Skip to contents

IMC-002 anti-CD47 antibody (Jeong 2025)

Source: vignettes/articles/Jeong_2025_imc002.Rmd
Jeong_2025_imc002.Rmd

Model and source 

ui <- rxode2::rxode(readModelDb("Jeong_2025_imc002"))
#> ℹ parameter labels from comments will be replaced by 'label()'
  • Citation: Jeong S, Lee SY, Kim SH, Kim HT, Yun H-y, Chae J-w, Lee S. Model-Informed Optimal Dosing of Anti-CD47 Antibody Using Target-Mediated Drug Disposition Model. Clin Transl Sci. 2025;18(8):e70321. doi:10.1111/cts.70321. Structure and fixed constants from Data S1 (NONMEM control stream); final parameter values from Table 2.
  • Description: Semi-mechanistic target-mediated drug disposition (TMDD) model for intravenous IMC-002, a fully human anti-CD47 IgG4 monoclonal antibody, in patients with advanced solid tumours. Free drug in the central compartment binds CD47 by full (non-QSS) second-order kinetics to form a complex that is catabolised, and is also taken up into a peripheral space where it equilibrates with FcRn by a quasi-steady-state quadratic and is recycled back to plasma. CD47 turns over by zero-order synthesis and first-order degradation. Every amount and concentration in the model is molar (umol and umol/L); the observed quantity is the FREE IMC-002 concentration.
  • Article: https://doi.org/10.1111/cts.70321
  • Supplement (Data S1, NONMEM control stream; Figure S1; Table S1): https://www.ebi.ac.uk/europepmc/webservices/rest/PMC12370379/supplementaryFiles

IMC-002 is a fully human anti-CD47 IgG4 monoclonal antibody. Jeong 2025 fitted a semi-mechanistic target-mediated drug disposition (TMDD) model to the phase Ia dose-escalation data and used it to recommend a phase Ib regimen of 20 mg/kg every three weeks (Q3W).

Population

Twelve patients with advanced solid tumours (9 hepatocellular carcinoma, 2 breast cancer, 1 gallbladder cancer) who had failed standard therapy were enrolled in a single-centre phase Ia open-label dose-escalation study in the Republic of Korea (NCT05276310). Three patients were assigned to each of four dose levels (5, 10, 20 and 30 mg/kg), given as a 3 h intravenous infusion every two weeks until progression or withdrawal, yielding 213 free-IMC-002 concentrations. Mean age was 58.8 years (SD 8.0) and mean weight 64.5 kg (SD 10.2); 8 patients were male and 4 female, and all 12 were Korean (Jeong 2025 Table 1; sampling schedule in Table S1). Weight, sex and age were screened as covariates by stepwise covariate modelling (forward p = 0.05, backward p = 0.01) and none was retained, so the packaged model carries no covariate effects.

The same information is available programmatically:

str(ui$population, max.level = 1)
#> List of 13
#>  $ species       : chr "human"
#>  $ n_subjects    : num 12
#>  $ n_studies     : num 1
#>  $ study_id      : chr "NCT05276310 (phase Ia open-label dose escalation, single centre)"
#>  $ age_mean      : chr "58.8 years (SD 8.0)"
#>  $ weight_mean   : chr "64.5 kg (SD 10.2)"
#>  $ sex_female_pct: num 33.3
#>  $ race_ethnicity: Named num 100
#>   ..- attr(*, "names")= chr "Asian"
#>  $ disease_state : chr "Adults with advanced solid tumours who had failed standard therapy: hepatocellular carcinoma (9), breast cancer"| __truncated__
#>  $ dose_range    : chr "5, 10, 20 or 30 mg/kg IMC-002 as a 3 h intravenous infusion every 2 weeks; 3 patients per dose level"
#>  $ regions       : chr "Republic of Korea"
#>  $ n_observations: num 213
#>  $ notes         : chr "Demographics from Jeong 2025 Table 1; the per-cohort sampling schedule is Table S1. Dosing is prescribed in mg/"| __truncated__

Model structure

Four states are solved. central and peripheral1 carry drug amounts (umol); target (CD47) and complex (the IMC-002-CD47 complex) carry concentrations (umol/L) referred to the central volume. That asymmetry is the published parameterisation and is why the volume factors differ between the drug and receptor equations.

