Skip to contents

Model and source

  • Citation: Mehta P, Soliman A, Rodriguez-Vera L, Schmidt S, Muniz P, Rodriguez M, Forcadell M, Gonzalez-Perez E, Vozmediano V. Interspecies brain PBPK modeling platform to predict passive transport through the blood-brain barrier and assess target site disposition. Pharmaceutics. 2024;16(2):226. doi:10.3390/pharmaceutics16020226.
  • Article: https://doi.org/10.3390/pharmaceutics16020226 (open access)
  • Supplementary material: “Supplementary_Model Code - Rats & Humans” (the Pumas 2.2.0 @model blocks for rat and human, with the Rodgers and Rowland, Poulin and Theil and Berezhkovskiy partition-coefficient functions and the human virtual-population equations) and “Supplementary figures and tables” (Tables S1-S3).
  • Model files:
    • Mehta_2024_acetaminophen_brain_pbpk: PBPK (whole-body, permeability-limited 4-compartment brain, Pumas 2.2.0). Acetaminophen in adults after IV dosing (Mehta et al. 2024, the human acetaminophen instantiation of an interspecies brain PBPK platform). Thirteen perfusion-limited body compartments (venous and arterial blood, lung, adipose, bone, heart, kidney, muscle, skin, spleen, gut, liver, rest of body) with Rodgers and Rowland tissue partition coefficients computed from adult tissue composition, plus the Verscheijden 2019 brain model (brain blood, brain mass, cranial CSF, spinal CSF) with passive blood-brain and blood-CSF barrier permeability. The blood-brain barrier permeability was optimised on rat neuropharmacokinetic data and scaled to humans by BBB surface area. Organ volumes and blood flows are age-, sex-, weight- and height-dependent adult equations. Total plasma clearance acts on venous blood. Forward-simulation model: all parameters are fixed; IIV is on clearance and the CSF flows; no residual error was reported.
    • Mehta_2024_oxycodone_brain_pbpk: PBPK (whole-body, permeability-limited 4-compartment brain, Pumas 2.2.0). Oxycodone in adults after IV dosing (Mehta et al. 2024, the human oxycodone instantiation of an interspecies brain PBPK platform). Thirteen perfusion-limited body compartments (venous and arterial blood, lung, adipose, bone, heart, kidney, muscle, skin, spleen, gut, liver, rest of body) with Poulin and Theil tissue partition coefficients computed from adult tissue composition, plus the Verscheijden 2019 brain model (brain blood, brain mass, cranial CSF, spinal CSF) with passive blood-brain and blood-CSF barrier permeability. The blood-brain barrier permeability was optimised on rat neuropharmacokinetic data and scaled to humans by BBB surface area. Organ volumes and blood flows are age-, sex-, weight- and height-dependent adult equations. Total plasma clearance acts on venous blood. Forward-simulation model: all parameters are fixed; IIV is on clearance and the CSF flows; no residual error was reported. With the published Poulin and Theil inputs the terminal half-life is about 1.3 h, shorter than the about 3 h of the paper’s Figure 5C (see the vignette).

Mehta et al. qualified a whole-body physiologically based pharmacokinetic (PBPK) platform with a four-compartment, permeability-limited brain (brain blood, brain mass, cranial CSF, spinal CSF; the Verscheijden 2019 framework) for drugs that cross the blood-brain barrier (BBB) by passive diffusion. For each of five drugs (acetaminophen, oxycodone, lacosamide, ibuprofen, levetiracetam) the BBB permeability was optimised on rat neuropharmacokinetic data and scaled to humans through the BBB surface area (Equation 1, Table 3). The human model then predicted plasma and CSF concentrations without further fitting (Table 4, Figure 5).

Two of the ten drug-species instantiations are fully reported: human acetaminophen and human oxycodone, both given intravenously. These are the two model files here. Each one carries the paper’s headline result on its own, because the rat-optimised permeability enters the human model as the tabulated human PSB (Table 3), not as a rat model output. The other instantiations cannot be rebuilt from the published material (see “Instantiations not reproduced” below).

Model structure

 venous blood --> lung --> arterial blood --> 10 perfused tissues --> venous blood
      |  CL x C_plasma             |                (spleen, gut -> liver)
      v                            +--> brain blood <--PSb--> brain mass
   eliminated                             ^   PSc        PSe  |  bulk flow
                                          +--> cranial CSF <--+
                                                    ^ |
                                         spinal CSF -+ +- sink flows to blood
  • 17 amount states (venous, arterial, lung, adipose, bone, heart, kidney, muscle, skin, spleen, other, gut, liver, brain_vascular, brain, brain_csf_sas_cranial, brain_csf_sas_spinal). The supplementary code writes the same equations in concentrations; here each is multiplied through by its volume.
  • IV doses go into venous (Methods 2.3.1).
  • Tissue:plasma partition coefficients are computed inside model() from the adult tissue composition of Table S2: Rodgers and Rowland for acetaminophen (monoprotic acid) and Poulin and Theil for oxycodone (monoprotic base), as in Table S3 and Methods 2.4. The rest-of-body Kp is the mean of the ten tissue values (code Kprest_human).
  • Outputs: Cc (plasma), Ccsf (cranial CSF), Ccsf_spinal (spinal CSF; the code’s Cspinal_csf) and Cecf (the code’s ECF approximation, Cbm x Vecf / Vbm with Vecf = 0.29 L).

