Skip to contents

Model and source

  • Citation: Shahraz A, Penney M, Candido J, Opoku-Ansah G, Neubauer M, Eyles J, Ojo O, Liu N, Luheshi NM, Phipps A, Vishwanathan K. A mechanistic PK/PD model of AZD0171 (anti-LIF) to support Phase II dose selection. CPT Pharmacometrics Syst Pharmacol. 2024;13(10):1670-1680. doi:10.1002/psp4.13204. Parameter values from Table 1 and Supplementary Table S4; equations from Supplementary Table S5 (fluxes, ODEs and rules) and Appendix S3 (Krogh tumor-penetration flux).
  • Description: QSP. Mechanistic PK/PD model of falbikitug (AZD0171), a humanized IgG1 anti-leukemia inhibitory factor (LIF) antibody, in adults with advanced solid tumors and metastatic pancreatic ductal adenocarcinoma (Shahraz 2024; Phase I NCT03490669 n = 41 and interim Phase II NCT04999969 n = 49). Three compartments: central (serum), a peripheral tissue compartment and a tumor compartment. AZD0171 has linear clearance from the central compartment, distributes to tissue by an intercompartmental clearance and to the tumor by a Krogh-cylinder permeability-diffusion flux. LIF is synthesized only in the tumor at a zero-order rate, eliminated from the central compartment with a 2 h half-life, and exchanges with tissue and tumor. In tissue and tumor LIF binds LIFR and gp130 independently to form the LIF:LIFR:gp130 signaling complex; AZD0171 binds free LIF and LIF:LIFR (in the central compartment free LIF only) and blocks gp130 recruitment. LIFR and gp130 turn over to fitted / fixed steady-state levels, and occupied receptors internalize. All states start receptor-only with no LIF: like the authors (first dose at 240 h in Figure 4), run the drug-free system for at least 240 h before the first dose so that LIF and the receptor complexes reach baseline. Outputs: serum AZD0171 (Cc, ug/mL), serum total LIF (Ctotal_target, pg/mL), and tumor free LIF and signaling complex for the target-engagement predictions.
  • Article: https://doi.org/10.1002/psp4.13204 (open access; PMC11494920)

Falbikitug (AZD0171) is a humanized IgG1 antibody that binds leukemia inhibitory factor (LIF) and blocks recruitment of gp130 to the LIF receptor (LIFR), so the LIF:LIFR:gp130 signaling complex cannot form. The model has a central (serum) compartment, a peripheral tissue compartment and a tumor compartment. LIF is made only in the tumor; LIF, AZD0171 and the AZD0171:LIF complex move between serum and tumor through a Krogh-cylinder permeability-diffusion flux. In tissue and tumor, LIF binds LIFR and gp130 independently, and AZD0171 binds free LIF and LIF:LIFR.

Population

The model was fitted in SimBiology to serum AZD0171 and serum total LIF from 90 patients: 41 adults with advanced, unresectable solid tumors in the Phase I dose-escalation study (NCT03490669; 75-1500 mg IV every 3 weeks), and 49 patients with first-line metastatic pancreatic ductal adenocarcinoma in an interim cut of the Phase II study (NCT04999969; 1500 mg IV every 2 weeks with durvalumab, gemcitabine and nab-paclitaxel). Each patient was fitted separately; Table 1 reports the median and interquartile range of the five fitted parameters. The paper has no demographic table; sampling schedules are in Appendix S1.

The same information is available programmatically via readModelDb("Shahraz_2024_falbikitug")()$population.

Source trace

Every ini() value carries an in-file comment pointing to its source. The equations come from Supplementary Table S5 (45 fluxes, 11 printed ODEs and 24 rules).

