Skip to contents

Nilotinib tablets vs capsules (Mauro 2025)

Source: vignettes/articles/Mauro_2025_nilotinib.Rmd
Mauro_2025_nilotinib.Rmd

Model and source

Mauro 2025 reports two sequential fits of the same structural model. Both are distributed, because each drove a different set of published simulations.

uiFinal <- rxode2::rxode(readModelDb("Mauro_2025_nilotinib"))
#> ℹ parameter labels from comments will be replaced by 'label()'
#> Warning: some etas defaulted to non-mu referenced, possible parsing error: etaiov_vc_1, etaiov_vc_2, etaiov_vc_3
#> as a work-around try putting the mu-referenced expression on a simple line
uiLowfat <- rxode2::rxode(readModelDb("Mauro_2025_nilotinib_lowfat"))
#> ℹ parameter labels from comments will be replaced by 'label()'
#> Warning: some etas defaulted to non-mu referenced, possible parsing error: etaiov_vc_1, etaiov_vc_2, etaiov_vc_3
#> as a work-around try putting the mu-referenced expression on a simple line
  • Citation: Mauro M, Radich J, Jain P, Sequeira D, Bellanti F, Douer D. Pharmacokinetic profile of novel reduced-dose Danziten (nilotinib tablets) versus Tasigna (nilotinib capsules): in vivo bioequivalence and population pharmacokinetic analysis. Cancer Chemother Pharmacol. 2025;95(1):56. doi:10.1007/s00280-025-04777-6. Parameters are the ‘Final Model’ of Supplemental Table 2, fitted to 13 of the 14 pooled studies and used for the published fasted steady-state bioequivalence simulations (main text Table 1). The companion ‘Updated Model’ that adds the modified-fasted low-fat-meal study is modellib(‘Mauro_2025_nilotinib_lowfat’).
  • Article: https://doi.org/10.1007/s00280-025-04777-6
  • Supplement: https://doi.org/10.1007/s00280-025-04777-6 (Supplementary Information, 280_2025_4777_MOESM1_ESM.docx; carries the model-development narrative, the F1 equations, and the two parameter tables)
Model file Source table Studies Prandial states covered Published result it drove
Mauro_2025_nilotinib Suppl. Table 2 (“Final Model”) 13 fasted; high-fat meal 0.5 h pre-dose; high-fat meal 2 h pre-dose Table 1 fasted steady-state bioequivalence
Mauro_2025_nilotinib_lowfat Suppl. Table 3 (“Updated Model”) 14 the three above, plus low-fat meal 2 h pre-dose Table 2 low-fat-meal steady-state exposures

The Updated Model adds study 052-23 (190 mg tablets vs 400 mg capsules under modified fasting with a low-fat meal) and keeps the structure of the Final Model, so its shared parameters shift only slightly (CL 34.5 -> 34.3 L/h, V1 557 -> 551 L). Per references/replicate-author-structure.md both are extracted, one file each, because both are final models for their own published purpose; the paper’s fasted bioequivalence conclusion rests on the first and its low-fat-meal food-effect conclusion on the second.

Population

The pooled analysis dataset comprises 30,089 nilotinib plasma concentrations from 502 healthy adult subjects across 14 single-dose, relative-bioavailability and food-effect crossover studies (Suppl. Table 1), with 1,311 below-limit-of-quantification records excluded. Nilotinib tablets were given at 50 to 300 mg and capsules at 50 to 400 mg, always as a single dose, with each dose delivered as two units (for example 190 mg = 2 x 95 mg tablets, 400 mg = 2 x 200 mg capsules). Only 39 participants were female, contributed by studies 006-23, 007-23 and 128-22; every other study enrolled healthy men only, which is why sex effects are identifiable only through those three studies. The trials were run in Phase 1 units in Hyderabad, India.

Baseline demographic distributions (age, weight, laboratory values) are not tabulated in the paper or in its Supplemental Information, so the virtual cohort below fixes body weight at the assumed 70 kg allometric reference and splits sex 50/50 to match the paper’s own simulation cohort (“A dataset of 50 healthy subjects (50% male/50% female) was randomly selected from the analysis dataset”).

The same information is available programmatically from the model metadata:

str(uiFinal$population)
#> List of 10
#>  $ species       : chr "human"
#>  $ n_subjects    : num 502
#>  $ n_studies     : num 13
#>  $ age_range     : chr "adults (not reported)"
#>  $ weight_range  : chr "not reported"
#>  $ sex_female_pct: num 7.8
#>  $ disease_state : chr "healthy"
#>  $ dose_range    : chr "50-300 mg single oral dose (nilotinib tablets); 50-400 mg single oral dose (nilotinib capsules)"
#>  $ regions       : chr "India (Phase 1 units, ZenRise, Hyderabad); analysis by GK Analytics, Hyderabad and Certara"
#>  $ notes         : chr "Healthy adult men and women in 14 single-dose relative-bioavailability and food-effect crossover studies (Suppl"| __truncated__

Model structure

A two-compartment model with linear elimination. Absorption has two parallel routes, both fed by the same dose amount:

  1. Primary, a zero-order input directly into central of duration D1 with lag ALAG1 and relative bioavailability F1. One-, mixed zero-then-first- order and transit-chain alternatives were tested and rejected (Suppl. Information, “Initial Model Development”).
  2. Secondary (depot2), an empirical first-order route with its own lag, carrying a fraction 1 - 0.888 = 0.112 of F1. It was added to describe an increase in concentrations at a time since last dose of about 12 and 24 h (delta OFV -575.6). Total absorbed drug is therefore 1.112 x F1 x dose.

In the event table each administration is two dose records with the same amt: one into central with rate = -2 (modelled duration) and one into depot2. The model files declare dosing <- c("central", "depot2").

Relative bioavailability follows the supplement’s equations verbatim:

F1_capsule = 1 * (DOSE/400)^theta10
F1_tablet  = 1 * (1 + theta6) * (DOSE/200)^theta10   for DOSE <= 200 mg
F1_tablet  = 1 * (1 + theta6)                        for DOSE >  200 mg

implemented as min(DOSE, 200) in the tablet branch so tablet exposure stays dose-proportional above 200 mg while capsule exposure remains less than dose proportional.