Population

The human verification studies (Table 1) were:

  • Acetaminophen: Singla et al., 1000 mg IV, n = 50, 19-73 y, plasma and CSF.
  • Oxycodone: Kokki et al., 0.092 mg/kg IV, n = 11, 26-60 y, plasma and CSF.

No human parameter was estimated. The human data served only to check the predictions. The supplementary code builds male and female virtual subjects from the age-to-height and height-to-weight equations of Table S2, with log-normal variability. The number of virtual subjects and the sex ratio are placeholders (N = ___, NF = ___) in the code.

str(readModelDb("Mehta_2024_acetaminophen_brain_pbpk")()$population)
#> List of 10
#>  $ species       : chr "human"
#>  $ n_subjects    : int 50
#>  $ n_studies     : int 1
#>  $ age_range     : chr "19-73 years (Singla 2012, Table 1)"
#>  $ weight_range  : chr "not reported; generated from age and height in the virtual population"
#>  $ sex_female_pct: num NA
#>  $ disease_state : chr "Adults with plasma and CSF sampling after a single IV acetaminophen dose (Singla et al. 2012, used for verification only)."
#>  $ dose_range    : chr "1000 mg IV, single dose"
#>  $ regions       : chr NA
#>  $ notes         : chr "Forward-simulation PBPK: no human parameter was estimated. Clearance and binding come from the literature (Tabl"| __truncated__
str(readModelDb("Mehta_2024_oxycodone_brain_pbpk")()$population)
#> List of 10
#>  $ species       : chr "human"
#>  $ n_subjects    : int 11
#>  $ n_studies     : int 1
#>  $ age_range     : chr "26-60 years (Kokki et al., Table 1)"
#>  $ weight_range  : chr "not reported; generated from age and height in the virtual population"
#>  $ sex_female_pct: num NA
#>  $ disease_state : chr "Adults with plasma and CSF sampling after a single IV oxycodone dose (Kokki et al., used for verification only)."
#>  $ dose_range    : chr "0.092 mg/kg IV, single dose"
#>  $ regions       : chr NA
#>  $ notes         : chr "Forward-simulation PBPK: no human parameter was estimated. Clearance and binding come from the literature (Tabl"| __truncated__

Source trace

Model element Acetaminophen Oxycodone Source
Plasma CL (L/h) 22.8 49.2 x 0.8 = 39.36 Table S3, ‘CL - Human’ and ‘CL-scalar’
IIV on CL (CV) 35.4% 30% Table S3; Methods 2.4
log P / pKa 0.46 / 9.38 (acid) 1.2 / 8.53 (base) Table S3
Blood:plasma ratio 1.09 1.3 Table S3, BP Humans
fu plasma / brain mass / cranial CSF 0.855 / 0.747 / 1 0.59 / 0.45 / 1 Table S3
Kp method Rodgers and Rowland Poulin and Theil Table S3; Methods 2.4
PSB (L/h) 5.98 14.47 Table 3, PSB in Humans
PSC 0.1 x PSB 0.1 x PSB Table S3 footnote beta; Methods 2.3.2
PSE (L/h) 300 300 Table S3
CSF production (L/h), CV 0.021, 10% same Table S2
Bulk, spinal sink, spinal outflow 0.25 Qprod (CV 8%), 0.38 (0.75 Qprod + Qbulk) (CV 30%), 0.9 Qssink (CV 100%) same Table S2; code
Cranial / spinal CSF volume Vbrain x 0.105 (male) or 0.092 (female) x 0.8 / 0.2 same Table S2
Organ volumes, cardiac output, tissue flows adult equations same Table S2; code, human section
Tissue composition Rodgers and Rowland column Poulin and Theil column Table S2
Kp equations calcKp_RR, acid branch calcKp_PT code, human section
ODEs and outputs @dynamics, @derived same code, human section
Vecf 0.29 L 0.29 L code (Vecf = 0.29)
Virtual-population height, weight, BSA Table S2 equations same Table S2; code