Two distinct binding processes act:

  • CD47, in the central compartment, as full second-order TMDD. The association rate is derived as Kon = Koff / KD,CD47 from two in-house fixed constants; CD47 turns over by zero-order synthesis Ksyn and first-order degradation Kdeg, so its drug-free baseline is Ksyn / Kdeg. The complex is catabolised, which is the nonlinear elimination route.
  • FcRn, in the peripheral space, at quasi-steady state. Drug taken up at Kup equilibrates with a fixed total FcRn pool by the QSS quadratic; the bound fraction is recycled back to plasma at Krec. Because the FcRn pool is finite, this limb saturates at high peripheral concentrations.
cat(vapply(ui$lstExpr, function(e) paste(deparse(e), collapse = "\n"), character(1)),
    sep = "\n")
#> vc <- exp(lvc + etalvc)
#> cl <- exp(lcl + etalcl)
#> vp <- exp(lvp)
#> kup <- exp(lkup)
#> ksyn <- exp(lksyn)
#> kdeg <- exp(lkdeg)
#> kd <- exp(lkd)
#> k2 <- exp(lk2)
#> kint <- exp(lkint)
#> kss <- exp(lkss)
#> c_fcrn_t <- exp(lc_fcrn_t)
#> krec <- exp(lkrec)
#> kel <- cl/vc
#> k1 <- k2/kd
#> rbase <- ksyn/kdeg
#> ctot_p <- peripheral1/vp
#> daa <- ctot_p - c_fcrn_t - kss
#> cfree_p <- 0.5 * (daa + sqrt(daa * daa + 4 * kss * ctot_p))
#> ccpx_p <- c_fcrn_t * cfree_p/(kss + cfree_p)
#> d/dt(central) <- -kup * central - kel * central - k1 * central * 
#>     target + k2 * complex * vc + krec * ccpx_p * vp
#> d/dt(target) <- ksyn - kdeg * target - k1 * (central/vc) * target + 
#>     k2 * complex
#> d/dt(complex) <- k1 * (central/vc) * target - k2 * complex - 
#>     kint * complex/vc
#> d/dt(peripheral1) <- kup * central - krec * ccpx_p * vp
#> target(0) <- rbase
#> Cc <- central/vc
#> Cc ~ lnorm(expSd)

Source trace

Per-parameter origin is recorded as an in-file comment beside each ini() entry in inst/modeldb/specificDrugs/Jeong_2025_imc002.R. Collected here for review. “Data S1” is the NONMEM control stream supplied as Supporting Information.

