Skip to contents

Furmonertinib (Zou 2022)

Source: vignettes/articles/Zou_2022_furmonertinib.Rmd
Zou_2022_furmonertinib.Rmd

Model and source 

#> ℹ parameter labels from comments will be replaced by 'label()'
  • Citation: Zou HX, Zhang YF, Zhong DF, Jiang Y, Liu F, Zhao QY, Zuo Z, Zhang YF, Yan XY (2022). Effect of autoinduction and food on the pharmacokinetics of furmonertinib and its active metabolite characterized by a population pharmacokinetic model. Acta Pharmacologica Sinica 43(7):1865-1874. doi:10.1038/s41401-021-00798-y. PMID 34789919.
  • Description: Semi-mechanistic joint parent-metabolite population PK model for oral furmonertinib (AST2818, a third-generation irreversible EGFR TKI) and its active metabolite AST5902 (Zou 2022). The parent has two-transit-compartment absorption (rate constant ka shared for depot and both transits) feeding a two-compartment parent disposition (CL/F, Vc/F, Q/F, Vp/F). The parent is eliminated through the central compartment; a fraction Fm of the parent elimination becomes AST5902, which is described by a two-compartment metabolite (Clm/(FFm), Vcm/(FFm), and inter-compartment rate constants k67 and k76). Because no intravenous data were available, the absolute parent bioavailability F and the fraction Fm are non-identifiable and are absorbed into the apparent parameters. Autoinduction of furmonertinib metabolism (mediated by CYP3A4) is modelled as an indirect-response (IDR III) enzyme pool with unity baseline: d(A_ENZ)/dt = kENZ * (1 + S * Cc) - kENZ * A_ENZ, and the apparent parent clearance is CLbase/F * A_ENZ. Covariates: alkaline phosphatase (ALP, U/L) via power effect on CLbase/F and on Clm/(FFm) (median 77.2 U/L); body weight (WT, kg) via power effect on Clm/(FFm) (median 65 kg); and a categorical food-with-a-high-fat-meal effect on parent oral bioavailability (+22.4%) and on the fraction converted to AST5902 (-33.5%).
  • Article: https://doi.org/10.1038/s41401-021-00798-y

The packaged model implements the Zou 2022 final semi-mechanistic parent / metabolite population PK model of furmonertinib (AST2818, also known as alflutinib) and its active metabolite AST5902. The parent has two-transit absorption (single rate constant ka shared for depot and both transits) feeding a two-compartment parent disposition (apparent CL/F, Vc/F, Q/F, Vp/F). A fraction Fm of the parent central-compartment elimination becomes AST5902, whose disposition is a two-compartment model with apparent CLm/(F*Fm), Vcm/(F*Fm), and inter-compartment rate constants k67 (central to peripheral) and k76 (peripheral to central). Because no intravenous data are available, the parent bioavailability F and the conversion fraction Fm are non-identifiable and are absorbed into the apparent parameters. Autoinduction of furmonertinib metabolism (mediated by CYP3A4) is described by an indirect-response (IDR III) enzyme pool with unity baseline: d(A_ENZ)/dt = kENZ * (1 + S * Cc) - kENZ * A_ENZ, and the apparent parent clearance is CLbase/F * A_ENZ. Covariates in the final model are ALP (alkaline phosphatase, U/L) on both CLbase/F and Clm/(F*Fm); body weight on Clm/(F*Fm); and a categorical food-with-a- high-fat-meal effect on parent F (+22.4%) and on Fm (-33.5%).