Equation / parameter Value Source location
lcl (CL) 0.0135 L/h Table 1; Table S4 CL_AZD0171
lvc (Vc) 3.72 L Table 1; Table S4 Vcentral
lq (Q, AZD0171 and AZD0171:LIF) 0.110 L/h Table 1; Table S4 Q_drug
lksyn_lif (ksynth_LIFt) 1.67 pmol/h Table 1; Table S4
llifr_ss (LIFR_ss) 5.926 nM Table 1; Table S4
lvp (Vtissue) 3.72 L, fixed Table S4; rule 21 (Vtissue = Vcentral)
v_tumor 0.258 L, fixed Table S4; Methods
kon_drug_lif, kon_lif_lifr, kon_lif_gp130 3.6 1/(nM h), fixed Table S4
kd_drug_lif 0.074 nM, fixed Table S4; koff = kon x KD = 0.2664 1/h (rule 23)
koff_lif_lifr 11.16 1/h, fixed Table S4
koff_lif_gp130 260.568 1/h, fixed Table S4
thalf_lif 2 h, fixed Table S4 (Gunawardana 2003)
gp130_ss 25 nM, fixed Table S4
kint_gp130 4.158 1/h, fixed Table S4
kint_lifr_slow, kint_lifr_fast 0.42, 1.89 1/h, fixed Table S4 (Hilton and Nicola 1992)
r_cap, r_krogh, void_frac 8 um, 75 um, 0.24, fixed Table S4; Appendix S3
mw_drug, mw_lif, mw_drug_lif 150, 25, 175 kDa, fixed Table S4
gamma_qlif, mu_qlif 2.25 L/h, 2.48, fixed Table S4 (Li 2017); Q_LIF by rule 7
Diffusion and permeability derived Table S5 rules 1-6 (Li 2019)
Tumor radius derived Table S5 rule 8
etalcl, etalvc, etalq, etalksyn_lif, etallifr_ss 0.0862 each, fixed Results: 30% CV log-normal virtual patients
d/dt(central) Table S5 ODE 10 (Reactions 3, 4, 6, 38)
d/dt(target), d/dt(complex) assembled from Table S5 Reactions 6, 8, 9, 10, 13, 36, 43
Tissue ODEs (9 states) Table S5 ODEs 1-9
Tumor ODEs (9 states) assembled from Table S5 Reactions 1, 20-25, 32-35, 37, 39, 44-47
Cc (ug/mL) rule 11
Ctotal_target (pg/mL) rule 10
Tumor free LIF, active LIFR fraction rules 18, 19, 24

Derived constants against Supplementary Table S4

The model computes the diffusion coefficients, permeabilities, LIF intercompartmental clearance, tumor radius and AZD0171 dissociation rate from the Table S5 rules. They should match the values that Table S4 prints.

mw_drug <- 150
mw_lif <- 25
mw_drug_lif <- 175
derived <- tibble::tribble(
  ~quantity, ~model, ~table_s4,
  "Diff_LIF (cm^2/h)", 114.58 * mw_lif^(-0.918) * 3600 / 1e7, 0.00215,
  "Diff_Ab (cm^2/h)", 114.58 * mw_drug^(-0.918) * 3600 / 1e7, 0.000415,
  "Diff_AbLIF (cm^2/h)", 114.58 * mw_drug_lif^(-0.918) * 3600 / 1e7, 0.00036,
  "Pcap_LIF (um/h)", (-10.43 * log(mw_lif) + 53.46) * 3600 / 1000, 71.59,
  "Pcap_Ab and Pcap_AbLIF (um/h)", (-10.43 * log(mw_drug) + 53.46) * 3600 / 1000, 4.32,
  "Q_LIF (L/h)", 2.25 * mw_lif^(-2.48) * (70 / 0.028)^0.75, 0.271,
  "Rtumor (cm)", (3 * 0.258 * 1000 / (4 * pi))^(1 / 3), 4,
  "koffAZD0171LIF (1/h)", 3.6 * 0.074, 0.2664
) |>
  mutate(pct_diff = 100 * (model - table_s4) / table_s4)

derived |>
  dplyr::rename(
    "Quantity" = quantity,
    "Model" = model,
    "Table S4" = table_s4,
    "Difference (%)" = pct_diff
  ) |>
  knitr::kable(digits = c(0, 5, 5, 1))
Quantity Model Table S4 Difference (%)
Diff_LIF (cm^2/h) 0.00215 0.00215 -0.1
Diff_Ab (cm^2/h) 0.00041 0.00042 -0.1
Diff_AbLIF (cm^2/h) 0.00036 0.00036 0.0
Pcap_LIF (um/h) 71.59365 71.59000 0.0
Pcap_Ab and Pcap_AbLIF (um/h) 4.31667 4.32000 -0.1
Q_LIF (L/h) 0.27149 0.27100 0.2
Rtumor (cm) 3.94921 4.00000 -1.3
koffAZD0171LIF (1/h) 0.26640 0.26640 0.0

