Skip to contents

Model and source

van der Meijden and colleagues characterised the pharmacokinetics and pharmacodynamics of polymyxin B in neutropenic-mouse thigh and lung infection models for Escherichia coli and Klebsiella pneumoniae. The paper contributes two kinds of model to nlmixr2lib:

  1. A population PK model (vanderMeijden_2023_polymyxinB): a one-compartment, first-order-absorption model for subcutaneous polymyxin B in which apparent clearance and the absorption rate constant both decrease with the dose given at each administration (non-linear PK).
  2. Ten PK/PD models that couple the typical PK to an inhibitory sigmoid Emax model of the unbound-AUC-over-MIC index (fAUC0-24/MIC) describing the 24 h change in thigh bacterial burden. Eight come from the dose-response study of Figure 3 (four E. coli and four K. pneumoniae isolates); two come from the dose-fractionation study of Figure 2 (E. coli ATCC 25922 and K. pneumoniae 104).

The eight dose-response files are suffixed by isolate (vanderMeijden_2023_polymyxinB_EcoliATCC25922, ..._Ecoli15, ..._Ecoli51, ..._Ecoli107, ..._KpneumoniaeATCC43816, ..._Kpneumoniae17, ..._Kpneumoniae58, ..._Kpneumoniae74); the two dose-fractionation files carry a _fractionation suffix (..._EcoliATCC25922_fractionation, ..._Kpneumoniae104_fractionation). The inhibitory sigmoid Emax coefficients differ per isolate, so each is a separate file (the same policy as the parallel CLSI / EUCAST fits of Beredaki_2023_micafungin_*).

Population

The study used outbred female CD-1 mice (20-25 g, 7-8 weeks old) rendered neutropenic with cyclophosphamide and infected 2 h before dosing, either intramuscularly in each thigh or intranasally (lung). Polymyxin B was given as a single subcutaneous dose over 0.5-64 mg/kg for the PK experiments, and as q6h, q12h or q24h regimens (total daily dose 16-512 mg/kg) for the dose-fractionation PD study. Plasma was sampled at ten time points per dose level (two mice per time point) and assayed by LC-MS/MS (LLOQ 0.2 mg/L); samples below the LLOQ were omitted. The unbound fraction in mouse plasma, determined by ultrafiltration with correction for non-specific binding, was 20% and independent of concentration (Results, “Plasma protein binding”). Model parameters are expressed for a virtual 1 kg mouse, so clearance and volume are per kg and doses are in mg/kg.

The isolates and their polymyxin B MICs are in Table S1 and Table 2: four E. coli (ATCC 25922, 15, 51, 107; MIC 1 mg/L) and five K. pneumoniae (ATCC 43816 MIC 0.5; 17 MIC 1; 58 and 74 MIC 2; 104 MIC 1 mg/L), carrying various ESBL / carbapenemase resistance mechanisms.

The population metadata of each model records these facts programmatically, e.g. readModelDb("vanderMeijden_2023_polymyxinB")()$population.

Source trace

Equation / parameter Value Source location
lka (ka at 10 mg/kg) 1.63 /h Table 1
lcl (CL/F at 10 mg/kg) 0.384 L/h/kg Table 1
lvc (Vc/F) 0.774 L/kg Table 1
e_dose_ka (power of dose on ka) -0.407 Table 1
e_dose_cl (power of dose on CL) -0.26 Table 1
etalka (BSV ka) 79% CV Table 1
etalcl (BSV CL) 45% CV Table 1
expSd (residual) variance 0.093 Table 1
ka, CL dose dependence ka = 1.63*(DOSE/10)^-0.407, CL = 0.384*(DOSE/10)^-0.26 Table 1 footnote
fu (unbound fraction) 0.20 Results, “Plasma protein binding”
MIC per isolate 0.5-2 mg/L Table 2, Table S1
inhibitory sigmoid Emax e0, emax, ec50, hill digitised Figure 2 (fractionation), Figure 3 (dose-response)
stasis / 1-log / 2-log fAUC/MIC targets per isolate Table 2

The inhibitory sigmoid Emax coefficients are not printed in the paper (only the per-panel R^2 is), so the maintainers recovered them from the fitted curves drawn in Figures 2 and 3. For Figure 3 the curves were read back from the PDF vector graphics (least-squares residual RMSE 0.002-0.004 log10 cfu over the drawn line); for Figure 2 the curves were digitised from the 1400-pixel raster. The recovered curves reproduce the paper’s own Table 2 targets (see “Reproducing the Table 2 targets” below), which is the independent check on the digitisation.