Encoding the four prandial states

The paper defines four mutually exclusive prandial states (Methods, “Prandial state definitions”), and all four carry different bioavailability effects, so all four must be separable. Meal composition and meal timing are orthogonal here: the paper’s “fed” and “modified fasting with a high-fat meal” states use the same high-fat meal and differ only in whether the dose followed the start of the meal by about 0.5 h or about 2 h. Composition indicators alone therefore cannot tell them apart, which is why this extraction registers the new canonical covariate MEAL_PREDOSE_2H (see inst/references/covariate-columns.md).

Paper’s prandial state FED FED_HIGHFAT FED_LOWFAT MEAL_PREDOSE_2H
fasted (overnight fast >= 10 h) 0 0 0 0
“fed” (high-fat meal started 0.5 h pre-dose) 1 1 0 0
“modified fasting” + high-fat meal (2 h pre-dose) 1 1 0 1
“modified fasting” + low-fat meal (2 h pre-dose) 1 0 1 1

MEAL_PREDOSE_2H is the mirror image of the existing MEAL_DELAY_1H canonical: MEAL_DELAY_<n> marks a meal taken after the dose, MEAL_PREDOSE_<n> a meal taken before it. The bare phrase “modified fasting” is deliberately not used to define the covariate, because it is not portable between papers – Mauro 2025 uses it for meal first, dose 2 h later, while other nilotinib-era papers use the same phrase for dose first, food withheld for 2 h afterwards.

Source trace

Every ini() entry in both model files carries an in-file comment naming its source row. The table below collects the Final Model (Suppl. Table 2); the Updated Model (Suppl. Table 3) is the identically-structured table on the following supplement page, plus four low-fat-meal rows.

Equation / parameter Value Source location
2-compartment, linear elimination, zero-order absorption with lag n/a Results, “Final population PK model”; Suppl. Info, “Initial Model Development”
Secondary depot with first-order absorption and lag n/a Suppl. Info, “Initial Model Development” (delta OFV -575.6)
F1 equations (formulation, dose power, 200 mg tablet cap) n/a Suppl. Info, “Initial Model Development”, displayed equations
lcl (CL) 34.5 L/h Suppl. Table 2, Typical Values
lvc (V1) 557 L Suppl. Table 2, Typical Values
lq (Q) 4.54 L/h Suppl. Table 2, Typical Values
lvp (V2) 223 L Suppl. Table 2, Typical Values
ld1 (D1) 2.32 h Suppl. Table 2, Typical Values
ltlag (ALAG1) 0.0988 h Suppl. Table 2, Typical Values
lka2 (secondary depot ka) 0.140 1/h Suppl. Table 2, Typical Values
ltlag2 (secondary depot lag) 5.35 h Suppl. Table 2, Typical Values
lfdepot2 (fraction of F1 re-absorbed) 1 - 0.888 Suppl. Table 2, Typical Values
lfdepot (capsule F1 anchor) 1 (fixed) Suppl. Info F1 equations
e_form_tablet_fdepot 0.988 Suppl. Table 2, Covariate Effects
e_dose_fdepot -0.701 Suppl. Table 2, Covariate Effects
e_form_tablet_d1 -0.423 Suppl. Table 2, Covariate Effects
e_wt_cl -0.727 Suppl. Table 2, Covariate Effects
e_hfmeal_cap_fdepot 1.81 Suppl. Table 2, “Food (modified fasted) effect on F1 for nilotinib capsules”
e_hfmeal05h_tab{50,142,190,240,300}_fdepot 0.166, 0.484, 0.617, 0.748, 0.614 Suppl. Table 2, “Food (fed) effect on F1 at for nilotinib tablets”
e_hfmeal2h_tab142_fdepot 0.585 Suppl. Table 2, “Food (modified fasted) effect on F1 at 142 mg for nilotinib tablets”
e_hfmeal2h_tab190_fdepot 0.603 Suppl. Table 2, row labelled “… at 190 mg for nilotinib capsules” (label typo; see Errata)
e_fed_tab_d1, e_fed_cap_d1 2.91, 0.746 Suppl. Table 2, “Food effect on D1 …”
e_fed_tab_tlag, e_fed_cap_tlag 2.98, 7.79 Suppl. Table 2, “Food effect on ALAG1 …”
e_fed_vc, e_fed_q -0.248, 2.07 Suppl. Table 2, “Food effect on V1 / Q”
e_sexf_cl -0.227 Suppl. Table 2, “Sex (female) covariate effect on CL”
e_sexf_fdepot 0.116 Suppl. Table 2, “Sex (female*) covariate effect on F1” + footnote
etalcl + etalvc block 0.108, 0.0628, 0.144 Suppl. Table 2, Between Subject Variability
etaiov_vc_* (IOV on V1) 0.157 Suppl. Table 2, Inter-Occasion Variability
propSd, addSd 0.230, 36.6 ng/mL Suppl. Table 2, Residual Error
e_lfmeal2h_tab_fdepot, e_lfmeal2h_cap_fdepot 0.294, 0.512 Suppl. Table 3, low-fat-meal rows (+ Suppl. Info text “29.4% and 51.2% for nilotinib tablets and capsules respectively”)
e_lfmeal2h_tab_d1, e_lfmeal2h_cap_tlag 0.948, 3.25 Suppl. Table 3, low-fat-meal rows (+ Suppl. Info text “an effect on D1 for nilotinib tablets and on lag-time for capsules”)

Virtual cohort

The paper’s steady-state simulations dosed twice daily for 10 days and computed AUCss, Cmaxss and Cminss over the 0-12 h interval after the last dose, in 50 resampled subjects split 50/50 by sex. The cohort below follows that design with 100 subjects per arm across 14 arms.