Virtual cohorts

Height and weight are drawn with the sex-specific equations of Table S2 (as coded in the supplement): height = (175.32 + 0.1113 age - 0.0025 age^2) x exp(N(0, 0.039)) for men and (161.66 + 0.1319 age - 0.0027 age^2) x exp(N(0, 0.039)) for women; weight = exp(2.643 + 0.0099 height) x exp(N(0, 0.15)) for men and exp(2.7383 + 0.0091 height) x exp(N(0, 0.188)) for women. Age is uniform over the study’s range (the code draws rand(Uniform(...))). Each arm has 100 subjects, half of them women.

make_cohort <- function(n, age_lo, age_hi, drug, id0 = 0) {
  female <- rep(c(0, 1), length.out = n)
  age <- runif(n, age_lo, age_hi)
  ht_typ <- ifelse(
    female == 1,
    161.66 + 0.1319 * age - 0.0027 * age^2,
    175.32 + 0.1113 * age - 0.0025 * age^2
  )
  ht <- ht_typ * exp(rnorm(n, 0, 0.039))
  wt_typ <- ifelse(female == 1, exp(2.7383 + 0.0091 * ht), exp(2.643 + 0.0099 * ht))
  wt <- wt_typ * exp(rnorm(n, 0, ifelse(female == 1, 0.188, 0.15)))
  data.frame(id = id0 + seq_len(n), drug = drug, AGE = age, WT = wt, HT = ht, SEXF = female)
}

# One IV bolus into venous blood per subject plus an observation grid on the
# venous state. The grid is fine over the first 0.1 h, where the bolus spike
# in the 2-3 L venous compartment decays, so that the trapezoidal AUC is
# accurate; after that it follows the code's 0.1-h step (obstimes = 0:0.1:24).
make_events <- function(cohort, dose_mg, t_end) {
  obs_times <- sort(unique(c(seq(0, 0.1, by = 0.0025), seq(0.1, t_end, by = 0.1))))
  dose <- cohort |>
    mutate(time = 0, amt = dose_mg, evid = 1L, cmt = "venous")
  obs <- tidyr::crossing(cohort, time = obs_times) |>
    mutate(amt = 0, evid = 0L, cmt = "venous")
  bind_rows(dose, obs) |> arrange(id, time, desc(evid))
}
rxode2::rxSetSeed(20240226)
set.seed(20240226)
n_arm <- 100
coh_apap <- make_cohort(n_arm, 19, 73, "Acetaminophen 1000 mg IV")
coh_oxy <- make_cohort(n_arm, 26, 60, "Oxycodone 0.092 mg/kg IV", id0 = 1000)

ev_apap <- make_events(coh_apap, 1000, 12)
ev_oxy <- make_events(coh_oxy, 0.092 * coh_oxy$WT, 24)

bind_rows(coh_apap, coh_oxy) |>
  group_by(drug) |>
  summarise(
    n = n(), `female (%)` = 100 * mean(SEXF),
    `age (y)` = sprintf("%.1f (%.1f-%.1f)", median(AGE), min(AGE), max(AGE)),
    `weight (kg)` = sprintf("%.1f (%.1f-%.1f)", median(WT), min(WT), max(WT)),
    `height (cm)` = sprintf("%.0f (%.0f-%.0f)", median(HT), min(HT), max(HT)),
    .groups = "drop"
  ) |>
  knitr::kable(caption = "Virtual cohorts: median (range).")
Virtual cohorts: median (range).
drug n female (%) age (y) weight (kg) height (cm)
Acetaminophen 1000 mg IV 100 50 48.0 (19.2-72.7) 75.5 (41.4-116.8) 168 (143-193)
Oxycodone 0.092 mg/kg IV 100 50 44.3 (26.4-59.2) 75.8 (37.2-125.9) 169 (150-189)

Simulation

mod_apap <- readModelDb("Mehta_2024_acetaminophen_brain_pbpk")
mod_oxy <- readModelDb("Mehta_2024_oxycodone_brain_pbpk")
out_cols <- c("Cc", "Ccsf", "Ccsf_spinal", "Cecf")

sim <- bind_rows(
  rxode2::rxSolve(mod_apap, ev_apap, keep = "drug", returnType = "data.frame") |>
    select(id, time, drug, all_of(out_cols), cl),
  rxode2::rxSolve(mod_oxy, ev_oxy, keep = "drug", returnType = "data.frame") |>
    select(id, time, drug, all_of(out_cols), cl)
)
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ parameter labels from comments will be replaced by 'label()'
stopifnot(!anyNA(sim$Cc), !anyNA(sim$Ccsf_spinal))

