen A (PAN-Biotech, Aidenbach, Germany)-coated culture flasks. LSECs were cultured in supplemented ECGM MV with 1X AAS as previously described (Kaden et al., 2023b). Medium exchange was performed every two to 3 days until seeding into biochips. LSECs were used up to passage 2.

The first-trimester (choriocarcinoma) trophoblastic cell line BeWo was purchased from the American Type Culture Collection (ATCC, Manassas, VA, USA) and was cultured in F-12K medium (Gibco/Thermo Fisher Scientific) containing 10% FCS, and 1% penicillin/streptomycin (Gibco/Thermo Fisher Scientific) and seeded at a density of 4 × 10⁴ cells/cm². Medium was exchanged every 3 days and prior to chip experiments, BeWo cells were gradually conditioned to the ECGM MV medium in proportions of 50%/50%, 75%/25%, and 100%/0% (ECGM MV/F-12K) every 24 h.

Biochips (BC001 for the placenta model, BC002 for gut and liver models) were manufactured by Dynamic42 GmbH (Jena, Germany) from injection-molded polybutylene terephthalate (PBT) bodies. Biochips contain two individual culture chambers used for experimental duplicates. Each culture chamber is composed of a top and a bottom cell culture channel with adjacent microfluidic channels occupying an area of 2.18 cm² top and 1.62 cm² bottom for BC002 and, 1.80 cm² top and 1.11 cm² bottom for BC001, respectively. Both channels are separated by a 12 µm polyethylene terephthalate (PET) membrane with a pore density of 1 × 10⁵ pores/cm² and a pore diameter of 8 µm (TRAKETCH Sabeu, Radeberg, Germany) for BC002 and a pore density of 0.6 × 10⁶ pores/cm² and a pore diameter of 3 µm for BC001, respectively. The total volume with the corresponding ports was 290 µL for the top channel and 270 µL for the bottom channel for BC002 and, 350 µL for the top channel and 300 µL for the bottom channel for BC001, respectively. For cell seeding, 200 µL of cell suspension was introduced in the top channel and 150 µL in the bottom channel for BC002 and, 220 µL in the top channel and 200 µL in the bottom channel for BC001, respectively.

Gut models were assembled as previously described (Maurer et al., 2019). Prior to cell seeding, biochips were sterilized with 70% ethanol (Nordbrand Nordhausen GmbH, Nordhausen, Germany) and washed twice with AQUA AD injectabilia (B.Braun, Melsungen, Germany) and once with phosphate buffered saline (PBS, Capricorn). HUVECs were seeded in top channels of BC002 coated with 50 μg/mL collagen A at a density of 1.85 × 10⁵ cells/cm² in ECGM MV with supplement and 1X AAS to form a confluent endothelial cell layer. Biochips were incubated for 24 h to facilitate attachment to the membrane. HUVECs were cultured statically for 48 h with daily medium exchange. For subsequent co-culture of HUVECs and MDMs, vascular perfusion medium (VPM) was prepared by addition of 10% autologous donor serum, 10 ng/mL M-CSF and 10 ng/mL GM-CSF to the supplemented ECGM MV with 1X AAS. MDMs were seeded at a density of 6.17 × 10⁴ cells/cm² (1 × 10⁵ cells in total) in VPM on top of the HUVECs to complete the vascular cell layer. Vascular cell layers were incubated statically for further 24 h. Subsequently, Caco-2 cells were seeded in the opposing bottom channel coated with 50 μg/mL collagen A at a density of 7.63 × 10⁴ cells/cm² in gut seeding medium (GSM) containing DMEM with 4.5 g/L glucose supplemented with 20% FCS, 1X MEM non-essential amino acids, 1 mM sodium pyruvate solution, 5 mg/mL holo-transferrin and 1X AAS. Biochips were incubated upside down for 24 h before perfusion was initiated. The medium in the Caco-2 channel was changed to gut perfusion medium (GPM) consisting of GSM with only 10% FCS 24 h post seeding.