Every pooled study was a crossover design, so the arms are simulated as a crossover too: each arm is solved separately with the random-number seed reset first, which gives subject k the same clearance, volume, occasion and residual-error draws in every arm. That is both the faithful design and the statistically sharper one – the published quantities below are all ratios between arms, and pairing removes the between-arm sampling noise that would otherwise dominate them at this cohort size.

simSeed <- 20250422L

nPerArm <- 100L
tau <- 12          # BID dosing interval (h)
nDose <- 20L       # 10 days BID, as in Mauro 2025 Methods, "Simulations"
tLast <- tau * (nDose - 1L)
obsGrid <- seq(0, tau, by = 0.5)

# A realistic bioequivalence sampling schedule inside the last interval; used
# only for the residual-error Cmax diagnostic further down. Every element is on
# obsGrid, so no extra solve is needed.
beGrid <- c(0, 0.5, 1, 1.5, 2, 2.5, 3, 4, 6, 8, 10, 12)
stopifnot(all(beGrid %in% obsGrid))

# Build one arm's event table. Each administration is two dose records with the
# same amt: the lagged zero-order input into central (rate = -2 selects the
# modelled duration D1) and the secondary depot. Only numeric covariate columns
# go into the event table; arm labels are attached after the solve.
makeArm <- function(dose, tablet, fed, highfat, lowfat, predose2h, n = nPerArm) {
  subj <- tibble(
    id = seq_len(n),
    SEXF = as.integer(seq_len(n) %% 2L == 0L),
    WT = 70
  )
  doses <- subj |>
    tidyr::crossing(time = seq(0, tLast, by = tau),
                    cmt = c("central", "depot2")) |>
    mutate(amt = dose, evid = 1L, rate = if_else(cmt == "central", -2, 0))
  obs <- subj |>
    tidyr::crossing(time = tLast + obsGrid) |>
    mutate(cmt = "central", amt = 0, evid = 0L, rate = 0)
  bind_rows(doses, obs) |>
    mutate(DOSE = dose, FORM_TABLET = tablet, FED = fed,
           FED_HIGHFAT = highfat, FED_LOWFAT = lowfat,
           MEAL_PREDOSE_2H = predose2h, OCC = 1L) |>
    arrange(id, time, desc(evid))
}

# Arm grid. Prandial-state labels use the paper's own wording.
armSpec <- tibble::tribble(
  ~product,                      ~state,        ~dose, ~tablet, ~fed, ~highfat, ~lowfat, ~predose2h,
  "142 mg tablets",              "fasted",        142,       1L,   0L,       0L,      0L,         0L,
  "300 mg capsules",             "fasted",        300,       0L,   0L,       0L,      0L,         0L,
  "190 mg tablets",              "fasted",        190,       1L,   0L,       0L,      0L,         0L,
  "400 mg capsules",             "fasted",        400,       0L,   0L,       0L,      0L,         0L,
  "142 mg tablets",              "high-fat 2 h",  142,       1L,   1L,       1L,      0L,         1L,
  "300 mg capsules",             "high-fat 2 h",  300,       0L,   1L,       1L,      0L,         1L,
  "190 mg tablets",              "high-fat 2 h",  190,       1L,   1L,       1L,      0L,         1L,
  "400 mg capsules",             "high-fat 2 h",  400,       0L,   1L,       1L,      0L,         1L,
  "142 mg tablets",              "high-fat 0.5 h",142,       1L,   1L,       1L,      0L,         0L,
  "190 mg tablets",              "high-fat 0.5 h",190,       1L,   1L,       1L,      0L,         0L,
  "142 mg tablets",              "low-fat 2 h",   142,       1L,   1L,       0L,      1L,         1L,
  "300 mg capsules",             "low-fat 2 h",   300,       0L,   1L,       0L,      1L,         1L,
  "190 mg tablets",              "low-fat 2 h",   190,       1L,   1L,       0L,      1L,         1L,
  "400 mg capsules",             "low-fat 2 h",   400,       0L,   1L,       0L,      1L,         1L
)
# The Final Model has no low-fat-meal estimates and does not declare FED_LOWFAT,
# so the low-fat arms come from the Updated Model and every other arm from the
# Final Model. That split is exactly how the paper assembled Table 2, whose
# fasted column came from the Final Model and whose low-fat column came from the
# Updated Model.
armSpec <- armSpec |>
  mutate(label = paste0(product, ", ", state),
         model = if_else(state == "low-fat 2 h", "Updated Model", "Final Model"),
         modelName = if_else(state == "low-fat 2 h",
                             "Mauro_2025_nilotinib_lowfat",
                             "Mauro_2025_nilotinib"))
armSpec |> select(label, model) |> knitr::kable()
label model
142 mg tablets, fasted Final Model
300 mg capsules, fasted Final Model
190 mg tablets, fasted Final Model
400 mg capsules, fasted Final Model
142 mg tablets, high-fat 2 h Final Model
300 mg capsules, high-fat 2 h Final Model
190 mg tablets, high-fat 2 h Final Model
400 mg capsules, high-fat 2 h Final Model
142 mg tablets, high-fat 0.5 h Final Model
190 mg tablets, high-fat 0.5 h Final Model
142 mg tablets, low-fat 2 h Updated Model
300 mg capsules, low-fat 2 h Updated Model
190 mg tablets, low-fat 2 h Updated Model
400 mg capsules, low-fat 2 h Updated Model

Simulation 

solveArm <- function(i) {
  s <- armSpec[i, ]
  ev <- makeArm(s$dose, s$tablet, s$fed, s$highfat, s$lowfat, s$predose2h)
  # Real duplicate guard: the two dose records of an administration share
  # (id, time, evid) by design, so the key must include cmt. A duplicate here
  # would silently double a dose.
  stopifnot(!anyDuplicated(ev[, c("id", "time", "evid", "cmt")]))
  if (s$model == "Final Model") ev <- select(ev, -FED_LOWFAT)
  set.seed(simSeed)   # pairs the random effects across arms; see above
  rxode2::rxSolve(readModelDb(s$modelName), events = ev, addDosing = FALSE) |>
    as.data.frame() |>
    mutate(t = round(time - tLast, 6), arm = s$label, product = s$product,
           state = s$state, model = s$model)
}

simAll <- bind_rows(lapply(seq_len(nrow(armSpec)), solveArm))
#> ℹ parameter labels from comments will be replaced by 'label()'
#> Warning: some etas defaulted to non-mu referenced, possible parsing error: etaiov_vc_1, etaiov_vc_2, etaiov_vc_3
#> as a work-around try putting the mu-referenced expression on a simple line
#> ℹ parameter labels from comments will be replaced by 'label()'
#> Warning: some etas defaulted to non-mu referenced, possible parsing error: etaiov_vc_1, etaiov_vc_2, etaiov_vc_3
#> as a work-around try putting the mu-referenced expression on a simple line
count(simAll, model, state, product) |> knitr::kable()
model state product n
Final Model fasted 142 mg tablets 2500
Final Model fasted 190 mg tablets 2500
Final Model fasted 300 mg capsules 2500
Final Model fasted 400 mg capsules 2500
Final Model high-fat 0.5 h 142 mg tablets 2500
Final Model high-fat 0.5 h 190 mg tablets 2500
Final Model high-fat 2 h 142 mg tablets 2500
Final Model high-fat 2 h 190 mg tablets 2500
Final Model high-fat 2 h 300 mg capsules 2500
Final Model high-fat 2 h 400 mg capsules 2500
Updated Model low-fat 2 h 142 mg tablets 2500
Updated Model low-fat 2 h 190 mg tablets 2500
Updated Model low-fat 2 h 300 mg capsules 2500
Updated Model low-fat 2 h 400 mg capsules 2500