# The paper plots the simulated mean, the 5th-95th percentile band and the
# minimum and maximum (Figure 5 caption).
summ <- sim |>
  pivot_longer(all_of(out_cols), names_to = "output", values_to = "conc") |>
  group_by(drug, output, time) |>
  summarise(
    mean = mean(conc), lo = quantile(conc, 0.05), hi = quantile(conc, 0.95),
    min = min(conc), max = max(conc),
    .groups = "drop"
  )

plot_fig5 <- function(d, title, ylab, log_y = FALSE) {
  p <- ggplot(d, aes(time, mean)) +
    geom_ribbon(aes(ymin = lo, ymax = hi), fill = "grey75") +
    geom_line(aes(y = min), linetype = "dashed") +
    geom_line(aes(y = max), linetype = "dashed") +
    geom_line(colour = "darkred", linewidth = 0.8) +
    labs(
      x = "Time (h)", y = ylab, title = title,
      caption = "Red: mean; band: 5th-95th percentile; dashed: minimum and maximum."
    )
  if (log_y) p <- p + scale_y_log10()
  p
}

# Log-scale panels: the integrator can return a value a few atol below zero
# in the far tail of the fastest-clearing subject. Values more than four
# decades below the peak mean are left out of the plot (a gap in the line).
floor_for_log <- function(d) {
  stopifnot(all(d$min >= -1e-6 * max(d$max)))
  cut <- 1e-4 * max(d$mean)
  mutate(d, min = ifelse(min > cut, min, NA_real_), lo = pmax(lo, cut))
}

Replicate published figures

The plots start at the first grid point after the bolus (0.1 h), as the venous concentration at time zero is the dose spread over venous blood only.

Figure 5A-B – acetaminophen 1000 mg IV

summ |>
  filter(drug == "Acetaminophen 1000 mg IV", output == "Cc", time >= 0.1) |>
  plot_fig5("Plasma acetaminophen (Figure 5A)", "Concentration (mg/L)")

summ |>
  filter(drug == "Acetaminophen 1000 mg IV", output == "Ccsf_spinal") |>
  plot_fig5("CSF acetaminophen (Figure 5B)", "Concentration (mg/L)")

Replicates Figure 5A-B of Mehta 2024. The CSF panel is the spinal CSF compartment, the code’s Cspinal_csf output; the clinical samples were lumbar CSF.

Figure 5C-D – oxycodone 0.092 mg/kg IV

summ |>
  filter(drug == "Oxycodone 0.092 mg/kg IV", output == "Cc", time >= 0.1) |>
  mutate(across(c(mean, lo, hi, min, max), ~ .x * 1000)) |>
  floor_for_log() |>
  plot_fig5("Plasma oxycodone (Figure 5C)", "Concentration (ug/L)", log_y = TRUE)
#> Warning: Removed 175 rows containing missing values or values outside the scale range
#> (`geom_line()`).

summ |>
  filter(drug == "Oxycodone 0.092 mg/kg IV", output == "Ccsf_spinal", time >= 0.1) |>
  mutate(across(c(mean, lo, hi, min, max), ~ .x * 1000)) |>
  floor_for_log() |>
  plot_fig5("CSF oxycodone (Figure 5D)", "Concentration (ug/L)", log_y = TRUE)
#> Warning: Removed 145 rows containing missing values or values outside the scale range
#> (`geom_line()`).

Replicates Figure 5C-D of Mehta 2024 (log scale, as published). The early part of both panels agrees with the publication, but the simulated concentrations fall faster after about 4 h. At 24 h the published mean is about 0.3 ug/L in plasma and in CSF (read from Figure 5C-D by the maintainers), and the mean here is one to two decades lower. The section “Volume of distribution and terminal phase” below gives the reason.

CNS compartments

summ |>
  filter(output %in% c("Cc", "Ccsf", "Ccsf_spinal", "Cecf"), time >= 0.1) |>
  mutate(output = factor(output,
    levels = c("Cc", "Ccsf", "Ccsf_spinal", "Cecf"),
    labels = c("Plasma", "Cranial CSF", "Spinal CSF", "ECF (Cecf)")
  )) |>
  ggplot(aes(time, mean, colour = output)) +
  geom_line() +
  facet_wrap(~drug, scales = "free") +
  scale_y_log10() +
  labs(x = "Time (h)", y = "Mean concentration (mg/L)", colour = NULL)

The cranial CSF follows plasma closely because the brain mass and the cranial CSF exchange freely (PSE = 300 L/h) and the brain mass equilibrates with brain blood through the BBB. The spinal CSF lags behind, as it is fed only by the CSF shuttle flow from the cranial CSF. Cecf is the code’s ECF approximation and is lower than the unbound brain-mass concentration by the factor Vecf / Vbm / fu_br.

