Skip to contents

Model and source

Wan et al. (2025) translated high-throughput toxicity data from human induced pluripotent stem cell-derived cardiomyocytes (hiPSC-CMs) into a predicted clinical incidence of left ventricular ejection fraction (LVEF) decline for 21 tyrosine kinase inhibitors (TKIs). Each TKI is a QSP-PK-TD model with three layers:

  1. PK. A published population PK model of the TKI, condensed to typical values and CVs in Table S2. The plasma concentration is converted to an effective myocardial concentration Cin = Cp x fu x Kp x H (Eqs. 17, 18, 21): fraction unbound (Table S5), heart-to-plasma partition coefficient (Table S7; observed where reported, otherwise the Poulin-Theil prediction) and a heterogeneous-distribution factor H = IC50,2D / IC50,3D (Table S8; 1 where no 3D value exists).
  2. TD. An in vitro toxicodynamic model fitted to 72-h hiPSC-CM viability, ATP and contractility data (Table 1; structure from Sang 2024): Cin above a threshold Cth injures cells, which die through three transit stages (Eqs. 1-8); Cin inhibits mitochondrial membrane potential (MMP) production, which drives ATP (Eqs. 9-11); and Cin raises the ATP needed for half-maximal average contractile force (AAC, Eqs. 12-14).
  3. QSP. The surviving fraction times AAC replaces the drug effect on stroke-volume production in the Sang 2021 human cardiovascular QSP model (Sang_2021_doxorubicin_human_qsp), with baselines and the LVESV feedback FB_LVESV for cardiovascular-healthy, hypertensive and dilated-cardiomyopathy patients (Table 2).

The package carries 19 of the 21 TKIs, one model file each. Ibrutinib and nilotinib are not included: Table S2 gives only a lag time for their absorption, and the zero-order absorption duration of the primary PK models (Al-Ghazawi 2021, Giles 2013) is needed to reproduce their exposure.

drugs <- c(
  "afatinib", "axitinib", "bosutinib", "cabozantinib", "crizotinib",
  "dabrafenib", "dasatinib", "erlotinib", "gefitinib", "imatinib",
  "lapatinib", "pazopanib", "ponatinib", "regorafenib", "sorafenib",
  "sunitinib", "trametinib", "vandetanib", "vemurafenib"
)
# Wan 2025 Figures 1-3, panel B = cardiotoxic, panel A = non-cardiotoxic
# (nilotinib, also cardiotoxic, is not packaged)
cardiotoxic <- c(
  "afatinib", "dasatinib", "imatinib", "lapatinib", "pazopanib",
  "ponatinib", "regorafenib", "sorafenib", "sunitinib", "trametinib",
  "vandetanib"
)
model_name <- function(drug) paste0("Wan_2025_", drug, "_qsp")
mods <- lapply(setNames(drugs, drugs), function(d) rxode2::rxode2(readModelDb(model_name(d))))
mods_typ <- lapply(mods, rxode2::zeroRe)

Dosing regimens are from Table S4. Ponatinib has two parallel absorption routes and takes the same dose record in both depot and depot2 (the model splits it by fdepot).

regimen <- tibble::tribble(
  ~drug, ~dose, ~ii,
  "afatinib", 40, 24, "axitinib", 5, 12, "bosutinib", 500, 24,
  "cabozantinib", 140, 24, "crizotinib", 250, 12, "dabrafenib", 150, 12,
  "dasatinib", 100, 24, "erlotinib", 150, 24, "gefitinib", 250, 24,
  "imatinib", 400, 24, "lapatinib", 1250, 24,
  "pazopanib", 800, 24, "ponatinib", 45, 24,
  "regorafenib", 160, 24, "sorafenib", 400, 12, "sunitinib", 50, 24,
  "trametinib", 2, 24, "vandetanib", 300, 24, "vemurafenib", 960, 24
)
dose_cmts <- function(drug) {
  switch(drug,
    ponatinib = c("depot", "depot2"),
    "depot"
  )
}
# Dosing records for `days` days of the Table S4 regimen
dose_events <- function(drug, days) {
  r <- regimen[regimen$drug == drug, ]
  tt <- seq(0, days * 24 - r$ii, by = r$ii)
  dplyr::bind_rows(lapply(dose_cmts(drug), function(cm) {
    data.frame(time = tt, amt = r$dose, evid = 1L, cmt = cm, dvid = NA_integer_)
  }))
}
obs_events <- function(times) {
  data.frame(time = times, amt = 0, evid = 0L, cmt = NA_character_, dvid = 1L)
}
# Cardiovascular-healthy typical patient (Table 2)
healthy_typ <- data.frame(
  LVEDV_BL = 113, LVESV_BL = 113 * (1 - 0.6), HR_BL = 70, MAP_BL = 94.92,
  DIS_HYPERT = 0L, DIS_DILATED_CARDIOMYOPATHY = 0L
)

Observation rows carry dvid = 1 and no compartment; every output (Cc, LVEF, LVEDV, LVESV, MAP) is computed on every row.