# Table S4 rounds to 2-3 significant figures (Rtumor to 4 cm), so allow the
# rounding error, and no more.
stopifnot(all(abs(derived$pct_diff[derived$quantity != "Rtumor (cm)"]) < 3))
stopifnot(abs(derived$pct_diff[derived$quantity == "Rtumor (cm)"]) < 5)

Baseline (drug-free) steady state

All states start with receptors only and no LIF. The authors started dosing at 240 h (Figure 4), so the drug-free system has time to settle; the same 240 h run-in is used for every simulation below.

mod <- readModelDb("Shahraz_2024_falbikitug")
mod_typical <- rxode2::zeroRe(mod)
#> ℹ parameter labels from comments will be replaced by 'label()'

run_in <- 240 # h before the first dose (Figure 4)

# Event table with dose rows on the central (serum) compartment and
# observation rows carrying dvid = 1 (the model has two error endpoints).
make_events <- function(ids, amt, tau, n_dose, grid, label) {
  dose_times <- run_in + tau * (seq_len(n_dose) - 1)
  dose <- expand.grid(id = ids, time = dose_times)
  dose$amt <- amt
  dose$dur <- 1
  dose$evid <- 1L
  dose$dvid <- NA_integer_
  obs <- expand.grid(id = ids, time = grid)
  obs$amt <- 0
  obs$dur <- 0
  obs$evid <- 0L
  obs$dvid <- 1L
  dplyr::bind_rows(dose, obs) |>
    dplyr::mutate(cmt = "central", regimen = label) |>
    dplyr::arrange(id, time, dplyr::desc(evid))
}
base_ev <- make_events(1L, 0, 336, 1, seq(0, run_in, by = 4), "no drug")
base_sim <- as.data.frame(rxode2::rxSolve(mod_typical, events = base_ev))
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc', 'etalq', 'etalksyn_lif', 'etallifr_ss'
at <- function(df, t, col) df[[col]][which.min(abs(df$time - t))]

baseline <- tibble::tibble(
  quantity = c(
    "Serum free LIF (pg/mL)", "Serum total LIF (pg/mL)",
    "Tumor free LIF (pg/mL)", "Tumor LIF:LIFR:gp130 (nM)"
  ),
  t120 = c(
    at(base_sim, 120, "lif_free_serum_pgml"), at(base_sim, 120, "Ctotal_target"),
    at(base_sim, 120, "lif_free_tumor_pgml"), at(base_sim, 120, "signal_tumor")
  ),
  t240 = c(
    at(base_sim, 240, "lif_free_serum_pgml"), at(base_sim, 240, "Ctotal_target"),
    at(base_sim, 240, "lif_free_tumor_pgml"), at(base_sim, 240, "signal_tumor")
  )
)
baseline |>
  dplyr::rename("Quantity" = quantity, "At 120 h" = t120, "At 240 h" = t240) |>
  knitr::kable(digits = 5)
Quantity At 120 h At 240 h
Serum free LIF (pg/mL) 22.36130 22.36130
Serum total LIF (pg/mL) 22.36130 22.36130
Tumor free LIF (pg/mL) 163.54773 163.54773
Tumor LIF:LIFR:gp130 (nM) 0.00398 0.00398

# The run-in reaches steady state: nothing moves between 120 h and 240 h.
stopifnot(all(abs(baseline$t240 / baseline$t120 - 1) < 1e-4))

The typical patient’s baseline tumor free LIF is about 164 pg/mL. Figure 4 does not agree with itself on this value: the reference curves in Figure 4a, 4e and 4f start near 55-60 pg/mL, while those in Figure 4b, 4c and 4d start near 300-700 pg/mL. The comparisons below therefore use the change relative to baseline.

Replicate Figure 4a: free LIF and signaling complex on 1500 mg Q2W

Eleven 1500 mg infusions every 2 weeks, from 240 h to 3600 h, followed by washout.