Deterministic checks

Partition coefficients

The paper does not print the Kp values. The table lists those of a typical man (40 y, 175 cm, 75 kg) so a reader can compare them with other implementations of the same methods.

typ <- data.frame(id = 1, AGE = 40, WT = 75, HT = 175, SEXF = 0)
ev_typ <- make_events(typ, 1, 1)
tissues <- c("adipose", "bone", "gut", "heart", "kidney", "liver", "lung", "muscle", "skin", "spleen", "other")
kp_row <- function(mod) {
  s <- rxode2::rxSolve(rxode2::zeroRe(mod), ev_typ, returnType = "data.frame")
  unlist(s[1, paste0("kp_", tissues)])
}
kp_tab <- data.frame(
  tissue = tissues,
  acetaminophen = kp_row(mod_apap),
  oxycodone = kp_row(mod_oxy)
)
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalq_csf_prod', 'etalf_bulk', 'etalf_ssink', 'etalf_sout'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalq_csf_prod', 'etalf_bulk', 'etalf_ssink', 'etalf_sout'
knitr::kable(kp_tab, digits = 3, row.names = FALSE,
  caption = "Tissue:plasma partition coefficients of the typical man.")
Tissue:plasma partition coefficients of the typical man.
tissue acetaminophen oxycodone
adipose 0.257 0.127
bone 0.569 1.270
gut 0.773 1.239
heart 0.717 0.809
kidney 0.756 0.941
liver 0.769 1.130
lung 0.738 0.712
muscle 0.718 0.923
skin 0.734 0.964
spleen 0.758 0.952
other 0.679 0.907
stopifnot(all(kp_tab$acetaminophen > 0), all(kp_tab$oxycodone > 0))

Mass balance

Clearance removes CL x C_plasma from venous blood. Over the simulation window the dose must equal the cleared mass plus what remains in the 17 states. A 1-h infusion keeps the venous concentration smooth enough for trapezoidal integration of the cleared mass.

states <- c("venous", "arterial", "lung", "adipose", "bone", "heart", "kidney", "muscle",
            "skin", "spleen", "other", "gut", "liver", "brain_vascular", "brain",
            "brain_csf_sas_cranial", "brain_csf_sas_spinal")
trap <- function(t, y) sum(diff(t) * (head(y, -1) + tail(y, -1)) / 2)
mass_balance <- function(mod, dose) {
  ev <- rxode2::et(amt = dose, cmt = "venous", dur = 1) |>
    rxode2::et(seq(0, 24, by = 0.01), cmt = "venous") |>
    as.data.frame() |>
    mutate(AGE = typ$AGE, WT = typ$WT, HT = typ$HT, SEXF = typ$SEXF)
  s <- rxode2::rxSolve(rxode2::zeroRe(mod), ev, rtol = 1e-10, atol = 1e-12, returnType = "data.frame")
  cleared <- trap(s$time, s$cl * s$Cc)
  left <- sum(unlist(s[nrow(s), states]))
  (cleared + left) / dose
}
mb <- c(acetaminophen = mass_balance(mod_apap, 1000), oxycodone = mass_balance(mod_oxy, 0.092 * 75))
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalq_csf_prod', 'etalf_bulk', 'etalf_ssink', 'etalf_sout'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalq_csf_prod', 'etalf_bulk', 'etalf_ssink', 'etalf_sout'
mb
#> acetaminophen     oxycodone 
#>             1             1
stopifnot(all(abs(mb - 1) < 0.002))

Volume of distribution and terminal phase

The steady-state volume of distribution of the typical man is the sum of the tissue volumes times their Kp values plus the blood volume times the blood:plasma ratio (the small brain compartments are left out). With the clearance it sets the terminal half-life.

vss_row <- function(mod, bp, label) {
  s <- rxode2::rxSolve(rxode2::zeroRe(mod), make_events(typ, 1, 24), returnType = "data.frame")
  r <- s[1, ]
  tis <- c("lung", "adipose", "bone", "heart", "kidney", "muscle", "skin", "spleen", "gut", "liver", "other")
  vss <- sum(unlist(r[paste0("v_", tis)]) * unlist(r[paste0("kp_", tis)])) + (r$v_venous + r$v_arterial) * bp
  late <- s[s$time >= 12, ]
  k <- -unname(coef(lm(log(Cc) ~ time, data = late))[2])
  data.frame(drug = label, `Vss (L)` = vss, `Vss (L/kg)` = vss / typ$WT, `CL (L/h)` = r$cl,
             `terminal t1/2 (h)` = log(2) / k, check.names = FALSE)
}
vss_tab <- bind_rows(
  vss_row(mod_apap, 1.09, "Acetaminophen"),
  vss_row(mod_oxy, 1.3, "Oxycodone")
)
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalq_csf_prod', 'etalf_bulk', 'etalf_ssink', 'etalf_sout'
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalq_csf_prod', 'etalf_bulk', 'etalf_ssink', 'etalf_sout'
knitr::kable(vss_tab, digits = 2, caption = "Typical man (40 y, 175 cm, 75 kg).")
Typical man (40 y, 175 cm, 75 kg).
drug Vss (L) Vss (L/kg) CL (L/h) terminal t1/2 (h)
Acetaminophen 46.81 0.62 22.80 1.65
Oxycodone 58.45 0.78 39.36 1.25
stopifnot(all(vss_tab$`Vss (L)` > 30), all(vss_tab$`Vss (L)` < 90))