Population

The pooled analysis dataset (Zou 2022 Table 1) comprises 54 subjects across three clinical trials in China: 38 non-small-cell lung cancer (NSCLC) patients with EGFR-sensitising / T790M-resistance mutations enrolled in Study 001 (dose escalation, NCT02973763, n = 14) and Study 002 (dose expansion, NCT03127449, n = 24), and 16 healthy adult male volunteers in Study 004 (food-effect crossover, NCT03926182). NSCLC patients received oral furmonertinib 20, 40, 80, 160, or 240 mg once daily for 21 days per cycle in the fasted state; healthy volunteers received a single 80 mg oral dose after an overnight fast and, in a second period separated by 22 days, immediately after a high-fat, high-calorie breakfast. Baseline demographics (Zou 2022 Table 1): age mean 48 years (SD 13.9, range 21-68), body weight mean 66.5 kg (SD 10.4, range 48-111), body mass index mean 24.0 kg/m^2, height mean 166.1 cm, sex 23 female (42.6%) / 31 male. Baseline clinical chemistry: alanine transaminase mean 21.3 U/L (SD 11.4), aspartate aminotransferase 22.0 U/L (SD 9.2), alkaline phosphatase 88.1 U/L (SD 44.3, median 77.2), total bilirubin 11.9 umol/L (SD 4.6), creatinine clearance 101.7 mL/min (SD 25.9). Hepatic-function status 48 normal / 6 mild dysfunction. The Bayesian concentration dataset contains 1,450 furmonertinib and 1,463 AST5902 plasma concentrations quantified by LC-MS/MS (LLOQ 0.20 ng/mL parent, 0.050 ng/mL metabolite). The same metadata is available programmatically via readModelDb("Zou_2022_furmonertinib")$meta$population.

Source trace

The per-parameter origin is recorded as an in-file comment next to each ini() entry in inst/modeldb/specificDrugs/Zou_2022_furmonertinib.R. The table below collects them in one place for review. All final-model estimates are from Zou 2022 Table 2 “Estimate” column.

Equation / parameter Value Source location
lcl_base (CLbase/F) log(70.5) -> 70.5 L/h Table 2, %RSE 6.41
lvc (Vc/F) log(2897) -> 2897 L Table 2, %RSE 6.35
lka (ka) log(1.34) -> 1.34 1/h Table 2, %RSE 5.63
lq (Q/F) log(12.4) -> 12.4 L/h Table 2, %RSE 34.2
lvp (Vp/F) log(1470) -> 1470 L Table 2, %RSE 26.7
lkenz (kENZ) log(0.00304) -> 0.00304 1/h Table 2, %RSE 6.56; log(2)/kENZ ~ 228 h CYP3A4 half-life
ls (S) log(0.0111) -> 0.0111 1/(ng/mL) Table 2, %RSE 18.1
lcl_ast5902 (Clm/(F*Fm)) log(119) -> 119 L/h Table 2, %RSE 3.56
lvc_ast5902 (Vcm/(F*Fm)) log(291) -> 291 L Table 2, %RSE 9.27
lk67 log(0.952) -> 0.952 1/h Table 2, %RSE 17.5
lk76 log(0.0542) -> 0.0542 1/h Table 2, %RSE 9.20
e_alp_cl_base fixed(-0.505) Table 2 theta_CLbase/F,ALP; equation below Table 2
e_alp_cl_ast5902 fixed(-0.278) Table 2 theta_CLm/(F*Fm),ALP; equation below Table 2
e_wt_cl_ast5902 fixed(0.622) Table 2 theta_CLm/(F*Fm),body weight; equation below Table 2
e_fed_f fixed(0.224) Table 2 theta_F,food; equation below Table 2
e_fed_fm fixed(-0.335) Table 2 theta_Fm,food; equation below Table 2
etalcl variance 0.0780 Table 2 omega^2 CL/F, shrinkage 3.47%
etalvc variance 0.144 Table 2 omega^2 Vc/F, shrinkage 3.39%
etalka variance 0.161 Table 2 omega^2 ka, shrinkage 1.11%
etalcl_ast5902 variance 0.0485 Table 2 omega^2 Clm/(F*Fm), shrinkage 3.25%
etalvc_ast5902 variance 0.0970 Table 2 omega^2 Vcm/(F*Fm), shrinkage 11.4%
propSd (parent) 0.336 Table 2 delta_ADD_ERR = 33.6% CV
propSd_ast5902 (metabolite) 0.275 Table 2 delta_ADD_ERR_AST5902 = 27.5% CV
ALP median (reference) 77.2 U/L Table 1 baseline demographics
Body weight median (reference) 65 kg Table 1 baseline demographics
ODE d/dt(depot) -ka * depot Methods “Structural model building” para 1
ODE d/dt(transit1..2) ka chain Methods “Structural model building” para 1; Fig 1
ODE d/dt(central) ka in, cl + Q/Vc out, Vp Methods “Structural model building” para 2
ODE d/dt(peripheral1) Q/Vc, Q/Vp Methods “Structural model building” para 2
ODE d/dt(central_ast5902) Fm-scaled formation, Clm out Methods “Structural model building” para 3
ODE d/dt(peripheral1_ast5902) k67 / k76 Methods “Structural model building” para 3
ODE d/dt(enz_pool) kENZ*(1 + S*Cc) - kENZ*enz Methods “Structural model building” Eq. 10-11
enz_pool(0) 1 Methods “Structural model building” para 5
f(depot) 1 + 0.224 * FED_HIGHFAT Covariate equation FTV below Table 2
Formation Fm scaling 1 - 0.335 * FED_HIGHFAT Covariate equation FmTV below Table 2
CLbase/F covariate (ALP/77.2)^(-0.505) CLbase_TV equation below Table 2
Clm/(F*Fm) covariate (ALP/77.2)^(-0.278) * (WT/65)^0.622 CLm_TV equation below Table 2