Replicate published figures

The published steady-state figures are Supplemental Figures 4-9 (prediction- corrected VPCs by dose and prandial state) and main-text Figure 3 (percent increases in Cmaxss, Cminss and AUCss under each meal condition). The panel below reproduces the underlying profiles: median with 5th-95th percentile band over the last dosing interval, by prandial state.

simAll |>
  group_by(state, product, t) |>
  summarise(Q05 = quantile(Cc, 0.05), Q50 = quantile(Cc, 0.50),
            Q95 = quantile(Cc, 0.95), .groups = "drop") |>
  mutate(state = factor(state, levels = c("fasted", "low-fat 2 h",
                                          "high-fat 0.5 h", "high-fat 2 h"))) |>
  ggplot(aes(t, Q50, colour = product, fill = product)) +
  geom_ribbon(aes(ymin = Q05, ymax = Q95), alpha = 0.15, colour = NA) +
  geom_line(linewidth = 0.7) +
  facet_wrap(~state) +
  labs(x = "Time since last dose (h)", y = "Nilotinib concentration (ng/mL)",
       colour = NULL, fill = NULL,
       title = "Simulated steady-state nilotinib, BID day 10",
       caption = paste("Reproduces the profiles behind Suppl. Figs 4-9 and",
                       "main-text Table 2 of Mauro 2025.")) +
  theme(legend.position = "bottom")

The paper’s headline food-effect numbers (main text, “Model predictions – fed conditions”, and Figure 3) are percentage increases in AUCss relative to the fasted state. Those come out of the same simulation:

The fasted denominator is the Final Model’s fasted arm, matching how the paper built Table 2: its fasted column comes from the Final Model and its low-fat column from the Updated Model.

aucById <- simAll |>
  group_by(state, product, id) |>
  arrange(t, .by_group = TRUE) |>
  summarise(auc = sum(diff(t) * (head(Cc, -1) + tail(Cc, -1)) / 2),
            .groups = "drop")

fastedRef <- aucById |>
  filter(state == "fasted") |>
  group_by(product) |>
  summarise(aucFasted = mean(auc), .groups = "drop")

ratios <- aucById |>
  filter(state != "fasted") |>
  group_by(state, product) |>
  summarise(aucFed = mean(auc), .groups = "drop") |>
  inner_join(fastedRef, by = "product") |>
  mutate(simPct = 100 * (aucFed / aucFasted - 1))

published <- tibble::tribble(
  ~state,           ~product,          ~pubPct,
  "low-fat 2 h",    "142 mg tablets",     26.0,
  "low-fat 2 h",    "190 mg tablets",     29.3,
  "low-fat 2 h",    "300 mg capsules",    56.8,
  "low-fat 2 h",    "400 mg capsules",    60.7,
  "high-fat 2 h",   "142 mg tablets",     48.6,
  "high-fat 2 h",   "190 mg tablets",     52.2,
  "high-fat 2 h",   "300 mg capsules",   180.6,
  "high-fat 2 h",   "400 mg capsules",   183.3
)

ratios |>
  inner_join(published, by = c("state", "product")) |>
  transmute(state, product,
            Published = round(pubPct, 1), Simulated = round(simPct, 1),
            Difference = round(simPct - pubPct, 1)) |>
  arrange(state, product) |>
  rename("Prandial state" = state, "Product" = product,
         "Published increase (%)" = Published,
         "Simulated increase (%)" = Simulated,
         "Difference (percentage points)" = Difference) |>
  knitr::kable(caption = paste("Percent increase in AUCss versus fasted.",
                               "Replicates Figure 3 and the 'Model predictions",
                               "- fed conditions' paragraph of Mauro 2025."))
Percent increase in AUCss versus fasted. Replicates Figure 3 and the ‘Model predictions - fed conditions’ paragraph of Mauro 2025.
Prandial state Product Published increase (%) Simulated increase (%) Difference (percentage points)
high-fat 2 h 142 mg tablets 48.6 50.2 1.6
high-fat 2 h 190 mg tablets 52.2 51.9 -0.3
high-fat 2 h 300 mg capsules 180.6 183.3 2.7
high-fat 2 h 400 mg capsules 183.3 183.3 0.0
low-fat 2 h 142 mg tablets 26.0 26.8 0.8
low-fat 2 h 190 mg tablets 29.3 26.9 -2.4
low-fat 2 h 300 mg capsules 56.8 57.4 0.6
low-fat 2 h 400 mg capsules 60.7 57.5 -3.2

PKNCA validation

Steady-state NCA over the last dosing interval, stratified by prandial state and product. cmin is used rather than ctrough, which PKNCA returns as all-NA on a steady-state multiple-dose interval.