Equation / parameter Value Source location
lvc (Vc) 4.19 L Table 2, RSE 4%; Data S1 $THETA 1
lksyn (Ksyn) 0.0027 umol/L/h Table 2, RSE 8%; Data S1 $THETA 2
lkdeg (Kdeg) 0.0213 1/h, fixed Table 2 (starred); sec. 3.2, from CD47 half-life (ref. 14); Data S1 $THETA 3 FIX
lkd (KD,CD47) 0.046 umol/L, fixed Table 2 (starred); sec. 3.2, in-house binding; Data S1 $THETA 4 FIX
lk2 (Koff) 90.4 1/h, fixed Table 2 (starred); sec. 3.2, in-house binding; Data S1 $THETA 5 FIX
lcl (CL) 0.0115 L/h Table 2, RSE 12%; Data S1 $THETA 6
lvp (Vp) 85.1 L Table 2, RSE 18%; Data S1 $THETA 7
lkup (Kup) 0.0097 1/h Table 2, RSE 5%; Data S1 $THETA 8
lkss (KD,FcRn) 0.117 umol/L, fixed Table 2 (starred); Data S1 $THETA 9 FIX (“KSS1”)
lc_fcrn_t (Rtot,FcRn) 0.291 umol/L, fixed Table 2 (starred); sec. 3.2, from Li 2018 (ref. 15); Data S1 $THETA 10 FIX
lkrec (Krec) 0.0142 1/h Table 2, RSE 11%; Data S1 $THETA 11
lkint (Kel,D-R complex) 0.0188 1/h Table 2, RSE 13%; Data S1 $THETA 12
expSd 0.119 log units Table 2 row “Proportional error”, RSE 9%; Data S1 $THETA 13
IIV variances 0.037 (Vc), 0.319 (CL) Table 2 “Interindividual variability” rows
IIV covariance 0.0199 Data S1 $OMEGA BLOCK(2) only; not printed in Table 2
Kon = Koff / KD,CD47 n/a Data S1 $PK; paper sec. 3.2 symbol list
target(0) = Ksyn / Kdeg n/a Data S1 $PK (BASE); paper sec. 3.2 initial conditions
QSS quadratic for free peripheral drug n/a Data S1 $DES (DAA / Dfree / Dcpx); paper sec. 3.2 equation for C
d/dt(central), d/dt(target), d/dt(complex), d/dt(peripheral1) n/a Data S1 $DES; paper sec. 3.2 differential equations

Unit conversion

The model is parameterised entirely in molar units because the fixed binding constants (KD,CD47, KD,FcRn, Rtot,FcRn) and the estimated Ksyn are all reported in umol/L. The paper prescribes doses in mg/kg and reports exposures in ug/mL, so a molecular weight is needed to move between the two. Jeong 2025 does not report the molecular weight of IMC-002; a representative human IgG4 value is assumed here and flagged in “Assumptions and deviations”.

MW  <- 146000      # g/mol, assumed human IgG4 molecular weight (NOT from the paper)
WT  <- 64.5        # kg, Table 1 mean body weight
TAU <- 21 * 24     # h, Q3W dosing interval
NCYC <- 8L         # cycles simulated; the paper reports steady state by cycle 5
MEC <- 24          # ug/mL, minimum effective concentration (sec. 2.3)

mgkg_to_umol <- function(mgkg) mgkg * WT * 1000 / MW   # mg/kg -> umol for a 64.5 kg patient
umolL_to_ugmL <- function(x) x * MW / 1000             # umol/L -> ug/mL

Structural checks

These are deterministic typical-value checks with no simulated variability, so they are exact rather than cohort-dependent and can be asserted tightly.

# Guard: the model must solve as an ODE system. A cl/vc pair can make rxode2
# auto-substitute a linear closed-form solution and silently discard the ODEs.
stopifnot(!isTRUE(ui$linCmt),
          identical(ui$state, c("central", "target", "complex", "peripheral1")))

mod_typ <- rxode2::zeroRe(readModelDb("Jeong_2025_imc002"))
#> ℹ parameter labels from comments will be replaced by 'label()'
ev_typ <- rxode2::et(amt = mgkg_to_umol(20), dur = 3, ii = TAU,
                     until = TAU * (NCYC - 1), cmt = "central") |>
  rxode2::et(seq(0, TAU * NCYC, by = 0.25), cmt = "central")
sim_typ <- rxode2::rxSolve(mod_typ, ev_typ, returnType = "data.frame",
                           atol = 1e-12, rtol = 1e-10, maxsteps = 1e6)
#> ℹ omega/sigma items treated as zero: 'etalvc', 'etalcl'

# 1. Drug-free CD47 baseline must equal Ksyn / Kdeg exactly (Data S1 $PK BASE).
base_expected <- 0.0027 / 0.0213
stopifnot(abs(sim_typ$target[1] / base_expected - 1) < 1e-8)