Virtual cohort

Zou 2022 does not publish per-subject observed concentration-time data. The vignette builds a synthetic virtual cohort at three of the five dose levels studied in Study 001 / 002 (20, 80, and 240 mg once daily for 21 days), using the covariate summary statistics for the pooled analysis dataset (Zou 2022 Table 1). Simulations are run at a fixed typical body weight (65 kg, the cohort median) and ALP (77.2 U/L, the cohort median) so that any differences between dose arms are driven by the autoinduction mechanism alone, matching the paper’s Fig 6 simulation setup (“the typical subject in the dataset (fasting, weighing 65 kg and the ALP level at 77.2 U/L)”). The cohort is 100 subjects per dose arm (well under the 200-per-arm skill cap); IIV is sampled from the omega block reported in Table 2. All doses are in the fasted state; the food-effect layer is exercised separately in the food-effect section below.

set.seed(20220308)

mod <- readModelDb("Zou_2022_furmonertinib")

n_per_arm <- 100L
cycle_days <- 21L
tau_h <- 24
sim_horizon_h <- cycle_days * tau_h

dose_levels_mg <- c(20, 80, 240)

# Build the event table for one dose arm at a time then bind.
make_dose_arm <- function(dose_mg, id_offset) {
  ids <- id_offset + seq_len(n_per_arm)
  # Once-daily oral doses for `cycle_days`.
  doses <- expand.grid(id = ids, dose_time = seq(0, sim_horizon_h - tau_h, by = tau_h)) |>
    dplyr::transmute(
      id, time = dose_time, evid = 1L, amt = dose_mg,
      cmt = "depot", dvid = NA_integer_,
      dose_mg = dose_mg
    )
  # Observation grid: dense over first 24 h (absorption / early elimination),
  # then daily trough plus mid-cycle points to characterise the autoinduction.
  obs_times <- sort(unique(c(
    seq(0, 24, by = 0.5),
    seq(24, 168, by = 6),
    seq(168, sim_horizon_h, by = 24),
    (cycle_days - 1) * tau_h + seq(0, 24, by = 0.5)
  )))
  obs <- expand.grid(id = ids, time = obs_times) |>
    dplyr::transmute(
      id, time, evid = 0L, amt = NA_real_,
      cmt = "central", dvid = 1L, dose_mg = dose_mg
    )
  dplyr::bind_rows(doses, obs) |>
    dplyr::mutate(
      ALP = 77.2, WT = 65, FED_HIGHFAT = 0L
    ) |>
    dplyr::arrange(id, time, dplyr::desc(evid))
}

events <- dplyr::bind_rows(
  make_dose_arm(20,  id_offset =    0L),
  make_dose_arm(80,  id_offset = 1000L),
  make_dose_arm(240, id_offset = 2000L)
)
stopifnot(!anyDuplicated(unique(events[, c("id", "time", "evid")])))

Simulation 

sim <- rxode2::rxSolve(
  mod,
  events = events,
  keep   = c("dose_mg", "ALP", "WT", "FED_HIGHFAT")
) |>
  as.data.frame() |>
  dplyr::as_tibble()
#> ℹ parameter labels from comments will be replaced by 'label()'