Population

No patient data were fitted. Wan et al. generated 5000 virtual patients for each of three cardiovascular populations by Monte Carlo sampling around the Table 2 typical values and CVs, and weighted the population incidences by prevalence (64.97% cardiovascular healthy, 35% hypertension, 0.03% dilated cardiomyopathy; Eq. 16). Hypertension was defined by a baseline mean arterial pressure of at least 106 mmHg and dilated cardiomyopathy by a baseline LVEDV of at least 146 mL. The TD parameters were estimated (Monolix 2021R2) on viability, ATP and contractile-force data digitised from the 21-TKI hiPSC-CM screen of Sharma et al. (2017) after 72 h of exposure.

readModelDb("Wan_2025_sunitinib_qsp")()$population
#> $species
#> [1] "human"
#> 
#> $n_subjects
#> [1] 15000
#> 
#> $disease_state
#> [1] "Virtual adult cancer patients treated with sunitinib (50 mg, QD), in three cardiovascular populations: cardiovascular healthy, hypertension (MAP0 >= 106 mmHg) and dilated cardiomyopathy (LVEDV0 >= 146 mL)"
#> 
#> $dose_range
#> [1] "50 mg, QD for one year (Wan 2025 Table S4)"
#> 
#> $notes
#> [1] "Virtual population only; no human data were fitted. Wan 2025 generated 5000 virtual patients per cardiovascular population by Monte Carlo sampling (Table 2 typical values and CVs) and weighted the incidence 64.97% healthy, 35% hypertension, 0.03% dilated cardiomyopathy (Eq. 16). The PK model is a published population model condensed to typical values and CVs in Table S2; the toxicodynamic parameters were fitted to digitised high-throughput hiPSC-CM viability, ATP and contractility data (Sharma 2017) after 72 h exposure."

Source trace

Every ini() value carries an in-file comment naming its source. The shared structure and constants:

Equation / parameter Value Source
PK structure, typical values, CVs per TKI Table S2
Regimen per TKI Table S4
fu per TKI Table S5 (printed in %)
mw per TKI Table S6
lkp_heart per TKI Table S7, observed Kp where reported, otherwise Kp_prediction
h_het = IC50,2D / IC50,3D per TKI Table S8 and Eq. 21; 1 without an IC50,3D
cin = Cc / MW x 1000 x fu x Kp x H – Eqs. 17, 18, 21
kinj, EC50inj, hill_inj, Cth, EC50MMP, EEC per TKI Table 1 (kinj x 1e-2, 1/h)
emax_inj 1 Section 4.2.1
Injury and death chain – Eqs. 1-8
tau_inj, kin_mmp, kin_atp, n_atp 4.9 h, 0.96 1/h, 0.032 1/h, 0.8 Sang 2024 Table I (DOX-TRZ fit), adopted per Section 4.2.4
kout_mmp = kin_mmp, kout_atp = kin_atp – Baselines MMP0 = ATP0 = 1 (Sang 2024)
MMP and ATP turnover – Eqs. 9-11
eaac, ATP50 = ATP50_ref / (1 - EAAC), AAC – Eqs. 12-14 (Eq. 13 corrected, see below)
atp50_ref 1.35 Back-solved from Figure 3 by the maintainers (see below)
etki = survival x AAC replacing E_EP_SV – Eq. 15
kout_sv, kout_edv, kout_hr, kout_tpr, FB_MAP_0, MAP exponent 0.0308, 0.0308, 2.83, 0.875 1/h, 0.0029, -1.98 Sang 2021 Table 1 and Eq. 8, per Section 4.3
FB_LVESV healthy / hypertension / DCM 0.0042 / 0.0045 / 0.0028 1/mL Table 2
etalfb_lvesv log(1 + 0.124^2) Table 2 CV_FB_LVESV
Baselines LVEDV0, Index_SV0, MAP0, HR0 covariates Table 2

The per-TKI values as stored in the model files:

theta_of <- function(m, nm) if (nm %in% names(m$theta)) m$theta[[nm]] else NA_real_
per_drug <- dplyr::bind_rows(lapply(drugs, function(d) {
  m <- mods_typ[[d]]
  data.frame(
    drug = d,
    CL = exp(theta_of(m, "lcl")), Vc = exp(theta_of(m, "lvc")),
    Q = exp(theta_of(m, "lq")), Vp = exp(theta_of(m, "lvp")),
    ka = exp(theta_of(m, "lka")), Tlag = exp(theta_of(m, "ltlag")),
    F = exp(theta_of(m, "lfdepot")), fu = theta_of(m, "fu"),
    Kp = exp(theta_of(m, "lkp_heart")), H = theta_of(m, "h_het"),
    kinj = exp(theta_of(m, "lkinj")), EC50inj = exp(theta_of(m, "lec50_inj")),
    Cth = exp(theta_of(m, "lcth")), EC50MMP = exp(theta_of(m, "lec50_mmp")),
    EEC = exp(theta_of(m, "leec"))
  )
}))
per_drug |>
  dplyr::mutate(dplyr::across(where(is.numeric), ~ signif(.x, 3))) |>
  dplyr::rename(
    "TKI" = drug, "CL (L/h)" = CL, "Vc (L)" = Vc, "Q (L/h)" = Q, "Vp (L)" = Vp,
    "ka (1/h)" = ka, "Tlag (h)" = Tlag, "kinj (1/h)" = kinj,
    "EC50inj (uM)" = EC50inj, "Cth (uM)" = Cth, "EC50MMP (uM)" = EC50MMP,
    "EEC (uM)" = EEC
  ) |>
  knitr::kable(caption = "Per-TKI parameters (NA = not part of that TKI's model).")
Per-TKI parameters (NA = not part of that TKI’s model).
TKI CL (L/h) Vc (L) Q (L/h) Vp (L) ka (1/h) Tlag (h) F fu Kp H kinj (1/h) EC50inj (uM) Cth (uM) EC50MMP (uM) EEC (uM)
afatinib 42.300 456.0 77.50 1130.0 0.25 NA 1.000 0.05000 1.770 1.000 0.1500 17.60 4.16 4.62 0.41
axitinib 17.000 45.3 1.74 45.9 0.53 0.46 0.465 0.01000 1.760 1.000 0.0032 4.00 4.70 50.10 NA
bosutinib 120.000 4920.0 56.40 388000.0 0.61 0.87 1.000 0.06000 2.060 1.000 0.1200 14.60 5.16 9.53 NA
cabozantinib 2.230 81.5 14.20 213.0 0.57 0.46 0.675 0.00300 0.500 1.000 0.0270 64.70 3.70 10.80 NA
crizotinib 102.000 2390.0 44.40 2200.0 0.82 0.65 1.000 0.09000 0.550 1.000 0.0440 12.00 1.70 10.80 44.20
dabrafenib 17.000 70.3 3.30 154.0 1.88 0.48 0.555 0.00300 1.980 1.000 0.0190 55.20 28.60 45.70 NA
dasatinib 126.000 166.0 11.10 1300.0 0.26 NA 1.000 0.04000 3.160 1.780 0.0310 20.00 9.60 13.30 2.76
erlotinib 3.800 166.0 NA NA 1.48 NA 1.000 0.07000 0.230 1.000 0.0190 93.50 5.96 5.54 NA
gefitinib 21.000 934.0 46.90 1500.0 0.39 NA 1.000 0.00100 0.950 1.000 0.0740 44.60 5.16 10.60 10.50
imatinib 7.290 202.0 NA NA 1.03 NA 1.000 0.05000 3.620 0.422 0.1800 72.70 19.80 16.40 33.00
lapatinib 19.400 124.0 NA NA 0.12 NA 1.000 0.01000 0.120 1.000 0.0130 27.40 24.20 69.60 1.80
pazopanib 0.458 22.3 NA NA 0.98 NA 1.000 0.01000 0.306 1.000 0.0150 72.40 13.00 21.50 NA
ponatinib 34.300 839.0 17.20 347.0 4.41 3.93 0.534 0.01000 5.360 1.000 0.1100 11.60 0.46 2.54 2.57
regorafenib 4.020 10.7 11.00 162.0 0.48 0.36 1.000 0.00500 0.560 0.731 0.0440 5.38 2.70 3.20 16.30
sorafenib 8.130 213.0 NA NA 2.53 NA 1.000 0.00500 0.770 0.650 0.0450 4.36 1.00 3.31 0.79
sunitinib 24.100 1070.0 0.28 63.8 0.38 0.64 1.000 0.05000 2.200 0.234 0.1000 22.00 2.20 5.01 1.42
trametinib 5.830 61.9 64.90 417.0 0.91 0.71 1.000 0.02600 3.260 1.000 0.0130 24.80 15.40 12.60 0.94
vandetanib 7.260 1280.0 10.70 616.0 0.83 NA 1.000 0.06330 3.600 1.530 0.0710 79.90 1.38 10.70 0.94
vemurafenib 1.300 106.0 NA NA 0.19 NA 0.850 0.00001 0.780 1.000 0.0260 44.70 2.20 15.50 31.30

Untreated steady state

With no drug every state must stay at its baseline: the TD states at 1 and the QSP at the covariate baselines. This checks the kin/kout balance in all 19 files.