# 2. Steady-state mass balance. Over one dosing interval at steady state the dose
#    must equal what leaves the system: linear clearance from central plus
#    catabolism of the drug-CD47 complex. This is the cheapest gate that pins the
#    whole ODE system, including the two fixed binding constants.
trap <- function(x, y) sum(diff(x) * (utils::head(y, -1) + utils::tail(y, -1)) / 2)
ss <- sim_typ |>
  dplyr::filter(time >= TAU * (NCYC - 1), time <= TAU * NCYC) |>
  dplyr::arrange(time)
elim_linear  <- trap(ss$time, ss$kel * ss$central)
elim_complex <- trap(ss$time, ss$kint * ss$complex)
elim_total   <- elim_linear + elim_complex
stopifnot(abs(elim_total / mgkg_to_umol(20) - 1) < 0.01)

tibble::tibble(
  Check = c("CD47 baseline = Ksyn/Kdeg",
            "Steady-state mass balance (dose vs eliminated)",
            "Share of elimination via linear CL",
            "Share of elimination via CD47-complex catabolism"),
  Value = c(sprintf("%.6f umol/L", sim_typ$target[1]),
            sprintf("%.4f%% error", 100 * abs(elim_total / mgkg_to_umol(20) - 1)),
            sprintf("%.1f%%", 100 * elim_linear / elim_total),
            sprintf("%.1f%%", 100 * elim_complex / elim_total))
) |>
  knitr::kable(caption = "Deterministic structural checks at 20 mg/kg Q3W steady state.")
Deterministic structural checks at 20 mg/kg Q3W steady state.
Check Value
CD47 baseline = Ksyn/Kdeg 0.126761 umol/L
Steady-state mass balance (dose vs eliminated) 0.0877% error
Share of elimination via linear CL 52.0%
Share of elimination via CD47-complex catabolism 48.0%

At 20 mg/kg roughly half the dose is cleared by catabolism of the drug-CD47 complex, consistent with the paper’s statement that IMC-002 “is mainly eliminated through drug-complex catabolism, not in the blood compartment” (sec. 4).

Virtual cohort

The published simulations drew 1000 subjects per regimen. Rather than resample randomly, the cohort below is a deterministic product-quadrature lattice over the bivariate log-normal IIV on Vc and CL: 14 stratified quantile points per eta, transformed by the Cholesky factor of the published OMEGA. This gives 196 subjects per arm (inside the 200-per-arm cap), reproduces the published variances exactly, and – unlike random sampling – is identical on every machine regardless of the number of solver threads. That removes the usual source of locally-passing / CI-failing assertions.

K <- 14L
z <- stats::qnorm((seq_len(K) - 0.5) / K)
z <- z / sqrt(mean(z^2))   # rescale so the lattice variance is exactly 1
etas <- as.matrix(expand.grid(z_vc = z, z_cl = z)) %*% t(t(chol(ui$omega)))
colnames(etas) <- c("etalvc", "etalcl")

# The lattice reproduces the published OMEGA diagonal exactly.
lattice_var <- diag(stats::var(etas) * (nrow(etas) - 1) / nrow(etas))
stopifnot(max(abs(lattice_var / diag(ui$omega) - 1)) < 1e-8)

doses <- c(5, 10, 20, 30)
arm_levels <- paste0(doses, " mg/kg")

subjects <- lapply(seq_along(doses), function(i) {
  tibble::tibble(
    id     = (i - 1L) * nrow(etas) + seq_len(nrow(etas)),  # disjoint id ranges per arm
    etalvc = etas[, "etalvc"],
    etalcl = etas[, "etalcl"],
    arm    = factor(arm_levels[i], levels = arm_levels),
    amt    = mgkg_to_umol(doses[i])
  )
}) |>
  dplyr::bind_rows()

obs_grid <- unique(c(seq(0, TAU * (NCYC - 1), by = 12),
                     seq(TAU * (NCYC - 1), TAU * NCYC, by = 3)))

events <- dplyr::bind_rows(
  tidyr::crossing(subjects, time = seq(0, TAU * (NCYC - 1), by = TAU)) |>
    dplyr::mutate(evid = 1L, dur = 3, cmt = "central"),
  tidyr::crossing(dplyr::select(subjects, -amt), time = obs_grid) |>
    dplyr::mutate(evid = 0L, amt = NA_real_, dur = NA_real_, cmt = "central")
) |>
  dplyr::arrange(id, time, dplyr::desc(evid))

stopifnot(!anyDuplicated(unique(events[, c("id", "time", "evid")])))