Replicate published figures

Figure 6a – Relative increase in furmonertinib CL/F over the 21-day cycle

Zou 2022 Fig 6a shows the model-predicted relative increase in furmonertinib apparent clearance CL/F from day 0 to day 21 for the typical subject (fasting, 65 kg, ALP 77.2 U/L), across the five dose levels 20-240 mg. The paper reports approximately 1.1-fold, 1.3-fold, and 1.8-fold increases at 20, 80, and 240 mg respectively (Results “Model-based simulation” paragraph 1). Here we plot the median enz_pool (which is exactly CL/F / CLbase/F) across the virtual cohort for each dose arm.

induction_by_time <- sim |>
  dplyr::filter(!is.na(enz_pool)) |>
  dplyr::group_by(dose_mg, time) |>
  dplyr::summarise(
    med = median(enz_pool, na.rm = TRUE),
    q05 = quantile(enz_pool, 0.05, na.rm = TRUE),
    q95 = quantile(enz_pool, 0.95, na.rm = TRUE),
    .groups = "drop"
  )

ggplot(induction_by_time, aes(time / 24, med, colour = factor(dose_mg),
                              fill = factor(dose_mg))) +
  geom_ribbon(aes(ymin = q05, ymax = q95), alpha = 0.15, colour = NA) +
  geom_line(linewidth = 0.9) +
  scale_colour_brewer(palette = "Dark2", name = "Dose (mg)") +
  scale_fill_brewer(palette = "Dark2", name = "Dose (mg)") +
  labs(
    x       = "Time after first dose (days)",
    y       = "Relative furmonertinib clearance CL/F (fold vs. baseline)",
    title   = "Fig 6a -- Autoinduction of CL/F across a 21-day cycle",
    caption = "Replicates the shape of Zou 2022 Figure 6a."
  )

Figure 6b – Furmonertinib, AST5902, and total active-compound concentrations

Zou 2022 Fig 6b shows the model-based simulation of the concentration-time profile of furmonertinib, AST5902, and the sum of the two (“total active compounds”) for the typical subject at each dose level. The paper’s key observation is that furmonertinib decreases over time (autoinduction) while AST5902 increases, and the two roughly counteract so the total active exposure is relatively stable. To reproduce the visual we use the median of the virtual cohort at each observation time.

med_conc <- sim |>
  dplyr::filter(!is.na(Cc), !is.na(Cc_ast5902)) |>
  dplyr::group_by(dose_mg, time) |>
  dplyr::summarise(
    Cc_med         = median(Cc, na.rm = TRUE),
    Cc_ast5902_med = median(Cc_ast5902, na.rm = TRUE),
    Cc_total_med   = median(Cc + Cc_ast5902, na.rm = TRUE),
    .groups = "drop"
  ) |>
  tidyr::pivot_longer(
    cols      = c(Cc_med, Cc_ast5902_med, Cc_total_med),
    names_to  = "analyte",
    values_to = "conc"
  ) |>
  dplyr::mutate(
    analyte = dplyr::recode(analyte,
                            Cc_med         = "Furmonertinib",
                            Cc_ast5902_med = "AST5902",
                            Cc_total_med   = "Total active")
  )

ggplot(med_conc, aes(time / 24, conc, colour = analyte)) +
  geom_line(linewidth = 0.7) +
  facet_wrap(~ paste0(dose_mg, " mg"), nrow = 1, scales = "free_y") +
  scale_colour_manual(values = c(Furmonertinib = "steelblue",
                                 AST5902       = "darkorange",
                                 `Total active` = "seagreen"),
                      name = NULL) +
  labs(
    x       = "Time after first dose (days)",
    y       = "Plasma concentration (ng/mL)",
    title   = "Fig 6b -- Furmonertinib / AST5902 / total active concentration",
    caption = "Replicates the shape of Zou 2022 Figure 6b."
  )

Figure 2 – Single-dose vs steady-state apparent clearance

Zou 2022 Fig 2 shows estimated furmonertinib CL/F on the first-dose occasion and after multiple-dose treatment (steady state), with paired-t-test p-values below 0.05 for the 160 and 240 mg groups. Here we compute the ratio of the day-21 CL/F (= CLbase/F * enz_pool(day 21)) to the day-0 CL/F (= CLbase/F when enz_pool = 1) per subject and dose, and show the distribution across the virtual cohort.