For acetaminophen the simulated profile follows Figure 5A over the whole 12 h. For oxycodone the Poulin and Theil Kp values, computed from the Table S2 composition with the printed calcKp_PT function, are all close to 1 except adipose. That gives a Vss of about 0.8 L/kg and a terminal half-life of about 1.3 h. The mean line of Figure 5C falls from about 50 ug/L just after the dose to about 0.3 ug/L at 24 h, a terminal half-life of about 3 h. With the same clearance that needs a Vss of roughly 2.3 L/kg, about three times larger. The supplementary code reads the human tissue composition for this method from a file named human_PT_Edited.csv, which is not part of the supplement. Its contents may differ from Table S2. The model here uses the published values unchanged. It therefore reproduces the Table 4 AUC and the CSF peak, which are dominated by the first hours, but not the late oxycodone concentrations.

PKNCA validation

Two NCA runs per drug. Per-subject NCA of plasma confirms that the plasma clearance from NCA equals the model clearance (both sides use the same drawn parameters, so this check is tight). NCA of the cohort-mean profile is then compared with the predicted values of Table 4, which the paper computed from its simulated profiles (Methods 2.6; AUClast over the study window, 12 h for acetaminophen and 24 h for oxycodone, the time axes of Figure 5).

conc_subj <- sim |>
  filter(!is.na(Cc)) |>
  transmute(id, drug, time, conc = Cc)
dose_subj <- bind_rows(ev_apap, ev_oxy) |>
  filter(evid == 1) |>
  select(id, drug, time, amt)
o_conc <- PKNCA::PKNCAconc(conc_subj, conc ~ time | drug + id)
o_dose <- PKNCA::PKNCAdose(dose_subj, amt ~ time | drug + id, route = "intravascular")
nca_subj <- PKNCA::pk.nca(PKNCA::PKNCAdata(o_conc, o_dose,
  intervals = data.frame(start = 0, end = Inf, aucinf.obs = TRUE, cl.obs = TRUE, half.life = TRUE)
))
cl_chk <- as.data.frame(nca_subj) |>
  filter(PPTESTCD == "cl.obs") |>
  select(id, drug, cl_nca = PPORRES) |>
  inner_join(distinct(sim, id, cl), by = "id") |>
  mutate(ratio = cl_nca / cl)
cl_chk |>
  group_by(drug) |>
  summarise(`median NCA CL / model CL` = median(ratio), `max abs deviation` = max(abs(ratio - 1)), .groups = "drop") |>
  knitr::kable(digits = 4)
drug median NCA CL / model CL max abs deviation
Acetaminophen 1000 mg IV 0.9978 0.0044
Oxycodone 0.092 mg/kg IV 0.9974 0.0047
stopifnot(abs(median(cl_chk$ratio) - 1) < 0.02, quantile(abs(cl_chk$ratio - 1), 0.9) < 0.04)
mean_prof <- summ |>
  filter(output %in% c("Cc", "Ccsf_spinal")) |>
  mutate(
    matrix = ifelse(output == "Cc", "Plasma", "CSF"),
    # Table 4 units: mg/L for acetaminophen, ug/L for oxycodone
    conc = ifelse(grepl("Oxycodone", drug), mean * 1000, mean),
    treatment = paste(drug, matrix, sep = ", ")
  ) |>
  select(treatment, drug, matrix, time, conc) |>
  filter(!is.na(conc))
dose_mean <- distinct(mean_prof, treatment, drug, matrix) |>
  mutate(time = 0, amt = 1)
t_end <- c("Acetaminophen 1000 mg IV" = 12, "Oxycodone 0.092 mg/kg IV" = 24)
o_conc_m <- PKNCA::PKNCAconc(mean_prof, conc ~ time | treatment)
o_dose_m <- PKNCA::PKNCAdose(dose_mean, amt ~ time | treatment, route = "intravascular")
# AUC from time zero; Cmax from the first post-bolus grid point (0.1 h), as
# the time-zero venous value is the undistributed bolus.
intervals_m <- bind_rows(
  distinct(mean_prof, treatment, drug) |>
    mutate(start = 0, end = t_end[drug], auclast = TRUE, cmax = FALSE),
  distinct(mean_prof, treatment, drug) |>
    mutate(start = 0.1, end = t_end[drug], auclast = FALSE, cmax = TRUE)
) |>
  select(-drug)
nca_mean <- PKNCA::pk.nca(PKNCA::PKNCAdata(o_conc_m, o_dose_m, intervals = intervals_m))