grid_long <- sort(unique(c(seq(0, 12000, by = 8), run_in - 4)))
ev_4a <- make_events(1L, 1500, 336, 11, grid_long, "1500 mg Q2W")
sim_4a <- as.data.frame(rxode2::rxSolve(mod_typical, events = ev_4a))
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc', 'etalq', 'etalksyn_lif', 'etallifr_ss'

sim_4a |>
  dplyr::select(time, lif_free_serum_pgml, lif_free_tumor_pgml, signal_tumor) |>
  tidyr::pivot_longer(-time) |>
  dplyr::mutate(name = dplyr::recode(name,
    lif_free_serum_pgml = "Serum free LIF (pg/mL)",
    lif_free_tumor_pgml = "Tumor free LIF (pg/mL)",
    signal_tumor = "Tumor LIF:LIFR:gp130 (nM)"
  )) |>
  # LIF starts at zero before the run-in; drop it from the log-scale plot
  dplyr::filter(value > 0) |>
  ggplot(aes(time, value, colour = name)) +
  geom_line() +
  scale_y_log10() +
  labs(
    x = "Time (h)", y = "Concentration", colour = NULL,
    caption = "Replicates Figure 4a of Shahraz 2024 (typical patient)."
  ) +
  theme(legend.position = "bottom")

bl_free <- at(sim_4a, run_in - 4, "lif_free_tumor_pgml")
last_dose <- 3600
inh_free <- function(t) 100 * (1 - at(sim_4a, t, "lif_free_tumor_pgml") / bl_free)
first_cycle <- sim_4a |> dplyr::filter(time > run_in, time < run_in + 336)
post <- sim_4a |> dplyr::filter(time > last_dose)
t90 <- post$time[which(post$lif_free_tumor_pgml >= 0.9 * bl_free)[1]] - last_dose

fig4a <- tibble::tribble(
  ~quantity, ~model, ~paper,
  "Inhibition of tumor free LIF 2 weeks after the last dose (%)", inh_free(last_dose + 336), 98,
  "Inhibition of tumor free LIF 4 weeks after the last dose (%)", inh_free(last_dose + 672), 97,
  "Time of the tumor free LIF nadir after the first dose (h)",
  first_cycle$time[which.min(first_cycle$lif_free_tumor_pgml)] - run_in, 50,
  "Time from the last dose to 90% of baseline tumor free LIF (h)", t90, 3360
)
fig4a |>
  dplyr::rename("Quantity" = quantity, "Model" = model, "Paper (Results text)" = paper) |>
  knitr::kable(digits = 1)
Quantity Model Paper (Results text)
Inhibition of tumor free LIF 2 weeks after the last dose (%) 99.6 98
Inhibition of tumor free LIF 4 weeks after the last dose (%) 99.3 97
Time of the tumor free LIF nadir after the first dose (h) 32.0 50
Time from the last dose to 90% of baseline tumor free LIF (h) 4704.0 3360

# The paper states 98% and 97% inhibition persisting 2 and 4 weeks after the
# last dose. A mis-scaled dose, KD or synthesis rate moves these by far more
# than a few percent.
stopifnot(inh_free(last_dose + 336) > 95, inh_free(last_dose + 672) > 95)

The model holds tumor free LIF more than 95% suppressed for 4 weeks after the last dose, as the paper reports. Recovery to baseline takes about 4704 h in the model against “~3360 h (140 days)” in the text; the reference curve in Figure 4b (same regimen, CL = 0.0135 L/h) plateaus near 8000 h, about 4400 h after the last dose, which is in line with the model.

Replicate Figure 4b-f: sensitivity of tumor free LIF

Each sensitive parameter is varied from -50% to +50% of its reference value, as in Figure 4b-f.

sens <- tibble::tribble(
  ~panel, ~param, ~ref, ~log_scale,
  "b: AZD0171 clearance (L/h)", "lcl", 0.0135, TRUE,
  "c: Tumor volume (L)", "v_tumor", 0.258, FALSE,
  "d: LIF synthesis rate (pmol/h)", "lksyn_lif", 1.67, TRUE,
  "e: gp130 steady state (nM)", "gp130_ss", 25, FALSE,
  "f: LIFR internalization (1/h)", "kint_lifr_slow", 0.42, FALSE
)
mult <- c(0.5, 0.75, 1, 1.25, 1.5)

run_sens <- function(param, value, log_scale) {
  theta <- mod_typical$theta
  theta[[param]] <- if (log_scale) log(value) else value
  sim <- rxode2::rxSolve(mod_typical, events = ev_4a, params = theta)
  as.data.frame(sim)[, c("time", "lif_free_tumor_pgml")]
}

sens_sim <- dplyr::bind_rows(lapply(seq_len(nrow(sens)), function(i) {
  dplyr::bind_rows(lapply(mult, function(m) {
    run_sens(sens$param[i], sens$ref[i] * m, sens$log_scale[i]) |>
      dplyr::mutate(panel = sens$panel[i], multiplier = m)
  }))
}))