ss_check <- dplyr::bind_rows(lapply(drugs, function(d) {
  ev <- cbind(id = 1L, obs_events(c(0, 24 * 365)), healthy_typ)
  s <- rxode2::rxSolve(mods_typ[[d]], ev, returnType = "data.frame")
  data.frame(drug = d, LVEF_0 = s$LVEF[1], LVEF_1y = s$LVEF[2], MAP_1y = s$MAP[2])
}))
stopifnot(
  all(abs(ss_check$LVEF_1y - 60) < 1e-6),
  all(abs(ss_check$MAP_1y - 94.92) < 1e-6)
)

In vitro toxicodynamics (Figure 3) and ATP50_ref

ATP50_ref, the relative ATP level giving half-maximal contractile force before treatment, is required by Eq. 14 but is reported neither in Wan 2025 nor in the cited Sang 2024 study. The maintainers back-solved it from the model curves of Figure 3, holding every Table 1 value fixed and using the contractile force as the observed AAC (as in Sang 2024 Eq. 15). The curves of all 21 panels were digitised pixel by pixel and fitted with one shared value: 1.35 (drug-bootstrap 95% interval 0.75-2.31; root-mean-square error 0.10). Individual-TKI optima mostly fall between 1 and 2.4; erlotinib (flat curve despite falling ATP) and axitinib and crizotinib (curves rising above 1) are not reproducible by Eqs. 9-14 at any value.

The packaged models reproduce the 72-h in vitro experiment when the PK layer is switched off: a bolus into central with clearance and distribution set to zero holds the concentration constant, fu = Kp = H = 1 makes cin the medium concentration, and the stroke-volume and LVEDV turnover is frozen.

concs <- 10^seq(-1, 2, by = 0.25)
invitro <- function(d) {
  m <- mods_typ[[d]]
  th <- m$theta
  vc <- exp(th[["lvc"]])
  # Constant medium concentration: no clearance or distribution, fu = Kp = H = 1.
  # The cardiovascular turnover is switched off as well (SV and LVEDV frozen),
  # since the in vivo QSP plays no part in the 72-h cell experiment.
  p <- c(lcl = -40, lfdepot = 0, fu = 1, lkp_heart = 0, h_het = 1, lkout_sv = -40, lkout_edv = -40)
  if ("lq" %in% names(th)) p <- c(p, lq = -40)
  if ("ltlag" %in% names(th)) p <- c(p, ltlag = log(1e-6))
  ev <- dplyr::bind_rows(lapply(seq_along(concs), function(i) {
    dplyr::bind_rows(
      data.frame(id = i, time = 0, amt = concs[i] * th[["mw"]] / 1000 * vc, evid = 1L, cmt = "central", dvid = NA_integer_),
      data.frame(id = i, time = 72, amt = 0, evid = 0L, cmt = NA_character_, dvid = 1L)
    )
  }))
  ev <- cbind(ev, healthy_typ)
  s <- suppressWarnings(rxode2::rxSolve(m, ev, params = p, returnType = "data.frame", maxsteps = 5e5))
  # `eaac` is the constant 0 in the six models without an EEC, and rxode2 does
  # not return constant model variables
  eaac <- if ("eaac" %in% names(s)) s$eaac else 0
  data.frame(drug = d, conc = concs, cin = s$cin, survival = s$survival, atp = s$atp, eaac = eaac, aac = s$aac)
}
iv <- dplyr::bind_rows(lapply(drugs, invitro))
stopifnot(all(abs(iv$cin / iv$conc - 1) < 1e-3))

Model curves of Figure 3 read at the tested concentrations for the 19 packaged TKIs (digitised by the maintainers; NA where a data point hides the curve):

fig3 <- tibble::tribble(
  ~drug, ~c0.3, ~c1, ~c3, ~c10, ~c30, ~c100,
  "afatinib", 0.62, 0.41, 0.21, 0.01, 0.00, 0.03,
  "axitinib", 0.97, 0.96, 0.94, NA, 0.67, NA,
  "bosutinib", 1.03, 1.08, 1.08, 0.78, 0.40, NA,
  "cabozantinib", 1.00, 0.97, 0.91, 0.74, NA, NA,
  "crizotinib", 1.06, 1.05, 0.82, 0.33, 0.08, NA,
  "dabrafenib", 0.99, 0.98, 0.97, NA, 0.78, 0.62,
  "dasatinib", 0.93, 0.77, 0.54, 0.26, 0.10, NA,
  "erlotinib", 0.96, 0.97, 0.97, NA, 0.99, NA,
  "gefitinib", 0.97, 0.91, 0.74, 0.43, 0.22, NA,
  "imatinib", 0.96, NA, 0.82, 0.72, 0.47, NA,
  "lapatinib", 0.90, 0.74, 0.50, 0.23, 0.10, NA,
  "pazopanib", 0.99, 0.97, 0.94, 0.83, NA, 0.49,
  "ponatinib", 0.85, 0.63, 0.36, 0.13, NA, NA,
  "regorafenib", 0.94, 0.81, 0.62, 0.36, 0.17, NA,
  "sorafenib", 0.75, 0.44, 0.19, NA, -0.03, NA,
  "sunitinib", 0.81, 0.62, 0.32, 0.10, 0.00, -0.01,
  "trametinib", 0.79, NA, 0.28, 0.09, -0.01, NA,
  "vandetanib", 0.79, 0.53, 0.32, 0.10, NA, NA,
  "vemurafenib", 1.00, 0.95, 0.87, 0.63, 0.39, NA
) |>
  tidyr::pivot_longer(-drug, names_to = "conc", values_to = "aac_fig3") |>
  dplyr::mutate(conc = as.numeric(sub("c", "", conc))) |>
  dplyr::filter(!is.na(aac_fig3))