sim_wide <- as.data.frame(nca_mean) |>
  filter(PPTESTCD %in% c("auclast", "cmax")) |>
  select(treatment, PPTESTCD, PPORRES)
ref_tab4 <- data.frame(
  treatment = c(
    "Acetaminophen 1000 mg IV, Plasma", "Acetaminophen 1000 mg IV, CSF",
    "Oxycodone 0.092 mg/kg IV, Plasma", "Oxycodone 0.092 mg/kg IV, CSF"
  ),
  auclast = c(37.81, 40.59, 201.14, 137.77),
  cmax = c(18.44, 6.72, 96.24, 22.35)
)
cmp <- nlmixr2lib::ncaComparisonTable(
  sim_wide, ref_tab4,
  by = "treatment",
  tolerance_pct = 20,
  units = c(auclast = "h x mg/L or h x ug/L", cmax = "mg/L or ug/L")
)
knitr::kable(cmp, caption = paste(
  "Cohort-mean NCA versus the predicted values of Table 4 (acetaminophen in",
  "mg and L, oxycodone in ug and L).", attr(cmp, "footnote")
))
Cohort-mean NCA versus the predicted values of Table 4 (acetaminophen in mg and L, oxycodone in ug and L). * differs from reference by more than ±20%.
NCA parameter treatment Reference Simulated % diff
Cmax (mg/L or ug/L) Acetaminophen 1000 mg IV, Plasma 18.4 28.9 +56.9%*
Cmax (mg/L or ug/L) Acetaminophen 1000 mg IV, CSF 6.72 7.05 +5.0%
Cmax (mg/L or ug/L) Oxycodone 0.092 mg/kg IV, Plasma 96.2 146 +51.4%*
Cmax (mg/L or ug/L) Oxycodone 0.092 mg/kg IV, CSF 22.4 25 +11.9%
AUClast (h x mg/L or h x ug/L) Acetaminophen 1000 mg IV, Plasma 37.8 46.5 +23.0%*
AUClast (h x mg/L or h x ug/L) Acetaminophen 1000 mg IV, CSF 40.6 38.2 -5.8%
AUClast (h x mg/L or h x ug/L) Oxycodone 0.092 mg/kg IV, Plasma 201 194 -3.7%
AUClast (h x mg/L or h x ug/L) Oxycodone 0.092 mg/kg IV, CSF 138 114 -17.4%
pct <- sim_wide |>
  inner_join(pivot_longer(ref_tab4, -treatment, names_to = "PPTESTCD", values_to = "ref"),
             by = c("treatment", "PPTESTCD")) |>
  mutate(pct_diff = 100 * (PPORRES / ref - 1))
stopifnot(nrow(pct) == 8L)
row_of <- function(trt, code) {
  r <- pct[pct$treatment == trt & pct$PPTESTCD == code, ]
  if (nrow(r) != 1L) stop("no unique row for ", trt, " / ", code)
  r
}
# CNS AUC and Cmax and oxycodone plasma AUC against the Table 4 predictions.
# AUC is set by dose / CL (plasma) and by the BBB and CSF parameters (CSF);
# a mis-transcribed clearance, PSB, unbound fraction or CSF volume moves it
# by well over 35%. Realised over four cohort seeds: oxycodone CSF AUC -17
# to -25%, every other gated row within 13%.
gated <- bind_rows(
  row_of("Acetaminophen 1000 mg IV, CSF", "auclast"),
  row_of("Acetaminophen 1000 mg IV, CSF", "cmax"),
  row_of("Oxycodone 0.092 mg/kg IV, CSF", "auclast"),
  row_of("Oxycodone 0.092 mg/kg IV, CSF", "cmax"),
  row_of("Oxycodone 0.092 mg/kg IV, Plasma", "auclast")
)
stopifnot(all(abs(gated$pct_diff) < 35))
# Acetaminophen plasma AUC: Table 4's predicted 37.81 mg*h/L is below the
# dose / CL of the paper's own clearance (1000 / 22.8 = 43.9 mg*h/L), so it
# is compared with the observed Table 4 value (45.58 mg*h/L) instead.
apap_auc <- row_of("Acetaminophen 1000 mg IV, Plasma", "auclast")$PPORRES
stopifnot(abs(apap_auc / 45.58 - 1) < 0.2)

The table compares the virtual cohort with the paper’s own predictions in Table 4. The CNS values (CSF AUC and Cmax for both drugs) and the oxycodone plasma AUC agree within about 25%; the oxycodone CSF AUC is the lowest of them (about 20% below the prediction, depending on the cohort draw). Two rows differ more:

  • Acetaminophen plasma AUC (+23%). After an IV dose the plasma AUC to infinity is dose / CL. For the paper’s clearance of 22.8 L/h that is 43.9 mg*h/L for a typical subject, and more for the cohort mean, because the mean of dose / CL over log-normal clearance exceeds dose over the typical CL. The 12-h AUC here (46.5 mg*h/L) is close to the observed 45.58 mg*h/L of Table 4. The predicted 37.81 mg*h/L implies either a shorter AUC window or a different clearance from the one in Table S3. The source does not say which.
  • Plasma Cmax (both drugs). It depends on the dose administration and the sampling grid, and the code leaves both as placeholders (DosageRegimen(...)). A bolus into the 2-3 L venous compartment gives a high first grid point. The paper’s acetaminophen plasma Cmax (18.44 mg/L) matches the bolus profile 0.2-0.25 h after dosing. Plasma Cmax is shown without a pass/fail check.

Instantiations not reproduced

The supplementary code leaves every drug-specific input as a placeholder (___); Table S3 fills them for the human IV drugs. The other eight instantiations need values that no published table or text gives:

  • Human lacosamide, ibuprofen and levetiracetam (oral): the absorption rate constant tvKa, the bioavailability tvfa, the lag time tvlag and their variances. Table S3 has no absorption row, and the text says only that the variability on Ka “was based on existing studies”.
  • All five rat models: the rat clearance tvCL was estimated by a naive-pooled fit and is not printed (Table S3 lists human CL only). The rat intraperitoneal (ibuprofen, levetiracetam) and oral (lacosamide) absorption parameters are not given, nor is the value or form of the BBB efflux term the Discussion says was added to the rat lacosamide model.

Assumptions and deviations

  • Oxycodone terminal phase. The published inputs give a terminal half-life of about 1.3 h, while Figure 5C-D implies about 3 h (see “Volume of distribution and terminal phase”). The Kp inputs were not adjusted to close the gap. Use the oxycodone model for the first hours after a dose, the window its Table 4 metrics cover. Treat later concentrations with caution.
  • Oxycodone CL-scalar. Table S3 lists a ‘CL-scalar’ of 0.8 (optimised) for oxycodone, while the code has only Kp_scalar = 1. The scalar is applied to clearance (49.2 x 0.8 = 39.36 L/h). After an IV dose the plasma AUC is dose / CL whatever the Kp values, so only a clearance scalar can move it. With 39.36 L/h a typical 70-80 kg adult has a 0-24 h plasma AUC of 165-189 ug*h/L against Table 4’s predicted 201.14 ug*h/L; with 49.2 L/h it is 133-152 ug*h/L.
  • IIV on CL. Table S3 gives a CV; the variance is log(1 + CV^2). The CSF-flow variabilities are the standard deviations of the code’s exp(rand(Normal(0, SD))) draws, used as variances SD^2.
  • Bulk-flow variability. In the code the bulk flow that carries drug from brain mass to cranial CSF is 0.25 x Qproduction without the 8% variability. The varied bulk flow enters only the spinal-sink and cranial-sink flows. The model keeps this split.
  • Independent virtual-population draws. The code resets Random.seed!(123) before each variability draw, so within one sex the height, weight and CSF-flow deviations share the same standard-normal vector. That is a side effect of re-seeding, not a stated correlation. The draws here are independent.
  • Haematocrit. Table S2 lists haematocrit (0.43 male, 0.38 female), and the code’s Rodgers and Rowland function fixes HCT = 0.45. Haematocrit enters only the base branch of Rodgers and Rowland, which neither model uses (acetaminophen is an acid; oxycodone uses Poulin and Theil). The blood:plasma ratio is the fixed Table S3 value.
  • Rest-of-body Kp. The code takes the mean of the ten tissue Kp values. The Poulin and Theil function also returns a pancreas Kp of 0, which the model does not use.
  • Unused inputs. The spinal-CSF unbound fraction (Fuscsf = 1) is declared in the code but enters no equation; the model has one CSF unbound fraction, for the cranial CSF.
  • ECF output. Cecf = Cbm x Vecf / Vbm is encoded as printed (Methods 2.3.2: “the ECF concentrations are approximated based on the volume of ECF”).
  • Dose administration. The code’s dosing regimen is a placeholder. Doses are simulated as an IV bolus into venous blood (Methods 2.3.1), and oxycodone is dosed per kg of the virtual subject’s weight.
  • Sex ratio and cohort size. Not reported; 50% women and 100 subjects per arm are used.
  • Observed data. The observed points of Figure 5 were not digitised. The comparison is with the paper’s predicted values in Table 4.
  • Residual error. None was reported. propSd is fixed to 0.