Virtual cohort and PK simulation

# rxode2's simulation RNG is partitioned per solver thread, so a stochastic
# cohort is not reproducible across thread counts; the PK figures below use the
# typical-value model (zeroRe) where exact reproducibility is wanted, and a
# modest stochastic cohort only for the VPC.
set.seed(20231)

pk <- rxode2::rxode(readModelDb("vanderMeijden_2023_polymyxinB"))
#> ℹ parameter labels from comments will be replaced by 'label()'

# Single subcutaneous dose over the eight studied dose levels, one subject
# (one `id`) per dose so the dose levels do not collapse into a single subject
# receiving all eight doses. DOSE (the dose per administration, a model
# covariate) must appear on every record; the event columns come before DOSE so
# rxode2 does not read DOSE as the amount.
pk_doses <- c(0.5, 1, 2, 4, 8, 16, 32, 64)
obs_times <- seq(0, 8, by = 0.1)
make_pk <- function(D, id) {
  dplyr::bind_rows(
    tibble(id = id, time = 0, amt = D, evid = 1, cmt = "depot"),
    tibble(id = id, time = obs_times, amt = NA_real_, evid = 0, cmt = "depot")
  ) |>
    dplyr::mutate(dose_mgkg = D, DOSE = D)
}
pk_events <- dplyr::bind_rows(
  lapply(seq_along(pk_doses), function(i) make_pk(pk_doses[i], id = i))
)
stopifnot(!anyDuplicated(unique(pk_events[pk_events$evid == 1, c("id", "dose_mgkg")])))
# Replicates Figure 1 of van der Meijden 2023: typical plasma polymyxin B
# concentration versus time by dose. Tmax <= 2 h and the peak scales
# supra-proportionally because CL/F falls with dose.
sim_pk <- rxode2::rxSolve(
  rxode2::zeroRe(pk), events = pk_events, keep = "dose_mgkg"
) |>
  as.data.frame()
#> ℹ omega/sigma items treated as zero: 'etalka', 'etalcl'
#> Warning: multi-subject simulation without without 'omega'

sim_pk |>
  dplyr::mutate(dose = factor(dose_mgkg, levels = pk_doses)) |>
  ggplot(aes(time, Cc, colour = dose)) +
  geom_line() +
  geom_hline(yintercept = 0.2, linetype = "dashed") +
  scale_y_log10(limits = c(0.1, 100)) +
  labs(
    x = "Time (h)", y = "Polymyxin B (mg/L)", colour = "Dose (mg/kg)",
    title = "Figure 1 - typical plasma polymyxin B by dose",
    caption = "Replicates Figure 1 of van der Meijden 2023; dashed line LLOQ 0.2 mg/L."
  )
#> Warning in scale_y_log10(limits = c(0.1, 100)): log-10 transformation
#> introduced infinite values.
#> Warning: Removed 157 rows containing missing values or values outside the scale range
#> (`geom_line()`).

# A small stochastic cohort per dose for a VPC-style envelope (<= 200/arm).
make_pop <- function(D, n = 60, id_offset = 0L) {
  base <- tibble(id = id_offset + seq_len(n), dose_mgkg = D, DOSE = D)
  ev <- base |>
    group_by(id) |>
    reframe(
      time = c(0, seq(0, 8, by = 0.25)),
      amt = c(D, rep(NA_real_, length(seq(0, 8, by = 0.25)))),
      evid = c(1, rep(0, length(seq(0, 8, by = 0.25)))),
      cmt = "depot",
      dose_mgkg = dose_mgkg[1], DOSE = DOSE[1]
    )
  ev
}
pop_events <- dplyr::bind_rows(
  lapply(seq_along(pk_doses), function(i) make_pop(pk_doses[i], id_offset = (i - 1L) * 60L))
)
stopifnot(!anyDuplicated(unique(pop_events[, c("id", "time", "evid")])))

sim_pop <- rxode2::rxSolve(pk, events = pop_events, keep = "dose_mgkg") |>
  as.data.frame()