Simulation

etalvc / etalcl are supplied per subject in the event table and omega = NA suppresses resampling, so the lattice is used exactly as constructed.

sim <- rxode2::rxSolve(
  readModelDb("Jeong_2025_imc002"), events, omega = NA,
  keep = "arm", returnType = "data.frame",
  atol = 1e-10, rtol = 1e-8, maxsteps = 1e5
)
#> ℹ parameter labels from comments will be replaced by 'label()'
#> Warning: multi-subject simulation without without 'omega'
sim <- sim |>
  dplyr::mutate(arm = factor(as.character(arm), levels = arm_levels),
                conc = umolL_to_ugmL(Cc))
stopifnot(!anyNA(sim$conc), all(sim$conc >= 0))

Replicate published figures 

sim |>
  dplyr::filter(time >= TAU * (NCYC - 1), time <= TAU * NCYC) |>
  dplyr::mutate(t_in_tau = time - TAU * (NCYC - 1)) |>
  dplyr::group_by(arm, t_in_tau) |>
  dplyr::summarise(Q05 = stats::quantile(conc, 0.05),
                   Q50 = stats::median(conc),
                   Q95 = stats::quantile(conc, 0.95), .groups = "drop") |>
  ggplot(aes(t_in_tau, Q50)) +
  geom_ribbon(aes(ymin = Q05, ymax = Q95), alpha = 0.25) +
  geom_line() +
  geom_hline(yintercept = MEC, linetype = "dashed", colour = "grey40") +
  facet_wrap(~arm) +
  scale_y_log10() +
  labs(x = "Time within the Q3W interval (h)", y = "IMC-002 (ug/mL)",
       title = "Steady-state Q3W profiles by dose level",
       caption = "Replicates Figure 3 of Jeong 2025. Dashed line: MEC 24 ug/mL.")
Replicates Figure 3 of Jeong 2025: simulated IMC-002 concentration over the steady-state Q3W dosing interval by dose level, with the 24 ug/mL MEC.

Replicates Figure 3 of Jeong 2025: simulated IMC-002 concentration over the steady-state Q3W dosing interval by dose level, with the 24 ug/mL MEC.

Figure 2 of Jeong 2025 is a visual predictive check overlaying the observed phase Ia concentrations, which are not publicly available (Data Availability Statement), so only the simulation-only Figure 3 is reproduced here.

The TMDD mechanism itself is not plotted in the paper, but the packaged model exposes it directly: free CD47 is driven far below its Ksyn / Kdeg baseline while drug is present and recovers between doses.

sim_typ |>
  dplyr::filter(time >= TAU * (NCYC - 1), time <= TAU * NCYC) |>
  dplyr::transmute(t_in_tau = time - TAU * (NCYC - 1),
                   `Free CD47` = target, `IMC-002-CD47 complex` = complex) |>
  tidyr::pivot_longer(-t_in_tau, names_to = "species", values_to = "conc") |>
  ggplot(aes(t_in_tau, conc, colour = species)) +
  geom_line() +
  geom_hline(yintercept = base_expected, linetype = "dashed", colour = "grey40") +
  labs(x = "Time within the Q3W interval (h)", y = "Concentration (umol/L)",
       colour = NULL,
       caption = "Dashed line: drug-free CD47 baseline Ksyn/Kdeg.") +
  theme(legend.position = "bottom")
Free CD47 and the IMC-002-CD47 complex over the steady-state interval (typical-value subject, 20 mg/kg Q3W). Not a published figure.