Sterilized and collagen A-coated (50 μg/mL) BC002 biochip channels were gradually seeded with LSECs, MDMs, and differentiated HepaRG cells as previously described (Kaden et al., 2023a). The following cell densities were selected based on cell number ratios found in the liver. Concisely, human LSECs were seeded in top channels at a density of 0.45 × 10⁵ cells/cm² and were cultured statically in supplemented ECGM MV with 1X AAS until reaching confluency. Medium was exchanged to vascular perfusion medium (VPM, see gut model) prior to the seeding of MDMs. Human MDMs were seeded on top of the LSECs at a density of 0.23 × 10⁵ cells/cm² in VPM. MDMs and LSECs were cultured statically for further 24 h. Following this, differentiated HepaRG cells were seeded in the opposite hepatic channel at a density of 1.85 × 10⁵ cells/cm² in hepatic thawing and seeding medium (HTSM) containing William’s Medium E, 5% FCS, 4% hepatocyte thawing and plating supplements CM3000 (Gibco/Thermo Fisher Scientific), 1 µM dexamethasone (Gibco/Thermo Fisher Scientific), 2 mM L-glutamine, 5 μg/mL insulin, 5 µM hydrocortisone-hemisuccinate (Merck) and 1X AAS. Adhesion of the cells to the membrane was facilitated by upside down incubation of the biochip for 24 h. HTSM was replaced by hepatic perfusion medium (HPM) composed of William’s Medium E, 10% FCS, 3.6% hepatocyte maintenance supplements CM4000 (Gibco/Thermo Fisher Scientific), 0.1 µM dexamethasone, 2 mM L-glutamine, 5 μg/mL insulin, 5 µM hydrocortisone-hemisuccinate, 0.1% DMSO and 1X AAS after 24 h. The model was cultured statically for further 24 h with medium exchange in both channels before applying vascular perfusion.

BC001 biochip chambers were also sterilized with 70% ethanol and washed twice with ultra-pure distilled water (Invitrogen/Thermo Fisher Scientific) and gradually seeded with BeWos and HUVECs as reported previously (Blundell et al., 2016). HUVECs were seeded at a density of 0.45 × 10⁵ cells/cm² in the bottom channel in ECGM MV with 1% penicillin/streptomycin (Gibco/Thermo Fisher Scientific) to recapitulate the fetal vasculature, and the biochip was incubated upside down for 5 h to facilitate cell attachment to the membrane. BeWo cells were seeded at a density of 0.23 × 10⁵ cells/cm² in ECGM MV with 1% penicillin/streptomycin into the top channel to mimic the trophoblast layer at the maternal side. The biochip was maintained under static culture conditions for 24 h and the medium was exchanged before connection to perfusion.

For perfusion of the biochips, microfluidic medium reservoirs (Mobicol, Göttingen, Germany and microfluidic ChipShop, Jena, Germany) were attached to the channel ports (for hepatic models to the upper channel with LSECs and MDMs, for gut models to the upper channel with HUVECS and MDMs and the bottom Caco-2 channel, and for the placenta models to the upper BeWo channel). Biochips were connected to REGLO ICC peristaltic pumps (Masterflex/Ismatec, VWR International, Bruchsal, Germany) by silicon tubing (Dynamic42 GmbH, Jena, Germany).

In gut models, top channels with HUVECs and MDMs were perfused with VPM and Caco-2 channels with GPM as described. After static assembly, biochips were connected to bidirectional perfusion in both channels with 50 μL/min matching shear stress rates of 0.013 dyn/cm² (0.0013 Pa) in the top channel and 0.006 dyn/cm² (0.0006 Pa) in the bottom channel. Gut models were pre-perfused for 5 days prior to triple-model connection with medium exchange every two to 3 days.

Both the liver and placenta models were perfused in channels with LSECs and MDMs (liver) or BeWo (placenta) with VPM at 25 μL/min (0.0065 dyn/cm², 0.00065 Pa). Placenta models were pre-perfused for 24 h prior to triple-model connection. Liver models were not pre-perfused prior to triple-model connection.

Medium was exchanged in all channels and reservoirs before triple-model connection. Using silicon tubing, the top channel of the gut model with HUVECs and MDMs was connected with the top channel of the liver model containing LSECs and MDMs and the latter with the top channel of the placenta model with BeWo cells creating a cyclic flow, which simulates the physiological blood circulation (Figure 1A). Top channels were perfused at a rate of 25 μL/min, while the gut Caco-2 channel was perfused at a rate of 50 μL/min. The bottom channels of the liver and placenta model were maintained in a static state.