# VPC-style envelope (median and 5th-95th percentile) per dose, the summary
# Figure 1 of van der Meijden 2023 plots as mean +- SD.
sim_pop |>
  dplyr::filter(!is.na(Cc)) |>
  dplyr::group_by(dose_mgkg, time) |>
  dplyr::summarise(
    lo = quantile(Cc, 0.05), md = median(Cc), hi = quantile(Cc, 0.95),
    .groups = "drop"
  ) |>
  dplyr::mutate(dose = factor(dose_mgkg, levels = pk_doses)) |>
  ggplot(aes(time, md)) +
  geom_ribbon(aes(ymin = lo, ymax = hi), alpha = 0.25) +
  geom_line() +
  geom_hline(yintercept = 0.2, linetype = "dashed") +
  facet_wrap(~dose, ncol = 4) +
  scale_y_log10(limits = c(0.01, 150)) +
  labs(
    x = "Time (h)", y = "Polymyxin B (mg/L)",
    title = "Figure 1 - simulated exposure envelope by dose (5th-95th percentile)",
    caption = "Compare Figure 1 of van der Meijden 2023; dashed line LLOQ 0.2 mg/L."
  )
#> Warning in scale_y_log10(limits = c(0.01, 150)): log-10 transformation introduced infinite values.
#> log-10 transformation introduced infinite values.
#> log-10 transformation introduced infinite values.
#> log-10 transformation introduced infinite values.
#> Warning: Removed 73 rows containing missing values or values outside the scale range
#> (`geom_ribbon()`).
#> Warning: Removed 21 rows containing missing values or values outside the scale range
#> (`geom_line()`).

NCA validation against the digitised Figure S1

The paper’s Figure S1 plots the observed Cmax and AUC for each of the eight studied doses. The maintainers digitised those points to provide a numeric check on the simulated typical-value PK. We compute Cmax and AUC of the typical-value profiles with PKNCA and compare.

# Only `!is.na(Cc)`; keep the time-zero row so AUC0-* anchors correctly.
sim_nca <- sim_pk |>
  dplyr::filter(!is.na(Cc)) |>
  dplyr::mutate(treatment = factor(dose_mgkg, levels = pk_doses)) |>
  dplyr::select(id = dose_mgkg, time, Cc, treatment)

# Guarantee a time=0 row (extravascular pre-dose Cc = 0).
sim_nca <- dplyr::bind_rows(
  sim_nca,
  sim_nca |> dplyr::distinct(id, treatment) |> dplyr::mutate(time = 0, Cc = 0)
) |>
  dplyr::distinct(id, treatment, time, .keep_all = TRUE) |>
  dplyr::arrange(id, treatment, time)

conc_obj <- PKNCA::PKNCAconc(sim_nca, Cc ~ time | treatment + id)

dose_df <- pk_events |>
  dplyr::filter(!is.na(amt)) |>
  dplyr::mutate(treatment = factor(dose_mgkg, levels = pk_doses)) |>
  dplyr::select(id = dose_mgkg, time, amt, treatment)
dose_obj <- PKNCA::PKNCAdose(dose_df, amt ~ time | treatment + id)

intervals <- data.frame(start = 0, end = Inf, cmax = TRUE, aucinf.obs = TRUE)
nca_res <- PKNCA::pk.nca(PKNCA::PKNCAdata(conc_obj, dose_obj, intervals = intervals))
# Figure S1 (digitised by the maintainers from the 1400-pixel raster). AUC is
# total plasma AUC0-inf in mg*h/L; Cmax in mg/L.
published <- tibble::tribble(
  ~treatment, ~cmax,  ~aucinf.obs,
  "0.5",      0.549,  1.24,
  "1",        0.847,  1.68,
  "2",        1.39,   2.66,
  "4",        3.69,   6.97,
  "8",        2.29,   4.31,
  "16",       17.9,   51.2,
  "32",       28.7,   106,
  "64",       42.5,   252
)