day0_end <- 0
day21_time <- (cycle_days - 1) * tau_h

day21_ratio <- sim |>
  dplyr::filter(time == day21_time) |>
  dplyr::group_by(id, dose_mg) |>
  dplyr::summarise(enz_ratio = enz_pool[1], .groups = "drop")

ggplot(day21_ratio, aes(factor(dose_mg), enz_ratio)) +
  geom_boxplot(fill = "lightblue", outlier.size = 0.6) +
  geom_hline(yintercept = 1, linetype = 2, colour = "grey40") +
  labs(
    x     = "Dose (mg)",
    y     = "CL/F ratio (day 21 / day 0)",
    title = "Fig 2 -- Single-dose vs steady-state CL/F ratio by dose",
    caption = "Replicates the shape of Zou 2022 Figure 2. Ratios of 1.1x, 1.3x, and 1.8x are reported at 20, 80, and 240 mg."
  )

Food effect

Zou 2022 Study 004 dosed 16 healthy males in a two-period crossover at 80 mg under fasted and fed (high-fat, high-calorie breakfast) conditions. The final model attributes the food effect to a +22.4% increase in parent bioavailability F and a -33.5% decrease in the conversion fraction Fm (covariate equations below Table 2). The paper’s NCA (Zou 2022 Discussion paragraph 3, citing ref [20]) reports the AUC of furmonertinib increased by approximately 32% under fed vs fasted, while the AUC of AST5902 decreased by approximately 8%. Here we simulate a single 80 mg dose under both conditions and compute the fed/fasted AUC ratio via PKNCA.

n_fe <- 100L

make_food_arm <- function(fed_flag, id_offset, label) {
  ids <- id_offset + seq_len(n_fe)
  dose_row <- data.frame(
    id = ids, time = 0, evid = 1L, amt = 80,
    cmt = "depot", dvid = NA_integer_,
    ALP = 77.2, WT = 65, FED_HIGHFAT = fed_flag,
    fed_label = label
  )
  obs_times <- sort(unique(c(
    seq(0, 12, by = 0.25),
    seq(12, 24, by = 0.5),
    seq(24, 72, by = 2),
    seq(72, 504, by = 12)
  )))
  obs_row <- expand.grid(id = ids, time = obs_times) |>
    dplyr::transmute(
      id, time, evid = 0L, amt = NA_real_,
      cmt = "central", dvid = 1L,
      ALP = 77.2, WT = 65, FED_HIGHFAT = fed_flag,
      fed_label = label
    )
  dplyr::bind_rows(dose_row, obs_row) |>
    dplyr::arrange(id, time, dplyr::desc(evid))
}

events_fe <- dplyr::bind_rows(
  make_food_arm(0L, id_offset =    0L, label = "Fasted"),
  make_food_arm(1L, id_offset = 1000L, label = "Fed")
)

sim_fe <- rxode2::rxSolve(
  mod,
  events = events_fe,
  keep   = c("fed_label", "ALP", "WT", "FED_HIGHFAT")
) |>
  as.data.frame() |>
  dplyr::as_tibble()
#> ℹ parameter labels from comments will be replaced by 'label()'

# PKNCA parent
sim_nca_p <- sim_fe |>
  dplyr::filter(!is.na(Cc)) |>
  dplyr::select(id, time, Cc, fed_label)
sim_nca_p <- dplyr::bind_rows(
  sim_nca_p,
  sim_nca_p |>
    dplyr::distinct(id, fed_label) |>
    dplyr::mutate(time = 0, Cc = 0)
) |>
  dplyr::distinct(id, fed_label, time, .keep_all = TRUE) |>
  dplyr::arrange(id, fed_label, time)

sim_nca_m <- sim_fe |>
  dplyr::filter(!is.na(Cc_ast5902)) |>
  dplyr::select(id, time, Cc_ast5902, fed_label)
sim_nca_m <- dplyr::bind_rows(
  sim_nca_m,
  sim_nca_m |>
    dplyr::distinct(id, fed_label) |>
    dplyr::mutate(time = 0, Cc_ast5902 = 0)
) |>
  dplyr::distinct(id, fed_label, time, .keep_all = TRUE) |>
  dplyr::arrange(id, fed_label, time)

