Skip to contents

Which parameters of a PK/PD model can the data determine? Structural and practical identifiability for ODE models written in rxode2.

pkident checks a pharmacokinetic or pharmacodynamic model before or after it is fitted:

  • Structure: with the chosen outputs and known doses, which parameters can be determined at all, and which combinations (such as CL/F) can?
  • Design: with these sampling times, doses and assay error, how precisely would each parameter be estimated, and which change to the design helps?
  • Fit: in an nlmixr2 fit, which parameters do the data actually determine, and is a flat profile a property of the model or of the data?

Every decision is local and says so. Anything the methods cannot decide is reported as unresolved rather than guessed, and two independent methods must agree before a structural decision is made.

Documentation: https://daikikumakura.github.io/pkident/ (function reference and three case studies)

Workflow

   model (rxode2 ODEs) + outputs + doses
                 │
   pkpd_model()                       describe what is observed and known
                 │
   structural_identifiability()       can the parameters be determined at all?
                 │
   practical_identifiability()        how precisely, with this design?
   compare_designs()                  which design is better?
                 │
   classify_profiles()                what did an nlmixr2 fit determine?
Question Function Answer Needs
Describe the model pkpd_model(), bolus_dose(), infusion_dose(), reference_model() a model object with outputs, doses and known quantities rxode2
Structure structural_identifiability() identifiable, combination_only, non_identifiable or unresolved per parameter, with the identifiable combinations symengine
Design practical_identifiability(), compare_designs(), experiment(), residual_error() expected RSE per parameter, condition number, weakest direction —
Fit classify_profiles() bounded, one-sided, flat or shallow profile per parameter, linked to the structural result nlmixr2est, nlmixr2extra
Report as_table() a data frame of any result —

Installation

install.packages("pkident", repos = c(
  "https://daikikumakura.r-universe.dev",
  "https://cloud.r-project.org"
))

Or from GitHub: remotes::install_github("DaikiKumakura/pkident"). pkident needs rxode2 and symengine; classify_profiles() also needs nlmixr2est and nlmixr2extra.

Quick start: structure

An oral one-compartment model with unknown bioavailability:

library(pkident)

m <- pkpd_model(
  "d/dt(depot)   = -ka*depot
   d/dt(central) = ka*depot - CL/V*central
   cp            = central/V",
  outputs = "cp",
  doses   = bolus_dose("depot", "F*DOSE"),
  known   = "DOSE"
)

structural_identifiability(m)
#> <pkident structural identifiability>
#> Decision is LOCAL (Lie derivatives at the known initial state up to order 8, checked by output sensitivities; 5 random points).
#> Methods agree at 5 of 5 points.
#>   CL           combination_only
#>   F            combination_only
#>   V            combination_only
#>   ka           identifiable
#> Identifiable combinations:
#>   CL/F
#>   CL/V

Only ka and the ratios CL/F and CL/V are determined. The two methods are Lie derivatives of the outputs at the known initial state (which keeps what a dose tells about volumes) and the rank and null space of output sensitivities solved with rxode2. Known input signals, such as a plasma concentration driving a PD model, are supported.

Design: how precise would the estimates be?

The same model with F known (reference_model("A3")), sampled three times in the first two hours, and then eight times up to 24 hours:

values <- c(ka = 1, CL = 2, V = 20, DOSE = 100)
error  <- residual_error(add = 0.05, prop = 0.1)

practical_identifiability(reference_model("A3"), design = c(0.5, 1, 2),
                          values = values, error = error)
#> Condition number (log scale): 36900
#>   CL           RSE    975.1%  unresolved
#>   V            RSE    224.2%  unresolved
#>   ka           RSE    257.2%  unresolved
#> Weakest direction: +0.94 log(CL) -0.25 log(ka) -0.22 log(V) (expected SD on log scale 10.3)

practical_identifiability(reference_model("A3"), design = c(0.5, 1, 2, 4, 6, 8, 12, 24),
                          values = values, error = error)
#> Condition number (log scale): 13.3
#>   CL           RSE      4.7%  identifiable
#>   V            RSE      7.1%  identifiable
#>   ka           RSE     14.6%  identifiable

Three early samples do not observe elimination. compare_designs() puts candidate designs side by side, with the structural status for the outputs each one observes. Doses can be event tables (for example repeated infusions), and residual error can be additive, proportional, combined or exponential.

Fit: what did the data determine?

classify_profiles(fit, structural = structural_identifiability(emax_model),
                  map = c(te0 = "E0", temax = "Emax", tec50 = "EC50"))   # excerpt
#>   temax  non_identifiable  open towards upper values (explored -0.298 to 3.7, 10 fits)
#>          -> structurally identifiable but not determined by these data (structural: identifiable)

Profiles are classified as bounded, one-sided, flat or shallow (Raue et al. 2009) and compared with the structural result, which tells a model problem from a data problem. A profile point is re-estimated from a neighbouring point when needed, because a stuck re-estimation can only overstate the objective function.

Validation

Check Result
Ten reference models with expected results fixed from the literature or by derivation (and one variant) All decisions as expected at four seeds; the two structural methods agree at every point
Fisher information against the analytic result Relative difference 1e-11 (also for repeated doses from an event table)
Expected RSEs against 1,000 repeated estimates on simulated data Ratio 0.96–1.08 in four scenarios; limitations at RSEs of 30–40% documented
Profile classification on three synthetic nlmixr2 fits All ten parameters classified as expected

Details are in validation/. R CMD check passes in CI.

Case studies

  • Structural identifiability of reference PK/PD models.
  • Nimotuzumab target-mediated disposition: with the public nimoData (nlmixr2data), clearance, volume and koff are determined, kon, kdeg, target baseline and kint are not, and measuring total target would determine six of the seven parameters.
  • PSA decline and regrowth: the regrowth rate becomes estimable only about twice the time to nadir after the start of treatment.

They are articles on the documentation site; the analysis scripts are in case-studies/.

Scope and limits

  • Decisions are local. A locally identifiable parameter can have a finite number of alternative values (for example the flip-flop of absorption and elimination).
  • Fixed effects of a typical individual. Identifiability of between-subject variability and population designs are not covered (see PopED for optimal design).
  • Closed-form models (linCmt()) must be written as ODEs.
  • The expected information is a local, linear approximation: reliable when RSEs are below about 30%; examine likelihood profiles otherwise.

Citation

Cite the archived version: https://doi.org/10.5281/zenodo.24210472. The citation metadata is in CITATION.cff.

License

GPL-3