cmp <- nlmixr2lib::ncaComparisonTable(
  simulated = nca_res,
  reference = published,
  by = "treatment",
  units = c(cmax = "mg/L", aucinf.obs = "mg*h/L"),
  tolerance_pct = 30
)

cmp |>
  dplyr::rename("Dose (mg/kg)" = treatment) |>
  knitr::kable(
    caption = paste(
      "Simulated typical-value NCA versus the digitised Figure S1 observed",
      "Cmax and AUC. * differs by more than 30%."
    ),
    digits = 3
  )
Simulated typical-value NCA versus the digitised Figure S1 observed Cmax and AUC. * differs by more than 30%.
NCA parameter Dose (mg/kg) Reference Simulated % diff
Cmax (mg/L) 0.5 0.549 0.433 -21.1%
Cmax (mg/L) 1 0.847 0.845 -0.3%
Cmax (mg/L) 2 1.39 1.65 +18.4%
Cmax (mg/L) 4 3.69 3.19 -13.4%
Cmax (mg/L) 8 2.29 6.2 +170.8%*
Cmax (mg/L) 16 17.9 12 -32.9%*
Cmax (mg/L) 32 28.7 23.2 -19.0%
Cmax (mg/L) 64 42.5 44.9 +5.6%
AUC0-∞ (obs) (mg*h/L) 0.5 1.24 0.594 -52.1%*
AUC0-∞ (obs) (mg*h/L) 1 1.68 1.43 -15.1%
AUC0-∞ (obs) (mg*h/L) 2 2.66 3.42 +28.6%
AUC0-∞ (obs) (mg*h/L) 4 6.97 8.2 +17.6%
AUC0-∞ (obs) (mg*h/L) 8 4.31 19.6 +355.8%*
AUC0-∞ (obs) (mg*h/L) 16 51.2 47.1 -8.0%
AUC0-∞ (obs) (mg*h/L) 32 106 113 +6.6%
AUC0-∞ (obs) (mg*h/L) 64 252 272 +8.0%

Figure S1 plots the observed Cmax and AUC of individual mice (two per dose), whose scatter is real, whereas the simulated values are the typical-value model. Seven of the eight doses agree to within about a third. The outlier is the 8 mg/kg group, whose observed Cmax (2.3 mg/L) sits below the 4 mg/kg group (3.7 mg/L) and well under the dose trend: this below-trend 8 mg/kg thigh profile is visible in the main-text Figure 1 as well, so it is a feature of the sparse data (two mice per dose, and a between-subject CV of 79% on the absorption rate), not an extraction error. The typical-value model is monotonic in dose and therefore cannot reproduce that single low group; it reproduces the overall exposure scale and the supra-proportional rise with dose. The comparison uses a 30% tolerance and is a sanity check on the exposure scale, not a precision claim.

Reproducing the Table 2 targets

Each PK/PD model couples the typical PK to the inhibitory sigmoid Emax model of fAUC0-24/MIC. The internal state fauc_0_24 integrates the unbound concentration over the first 24 h of treatment; fauc_mic divides it by the isolate MIC; log_cfu is the 24 h bacterial burden (start-of-therapy burden plus the Emax change dlog10cfu). For the dose-response isolates the paper reports the fAUC0-24/MIC producing stasis, a 1-log kill and a 2-log kill (Table 2); feeding the recovered Emax curve those index values must return 0, -1 and -2 log10 cfu.

targets <- tibble::tribble(
  ~model,                                             ~stasis, ~kill1, ~kill2,
  "vanderMeijden_2023_polymyxinB_EcoliATCC25922",     95.3,    NA,     NA,
  "vanderMeijden_2023_polymyxinB_Ecoli15",            102.6,   NA,     NA,
  "vanderMeijden_2023_polymyxinB_Ecoli51",            31.6,    46.8,   65.3,
  "vanderMeijden_2023_polymyxinB_Ecoli107",           27.5,    54.5,   NA,
  "vanderMeijden_2023_polymyxinB_KpneumoniaeATCC43816", 60.5,  NA,     NA,
  "vanderMeijden_2023_polymyxinB_Kpneumoniae17",      5.9,     NA,     NA,
  "vanderMeijden_2023_polymyxinB_Kpneumoniae58",      11.6,    25.2,   NA,
  "vanderMeijden_2023_polymyxinB_Kpneumoniae74",      11.6,    NA,     NA
)