dose_fe <- events_fe |>
  dplyr::filter(evid == 1) |>
  dplyr::select(id, time, amt, fed_label)

conc_obj_p <- PKNCA::PKNCAconc(sim_nca_p,     Cc         ~ time | fed_label + id,
                               concu = "ng/mL", timeu = "hr")
conc_obj_m <- PKNCA::PKNCAconc(sim_nca_m,     Cc_ast5902 ~ time | fed_label + id,
                               concu = "ng/mL", timeu = "hr")
dose_obj   <- PKNCA::PKNCAdose(dose_fe,       amt        ~ time | fed_label + id,
                               doseu = "mg")

intervals_fe <- data.frame(
  start   = 0, end = 504,
  cmax    = TRUE, tmax = TRUE, auclast = TRUE, aucinf.obs = TRUE
)

nca_p <- PKNCA::pk.nca(PKNCA::PKNCAdata(conc_obj_p, dose_obj, intervals = intervals_fe))
nca_m <- PKNCA::pk.nca(PKNCA::PKNCAdata(conc_obj_m, dose_obj, intervals = intervals_fe))

gm <- function(x) exp(mean(log(x[x > 0]), na.rm = TRUE))
food_gmr <- function(nca, label) {
  as.data.frame(nca$result) |>
    dplyr::filter(PPTESTCD %in% c("cmax", "auclast", "aucinf.obs")) |>
    dplyr::group_by(fed_label, PPTESTCD) |>
    dplyr::summarise(gm = gm(PPORRES), .groups = "drop") |>
    tidyr::pivot_wider(names_from = fed_label, values_from = gm) |>
    dplyr::mutate(analyte = label, ratio_fed_over_fasted = Fed / Fasted) |>
    dplyr::select(analyte, PPTESTCD, Fasted, Fed, ratio_fed_over_fasted)
}

food_summary <- dplyr::bind_rows(
  food_gmr(nca_p, "Furmonertinib"),
  food_gmr(nca_m, "AST5902")
)

published <- tibble::tribble(
  ~analyte,        ~PPTESTCD, ~published_ratio,
  "Furmonertinib", "auclast", 1.32,   # Zou 2022 Discussion paragraph 3: 32% increase (NCA)
  "AST5902",       "auclast", 0.92    # Zou 2022 Discussion paragraph 3: 8% decrease (NCA)
)

food_cmp <- food_summary |>
  dplyr::left_join(published, by = c("analyte", "PPTESTCD"))

knitr::kable(
  food_cmp |>
    dplyr::rename(
      Analyte              = analyte,
      `NCA parameter`      = PPTESTCD,
      `Fasted (sim)`       = Fasted,
      `Fed (sim)`          = Fed,
      `Fed/Fasted (sim)`   = ratio_fed_over_fasted,
      `Fed/Fasted (paper NCA)` = published_ratio
    ),
  digits = 3,
  caption = paste0(
    "Simulated fed/fasted geometric-mean ratios vs the paper's NCA-derived ratios ",
    "(Zou 2022 Discussion paragraph 3, citing ref [20]). The model predicts a larger ",
    "downward shift in AST5902 exposure than the NCA reports; see Assumptions and deviations."
  )
)
Simulated fed/fasted geometric-mean ratios vs the paper’s NCA-derived ratios (Zou 2022 Discussion paragraph 3, citing ref [20]). The model predicts a larger downward shift in AST5902 exposure than the NCA reports; see Assumptions and deviations.
Analyte NCA parameter Fasted (sim) Fed (sim) Fed/Fasted (sim) Fed/Fasted (paper NCA)
Furmonertinib aucinf.obs 1117.971 1333.341 1.193 NA
Furmonertinib auclast 1105.481 1319.101 1.193 1.32
Furmonertinib cmax 24.292 29.722 1.224 NA
AST5902 aucinf.obs 650.884 542.291 0.833 NA
AST5902 auclast 639.294 532.418 0.833 0.92
AST5902 cmax 5.098 4.300 0.843 NA