simNca <- simAll |>
  dplyr::filter(!is.na(Cc)) |>
  dplyr::select(id, arm, time, Cc)

doseNca <- armSpec |>
  tidyr::crossing(id = seq_len(nPerArm), time = seq(0, tLast, by = tau)) |>
  transmute(id, arm = label, time, amt = dose)

concObj <- PKNCA::PKNCAconc(simNca, Cc ~ time | arm + id,
                            concu = "ng/mL", timeu = "h")
doseObj <- PKNCA::PKNCAdose(doseNca, amt ~ time | arm + id, doseu = "mg")

intervals <- data.frame(
  start = tLast, end = tLast + tau,
  cmax = TRUE, tmax = TRUE, cmin = TRUE, auclast = TRUE, cav = TRUE
)

ncaRes <- PKNCA::pk.nca(PKNCA::PKNCAdata(concObj, doseObj, intervals = intervals))

ncaByArm <- as.data.frame(ncaRes) |>
  filter(!is.na(PPORRES)) |>
  group_by(arm, PPTESTCD) |>
  summarise(value = mean(PPORRES), .groups = "drop")
head(ncaByArm)
#> # A tibble: 6 × 3
#>   arm                            PPTESTCD   value
#>   <chr>                          <chr>      <dbl>
#> 1 142 mg tablets, fasted         auclast  15811. 
#> 2 142 mg tablets, fasted         cav       1318. 
#> 3 142 mg tablets, fasted         cmax      1707. 
#> 4 142 mg tablets, fasted         cmin      1024. 
#> 5 142 mg tablets, fasted         tmax         1.5
#> 6 142 mg tablets, high-fat 0.5 h auclast  22229.

Comparison against published steady-state exposures (Table 2)

Main-text Table 2 reports mean Cminss, Cmaxss and AUCss for every simulated arm. Capsules under the standard fed condition are reported as “nc” (not calculated), so those arms are not simulated here.

publishedSs <- tibble::tribble(
  ~arm,                                    ~cmin, ~cmax, ~auclast,
  "142 mg tablets, fasted",                  926,  2070,    14600,
  "300 mg capsules, fasted",                 900,  1950,    13900,
  "190 mg tablets, fasted",                 1000,  2250,    15700,
  "400 mg capsules, fasted",                 970,  2120,    15000,
  "142 mg tablets, low-fat 2 h",            1130,  2680,    18400,
  "300 mg capsules, low-fat 2 h",           1350,  3230,    21800,
  "190 mg tablets, low-fat 2 h",            1230,  2980,    20300,
  "400 mg capsules, low-fat 2 h",           1470,  3600,    24100,
  "142 mg tablets, high-fat 2 h",           1460,  2980,    21700,
  "300 mg capsules, high-fat 2 h",          2590,  5500,    39000,
  "190 mg tablets, high-fat 2 h",           1610,  3290,    23900,
  "400 mg capsules, high-fat 2 h",          2830,  6010,    42500,
  "142 mg tablets, high-fat 0.5 h",         1360,  2790,    20300,
  "190 mg tablets, high-fat 0.5 h",         1620,  3320,    24100
)

ncaForCompare <- as.data.frame(ncaRes) |>
  filter(!is.na(PPORRES), PPTESTCD %in% c("cmin", "cmax", "auclast"))

cmp <- nlmixr2lib::ncaComparisonTable(
  simulated = ncaForCompare,
  reference = publishedSs,
  by = "arm",
  units = c(cmin = "ng/mL", cmax = "ng/mL", auclast = "ng*h/mL"),
  tolerance_pct = 20
)

knitr::kable(
  cmp,
  caption = paste("Simulated versus published steady-state exposures",
                  "(Mauro 2025 Table 2). * differs by more than 20%."),
  align = c("l", "l", "r", "r", "r")
)
Simulated versus published steady-state exposures (Mauro 2025 Table 2). * differs by more than 20%.
NCA parameter arm Reference Simulated % diff
Cmax (ng/mL) 142 mg tablets, fasted 2070 1620 -21.8%*
Cmax (ng/mL) 300 mg capsules, fasted 1950 1460 -24.9%*
Cmax (ng/mL) 190 mg tablets, fasted 2250 1770 -21.5%*
Cmax (ng/mL) 400 mg capsules, fasted 2120 1600 -24.7%*
Cmax (ng/mL) 142 mg tablets, low-fat 2 h 2680 2030 -24.4%*
Cmax (ng/mL) 300 mg capsules, low-fat 2 h 3230 2460 -24.0%*
Cmax (ng/mL) 190 mg tablets, low-fat 2 h 2980 2210 -25.8%*
Cmax (ng/mL) 400 mg capsules, low-fat 2 h 3600 2680 -25.6%*
Cmax (ng/mL) 142 mg tablets, high-fat 2 h 2980 2180 -26.9%*
Cmax (ng/mL) 300 mg capsules, high-fat 2 h 5500 4150 -24.6%*
Cmax (ng/mL) 190 mg tablets, high-fat 2 h 3290 2400 -26.9%*
Cmax (ng/mL) 400 mg capsules, high-fat 2 h 6010 4520 -24.8%*
Cmax (ng/mL) 142 mg tablets, high-fat 0.5 h 2790 2040 -26.9%*
Cmax (ng/mL) 190 mg tablets, high-fat 0.5 h 3320 2430 -26.9%*
Cmin (ng/mL) 142 mg tablets, fasted 926 1010 +8.6%
Cmin (ng/mL) 300 mg capsules, fasted 900 1000 +11.6%
Cmin (ng/mL) 190 mg tablets, fasted 1000 1100 +9.7%
Cmin (ng/mL) 400 mg capsules, fasted 970 1090 +12.8%
Cmin (ng/mL) 142 mg tablets, low-fat 2 h 1130 1260 +11.1%
Cmin (ng/mL) 300 mg capsules, low-fat 2 h 1350 1510 +11.5%
Cmin (ng/mL) 190 mg tablets, low-fat 2 h 1230 1370 +11.4%
Cmin (ng/mL) 400 mg capsules, low-fat 2 h 1470 1640 +11.7%
Cmin (ng/mL) 142 mg tablets, high-fat 2 h 1460 1600 +9.3%
Cmin (ng/mL) 300 mg capsules, high-fat 2 h 2590 2840 +9.6%
Cmin (ng/mL) 190 mg tablets, high-fat 2 h 1610 1760 +9.4%
Cmin (ng/mL) 400 mg capsules, high-fat 2 h 2830 3090 +9.3%
Cmin (ng/mL) 142 mg tablets, high-fat 0.5 h 1360 1490 +9.9%
Cmin (ng/mL) 190 mg tablets, high-fat 0.5 h 1620 1780 +9.6%
AUClast (ng*h/mL) 142 mg tablets, fasted 14600 15300 +4.6%
AUClast (ng*h/mL) 300 mg capsules, fasted 13900 14600 +5.4%
AUClast (ng*h/mL) 190 mg tablets, fasted 15700 16700 +6.1%
AUClast (ng*h/mL) 400 mg capsules, fasted 15000 16000 +6.4%
AUClast (ng*h/mL) 142 mg tablets, low-fat 2 h 18400 19400 +5.2%
AUClast (ng*h/mL) 300 mg capsules, low-fat 2 h 21800 23100 +5.9%
AUClast (ng*h/mL) 190 mg tablets, low-fat 2 h 20300 21100 +4.1%
AUClast (ng*h/mL) 400 mg capsules, low-fat 2 h 24100 25200 +4.5%
AUClast (ng*h/mL) 142 mg tablets, high-fat 2 h 21700 22900 +5.6%
AUClast (ng*h/mL) 300 mg capsules, high-fat 2 h 39000 41500 +6.5%
AUClast (ng*h/mL) 190 mg tablets, high-fat 2 h 23900 25300 +5.8%
AUClast (ng*h/mL) 400 mg capsules, high-fat 2 h 42500 45300 +6.5%
AUClast (ng*h/mL) 142 mg tablets, high-fat 0.5 h 20300 21500 +5.7%
AUClast (ng*h/mL) 190 mg tablets, high-fat 0.5 h 24100 25500 +5.8%