Prednisone (CAS: 53-03-2, Merck) was dissolved in DMSO to obtain a 50 mM stock solution. The stock solution was diluted in VPM or GPM to the respective working concentrations, not exceeding a DMSO concentration of 0.1% to avoid non-specific solvent toxicity. Diluted prednisone was administered at a concentration of 55 nM in the Caco-2 channel of the single-gut model, the LSECs channel of the single-liver model and the BeWo channel of the single-placenta model. To achieve a prednisone concentration of 55 nM in the triple model, it was administered at a concentration of 142 nM in the Caco-2 channel (gut model), taking into account the dilution by the additional volume of the liver and placenta compartments. This increase in concentration corresponds to the increase in volume in the triple model compared to the single model. Prednisone administration was initiated 24 h post triple-model connection for 24 h. Control chips for single and connected perfusion were left untreated and were run with medium containing 0.1% DMSO. Supernatants of 80 µL were collected directly from the reservoirs of perfused channels or medium exchange of statically cultured channels after 10 h, 20 h, and 24 h of treatment for downstream analysis of lactate dehydrogenase (LDH) and for HPLC-MS/MS measurements. HPLC-MS/MS samples were directly stored at −20°C, while samples for LDH analysis were stored at 2°C–8°C until measurement.

Drug adsorption to the biochip periphery was assessed in cell-free biochip setups. Different concentrations of prednisone (5.5 nM, 55 nM and 550 nM) in supplemented ECGM MV with AAS were transferred to the single setups as well as to the triple setup, filling the total volume of the system, including reservoirs, tubing, and top and bottom channels of the biochip. Samples were collected after 24 h of drug perfusion and stored at −20°C until HPLC-MS/MS analysis. Samples from each compartment of the triple model were pooled before storing.

All steps were performed using PBS containing Ca²⁺/Mg²⁺ (Capricorn). Membranes were excised from the biochips using a scalpel and transferred into a PBS-loaded 24-well plate. Membranes were washed once with PBS and fixed with either Histofix 4% (Carl Roth, Karlsruhe, Germany) for 15 min at RT or ice-cold methanol (Carl Roth) for 15 min at −20°C. After fixation, membranes were washed with PBS and incubated in permeabilizing/blocking solution containing PBS with 0.1% saponin (Carl Roth) and 3% normal donkey serum (NDS, Abcam, Amsterdam, Netherlands) for 30 min at RT. Upon permeabilization and blocking, membranes were incubated in permeabilizing/blocking solution containing diluted primary antibody (Supplementary Table S1) at 2°C–8°C overnight. Stained membranes were washed in PBS with 0.1% saponin and incubated in permeabilizing/blocking solution containing diluted secondary antibody (Supplementary Table S1) for 1 h at RT. Membranes were washed again with PBS with 0.1% saponin, PBS, and water and were mounted between two glass coverslips using fluorescent mounting medium (Agilent, Waldbronn, Germany). For the staining of BeWo cells the membranes were fixed with ice-cold methanol for 5 min at −20°C. After washing with fresh PBS, the membranes were incubated with a blocking solution (0.1% FCS-PBS) for 20 min at room temperature. After blocking, the membranes were incubated with fluorophore-coupled antibodies (Supplementary Table S1) for 2 h at 37°C with humidity. Nuclei were stained with 1 μg/mL of 40,6-Diamidine-20-phenylindole dihydrochloride (DAPI, Life Technologies, Karlsruhe, Germany).

Fluorescence images were acquired as Z-stacks using the AxioObserver Z1 fluorescence microscope with the ApoTome-2 (Carl Zeiss AG, Jena, Germany) and the AxioObserver 7 fluorescence microscope with the ApoTome-3 (Carl Zeiss AG). All images were taken with a PlanApochromat 20x/0.8 M27 objective (Carl Zeiss AG). Exposure times were set analogues for all images. The ZEN 2 Pro software was used for manual control of the microscope, image acquisition and image compression using the ApoTome Raw Convert function. ZEN 3.5 (blue edition) software was used for subsequent orthogonal projection and image processing.