PKNCA validation – single-dose NCA at 80 mg fasted

The Study 004 fasted arm underpins the paper’s food-effect NCA (ref [20]). The Zou 2022 Discussion (paragraph 3) does not print the fasted NCA metrics themselves, but Cmax and AUC(0-inf) values for the 80 mg fasted arm are routinely reported in the underlying single-dose PK study. Here we run PKNCA on the simulated 80 mg fasted arm to document that the model reproduces reasonable NCA metrics; the primary quantitative comparison against the paper is the fed/fasted AUC ratio table above.

sim_sd <- sim_fe |>
  dplyr::filter(fed_label == "Fasted", !is.na(Cc)) |>
  dplyr::select(id, time, Cc)

sim_sd <- dplyr::bind_rows(
  sim_sd,
  sim_sd |> dplyr::distinct(id) |> dplyr::mutate(time = 0, Cc = 0)
) |>
  dplyr::distinct(id, time, .keep_all = TRUE) |>
  dplyr::arrange(id, time)

dose_sd <- events_fe |>
  dplyr::filter(evid == 1, fed_label == "Fasted") |>
  dplyr::select(id, time, amt)

conc_sd <- PKNCA::PKNCAconc(sim_sd, Cc ~ time | id, concu = "ng/mL", timeu = "hr")
dose_sd_obj <- PKNCA::PKNCAdose(dose_sd, amt ~ time | id, doseu = "mg")

intervals_sd <- data.frame(
  start = 0, end = 504,
  cmax = TRUE, tmax = TRUE, auclast = TRUE, aucinf.obs = TRUE, half.life = TRUE
)

nca_sd <- PKNCA::pk.nca(PKNCA::PKNCAdata(conc_sd, dose_sd_obj, intervals = intervals_sd))

nca_summary <- as.data.frame(nca_sd$result) |>
  dplyr::filter(PPTESTCD %in% c("cmax", "tmax", "auclast", "aucinf.obs", "half.life")) |>
  dplyr::group_by(PPTESTCD) |>
  dplyr::summarise(
    median = median(PPORRES, na.rm = TRUE),
    q05    = quantile(PPORRES, 0.05, na.rm = TRUE),
    q95    = quantile(PPORRES, 0.95, na.rm = TRUE),
    .groups = "drop"
  ) |>
  dplyr::mutate(
    parameter = dplyr::recode(PPTESTCD,
                              cmax       = "Cmax (ng/mL)",
                              tmax       = "Tmax (h)",
                              auclast    = "AUC(0-tlast) (ng*h/mL)",
                              aucinf.obs = "AUC(0-inf) (ng*h/mL)",
                              half.life  = "t1/2 (h)")
  ) |>
  dplyr::select(parameter, median, q05, q95)

knitr::kable(
  nca_summary |>
    dplyr::rename(
      `NCA parameter` = parameter,
      `Median (sim)`  = median,
      `5% (sim)`      = q05,
      `95% (sim)`     = q95
    ),
  digits = 2,
  caption = "Simulated single-dose (80 mg fasted) NCA summary for furmonertinib across the 100-subject virtual cohort."
)
Simulated single-dose (80 mg fasted) NCA summary for furmonertinib across the 100-subject virtual cohort.
NCA parameter Median (sim) 5% (sim) 95% (sim)
AUC(0-inf) (ng*h/mL) 1107.51 740.09 1720.71
AUC(0-tlast) (ng*h/mL) 1099.94 737.78 1684.12
Cmax (ng/mL) 23.75 14.50 45.24
t1/2 (h) 100.92 93.18 116.53
Tmax (h) 5.62 3.50 9.01