Free CD47 and the IMC-002-CD47 complex over the steady-state interval (typical-value subject, 20 mg/kg Q3W). Not a published figure.

PKNCA validation

NCA is taken over the final simulated dosing interval, which is at steady state (the paper reports steady state by cycle 5; cycle 8 is simulated). auclast over that interval is AUCtau and cmin is the steady-state trough: the profile declines monotonically after each infusion, so the minimum sits at one of the two interval endpoints, and at steady state both are troughs. (The MEC section below uses the end-of-interval trough explicitly; the two agree to within 2%.)

sim_nca <- sim |>
  dplyr::filter(!is.na(Cc)) |>
  dplyr::transmute(id, time, Cc = conc, arm)

# Guarantee a time-zero record per subject (Cc = 0 pre-dose) so PKNCA has an anchor.
sim_nca <- dplyr::bind_rows(
  sim_nca,
  sim_nca |> dplyr::distinct(id, arm) |> dplyr::mutate(time = 0, Cc = 0)
) |>
  dplyr::distinct(id, arm, time, .keep_all = TRUE) |>
  dplyr::arrange(id, arm, time)

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

dose_df <- events |>
  dplyr::filter(evid == 1L) |>
  dplyr::select(id, time, amt, arm)
dose_obj <- PKNCA::PKNCAdose(dose_df, amt ~ time | arm + id)

intervals <- data.frame(
  start   = TAU * (NCYC - 1),
  end     = TAU * NCYC,
  auclast = TRUE,
  cmax    = TRUE,
  cmin    = TRUE,
  tmax    = TRUE
)

nca_res <- PKNCA::pk.nca(PKNCA::PKNCAdata(conc_obj, dose_obj, intervals = intervals))
nca_long <- as.data.frame(nca_res)

# A gate that cannot go red is worse than none: confirm every arm/parameter cell
# actually produced a value.
stopifnot(nrow(nca_long) > 0, !anyNA(nca_long$PPORRES))
stopifnot(setequal(unique(nca_long$PPTESTCD), c("auclast", "cmax", "cmin", "tmax")))

Comparison against published NCA

Jeong 2025 Table 3 reports the mean of each parameter over its 1000 simulated subjects. ncaComparisonTable() aggregates per-subject results with the median by default, so the simulated side is pre-aggregated to arm means here to match the published statistic.

sim_wide <- nca_long |>
  dplyr::filter(PPTESTCD %in% c("auclast", "cmin")) |>
  dplyr::mutate(arm = factor(as.character(arm), levels = arm_levels)) |>
  dplyr::group_by(arm, PPTESTCD) |>
  dplyr::summarise(value = mean(PPORRES), .groups = "drop") |>
  tidyr::pivot_wider(names_from = PPTESTCD, values_from = value)

published <- tibble::tibble(
  arm     = factor(arm_levels, levels = arm_levels),
  auclast = c(8689, 23760, 66036, 116776),   # Table 3, AUCtau (ug*h/mL)
  cmin    = c(5.2, 15.6, 58.8, 122.4)        # Table 3, Ctrough (ug/mL)
)