sens_sim |>
  dplyr::filter(lif_free_tumor_pgml > 0) |>
  ggplot(aes(time, lif_free_tumor_pgml, colour = factor(multiplier))) +
  geom_line() +
  scale_y_log10() +
  facet_wrap(~panel, scales = "free_y") +
  labs(
    x = "Time (h)", y = "Tumor free LIF (pg/mL)", colour = "Multiplier",
    caption = "Replicates Figure 4b-f of Shahraz 2024."
  )

sens_summary <- sens_sim |>
  dplyr::group_by(panel, multiplier) |>
  dplyr::summarise(
    baseline = lif_free_tumor_pgml[time == run_in - 4],
    on_drug = lif_free_tumor_pgml[time == last_dose - 8],
    t50 = time[time > last_dose & lif_free_tumor_pgml >= 0.5 * baseline][1] - last_dose,
    .groups = "drop"
  ) |>
  dplyr::mutate(inhibition = 100 * (1 - on_drug / baseline))

sens_summary |>
  dplyr::rename(
    "Panel" = panel, "Multiplier" = multiplier,
    "Baseline (pg/mL)" = baseline, "Before last dose (pg/mL)" = on_drug,
    "Inhibition (%)" = inhibition, "Last dose to 50% recovery (h)" = t50
  ) |>
  knitr::kable(digits = c(0, 2, 1, 2, 0, 1))
Panel Multiplier Baseline (pg/mL) Before last dose (pg/mL) Last dose to 50% recovery (h) Inhibition (%)
b: AZD0171 clearance (L/h) 0.50 163.5 0.28 7656 99.8
b: AZD0171 clearance (L/h) 0.75 163.5 0.44 4952 99.7
b: AZD0171 clearance (L/h) 1.00 163.5 0.62 3632 99.6
b: AZD0171 clearance (L/h) 1.25 163.5 0.83 2864 99.5
b: AZD0171 clearance (L/h) 1.50 163.5 1.07 2360 99.3
c: Tumor volume (L) 0.50 326.6 1.25 3504 99.6
c: Tumor volume (L) 0.75 217.8 0.83 3584 99.6
c: Tumor volume (L) 1.00 163.5 0.62 3632 99.6
c: Tumor volume (L) 1.25 131.0 0.50 3664 99.6
c: Tumor volume (L) 1.50 109.3 0.41 3696 99.6
d: LIF synthesis rate (pmol/h) 0.50 81.7 0.31 3728 99.6
d: LIF synthesis rate (pmol/h) 0.75 122.6 0.46 3680 99.6
d: LIF synthesis rate (pmol/h) 1.00 163.5 0.62 3632 99.6
d: LIF synthesis rate (pmol/h) 1.25 204.6 0.78 3592 99.6
d: LIF synthesis rate (pmol/h) 1.50 245.6 0.93 3552 99.6
e: gp130 steady state (nM) 0.50 235.3 0.62 3784 99.7
e: gp130 steady state (nM) 0.75 192.9 0.62 3704 99.7
e: gp130 steady state (nM) 1.00 163.5 0.62 3632 99.6
e: gp130 steady state (nM) 1.25 142.0 0.62 3568 99.6
e: gp130 steady state (nM) 1.50 125.5 0.62 3512 99.5
f: LIFR internalization (1/h) 0.50 179.0 0.75 3528 99.6
f: LIFR internalization (1/h) 0.75 170.8 0.67 3600 99.6
f: LIFR internalization (1/h) 1.00 163.5 0.62 3632 99.6
f: LIFR internalization (1/h) 1.25 157.0 0.59 3648 99.6
f: LIFR internalization (1/h) 1.50 151.0 0.57 3656 99.6

cl_rows <- sens_summary |> dplyr::filter(grepl("^b:", panel))
other_rows <- sens_summary |> dplyr::filter(!grepl("^b:", panel))
# The paper's qualitative findings (Results), which hold exactly for this
# deterministic typical-value model:
stopifnot(
  # clearance does not change the baseline ...
  diff(range(cl_rows$baseline)) / mean(cl_rows$baseline) < 1e-3,
  # ... but lower clearance delays the rebound
  all(diff(cl_rows$t50) < 0),
  # tumor volume, LIF synthesis, gp130 and LIFR internalization move the
  # baseline but not the relative suppression at the clinical dose
  all(other_rows$inhibition > 98)
)
cl_delay_days <- (cl_rows$t50[cl_rows$multiplier == 0.5] -
  cl_rows$t50[cl_rows$multiplier == 1]) / 24

Halving clearance delays 50% recovery of tumor free LIF by 168 days in the model; the paper reports a 292-day increase in “rebound time” for the same change without defining the recovery level. The other four parameters change the baseline but leave the relative suppression above 98%, which is the paper’s conclusion for Figure 4c-f.