Refit ATP50_ref to this coarser table with the packaged models’ own 72-h ATP and EAAC. AAC depends on ATP50_ref only through Eq. 14, so the refit needs no further simulation:

iv_pts <- dplyr::bind_rows(lapply(drugs, function(d) {
  x <- fig3[fig3$drug == d, ]
  iv_d <- iv[iv$drug == d, ]
  data.frame(
    drug = d, conc = x$conc, aac_fig3 = x$aac_fig3,
    atp = approx(log(iv_d$conc), iv_d$atp, log(x$conc))$y,
    eaac = approx(log(iv_d$conc), iv_d$eaac, log(x$conc))$y
  )
}))
aac_eq14 <- function(a, atp, eaac) atp * (1 + a) * (1 - eaac) / (atp * (1 - eaac) + a)
sse <- function(la) sum((iv_pts$aac_fig3 - aac_eq14(exp(la), iv_pts$atp, iv_pts$eaac))^2)
refit <- exp(optimize(sse, log(c(1e-3, 100)))$minimum)
shipped <- exp(mods_typ$afatinib$theta[["latp50_ref"]])
c(refit_table = refit, shipped = shipped)
#> refit_table     shipped 
#>    1.506024    1.350000
# The coarse table must land inside the interval of the pixel-level fit.
stopifnot(refit > 0.75, refit < 2.31)
ggplot(iv, aes(conc, aac)) +
  geom_line() +
  geom_point(data = fig3, aes(conc, aac_fig3), colour = "firebrick") +
  facet_wrap(~drug, ncol = 5) +
  scale_x_log10() +
  labs(
    x = "Medium concentration (uM)", y = "Average contractile force (AAC)",
    caption = "Lines: packaged models, 72 h, ATP50_ref = 1.35. Points: Wan 2025 Figure 3 model curves (digitised).\nReplicates Figure 3 of Wan 2025."
  )

Effective concentration at steady state (Table S1)

Table S1 lists the effective myocardial concentration of the 12 cardiotoxic TKIs, 11 of which are packaged. It is the steady-state peak of Cc x fu x Kp x H; PKNCA computes the peak of cin over the fourth-week dosing interval of each typical patient.

ss_sim <- dplyr::bind_rows(lapply(drugs, function(d) {
  r <- regimen[regimen$drug == d, ]
  t_start <- 27 * 24
  ev <- dplyr::bind_rows(dose_events(d, 28), obs_events(seq(t_start, t_start + r$ii, by = 0.1))) |>
    dplyr::arrange(time, dplyr::desc(evid))
  ev <- cbind(id = 1L, ev, healthy_typ)
  s <- rxode2::rxSolve(mods_typ[[d]], ev, returnType = "data.frame", maxsteps = 5e5)
  # A single-subject solve returns no id column; models without an EEC have no
  # aeaac state
  s$id <- 1L
  s$drug <- d
  s$aeaac <- if ("aeaac" %in% names(s)) s$aeaac else 0
  s
}))

nca_in <- ss_sim |>
  dplyr::filter(!is.na(cin)) |>
  dplyr::group_by(drug) |>
  dplyr::mutate(tad = time - min(time)) |>
  dplyr::ungroup() |>
  dplyr::select(id, drug, tad, cin)
conc_obj <- PKNCA::PKNCAconc(nca_in, cin ~ tad | drug + id)
ivl <- regimen |>
  dplyr::transmute(drug, start = 0, end = ii, cmax = TRUE, cav = TRUE)
nca_res <- as.data.frame(PKNCA::pk.nca(PKNCA::PKNCAdata(conc_obj, intervals = ivl)))