AUCss and Cminss agree with the published values throughout, and – importantly – the residual offset is uniform across all 14 arms rather than arm-specific: AUCss high by 4.1 to 6.5%, Cminss high by 8.6 to 12.8%. A uniform offset is a property of the comparison cohort, not of the model’s structure or its covariate effects, which would produce arm-dependent errors.

Two cohort-level effects contribute, and neither is a model defect. First, ncaComparisonTable() aggregates the simulated subjects with the median while the paper’s Table 2 reports arithmetic means, so the two sides are not the same statistic. Second, AUCss over a dosing interval at steady state is exactly dose * F / CL, so an arithmetic mean over log-normal clearance carries a Jensen inflation of exp(omega^2 / 2) - 1 = 5.5% relative to the typical-value prediction, whereas the paper resampled 50 real subjects whose realised clearance spread need not match a full draw from the estimated omega. The three statistics differ by about the expected amount:

as.data.frame(ncaRes) |>
  filter(PPTESTCD == "auclast", !is.na(PPORRES)) |>
  group_by(arm) |>
  summarise(Median = median(PPORRES), Geometric = exp(mean(log(PPORRES))),
            Arithmetic = mean(PPORRES), .groups = "drop") |>
  inner_join(publishedSs |> select(arm, aucPub = auclast), by = "arm") |>
  tidyr::pivot_longer(c(Median, Geometric, Arithmetic),
                      names_to = "statistic", values_to = "value") |>
  group_by(statistic) |>
  summarise(arms = n(),
            meanPct = sprintf("%+.1f%%", mean(100 * (value / aucPub - 1))),
            worstPct = sprintf("%.1f%%", max(abs(100 * (value / aucPub - 1)))),
            .groups = "drop") |>
  rename("Simulated central tendency" = statistic, "Arms compared" = arms,
         "Mean % difference vs published" = meanPct,
         "Largest absolute % difference" = worstPct) |>
  knitr::kable(caption = paste("AUCss versus Mauro 2025 Table 2 under three",
                               "choices of simulated central tendency."))
AUCss versus Mauro 2025 Table 2 under three choices of simulated central tendency.
Simulated central tendency Arms compared Mean % difference vs published Largest absolute % difference
Arithmetic 14 +9.4% 10.4%
Geometric 14 +3.4% 4.6%
Median 14 +5.6% 6.5%

The published quantities that do not depend on the choice of cohort or statistic are the ratios – the Table 1 bioequivalence GMRs and the Figure 3 food-effect percentages – and those are reproduced to within 0.03 and 3.2 percentage points respectively (see above and below). That is the sharper test of the extraction, because a cohort-level offset cancels in a ratio while any error in a formulation, dose or prandial-state effect would not.

Cmaxss is systematically low, by roughly 20-27%, in every arm. That is also not a structural error: it is a difference in how the statistic is defined. The paper’s simulations “included both fixed and random effects (inter-individual as well as residual variability)”, so its Cmaxss is the maximum of residual-error-perturbed simulated concentrations over a sampling schedule, and the maximum of a noisy series is biased upward – whereas Cminss taken at the end of the interval is a single time point and so carries no such bias, and the trapezoidal AUC averages the noise away. That asymmetry across the three metrics is what identifies the mechanism. Recomputing Cmax the same way, from the sim column on a realistic 12-point bioequivalence sampling schedule, closes the gap almost completely:

cmaxResid <- simAll |>
  filter(t %in% beGrid) |>
  group_by(arm, id) |>
  summarise(cmaxObs = max(sim), cmaxPred = max(Cc), .groups = "drop") |>
  group_by(arm) |>
  summarise(cmaxObs = mean(cmaxObs), cmaxPred = mean(cmaxPred), .groups = "drop") |>
  inner_join(publishedSs |> select(arm, cmaxPub = cmax), by = "arm") |>
  mutate(pctPred = round(100 * (cmaxPred / cmaxPub - 1), 1),
         pctObs = round(100 * (cmaxObs / cmaxPub - 1), 1),
         across(c(cmaxObs, cmaxPred), round))