Replicate Figure 3 and Table 2: virtual-patient target engagement

The paper simulated 1350 virtual patients with the five fitted parameters drawn log-normally with 30% CV. Here each regimen gets 200 virtual patients and 12 weeks (about 3 months) of treatment. Inhibition is 100 * (1 - max(signal complex) / baseline signal complex), with the maximum taken over the last dosing interval of the 12 weeks (the paper gives the formula with “maximum post-baseline” and no window).

regimens <- tibble::tribble(
  ~regimen, ~amt, ~tau,
  "75 mg Q3W", 75, 504,
  "225 mg Q3W", 225, 504,
  "750 mg Q3W", 750, 504,
  "1125 mg Q3W", 1125, 504,
  "1500 mg Q3W", 1500, 504,
  "750 mg Q2W", 750, 336,
  "1125 mg Q2W", 1125, 336,
  "1500 mg Q2W", 1500, 336
)
n_per_arm <- 200
treat_end <- run_in + 12 * 168 # 12 weeks of treatment

events <- dplyr::bind_rows(lapply(seq_len(nrow(regimens)), function(i) {
  tau <- regimens$tau[i]
  grid <- sort(unique(c(
    run_in - 4,
    seq(run_in, treat_end - tau, by = 24),
    seq(treat_end - tau, treat_end, by = 4)
  )))
  make_events(
    ids = (i - 1) * n_per_arm + seq_len(n_per_arm),
    amt = regimens$amt[i], tau = tau,
    n_dose = (treat_end - run_in) / tau, grid = grid,
    label = regimens$regimen[i]
  ) |>
    dplyr::mutate(tau = tau)
}))
stopifnot(!anyDuplicated(events[, c("id", "time", "evid")]))

rxode2::rxSetSeed(20241015)
sim <- as.data.frame(rxode2::rxSolve(mod, events = events, keep = c("regimen", "tau")))
#> ℹ parameter labels from comments will be replaced by 'label()'
sim$regimen <- factor(sim$regimen, levels = regimens$regimen)
engagement <- sim |>
  dplyr::group_by(id, regimen) |>
  dplyr::summarise(
    baseline = signal_tumor[time == run_in - 4],
    peak = max(signal_tumor[time >= treat_end - tau[1]]),
    .groups = "drop"
  ) |>
  dplyr::mutate(inhibition = 100 * (1 - peak / baseline))

ggplot(engagement, aes(regimen, inhibition)) +
  geom_boxplot() +
  labs(
    x = NULL, y = "Inhibition of active signaling complex (%)",
    caption = "Replicates Figure 3 of Shahraz 2024 (200 virtual patients per regimen)."
  ) +
  theme(axis.text.x = element_text(angle = 30, hjust = 1))

published_t2 <- tibble::tribble(
  ~regimen, ~pub85, ~pub90, ~pub95,
  "75 mg Q3W", 38, 24, 10,
  "225 mg Q3W", 71, 61, 41,
  "750 mg Q3W", 89, 84, 68,
  "1125 mg Q3W", 93, 88, 76,
  "1500 mg Q3W", 95, 90, 80,
  "750 mg Q2W", 95, 91, 80,
  "1125 mg Q2W", 97, 94, 85,
  "1500 mg Q2W", 98, 95, 90
)
table2 <- engagement |>
  dplyr::group_by(regimen) |>
  dplyr::summarise(
    median_inh = median(inhibition),
    sim85 = 100 * mean(inhibition >= 85),
    sim90 = 100 * mean(inhibition >= 90),
    sim95 = 100 * mean(inhibition >= 95),
    .groups = "drop"
  ) |>
  dplyr::mutate(regimen = as.character(regimen)) |>
  dplyr::left_join(published_t2, by = "regimen")