table_s1 <- tibble::tribble(
  ~drug, ~ceff_paper,
  "afatinib", 0.01, "dasatinib", 0.05, "imatinib", 0.47, "lapatinib", 0.008,
  "pazopanib", 0.61, "ponatinib", 0.0076,
  "regorafenib", 0.01, "sorafenib", 0.077, "sunitinib", 0.047,
  "trametinib", 0.0003, "vandetanib", 1.32
)
s1_cmp <- nca_res |>
  dplyr::filter(PPTESTCD == "cmax") |>
  dplyr::select(drug, cmax_sim = PPORRES) |>
  dplyr::inner_join(table_s1, by = "drug") |>
  dplyr::mutate(ratio = cmax_sim / ceff_paper)
s1_cmp |>
  dplyr::mutate(dplyr::across(where(is.numeric), ~ signif(.x, 3))) |>
  dplyr::rename(
    "TKI" = drug, "Simulated steady-state Cmax of Cin (uM)" = cmax_sim,
    "Table S1 Ceff (uM)" = ceff_paper, "Simulated / Table S1" = ratio
  ) |>
  knitr::kable()
TKI Simulated steady-state Cmax of Cin (uM) Table S1 Ceff (uM) Simulated / Table S1
afatinib 0.01020 0.0100 1.020
dasatinib 0.05280 0.0500 1.060
imatinib 0.47700 0.4700 1.020
lapatinib 0.00792 0.0080 0.990
pazopanib 0.60500 0.6100 0.992
ponatinib 0.00796 0.0076 1.050
regorafenib 0.01060 0.0100 1.060
sorafenib 0.02620 0.0770 0.340
sunitinib 0.00639 0.0470 0.136
trametinib 0.00311 0.0003 10.400
vandetanib 1.21000 1.3200 0.919
# Deterministic typical-value solves: eight of the eleven TKIs reproduce Table S1.
matched <- c("afatinib", "dasatinib", "imatinib", "lapatinib", "pazopanib", "ponatinib", "regorafenib", "vandetanib")
stopifnot(all(abs(s1_cmp$ratio[s1_cmp$drug %in% matched] - 1) < 0.15))

Eight TKIs reproduce Table S1 within 10%, which confirms the exposure conventions: the observed Kp where one is reported, H = IC50,2D / IC50,3D, total ponatinib bioavailability 1 split over its two absorption routes, and the steady-state peak as the summary. Imatinib matches only with its two Table S8 columns transposed (see below). Three do not match: sorafenib (0.34x) and sunitinib (0.14x), for which no combination of the printed fu, Kp and H reproduces the table (sunitinib matches the predicted Kp without H), and trametinib (10x, consistent with a fraction unbound of 0.26% rather than the printed 2.6%). The model files keep the printed values.

AEMMP and AEAAC (Figure 6)

The area under the EMMP and EAAC effect curves over one steady-state day (the aemmp and aeaac accumulators differenced over the fourth-week interval; twice-daily drugs cover their 12-h interval, as Wan 2025 integrates “over a dosing cycle”):

fig6 <- ss_sim |>
  dplyr::group_by(drug) |>
  dplyr::summarise(
    AEMMP = dplyr::last(aemmp) - dplyr::first(aemmp),
    AEAAC = dplyr::last(aeaac) - dplyr::first(aeaac),
    .groups = "drop"
  ) |>
  dplyr::mutate(cardiotoxicity = ifelse(drug %in% cardiotoxic, "Positive", "Negative"))
floor_val <- 1e-4
ggplot(fig6, aes(pmax(AEMMP, floor_val), pmax(AEAAC, floor_val), colour = cardiotoxicity)) +
  geom_point(size = 3) +
  geom_text(aes(label = drug), size = 2.5, vjust = -1, show.legend = FALSE) +
  scale_x_log10() +
  scale_y_log10() +
  scale_colour_manual(values = c(Negative = "goldenrod", Positive = "palevioletred")) +
  labs(
    x = "AEMMP (h)", y = "AEAAC (h)", colour = "Cardiotoxicity",
    caption = "Typical patients; values below 1e-4 (EAAC fixed to 0) plotted at 1e-4.\nReplicates Figure 6H of Wan 2025 (Cp x fu x Kp x H)."
  )

# Deterministic: cardiotoxic TKIs sit above non-cardiotoxic ones on AEAAC.
stopifnot(
  median(fig6$AEAAC[fig6$cardiotoxicity == "Positive"]) >
    10 * max(median(fig6$AEAAC[fig6$cardiotoxicity == "Negative"]), 1e-6)
)

Cardiovascular populations

The three Table 2 populations for a typical sunitinib patient. The dilated-cardiomyopathy patient has the largest baseline LVEDV, a lower baseline LVEF and the strongest response.