Assumptions and deviations

  • Indirect-response form for autoinduction. The paper describes a linear-slope IDR III form (“S is the slope parameter describing the relationship between drug concentration and the enzyme formation rate”), arrived at after an initial Emax attempt yielded extremely large Emax / SC50 (Discussion paragraph 2). The packaged model encodes the linearised Emax stimulation as d(A_ENZ)/dt = kENZ * (1 + S * Cc) - kENZ * A_ENZ, which matches the paper’s Eq. 10-11 language and reproduces the reported 1.1 / 1.3 / 1.8-fold CL/F increases at 20 / 80 / 240 mg (paper Results “Model-based simulation” paragraph 1; verified in Figure 6a chunk above).
  • Non-identifiability of F and Fm. The parent bioavailability F and the fraction of parent metabolised to AST5902 (Fm) are absorbed into the apparent parameters CL/F, Vc/F, Q/F, Vp/F, Clm/(F*Fm), and Vcm/(F*Fm) because no intravenous data are available (Zou 2022 Methods “Structural model building” paragraphs 2-3). The absolute bioavailability, absolute conversion fraction, and absolute metabolite distribution volume cannot be recovered from the model. The food effect layer is implemented as: (a) f(depot) = 1 + 0.224 * FED_HIGHFAT on parent bioavailability, and (b) a (1 - 0.335 * FED_HIGHFAT) multiplicative modifier on the parent-to-metabolite formation flux, which is the direct encoding of FTV = 1 + 0.224 * FOOD and `FmTV = 1 - 0.335
    • FOOD` (Zou 2022 covariate equations below Table 2).
  • Food-effect NCA gap. The paper’s model estimates a food effect that produces approximately +22.4% parent bioavailability and -33.5% Fm; the combined effect on the metabolite AUC is a decrease of ~18-19%, larger in magnitude than the NCA-reported -8%. This apparent gap reflects a known structural simplification acknowledged in the Discussion: the paper’s model does not separately identify the first-pass and systemic- circulation fractions of AST5902 formation (“The result of the food effect implies an alternative model structure … However, such a model is not supported by current data”). The packaged model faithfully encodes the paper’s structural equations; the divergence in the metabolite fed/fasted ratio is not a deviation from the model but a known feature of the simplified single-pathway parameterisation.
  • Residual error implementation. Zou 2022 fit an additive residual error on log-transformed data; Table 2 reports each delta as a coefficient of variation (33.6% parent, 27.5% metabolite). The packaged model encodes these as proportional residuals on the linear-scale plasma concentration (Cc ~ prop(propSd) and Cc_ast5902 ~ prop(propSd_ast5902)), which is the equivalent linear-space representation of the paper’s log-scale additive form for small residual SDs. For larger SDs an additive-on-log or lognormal parameterisation would be a closer match.
  • Covariate coefficient fixing. The five covariate exponents (e_alp_cl_base, e_alp_cl_ast5902, e_wt_cl_ast5902, e_fed_f, e_fed_fm) are encoded as fixed() in the packaged model. Zou 2022 reports their point estimates with %RSE and bootstrap 95% CIs (Table 2) but does not report IIV on any of them, so they enter the model as typical-value structural quantities rather than as random-effect-carrying parameters.
  • Body weight covariate implementation. The paper’s covariate equation is CLm_TV = 119 * (ALP/77.2)^(-0.278) * (WT/65)^0.622. The packaged model applies these both to the typical value and to the individual parameter via the multiplicative form exp(lcl_ast5902 + etalcl_ast5902) * (ALP/77.2)^e_alp_cl_ast5902 * (WT/65)^e_wt_cl_ast5902, which matches the paper’s power-model expression exactly.
  • ALP sensitivity to two outliers. Zou 2022 Discussion paragraph 3 reports that the ALP covariate significance is driven by two subjects with unusually high baseline ALP (Supplementary Fig. S1). Removing these two subjects rendered ALP non-significant on parent CL/F (p = 0.03 vs 0.61). The packaged model retains the ALP covariate as-published because the primary paper’s final model includes it; users needing a covariate-simplified version can set e_alp_cl_base and e_alp_cl_ast5902 to 0 in a modified copy of the model file.
  • FED vs FED_HIGHFAT canonicalisation. The Study 004 food-effect arm used an FDA-style high-fat, high-calorie breakfast (Methods “Study design” paragraph 3). Because the effect coefficient quantifies the high-fat food-specific shift, the packaged model uses the FED_HIGHFAT canonical covariate rather than the generic FED (inst/references/covariate-columns.md). Users simulating a light or moderate meal should either set FED_HIGHFAT = 0 or supply an interpolated fractional value with clear documentation.
  • No erratum / corrigendum identified. A search of the Acta Pharmacologica Sinica corrections feed and PubMed for the paper’s DOI (10.1038/s41401-021-00798-y) on 2026-07-10 found no published correction. If a correction is later posted that revises any of the Table 2 estimates or the covariate equations, the packaged file’s reference field must be extended to cite the erratum, and the affected in-file source-trace comments repointed accordingly.