Fluorescein isothiocyanate–dextran (FITC-dextran, CAS: 60842-46-8, Merck) with an average molecular weight of 3–5 kDa was used. Medium in both channels was replaced by pre-warmed phenol red-free William’s Medium E. A volume of 250 µL 1 mg/mL FITC-dextran solution in DMEM without phenol red was added to the gut top channel. This step was repeated to minimize dilution of the FITC-dextran solution within the channel. The biochips were incubated for 1 h at 37 C and 5% CO₂ and protected from light. The 40 kDa FITC-dextran was used to test the POC model. Similarly, 250 µL (1 mg/mL) of FITC-dextran solution was added to the maternal top channel and the biochips were incubated for 30 min at 37°C and 5% CO₂ and protected from light. Both solutions in top and bottom channel were collected and transferred to a black 96-well microplate (Corning Incorporated, New York, USA). Fluorescence (wavelengths: excitation 492 nm, emission 518 nm) was measured in a microplate reader (INFINITE 200 PRO, Tecan). To avoid values over measuring range, automatic range finder was selected. FITC-dextran concentrations were calculated from a standard curve obtained from the 1 mg/mL stock solution.

LDH concentrations were measured in medium supernatants using the Cytotoxicity Detection Kit PLUS (Roche, Basel, Switzerland). In the case of 2D assays, cell culture plates were centrifuged at 250 g for 10 min at RT to remove cell debris. Supernatants were diluted 1:2 in PBS prior to the assay. The assay was performed as described in the protocol provided by the manufacturer. Absorption was measured at 490 nm with a reference wavelength of 620 nm in a microplate reader (INFINITE 200 PRO, Tecan). LDH concentrations were calculated from a LDH standard (Merck) curve.

Samples were analyzed with HPLC-MS/MS via direct injection as formate aducts by JenaBios GmbH, Jena, Germany.

All reported experiments were performed at least 3 times, with at least 2 technical repeats in each group. Data is represented as mean ± S.D. Graph plotting and statistical analysis of biological readouts was performed on GraphPad Prism v10.3.0 (GraphPad Software, La Jolla, CA, USA). Statistical significance for 2 group comparisons were determined using two-tailed Student’s t-test, using Welch’s correction for unpaired t-test. For multiple comparisons comparing to a single control group, one way ANOVA was used with Dunnett’s multiple comparison test. All n, p-values and F, df values are mentioned in the respective figure or the associated legends.

Our approach maps the chip architecture to a compartmental model to describe the time-dependent distribution of a compound on-chip. The compartment model uses time-dependent ordinary differential equations and assume well-mixing within compartments. These equations are generally accepted to describe the distribution of exogenous and endogenous compounds and molecules (Aravindakshan et al., 2025). A physical chamber separated by a membrane or connected by flow to another chamber is represented by a compartment in the software. Serial compartments are connected via concentration-dependent flow rates (typically in mL/min) between the compartments and normalized by the volume of the originating compartment. Movement across a membrane is described by permeation and the corresponding surface area. To predict the PK of prednisone and its main metabolite prednisolone on-chip, a digital twin of each of the individual MPS was developed. Clearance of prednisone (liver) and production of prednisolone (liver) were estimated by least-square fitting of model parameters to the observed PK data. For gut and placenta-chips, the permeability across the tissue barrier was estimated to best describe the PK data. The least-square approach was implemented in R and minimized the squared weighted difference (ssq) between the model prediction (pred) and the experimental observation (obs) (Equation 1): s s q = min ∑ p r e d − o b s o b s (1)

Following, a digital twin of the 3-way interactome was developed (Figure 6). Here, the distribution after ‘oral’ administration of prednisone was predicted using the key PK-related parameters (clearance, metabolite formation, permeability across tissue barrier) from the single MPS experiments.

To predict the human PK of prednisone, first, a PBPK was developed using qualified installations of the PBPK software PK-Sim (Frechen et al., 2021). A whole-body PBPK model (Supplementary Figure 2.1) includes an explicit representation of the organs most relevant to the uptake, distribution, excretion, and metabolism of the drug. These typically include the heart, lungs, brain, stomach, spleen, pancreas, intestine, liver, kidney, gonads, thymus, adipose tissue, muscles, bones, and skin. More information can be found in the supplementary material (Supplementary Figure 2.2).

The tissues are interconnected by arterial and venous blood compartments, and each is characterized by an associated blood flow rate, volume, tissue partition coefficient, and permeability (Kuepfer et al., 2016). The analytical approach is based on the principles set out in the guidelines of the EMA, FDA, and/or OECD for reporting on PBPK M&S (EMA, 2024).