# Evaluate the Emax curve directly from each model's ini() so the check does
# not depend on the ODE solve: dlog10cfu(index) = e0 - emax*I^h/(I^h+ec50^h).
# Read each model's thetas ONCE (parsing a model is expensive) and cache them.
emax_pars <- lapply(targets$model, function(mn) {
  th <- rxode2::rxode(readModelDb(mn))$theta
  list(e0 = exp(th[["le0"]]), emax = exp(th[["lemax"]]),
       ec50 = exp(th[["lec50"]]), h = exp(th[["lhill"]]))
})
names(emax_pars) <- targets$model

emax_change <- function(model_name, index) {
  p <- emax_pars[[model_name]]
  p$e0 - p$emax * index^p$h / (index^p$h + p$ec50^p$h)
}

check <- targets |>
  rowwise() |>
  mutate(
    at_stasis = emax_change(model, stasis),
    at_kill1 = if (is.na(kill1)) NA_real_ else emax_change(model, kill1),
    at_kill2 = if (is.na(kill2)) NA_real_ else emax_change(model, kill2)
  ) |>
  ungroup()

# Each attained change should match the nominal effect size (0, -1, -2) within
# the digitisation error of the Figure 3 curves.
stopifnot(
  max(abs(check$at_stasis - 0), na.rm = TRUE) < 0.2,
  max(abs(check$at_kill1 - (-1)), na.rm = TRUE) < 0.2,
  max(abs(check$at_kill2 - (-2)), na.rm = TRUE) < 0.2
)

check |>
  transmute(
    Isolate = sub("vanderMeijden_2023_polymyxinB_", "", model),
    `Stasis fAUC/MIC` = stasis, `dlog10cfu @ stasis` = round(at_stasis, 3),
    `1-log fAUC/MIC` = kill1, `dlog10cfu @ 1-log` = round(at_kill1, 3),
    `2-log fAUC/MIC` = kill2, `dlog10cfu @ 2-log` = round(at_kill2, 3)
  ) |>
  knitr::kable(
    caption = paste(
      "The recovered inhibitory sigmoid Emax curve returns the nominal effect",
      "size at each Table 2 fAUC/MIC target, confirming the digitisation."
    )
  )
The recovered inhibitory sigmoid Emax curve returns the nominal effect size at each Table 2 fAUC/MIC target, confirming the digitisation.
Isolate Stasis fAUC/MIC dlog10cfu @ stasis 1-log fAUC/MIC dlog10cfu @ 1-log 2-log fAUC/MIC dlog10cfu @ 2-log
EcoliATCC25922 95.3 -0.008 NA NA NA NA
Ecoli15 102.6 -0.007 NA NA NA NA
Ecoli51 31.6 -0.006 46.8 -1.013 65.3 -2.007
Ecoli107 27.5 -0.004 54.5 -1.009 NA NA
KpneumoniaeATCC43816 60.5 -0.004 NA NA NA NA
Kpneumoniae17 5.9 -0.001 NA NA NA NA
Kpneumoniae58 11.6 -0.002 25.2 -1.004 NA NA
Kpneumoniae74 11.6 -0.003 NA NA NA NA

Replicating the exposure-response curves (Figures 2 and 3)

# Replicates Figure 3: dlog10cfu versus fAUC/MIC for each dose-response isolate.
dose_response_models <- check$model
index_grid <- 10^seq(-1, 3, length.out = 200)

curves <- dplyr::bind_rows(lapply(dose_response_models, function(mn) {
  tibble(
    isolate = sub("vanderMeijden_2023_polymyxinB_", "", mn),
    fauc_mic = index_grid,
    dlog10cfu = vapply(index_grid, function(i) emax_change(mn, i), numeric(1))
  )
}))

ggplot(curves, aes(fauc_mic, dlog10cfu)) +
  geom_line() +
  geom_hline(yintercept = 0, linetype = "dotted") +
  facet_wrap(~isolate, ncol = 4) +
  scale_x_log10() +
  labs(
    x = "fAUC/MIC", y = expression(Delta ~ log[10] ~ cfu),
    title = "Figure 3 - exposure-response by isolate",
    caption = "Replicates Figure 3 of van der Meijden 2023 (recovered curves)."
  )