table2 |>
  dplyr::select(regimen, median_inh, sim85, pub85, sim90, pub90, sim95, pub95) |>
  dplyr::rename(
    "Regimen" = regimen, "Median inhibition (%)" = median_inh,
    ">=85% model" = sim85, ">=85% Table 2" = pub85,
    ">=90% model" = sim90, ">=90% Table 2" = pub90,
    ">=95% model" = sim95, ">=95% Table 2" = pub95
  ) |>
  knitr::kable(digits = 1, caption = "Percent of patients reaching each level of signaling-complex inhibition after about 3 months.")
Percent of patients reaching each level of signaling-complex inhibition after about 3 months.
Regimen Median inhibition (%) >=85% model >=85% Table 2 >=90% model >=90% Table 2 >=95% model >=95% Table 2
75 mg Q3W 94.0 95.5 38 90 24 36.5 10
225 mg Q3W 98.2 100.0 71 100 61 98.5 41
750 mg Q3W 99.6 100.0 89 100 84 100.0 68
1125 mg Q3W 99.7 100.0 93 100 88 100.0 76
1500 mg Q3W 99.8 100.0 95 100 90 100.0 80
750 mg Q2W 99.8 100.0 95 100 91 100.0 80
1125 mg Q2W 99.8 100.0 97 100 94 100.0 85
1500 mg Q2W 99.9 100.0 98 100 95 100.0 90

# Gates on what the model and paper agree on. Robust to the cohort draw: the
# medians sit far from these bounds.
stopifnot(
  # inhibition rises with dose within the Q3W regimens
  all(diff(table2$median_inh[1:5]) > 0),
  # the selected Phase II dose reaches >=95% inhibition in >=90% of patients
  table2$sim95[table2$regimen == "1500 mg Q2W"] >= 90
)

The model reproduces the dose dependence of Figure 3 and the paper’s dose selection: 1500 mg Q2W inhibits the signaling complex by at least 95% in at least 90% of patients. It does not reproduce the low-dose end of Table 2. At 75 and 225 mg Q3W the model predicts far more inhibition, and a much narrower spread across patients, than Figure 3 and Table 2 show (Figure 3 has medians near 80% and 93%, with individual values down to about 10% and 25%). The same gap appears for every reading of the inhibition window the maintainers tried (trough of the last interval, maximum after the first 24 h, value at 12 weeks), so it is not a matter of the metric. Possible sources are recorded under Assumptions and deviations; no parameter was adjusted.

PKNCA: steady-state serum AZD0171

The paper reports no NCA values. For reference, the steady-state exposure over the last dosing interval of the 12-week cohort is summarised with PKNCA (the paper notes that steady state is reached after four to five cycles).

sim_nca <- sim |>
  dplyr::filter(!is.na(Cc), time >= treat_end - tau) |>
  dplyr::select(id, time, Cc, regimen)

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

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

intervals <- regimens |>
  dplyr::transmute(
    regimen,
    start = treat_end - tau,
    end = treat_end,
    cmax = TRUE,
    cmin = TRUE,
    auclast = TRUE
  )

nca_res <- PKNCA::pk.nca(PKNCA::PKNCAdata(conc_obj, dose_obj, intervals = intervals))
nca_summary <- as.data.frame(nca_res$result) |>
  dplyr::filter(PPTESTCD %in% c("cmax", "cmin", "auclast")) |>
  dplyr::group_by(regimen, PPTESTCD) |>
  dplyr::summarise(median = median(PPORRES), .groups = "drop") |>
  tidyr::pivot_wider(names_from = PPTESTCD, values_from = median) |>
  dplyr::mutate(regimen = factor(regimen, levels = regimens$regimen)) |>
  dplyr::arrange(regimen)

nca_summary |>
  dplyr::select(regimen, cmax, cmin, auclast) |>
  dplyr::rename(
    "Regimen" = regimen,
    "Cmax (ug/mL)" = cmax,
    "Cmin (ug/mL)" = cmin,
    "AUCtau (h*ug/mL)" = auclast
  ) |>
  knitr::kable(digits = 1, caption = "Median steady-state serum AZD0171 over the last dosing interval of 12 weeks.")
Median steady-state serum AZD0171 over the last dosing interval of 12 weeks.
Regimen Cmax (ug/mL) Cmin (ug/mL) AUCtau (h*ug/mL)
75 mg Q3W 24.1 6.2 5310.4
225 mg Q3W 71.5 18.1 15676.4
750 mg Q3W 241.4 64.8 54966.6
1125 mg Q3W 369.8 91.9 80887.3
1500 mg Q3W 493.7 122.5 107569.3
750 mg Q2W 289.6 115.8 55057.1
1125 mg Q2W 443.7 163.3 77643.1
1500 mg Q2W 563.5 222.3 107416.0