cmaxResid |>
  select(arm, cmaxPub, cmaxPred, pctPred, cmaxObs, pctObs) |>
  rename("Arm" = arm, "Published Cmaxss (ng/mL)" = cmaxPub,
         "Model prediction, no residual error" = cmaxPred,
         "% difference (prediction)" = pctPred,
         "With residual error, 12-point schedule" = cmaxObs,
         "% difference (with residual error)" = pctObs) |>
  knitr::kable(caption = paste("Cmaxss depends on whether residual variability",
                              "is included, because the maximum of a noisy",
                              "series is biased upward."))
Cmaxss depends on whether residual variability is included, because the maximum of a noisy series is biased upward.
Arm Published Cmaxss (ng/mL) Model prediction, no residual error % difference (prediction) With residual error, 12-point schedule % difference (with residual error)
142 mg tablets, fasted 2070 1707 -17.5 2069 -0.1
142 mg tablets, high-fat 0.5 h 2790 2175 -22.1 2576 -7.7
142 mg tablets, high-fat 2 h 2980 2323 -22.1 2751 -7.7
142 mg tablets, low-fat 2 h 2680 2190 -18.3 2642 -1.4
190 mg tablets, fasted 2250 1862 -17.2 2256 0.3
190 mg tablets, high-fat 0.5 h 3320 2585 -22.1 3062 -7.8
190 mg tablets, high-fat 2 h 3290 2563 -22.1 3035 -7.7
190 mg tablets, low-fat 2 h 2980 2391 -19.8 2885 -3.2
300 mg capsules, fasted 1950 1593 -18.3 1944 -0.3
300 mg capsules, high-fat 2 h 5500 4220 -23.3 5055 -8.1
300 mg capsules, low-fat 2 h 3230 2645 -18.1 3157 -2.3
400 mg capsules, fasted 2120 1736 -18.1 2118 -0.1
400 mg capsules, high-fat 2 h 6010 4599 -23.5 5509 -8.3
400 mg capsules, low-fat 2 h 3600 2885 -19.9 3443 -4.4 

# Under the paper's own statistic every fasted arm lands within 1%, so the
# diagnosis is verified rather than merely plausible.
stopifnot(all(abs(cmaxResid$pctObs[grepl("fasted", cmaxResid$arm)]) < 2))

Comparison against the published fasted bioequivalence result (Table 1)

Table 1 of the paper reports geometric-mean ratios of tablet versus capsule steady-state exposure under fasted conditions, which is the study’s primary conclusion. Reproducing it needs geometric means of the individual exposures.

fastedInd <- as.data.frame(ncaRes) |>
  filter(grepl("fasted", arm),
         PPTESTCD %in% c("cmin", "cmax", "auclast"), !is.na(PPORRES)) |>
  select(arm, id, PPTESTCD, PPORRES)

# Fail loudly rather than returning a silently-empty geometric mean: every
# (arm, metric) cell must hold one value per simulated subject.
pull_metric <- function(armLabel, code) {
  v <- fastedInd$PPORRES[fastedInd$arm == armLabel & fastedInd$PPTESTCD == code]
  if (length(v) != nPerArm) {
    stop("expected ", nPerArm, " values for '", armLabel, "' / ", code,
         " but found ", length(v))
  }
  v
}
gmr <- function(testArm, refArm, code) {
  exp(mean(log(pull_metric(testArm, code))) -
        mean(log(pull_metric(refArm, code))))
}

beCompare <- tibble::tribble(
  ~comparison,                          ~test,                    ~ref,                      ~code,     ~metric,          ~pubGmr, ~pubLo, ~pubHi,
  "190 mg tablets vs 400 mg capsules", "190 mg tablets, fasted", "400 mg capsules, fasted", "auclast", "AUCss (0-12 h)",   1.04,   0.95,   1.17,
  "190 mg tablets vs 400 mg capsules", "190 mg tablets, fasted", "400 mg capsules, fasted", "cmax",    "Cmaxss",           1.06,   0.97,   1.20,
  "190 mg tablets vs 400 mg capsules", "190 mg tablets, fasted", "400 mg capsules, fasted", "cmin",    "Cminss",           1.02,   0.89,   1.18,
  "142 mg tablets vs 300 mg capsules", "142 mg tablets, fasted", "300 mg capsules, fasted", "auclast", "AUCss (0-12 h)",   1.03,   0.95,   1.13,
  "142 mg tablets vs 300 mg capsules", "142 mg tablets, fasted", "300 mg capsules, fasted", "cmax",    "Cmaxss",           1.06,   0.97,   1.17,
  "142 mg tablets vs 300 mg capsules", "142 mg tablets, fasted", "300 mg capsules, fasted", "cmin",    "Cminss",           1.00,   0.89,   1.15
)

beResult <- beCompare |>
  mutate(simGmr = mapply(gmr, test, ref, code),
         inside = simGmr >= pubLo & simGmr <= pubHi)

beResult |>
  transmute(comparison, metric,
            published = sprintf("%.2f (%.2f, %.2f)", pubGmr, pubLo, pubHi),
            simulated = sprintf("%.3f", simGmr),
            insideCi = if_else(inside, "yes", "no")) |>
  rename("Comparison" = comparison, "Metric" = metric,
         "Published GMR (90% CI)" = published,
         "Simulated GMR" = simulated,
         "Inside the published CI" = insideCi) |>
  knitr::kable(caption = paste("Fasted steady-state bioequivalence.",
                              "Replicates Table 1 of Mauro 2025."))
Fasted steady-state bioequivalence. Replicates Table 1 of Mauro 2025.
Comparison Metric Published GMR (90% CI) Simulated GMR Inside the published CI
190 mg tablets vs 400 mg capsules AUCss (0-12 h) 1.04 (0.95, 1.17) 1.034 yes
190 mg tablets vs 400 mg capsules Cmaxss 1.06 (0.97, 1.20) 1.066 yes
190 mg tablets vs 400 mg capsules Cminss 1.02 (0.89, 1.18) 1.011 yes
142 mg tablets vs 300 mg capsules AUCss (0-12 h) 1.03 (0.95, 1.13) 1.033 yes
142 mg tablets vs 300 mg capsules Cmaxss 1.06 (0.97, 1.17) 1.065 yes
142 mg tablets vs 300 mg capsules Cminss 1.00 (0.89, 1.15) 1.010 yes 

# The reproduction of the paper's primary conclusion is a hard check.
stopifnot(nrow(beResult) == 6L, all(beResult$inside))

Every simulated geometric-mean ratio falls inside the published 90% confidence interval, so the packaged Final Model reproduces the paper’s primary bioequivalence conclusion: 190 mg tablets are bioequivalent to 400 mg capsules, and 142 mg tablets to 300 mg capsules, under fasted conditions.

Because the arms are paired (same subjects, same random-effect draws), these ratios carry very little Monte Carlo noise: the between-subject variability on clearance cancels in the ratio, leaving only the residual-error and formulation-specific contributions. For contrast, an unpaired design at this cohort size would give a relative standard error of about 4.6% on each ratio – wide enough to swamp the 1-11% deviations seen above. The remaining deviations are therefore attributable to the model and to the difference between this cohort’s fixed 70 kg weight and the paper’s resampled subjects, not to sampling noise.