pops <- data.frame(
  id = 1:3,
  population = c("Healthy", "Hypertension", "Dilated cardiomyopathy"),
  LVEDV_BL = c(113, 100.95, 146),
  LVESV_BL = c(113 * 0.4, 100.95 * 0.4, 146 * 0.45),
  HR_BL = c(70, 70, 57),
  MAP_BL = c(94.92, 106, 91.54),
  DIS_HYPERT = c(0L, 1L, 0L),
  DIS_DILATED_CARDIOMYOPATHY = c(0L, 0L, 1L)
)
# Table 2 SV0 and TPR0 follow from the baselines
pops |>
  dplyr::mutate(SV0 = LVEDV_BL - LVESV_BL, TPR0 = MAP_BL / (HR_BL * SV0)) |>
  dplyr::select(population, SV0, TPR0) |>
  dplyr::mutate(dplyr::across(where(is.numeric), ~ signif(.x, 3))) |>
  knitr::kable()
population SV0 TPR0
Healthy 67.8 0.020
Hypertension 60.6 0.025
Dilated cardiomyopathy 80.3 0.020
ev_pop <- dplyr::bind_rows(dose_events("sunitinib", 365), obs_events(seq(0, 365 * 24, by = 24 * 7))) |>
  dplyr::arrange(time, dplyr::desc(evid))
ev_pop <- merge(pops, ev_pop) |> dplyr::arrange(id, time, dplyr::desc(evid))
pop_sim <- rxode2::rxSolve(mods_typ$sunitinib, ev_pop, keep = "population", returnType = "data.frame", maxsteps = 5e5)
#> Warning: multi-subject simulation without without 'omega'
ggplot(pop_sim, aes(time / 24 / 7, LVEF, colour = population)) +
  geom_line() +
  labs(x = "Week", y = "LVEF (%)", colour = NULL, caption = "Typical sunitinib patients, 50 mg once daily.")

Virtual trial: incidence of cardiac dysfunction (Figure 4H)

The virtual trial follows Section 4.4.1: one year of the Table S4 regimen, and cardiac dysfunction defined as an LVEF fall of more than 10 points to below 50%, or a fall of 20 points or more. Each cardiotoxic TKI is simulated in 50 cardiovascular-healthy and 50 hypertensive virtual patients (the paper used 5000 per population; a year of daily dosing makes larger cohorts too slow for this article); the 0.03% dilated-cardiomyopathy stratum is omitted and the two strata are weighted 64.97 : 35. Baselines are log-normal around the Table 2 typical values with the Table 2 CVs; hypertensive MAP is sampled until it is at least 106 mmHg. With 50 patients per stratum one case moves a stratum’s incidence by 2 percentage points, so the comparison is qualitative.

set.seed(2025)
rxode2::rxSetSeed(2025)
n_per <- 50
rlnorm_cv <- function(n, m, cv) m * exp(rnorm(n, -0.5 * log(1 + cv^2), sqrt(log(1 + cv^2))))
make_population <- function(n, population) {
  if (population == "Healthy") {
    map <- rlnorm_cv(n, 94.92, 0.1)
    lvedv <- rlnorm_cv(n, 113, 0.1)
  } else {
    map <- rlnorm_cv(10 * n, 106, 0.2)
    map <- head(map[map >= 106], n)
    lvedv <- rlnorm_cv(n, 100.95, 0.2)
  }
  index_sv <- rlnorm_cv(n, 0.6, 0.05)
  data.frame(
    population = population, LVEDV_BL = lvedv, LVESV_BL = lvedv * (1 - index_sv),
    HR_BL = 70, MAP_BL = map, DIS_HYPERT = as.integer(population == "Hypertension"),
    DIS_DILATED_CARDIOMYOPATHY = 0L
  )
}
cohort <- dplyr::bind_rows(make_population(n_per, "Healthy"), make_population(n_per, "Hypertension")) |>
  dplyr::mutate(id = dplyr::row_number())
stopifnot(nrow(cohort) == 2 * n_per)

