Gut–Liver Chip Simulations Show Why Sampling Design Changes Estimated Drug Parameters
A Roche–Manchester tutorial uses mechanistic simulations to plan gut–liver chip experiments. Its midazolam example shows that sampling sufficient for parent-drug parameters can still give poor estimates of metabolite clearance, with verification performed on simulated experiments.
Experimental Geometry Becomes Part of the Pharmacokinetic Model
A tutorial by Roche and University of Manchester researchers examines how to obtain interpretable drug absorption and metabolism parameters from gut–liver organ chips. The authors use a compartmental mathematical model to connect intestinal transport, hepatic metabolism and recirculating medium. Their examples concern planning and analyzing these experiments, with performance evaluated through simulation rather than a new clinical prediction study.
Chip geometry affects what can be measured. In a low-permeability example, with apparent permeability of 0.4 nm/s, increasing the intestinal surface area accelerated appearance of the compound in the basal compartment. Using the corresponding manufacturer-recommended volumes, the modeled concentration exceeded a 0.001 μM quantification limit after one hour with a 4.67 cm² insert, whereas a 0.143 cm² insert required at least 24 hours. The comparison changed surface area and its associated volume together.
A separate analysis varied apical volume at a fixed surface area. A smaller volume made apical depletion easier to observe. However, a highly permeable compound can disappear rapidly from that compartment, and residual liquid near cells can complicate estimation of very small volumes. The useful dimensions therefore depend on the drug and the measurement question.
Repeated sampling and evaporation create different disturbances. Sampling removes both drug and medium, changing the amount present without an immediate concentration change under the model’s mixing assumptions. Evaporation removes medium continuously and can raise concentration. Both increase the cell-to-medium ratio, potentially biasing apparent clearance if omitted from the analysis. The tutorial uses mycophenolic acid simulations to illustrate these effects rather than assigning a universal correction.
Donor Activity and Prior Data Define the Uncertainty
The authors examined metabolic-activity information from a commercial database containing 128 hepatocyte donors and a separate published enterocyte dataset covering 24 donors. Not every enzyme had measurements for every donor. These were reference activity datasets, not 128 or 24 newly performed chip experiments.
The proposed design workflow distinguishes prior measurements made with the same donor, different donors with known enzyme activities, microsomal data requiring unit conversion, and different donors with unknown activity. The last scenario carries additional uncertainty from the activity distribution. These distinctions determine how previous clearance estimates can inform the first chip experiment.
Before selecting sampling times, the mathematical model must also be structurally identifiable from the compartments that can be dosed and sampled. DAISY, a structural-identifiability analysis program, was used for this assessment. Structural identifiability alone does not guarantee precise parameter estimates from noisy experimental data.
Midazolam and Its Metabolite Need Information at Different Times
The worked example tracks midazolam and 1′-hydroxymidazolam in apical, basal and liver compartments. Five parameters are estimated: permeability of the parent and metabolite, intestinal and hepatic parent-drug clearance, and hepatic metabolite clearance. The design permits three samples per compartment, with different sampled volumes for the apical and other compartments.
The investigators generated 2,500 possible concentration profiles using prior drug information and uncertainty in donor activity, cell numbers, flow and other inputs. A dynamic sensitivity analysis identified when each concentration measurement was informative about each parameter. Early measurements helped define parent-drug handling, while later measurements supplied more information about metabolite clearance. Analytical quantification limits also constrained usable sampling times.
Verification selected 100 simulated experiments and fitted them using three replicate analyses with added residual variability. With optimized sampling, the reported relative root mean squared errors were no greater than 32%. The estimated hepatic clearance of 1′-hydroxymidazolam was particularly sensitive to the final liver sampling time.
Moving that sample from 24 hours to eight hours, with the other sampling times unchanged, increased the metabolite-clearance relative error from 32% to 75%. Its absolute average fold error increased from 1.2 to 3.2. Thus, an apparently shorter and simpler experiment could retain useful parent-drug information while weakening the estimate for a different biological process.
These comparisons test recovery of known parameters from synthetic observations generated with the same model structure. They do not establish the same accuracy for an independent wet-lab experiment. The tutorial calls for a separate design assessment for each drug, donor and device combination, followed by model selection and uncertainty evaluation against the actual experiment. The reported R-based workflow used RxODE for simulation and Phoenix/RsNLME for fitting.