Running the full PK/PD model end to end

The PK/PD files are driven by a dosing regimen: the PK layer produces the index fauc_mic, and log_cfu at t = 24 h is the predicted 24 h burden. For a single dose-response arm (q6h x 4 doses of D mg/kg), the attained fAUC/MIC and burden change are:

pd <- rxode2::rxode(readModelDb("vanderMeijden_2023_polymyxinB_Ecoli51"))
run_regimen <- function(D) {
  ev <- as.data.frame(
    et(amt = D, cmt = "depot", time = seq(0, 18, by = 6)) |> et(24)
  ) |>
    dplyr::mutate(DOSE = D)
  s <- rxode2::rxSolve(pd, ev, atol = 1e-10, rtol = 1e-10, returnType = "data.frame")
  tail(s, 1)
}
end <- dplyr::bind_rows(lapply(c(1, 4, 16, 32), function(D) {
  r <- run_regimen(D)
  tibble(dose_q6h = D, fauc_mic_24h = round(r$fauc_mic, 2), dlog10cfu = round(r$log_cfu - 7.6, 3))
}))
knitr::kable(end, caption = "E. coli 51: fAUC/MIC and 24 h burden change by q6h dose.")
E. coli 51: fAUC/MIC and 24 h burden change by q6h dose.
dose_q6h fauc_mic_24h dlog10cfu
1 1.14 0.329
4 6.51 0.294
16 36.59 -1.140
32 85.84 -3.517

The 32 mg/kg q6h regimen gives fAUC/MIC near 86, past the isolate’s 2-log-kill target of 65.3, so the predicted change approaches the -4 log10 cfu floor.

Assumptions and deviations

  • Inhibitory sigmoid Emax coefficients are figure-derived. The paper prints only the per-panel R^2, so the maintainers recovered e0, emax, ec50 and hill from the fitted curves drawn in Figures 2 and 3 (Figure 3 from the PDF vector paths, RMSE 0.002-0.004 log10 cfu; Figure 2 from the 1400-pixel raster). The recovered curves reproduce the paper’s Table 2 fAUC/MIC targets to within 0.7% (checked in “Reproducing the Table 2 targets”).
  • Lower plateau fixed at -4 for the E. coli dose-response curves. The free four-parameter read-back returned a bottom of -3.97 to -4.01 log10 cfu for all four E. coli isolates, consistent with the GraphPad fit having constrained the bottom to -4 (the paper does not state this). The K. pneumoniae dose-response curves plateau above -4 and their bottoms were left free.
  • Start-of-therapy burden. log_cfu adds the digitised Emax change to the Table 2 start-of-therapy burden for each dose-response isolate (6.8-7.6 log10 cfu/thigh); the two dose-fractionation models use the 7.18 log10 cfu/thigh average burden reported in Results (“Dose fractionation study”). These anchor the absolute burden but do not affect the change dlog10cfu.
  • No between-subject variability on the PK/PD models. The authors computed the fAUC/MIC indices from the typical population PK, so the PK/PD files carry no etas; only the standalone population PK model carries the Table 1 BSV on ka and
    1. Residual error on log_cfu is held at 0 because the paper reports only R^2.
  • Infection-site covariate omitted. The paper tested infection site (thigh vs lung) on PK and found no effect (Results, “Pharmacokinetic modelling”), so it is not a model covariate.
  • PK/PD restricted to the thigh model. The dose-response and dose-fractionation PD studies were done only in the thigh model; polymyxin B ELF penetration was too low to support a lung-model PD fit (Results, “PK”).
  • No correction notice for this article was found on the journal landing page or in PubMed/EuropePMC as of 2026-10-26.