# Linear PK in the drug: exposure scales with dose within one schedule.
q3w <- nca_summary |> dplyr::filter(grepl("Q3W", regimen))
ratio <- (q3w$auclast / q3w$auclast[1]) / (c(75, 225, 750, 1125, 1500) / 75)
stopifnot(all(abs(ratio - 1) < 0.1))

The median steady-state peak on 1500 mg Q2W, 564 ug/mL, lies within the range of observed concentrations in Figure 2a (mostly below 1000 ug/mL). The ratio check confirms that exposure is proportional to dose: LIF is present at picomolar levels, so binding does not measurably change AZD0171 clearance.

Assumptions and deviations

  • Tumor and serum LIF ODEs reconstructed from the flux list. Supplementary Table S5 prints all 45 fluxes, the 9 tissue ODEs and the serum AZD0171 ODE, but not the ODEs for serum free LIF, serum AZD0171:LIF or the 9 tumor species. The maintainers assembled them from the printed fluxes. Each flux names its reactants and appears in exactly one place, and the tumor block mirrors the printed tissue block, so the assembly has only one possible form.
  • LIF synthesis in tissue omitted. Reaction_40 (ksynth_LIFt * alpha / Vtissue) uses a coefficient alpha that Table S4 does not define. The Methods state that LIF is synthesized only in the tumor, so the term is set to zero.
  • Synthesis-rate units. Table S4 gives ksynth_LIFt = 1.67 pmol/h, while the model runs in nmol and nM, so the model divides by 1000. Figure 4d labels the same rate 0.0017 nmol/h, which confirms the conversion.
  • Tumor concentrations and LIF synthesis. Tumor species are concentrations per total tumor volume (the Krogh flux is multiplied by Vtumor where it leaves serum), but LIF synthesis enters per void volume, ksynth_LIFt / (epsilon_tumor * Vtumor) (Reaction_39). Both are kept as printed.
  • Total LIF uses the complex’s full molecular weight. Rule 10 converts serum AZD0171:LIF to pg/mL with MW_DrugLIF (175 kDa) rather than the LIF mass (25 kDa). The fitted parameters were estimated with this rule, so it is kept as printed.
  • Serum. The paper says “plasma”, but the Phase I and Phase II assays measured AZD0171 and total LIF in serum.
  • Initial conditions. Free LIFR and gp130 start at LIFR_ss and gp130_ss, and all LIF species start at zero. A 240 h drug-free run-in, the same as the first-dose time in Figure 4, brings the system to its baseline before the first dose.
  • The serum AUC state (Table S5 ODE 11) is a bookkeeping integral and is not carried; use PKNCA on Cc instead.
  • Between-patient variability is the 30% CV log-normal spread the paper used for virtual patients, applied to all five fitted parameters (CL, Vc, ksynth_LIFt, LIFR_ss and Q). The Results sentence counts five parameters but lists only four in its parenthesis, omitting CL; CL is the fifth fitted parameter in Table 1. The variances are fixed, not estimated.
  • Residual error is not reported. The patients were fitted individually in SimBiology, so both proportional residual SDs are fixed at zero.
  • Infusion duration is not reported; the simulations use 1-hour infusions, which has no visible effect at 2- to 3-week dosing intervals.
  • Table 2 and Figure 3 are not reproduced at low doses (see above). The published virtual population is described only as “30% CV on five parameters”, and the extracted equations give much less spread than Figure 3 shows. Possible reasons are a wider sampling distribution than stated, a different inhibition window, or a model version different from the one in Table S5. The internally inconsistent baselines across Figure 4 panels support the last possibility. No parameter was tuned.
  • Preclinical CT26 mouse model not extracted. The simplified mouse model (Figure S2) re-estimated AZD0171 clearance and a serum-to-tumor partition coefficient, but the paper reports neither value, nor the mouse tumor volume or LIF synthesis rate, so it cannot be built from the published material.
  • No correction notice for this article was found in Europe PMC as of 2026-10-30.