cmp <- nlmixr2lib::ncaComparisonTable(
  simulated = as.data.frame(sim_wide),
  reference = as.data.frame(published),
  by        = "arm",
  units     = c(auclast = "ug*h/mL", cmin = "ug/mL"),
  tolerance_pct = 20
)
knitr::kable(cmp, align = c("l", "l", "r", "r", "r"),
             caption = "Simulated vs. Jeong 2025 Table 3 (Q3W steady state). * differs by >20%.")
Simulated vs. Jeong 2025 Table 3 (Q3W steady state). * differs by >20%.
NCA parameter arm Reference Simulated % diff
Cmin (ug/mL) 5 mg/kg 5.2 5.1 -1.9%
Cmin (ug/mL) 10 mg/kg 15.6 14.5 -7.3%
Cmin (ug/mL) 20 mg/kg 58.8 59.7 +1.5%
Cmin (ug/mL) 30 mg/kg 122 129 +5.6%
AUClast (ug*h/mL) 5 mg/kg 8690 8060 -7.2%
AUClast (ug*h/mL) 10 mg/kg 23800 21700 -8.8%
AUClast (ug*h/mL) 20 mg/kg 66000 64900 -1.8%
AUClast (ug*h/mL) 30 mg/kg 117000 119000 +2.1% 
attr(cmp, "footnote")
#> NULL
# ncaComparisonTable's "% diff" column is formatted character, so the gate is
# computed numerically from the underlying values.
gate <- published |>
  dplyr::rename(ref_auc = auclast, ref_cmin = cmin) |>
  dplyr::left_join(sim_wide, by = "arm") |>
  dplyr::mutate(pct_auc  = 100 * (auclast / ref_auc - 1),
                pct_cmin = 100 * (cmin / ref_cmin - 1))

# The cohort is a deterministic lattice, so these numbers do not move between
# machines or thread counts. Realised: AUCtau -7.2 / -8.8 / -1.8 / +2.1 %,
# Ctrough -1.9 / -7.1 / +3.0 / +7.8 %. The 20% bound leaves headroom for the
# assumed molecular weight (the one non-paper input) while still going red on a
# mis-transcribed volume, dose or unit -- encoding the drug in mass rather than
# molar units, for instance, moves these by 30 to 95%.
stopifnot(max(abs(gate$pct_auc)) < 20, max(abs(gate$pct_cmin)) < 20)

gate |>
  dplyr::transmute(
    `Dose level`             = arm,
    `AUCtau % diff`          = round(pct_auc, 1),
    `Ctrough % diff`         = round(pct_cmin, 1)
  ) |>
  knitr::kable(caption = "Numeric deviation from Jeong 2025 Table 3.")
Numeric deviation from Jeong 2025 Table 3.
Dose level AUCtau % diff Ctrough % diff
5 mg/kg -7.2 -1.9
10 mg/kg -8.8 -7.3
20 mg/kg -1.8 1.5
30 mg/kg 2.1 5.6

Proportion of patients above the MEC

Table 3 also reports the percentage of simulated patients whose steady-state trough exceeds the 24 ug/mL MEC. This is the criterion the 20 mg/kg Q3W recommendation rests on.

mec_sim <- sim |>
  dplyr::filter(time >= TAU * (NCYC - 1), time <= TAU * NCYC) |>
  dplyr::group_by(arm, id) |>
  dplyr::summarise(ctrough = conc[which.max(time)], .groups = "drop") |>
  dplyr::group_by(arm) |>
  dplyr::summarise(simulated = 100 * mean(ctrough > MEC), .groups = "drop")

mec_cmp <- mec_sim |>
  dplyr::mutate(published = c(NA, 15.6, 74.6, 90.5),   # Table 3 "C trough over MEC (%)"
                difference = simulated - published)

# Realised absolute differences: 6.4 / 3.5 / 2.4 percentage points (deterministic).
stopifnot(max(abs(mec_cmp$difference), na.rm = TRUE) < 15)

mec_cmp |>
  dplyr::transmute(`Dose level` = arm,
                   `Simulated % above MEC` = round(simulated, 1),
                   `Published % above MEC` = published,
                   `Difference (points)`   = round(difference, 1)) |>
  knitr::kable(caption = "Percentage of patients with steady-state Q3W trough above 24 ug/mL, vs Jeong 2025 Table 3.")
Percentage of patients with steady-state Q3W trough above 24 ug/mL, vs Jeong 2025 Table 3.
Dose level Simulated % above MEC Published % above MEC Difference (points)
5 mg/kg 0.0 NA NA
10 mg/kg 9.2 15.6 -6.4
20 mg/kg 78.1 74.6 3.5
30 mg/kg 92.9 90.5 2.4

The model reproduces the paper’s decision: at 20 mg/kg Q3W roughly three quarters of patients hold a trough above the MEC, against roughly one in ten at 10 mg/kg. Table 3 has no entry for 5 mg/kg, and the model puts that arm at 0%.