The developed PBPK model is used to describe the human kinetics of prednisone. Key kinetic parameters are informed by either clinical data, literature values or on-chip predictions (Table 2). In short, fraction unbound in plasma of prednisone was reported to be around 0.29 (van Runnard Heimel et al., 2005), lipophilicity at 1.5, total liver clearance was reported to be 3.2 mL/min/kg in non-pregnant and around 9 mL/min/kg in pregnant women (Ryu et al., 2018) both at 10 mg single dose, and renal clearance was neglected at 5% (Rose et al., 1981), (Jusko, 2013); solubility was 0.1 mg/mL. Clinical PK data were digitized from (Ryu et al., 2018), (Xu et al., 2007).

In addition to the base PBPK model, a pregnancy (fetal-maternal) PBPK model was implemented as outlined before (Dallmann et al., 2018a) which is publicly available, and used to describe the systemic exposure of prednisone in the fetus. For a more detailed description of the implementation workflow within PK-Sim^® and MoBi, we refer to the tutorial (Dallmann et al., 2018a). Reported fetal/maternal plasma ratio of prednisone is 1.67 (van Runnard Heimel et al., 2005). In short, literature, fitted, and MPS-based PK-related parameters were implemented in the pregnancy PBPK model and partition coefficients from main plasma to placenta were adjusted to match observed plasma ratios. The parameter combinations were then used to simulate maternal and fetal prednisone kinetics in n = 100 virtual women for up to 3 months (100 days).

The centerpiece of this study is a dynamic and physiologically relevant MPS that integrates the gut, liver, and placenta to investigate drug absorption, distribution, and metabolism. An MPS using Caco-2 cells (intestinal epithelial cells, grown hanging in bottom channel), HUVECs (human umbilical vein endothelial cells, grown together with MDMs in top channel), and MDMs (monocyte-derived macrophages) replicates the intestinal barrier’s microenvironment (Blanford and Murphy, 1977; Shroff et al., 2022). Caco-2 cells form an intestinal epithelium with villus-like structures under perfusion (Kim and Ingber, 2013), allowing the study of drug absorption and first pass metabolism. HUVECs simulate the endothelial barrier of blood vessels, facilitating studies on vascular permeability and MDMs are incorporated to represent immune cells, enabling the investigation of immune responses and inflammation. They further regulate tissue homeostasis by improving enzyme expression and barrier functionality in OoC models (Rennert et al., 2015). The liver model with HepaRG (hepatocytes, grown hanging in bottom channel), LSECs (liver sinusoidal endothelial cells, grown together with MDMs in top channel), and MDMs simulate the liver’s microenvironment (Rennert et al., 2015), (Gröger et al., 2016). It enables the study of hepatic functions like drug metabolism. The placenta is modeled using the human trophoblastic BeWo cells (grown in top channel) and HUVECs (grown hanging in bottom channel) to replicate the maternal-fetal interface. BeWo cells are the second most used cell line that retains the key features of trophoblast cells allowing the study of drug transfer across the placental barrier (Pastuschek et al., 2021). It provides a realistic platform for understanding placental physiology and how drugs are processed and transported in pregnant women.

Each model was established separately through sequential cell seeding and pre-perfusion for the gut model, before connecting them via perfusion and drug application the day after (Figure 1B). To ensure the robustness and functionality of the connected three-organ model and to rule out impairments caused by the presence of the other organ models, we first assessed the viability, barrier integrity and marker expression of each individual organ model in comparison to the connected perfusion set up.

The viability of the models was evaluated using an LDH cytotoxicity assay. Only upon cell damage the intracellular LDH (lactate dehydrogenase) can be measured in perfusates. Apical and basolateral compartments of each organ, singly perfused or connected, were monitored individually (Figure 1C). LDH values vary normally between the biological replicates of the experiment, but no difference was measured within an experiment when models were cultured singly or in interconnection. Overall, the integrated gut-liver-placenta MPS maintained high cellular viability across all culture conditions.

The barrier integrity of the gut, liver and placenta models was measured by a fluorescein-isothiocyanate-labeled (FITC) dextran permeability assay (Figure 1D). When models were connected, permeability remained the same or tended to decrease (POC), which confirms that the process of connecting these organ chips did not adversely affect their barrier functions.

Similarly, immunostaining (Figure 1E) did not show any effect on cell integrity or marker expression when looking at the individual cell layers of each organ after single or connected perfusion.

These results validate the stability and functionality of our integrated MPS platform, providing a reliable foundation for further pharmacokinetic and pharmacodynamic studies.