Assumptions and deviations

Encoder decisions the paper did not specify.

  • Reference weight of 70 kg for the (WT/70)^-0.727 clearance term. The paper gives the exponent but not the normalising weight, so the rounded standard is used. Note that the exponent is genuinely negative in both parameter tables, with confidence intervals entirely below zero (-0.934 to -0.532 in Suppl. Table 2), so heavier subjects have lower nilotinib clearance in this model. That is unusual for an allometric term but it is what the paper reports; it is not a sign transcription error.
  • Three occasions for the inter-occasion variability on V1. The paper reports one IOV variance (0.157) but never states how many occasions it spans. Three are implemented, matching the largest crossover design among the pooled studies (three-way, three-period; Suppl. Table 1), using the NONMEM $OMEGA BLOCK(1) SAME idiom so all three occasions share the single published variance.
  • Dose-step boundaries for the tablet food effects. Mauro 2025 estimated the standard-fed tablet effect on F1 only at 50, 142, 190, 240 and 300 mg, and the high-fat-2 h tablet effect only at 142 and 190 mg. The model files apply each estimate over an interval bounded by the midpoints between published anchors, so any dose returns a defined effect. The boundaries are this extraction’s choice. Interpolating between anchors would not be defensible, because the published effects are non-monotonic in dose (0.617 at 190 mg vs 0.614 at 300 mg).
  • Capsule bioavailability under the standard fed condition. Suppl. Tables 2 and 3 report a single capsule food effect on F1, labelled “(modified fasted)”. Suppl. Table 1 shows capsules were dosed under both the 0.5 h and the 2 h high-fat condition, so one estimate must cover both, and it is applied to both timings here. The paper does not simulate capsules under the standard fed condition (Table 2 reports “nc”), so this has no published comparator.
  • Which food effects survive the low-fat condition (Updated Model). The supplement re-describes the low-fat condition’s absorption from scratch and retains only a tablet D1 effect and a capsule lag-time effect. In the model file each absorption parameter therefore takes either its own low-fat estimate (tablet D1, capsule ALAG1) or its fasted value (capsule D1, tablet ALAG1) – a covariate effect that was tested and not retained means no change from the reference, not the general fed value. The V1 and Q food effects are structural effects carried over unchanged from the Final Model and still apply. Both readings were simulated and scored against the paper’s own Table 2: this one reproduces the published low-fat Cmaxss within 5% (see the Cmaxss table above), whereas letting the general D1 and ALAG1 food effects persist under the low-fat condition roughly triples the capsule Cmaxss shortfall. The paper’s answer key was used to choose between two literal readings of ambiguous text; no parameter value was altered.
  • Secondary-depot dose placement. The supplement says the late re-absorption was described by “introducing additional doses via a secondary depot compartment” but does not say at what times those extra records sat. They are placed at the same time as the primary dose here, with the published 5.35 h (Final) / 4.72 h (Updated) lag doing the delaying. At steady state on a regular BID schedule this choice is immaterial: shifting a secondary dose by any whole number of dosing intervals leaves the steady-state profile unchanged. It does matter for single-dose simulations, where the paper’s own motivation (bumps at a time since last dose of about 12 and 24 h) is the guide.
  • Cohort covariates. Body weight is fixed at the 70 kg reference and sex is split 50/50, matching the paper’s own simulation cohort description. The paper and its supplement contain no demographic table, so no weight or age distribution could be reproduced.

Errata and internal inconsistencies in the source.

  • Suppl. Table 2 formulation-label typo. The row “Food (modified fasted) effect on F1 at 190 mg for nilotinib capsules = 0.603” is mislabelled. The identically-placed row in Suppl. Table 3 reads “for nilotinib tablets”, and capsules already carry a separate, non-dose-specific effect of 1.81 in the same table. The value is encoded as a tablet effect (e_hfmeal2h_tab190_fdepot). Reading it as a capsule effect would give capsules two conflicting high-fat estimates and would leave the 190 mg tablet 2 h condition with none, and would break the reproduction of the published +52.2% AUCss increase for 190 mg tablets.
  • Residual-error form. The supplement narrative says “A combined (additive and proportional) error model was used to describe the residual variability”, while Suppl. Tables 2 and 3 label the multiplicative term “Exponential Error”. The narrative is followed: the model uses Cc ~ add(addSd) + prop(propSd).
  • Study count per model. Methods says 14 studies; the supplement says the Final Model was fitted across 13 and that study 052-23 was added afterwards. The pooled figures the paper quotes (502 subjects, 30,089 observations) are for the full 14-study dataset and are not broken down per model, so both model files carry those figures with n_studies set to 13 and 14 respectively and a note in population$notes.
  • Q units. Suppl. Tables 2 and 3 give units for CL (L/h), V1 (L), V2 (L), D1 (h) and ALAG1 (h) but omit them for Q. L/h is assumed, which is the only reading consistent with the reproduced exposures.
  • Missing NCA table. The main text cites “Supplemental information, Table 2” for observed single-dose Cmax, Tmax and half-life, but Supplemental Table 2 is the population-PK parameter table. No observed NCA table appears in the supplement, so the single-dose bioequivalence results reported in the Results section cannot be checked numerically here; the validation above uses the paper’s simulated steady-state tables instead.

Not deviations, but worth noting for reuse.

  • Building either model emits some etas defaulted to non-mu referenced ... etaiov_vc_1, etaiov_vc_2, etaiov_vc_3. This is inherent to the $OMEGA BLOCK(1) SAME inter-occasion-variability idiom, in which the per-occasion etas are selected by (OCC == k) indicators and so cannot be mu-referenced. The already-registered Pejcic_2024_clopidogrel model emits the same warning. It affects mu-referencing during estimation only, not simulation.

  • The two models share compartment and parameter names, so modellib("Mauro_2025_nilotinib") and modellib("Mauro_2025_nilotinib_lowfat") can be swapped in a simulation script by changing the model name alone – except that the Final Model does not declare FED_LOWFAT and will error if the column is supplied.

  • Every administration needs two dose records (central with rate = -2, and depot2). Simulating with only the central record loses 11% of the absorbed dose and the whole late-re-absorption phase.