Assumptions and deviations

  • Molecular weight (the only non-paper input). The model is molar throughout, but doses are prescribed in mg/kg and exposures reported in ug/mL. Jeong 2025 never states the molecular weight of IMC-002, so MW = 146000 g/mol – a representative value for a human IgG4 monoclonal antibody – is assumed here to convert. Every ug/mL figure in this vignette therefore carries that assumption; the packaged model file itself does not, and is dosed in umol.
  • The printed equation for the free peripheral concentration has a typo; Data S1 settles it. Section 3.2 prints C = 0.5 * [(Ctot,p - Ap - KD,FcRn) + sqrt((Ctot,p - Ap - KD,FcRn)^2 + 4*KD,FcRn*Ctot,p)], in which Ap is defined two lines later as the amount of IMC-002 in the peripheral compartment – subtracting an amount from a concentration. Data S1 $DES writes the same line as DAA = A(4)/VP - FcRn - KSS1, i.e. the middle term is the total FcRn concentration Rtot,FcRn, which is also what the standard quasi-steady-state root requires. The control-stream form is encoded.
  • Complex elimination is reproduced with its published volume factor. Both the printed equation and Data S1 write the complex catabolism term as Kel,D-R complex * ADR47 / Vc, which is dimensionally inconsistent with the rest of that equation (the other terms are umol/L/h; this one is umol/L/h/L). The estimate 0.0188 1/h was obtained under that form, so it is reproduced verbatim: rescaling it to remove the stray volume would change the model rather than correct it. The effective first-order rate on the complex is therefore 0.0188 / 4.19 = 0.00449 1/h.
  • Kup is a rate constant, not a clearance. Table 2 tags Kup as “(L/h)”, but the equations use it as Kup * AD on an amount, which requires 1/h. The label in the model file records 1/h.
  • Residual error is log-additive, not proportional. Table 2 labels the row “Proportional error”, but Data S1 $ERROR fits IPRED = LOG(A(1)/VC) with Y = IPRED + W*EPS(1), W = THETA(13) and $SIGMA 1 FIX. That is a log-normal residual, encoded as Cc ~ lnorm(expSd). At 0.119 the two forms are numerically close. The Table 2 row “Residual variability 1*” is the fixed $SIGMA scaffolding, not a separate parameter.
  • The IIV covariance comes from the supplement only. Table 2 prints the two variances but no correlation; Data S1 $OMEGA BLOCK(2) supplies the off-diagonal 0.0199 (correlation 0.183). Its diagonals reproduce Table 2 exactly, which is what supports treating the block as final rather than initial estimates.
  • Where Table 2 and Data S1 disagree slightly, Table 2 wins. Data S1 carries Ksyn 0.00273 vs Table 2 0.0027, Vp 85 vs 85.1, and Kup 0.00968 vs 0.0097; the published final table is used for all three, with the control stream supplying structure.
  • Cohort construction. The published simulations used 1000 randomly sampled subjects per regimen; this vignette uses a 196-point deterministic quadrature lattice per arm (14 x 14 stratified quantiles of the two etas, rescaled to reproduce the published variances exactly). Means over a lattice and over 1000 random draws are not identical, which contributes part of the residual deviation in the comparison tables.
  • Body weight is held at the Table 1 mean (64.5 kg). Weight was screened and not retained as a covariate, so it enters only through the mg/kg dose conversion; using the mean weight matches the paper’s per-dose-level reporting.
  • No observed data. The phase Ia concentrations are not publicly available, so the visual predictive check of Figure 2 and the goodness-of-fit plots of Figure S1 cannot be reproduced. Validation here is against the published simulation summary (Table 3) and the model’s own internal mass balance.