As mentioned in the introduction, prednisone was used as a model drug in this study. It was designed to mimic the effects of cortisol, a hormone that regulates inflammation, immune response, and metabolism in the body. Prednisone is commonly used in medical treatments for its anti-inflammatory and immunosuppressive properties (Chaphekar et al., 2021), (Corticosteroids, 2012). Understanding its pharmacokinetics—how it is taken up, distributed, and metabolized in the human body—is crucial for its effective clinical use. After confirming that the connection of the three organ models did not affect their integrity, the next step was the administration of prednisone to the luminal side of the gut model 24 h after starting triple-perfusion. Preliminary experiments showed that the treatment of all relevant cell types with prednisone in a concentration range of 0.5 nM–50 µM did not result in any damage or activation of the immune cells (Supplementary Figure S1). Hence, since prednisone was not expected to be harmful to any of the tissues, we monitored the model integrity by assessing cell viability/cytotoxicity, barrier function, and performing immunostaining for tissue markers.

Immunostainings (Figure 2A) revealed no changes in cell integrity and marker expression, irrespective of whether the organ models were perfused individually or connected in series, nor did the treatment with prednisone induce any noticeable alterations.

Compared to untreated, single-perfused control setup, the permeability assay (Figure 2C) showed that the barriers remained intact or even improved under all tested conditions. Additionally, LDH measurements in the perfusates (Figure 2D) indicated that prednisone treatment did not affect cell viability in any of the MPS setups. The higher variability observed in samples from the gut model is anticipated due to the high cell shedding rates characteristic of gut epithelial tissue.

Thus, the distribution of prednisone and its metabolite, prednisolone, throughout the entire three-organ system occurred across intact cellular barriers. Any influence from compromised barriers or technical difficulties can be excluded. This is crucial for the reliability of the system and the interpretability of the PK/PD data.

Furthermore, the measurement of three prednisone concentrations before and after perfusion within the cell-free triple setup and all cell-free single setups (Figure 2B) revealed no significant binding of the drug to the plastic surface of the biochip or the perfusion equipment. Measuring drug adsorption to the biochip periphery is particularly relevant for prednisone. Due to its hydrophobic nature, prednisone is prone to adsorption onto the materials of commonly used in biochips. This can result in lower drug concentrations, potentially leading to underestimation of drug effects or efficacy.

Additionally, no binding to serum components of the cell culture medium was detected (data not shown). Therefore, it can be assumed that the intended concentration of 55 nM (Frey and Frey, 1990) (corresponds to the C_max, the highest concentration of a drug in blood plasma that can be reached after administration) is available to the cells.

The prednisone distribution dynamics across the different compartments of the three-organ model as well as of single-perfused models were analyzed by sampling the supernatant at 10, 20, and 24 h, with its concentration and metabolization to prednisolone measured via HPLC-MS/MS (Figure 1B). Due to the higher medium volume in the triple system compared to single-perfused models with a constant volume at substance application, the concentration of prednisone was increased from 55 nM applied in single models to 142 nM in the triple model. The percentage distribution of prednisone and prednisolone in the individual compartments of the models is shown over time in Figure 3, the corresponding concentrations of each experiment are listed in Supplementary Table S2.

In the connected model prednisone was administered orally on the luminal side of the gut model. Ten hours after administration, 63% of the initial prednisone was still present in the gut compartment. A fraction of the prednisone had entered the circulation, and hepatic metabolism had already begun converting prednisone to prednisolone.

After an additional 10 h, the circulating prednisone levels had slightly increased, and the amount of prednisolone in the hepatic compartment had also risen, with prednisolone being transferred into the circulation.

Twenty-four hours post-oral administration, the prednisone level in the gut lumen further decreased, while its circulating levels continued to rise, along with the amount of prednisolone. However, prednisolone did not transfer into the fetal compartment of the placenta model and was only detected in minor quantities on the luminal side of the gut.

The single-gut model displayed similar kinetics in prednisone transfer from the apical to basolateral side without metabolism to prednisolone. Metabolism was exclusively detected in the liver model, where prednisone was applied on the basolateral side of the liver sinusoidal endothelial cells (LSECs). Only 22% of the initial prednisone was detected after 10 h, with further reduction over time. In the hepatocyte compartment, approximately one-quarter of the initial prednisone was metabolized to prednisolone, with another quarter found in the vascular area of the LSECs, increasing to 35% with longer incubation.

In the single-placenta model, prednisone was applied to the maternal side of BeWo cells, with the majority remaining in this area and only 10%–12% was detected on the fetal side. Although trophoblast cells can metabolize prednisone, the presence of 11β-hydroxysteroid dehydrogenase (11β-HSD2) is reported to convert prednisolone back to prednisone (Sarkar et al., 2001), (Ince-Askan et al., 2019). As expected, no detectable levels of prednisolone were found in our placenta model, this was observed exclusively in the liver model.

These findings indicate effective metabolism and transport of prednisone in the MPS, with distinct distribution patterns observed across the different compartments over time.

We successfully developed a digital twin to simulate the pharmacokinetics (PK) of prednisone using data derived from the MPS studies. Initially, model parameters were fitted to the observed kinetic data from MPS models representing biological functions—gut absorption, liver metabolism, and placental transfer (Figure 4). The digital twin simulations accurately captured the on-chip pharmacokinetics across all three single MPS.

Subsequently, we integrated these model parameters into the digital twin of the three-organ interactome (gut-liver-placenta). This allowed us to predict the overall pharmacokinetics of prednisone within the interconnected system and evaluate potential organ-organ crosstalk. The digital twin simulations closely aligned with the experimental measurements, though some discrepancies indicated that organ-organ interactions influenced the pharmacokinetics more than expected.

Following, the model parameters describing the key biological functions were integrated within the digital twin of the 3-way interactome to predict the on-chip kinetics and evaluate the potential for organ-organ crosstalk. When compared to the actual measurements, the 3-way digital twin captured the main kinetics in all three organs well, although some further improvements are possible, indicating that organ-organ crosstalk is changing the PK (Figure 5).

A whole-body human digital twin was implemented using PK-Sim^® to simulate prednisone pharmacokinetics in pregnancy. Initially, this model was qualified using published clinical data for a single 10 mg dose of prednisone. The reported liver clearance (Cl = 9 mL/min/kg) and gut permeability (P = 22E-6 cm/min) resulted in an overprediction of plasma concentrations. After refining these parameters (Cl = 10.1 mL/min/kg; P = 4.2E-6 cm/min), the simulation better matched the observed clinical data (red curve, Supplementary Figure 2.1).

Next, the parameters derived from the single MPS models were translated to human physiology and integrated into the pregnant-woman digital twin (Cl = 12.4 mL/min/kg; P = 2.9E-6 cm/min). The prediction based solely on MPS data produced a good approximation of observed plasma levels in pregnant women. We calculated the fetal/maternal maximum concentration ratio as a quality indicator, with literature reporting a ratio of 1.6 (range 0.9–2.5). Our MPS-derived data yielded a ratio of 1.3, which closely aligns with clinical reports, though the digital twin underpredicted this ratio at 0.5.

To further investigate fetal exposure, we simulated daily prednisone dosing of 10 mg over 3 months. In the maternal compartment, there was no accumulation of prednisone, with residual levels clearing within 24 h (Figure 6). On the fetal side, prednisone plasma levels showed some accumulation, reaching a steady-state concentration of approximately 38 nM, which remained below the reported clinical maximum of 55 nM. Likewise, the steady-state concentration of prednisolone in the simulated fetus is estimated to be around 47 nM, with a concentration maximum of around 62 nM. This is still an order of magnitude lower than reported averaged plasma concentration maxima of around 400 nM (Table).

These findings suggest that fetal exposure to prednisone and prednisolone is non-toxic at typical therapeutic doses, supporting the utility of this integrated system for early-stage drug safety assessments during pregnancy.

In conclusion, our integrated gut-liver-placenta model provides a powerful tool for investigating drug pharmacokinetics during pregnancy. By simulating the key processes of drug absorption, metabolism, and placental transfer, we demonstrated the system’s capability to predict fetal exposure to prednisone and its active metabolite prednisolone. Our results indicate the importance of selecting a context-of-use specific MPS, sampling schedule, and approach for translating to human pharmacology to accurately predict fetal compound exposures.

Looking ahead, the application of MPS technology—especially integrated multi-organ systems as demonstrated here—holds considerable potential for reducing the reliance on animal models in drug development. As more MPS are validated, their ability to accurately predict human outcomes will likely reduce the need for animal experimentation, addressing both ethical concerns and the high costs associated with traditional in vivo testing.