incidence <- dplyr::bind_rows(lapply(cardiotoxic, function(d) {
  ev <- dplyr::bind_rows(dose_events(d, 365), obs_events(c(0, 365 * 24))) |>
    dplyr::arrange(time, dplyr::desc(evid))
  ev <- merge(cohort, ev) |> dplyr::arrange(id, time, dplyr::desc(evid))
  s <- rxode2::rxSolve(mods[[d]], ev, keep = "population", returnType = "data.frame", maxsteps = 5e5)
  # A patient whose solve fails (NA LVEF) is counted as unsolved, not as a
  # non-event
  s |>
    dplyr::group_by(id, population) |>
    dplyr::summarise(lvef0 = dplyr::first(LVEF), lvef1 = dplyr::last(LVEF), .groups = "drop") |>
    dplyr::mutate(drop = lvef0 - lvef1, event = (drop > 10 & lvef1 < 50) | drop >= 20) |>
    dplyr::group_by(population) |>
    dplyr::summarise(incidence = 100 * mean(event, na.rm = TRUE), n_solved = sum(!is.na(event)), .groups = "drop") |>
    dplyr::mutate(drug = d)
}))
#> Warning: some ID(s) could not solve the ODEs correctly; These values are
#> replaced with 'NA'
fig4h <- tibble::tribble(
  ~drug, ~paper_pred, ~clinical,
  "vandetanib", NA, "0.9",
  "lapatinib", 1.74, "0.2-1.76",
  "imatinib", 4.36, "0.04-2.7",
  "dasatinib", 2.41, "2-4",
  "afatinib", 3.82, "0-5.3",
  "regorafenib", 1.61, "0.3-5.6",
  "ponatinib", 1.88, "7",
  "sorafenib", 8.54, "3.3-18.6",
  "trametinib", 1.70, "3-11",
  "pazopanib", 4.43, "7-13",
  "sunitinib", 1.79, "2.7-19"
)
stopifnot(sum(incidence$n_solved) >= 0.95 * nrow(cohort) * length(cardiotoxic))
inc_tab <- incidence |>
  dplyr::select(-n_solved) |>
  tidyr::pivot_wider(names_from = population, values_from = incidence) |>
  dplyr::mutate(weighted = (64.97 * Healthy + 35 * Hypertension) / 99.97) |>
  dplyr::left_join(fig4h, by = "drug") |>
  dplyr::arrange(match(drug, fig4h$drug))
inc_tab |>
  dplyr::mutate(dplyr::across(c(Healthy, Hypertension, weighted), ~ round(.x, 1))) |>
  dplyr::select(drug, Healthy, Hypertension, weighted, paper_pred, clinical) |>
  dplyr::rename(
    "TKI" = drug, "Healthy (%)" = Healthy, "Hypertension (%)" = Hypertension,
    "Weighted (%)" = weighted, "Wan 2025 Fig. 4H predicted (%)" = paper_pred,
    "Clinical, Table S3 (%)" = clinical
  ) |>
  knitr::kable(caption = "Fig. 4H predicted vandetanib incidence is printed as '> 20'.")
Fig. 4H predicted vandetanib incidence is printed as ‘> 20’.
TKI Healthy (%) Hypertension (%) Weighted (%) Wan 2025 Fig. 4H predicted (%) Clinical, Table S3 (%)
vandetanib 100 100 100.0 NA 0.9
lapatinib 0 2 0.7 1.74 0.2-1.76
imatinib 0 6 2.1 4.36 0.04-2.7
dasatinib 0 4 1.4 2.41 2-4
afatinib 2 4 2.7 3.82 0-5.3
regorafenib 0 2 0.7 1.61 0.3-5.6
ponatinib 0 0 0.0 1.88 7
sorafenib 4 2 3.3 8.54 3.3-18.6
trametinib 2 2 2.0 1.70 3-11
pazopanib 2 4 2.7 4.43 7-13
sunitinib 0 4 1.4 1.79 2.7-19

ggplot(inc_tab, aes(factor(drug, levels = fig4h$drug))) +
  geom_point(aes(y = weighted, colour = "Simulated (this package)"), size = 3) +
  geom_point(aes(y = pmin(paper_pred, 20), colour = "Wan 2025 Fig. 4H"), size = 3, shape = 17) +
  scale_colour_manual(values = c("Simulated (this package)" = "steelblue", "Wan 2025 Fig. 4H" = "palevioletred")) +
  labs(
    x = NULL, y = "Incidence of cardiac dysfunction (%)", colour = NULL,
    caption = "Fig. 4H vandetanib (> 20%) plotted at 20. Replicates Figure 4H of Wan 2025."
  ) +
  theme(axis.text.x = element_text(angle = 45, hjust = 1))
#> Warning: Removed 1 row containing missing values or values outside the scale range
#> (`geom_point()`).

others <- inc_tab$weighted[inc_tab$drug != "vandetanib"]
stopifnot(
  # The paper's one outlier: vandetanib's predicted incidence exceeds 20%.
  inc_tab$weighted[inc_tab$drug == "vandetanib"] > 20,
  # The other ten sit in the low single digits, as in Fig. 4H
  # (median of the paper's ten values: 2.1%).
  median(others) < 8
)

The simulated incidences fall in the same low range as the paper’s Figure 4H predictions, and vandetanib is the clear outlier in both (its effective concentration exceeds its EEC). The back-solved ATP50_ref, which no in-vivo number was used to choose, therefore reproduces the paper’s translation to clinical incidence.

Assumptions and deviations

  • Eq. 13 is corrected. The paper prints ATP50 = ATP50_ref x (1 - EAAC), which lowers ATP50 as the drug concentration rises and so raises contractility, contradicting the text (“a higher ATP demand is needed to achieve the same AAC … ATP50 is considered proportional to drug exposure”) and the Results (lower EEC means greater cardiotoxicity). The models use ATP50 = ATP50_ref / (1 - EAAC), the only form that reproduces the Figure 3 contractility curves of the TKIs with an EEC; Eq. 14 is multiplied through by (1 - EAAC) so it stays finite.
  • ATP50_ref is back-solved. It is reported in neither Wan 2025 nor the cited Sang 2024; the value 1.35 was fitted by the maintainers to the digitised Figure 3 curves of all 21 TKIs (see above), with the Table 1 values fixed. The figure’s curves are not fully reproducible from Table 1 (erlotinib, axitinib and crizotinib), so the value is approximate.
  • kout_MMP and kout_ATP are not printed; they equal kin_MMP and kin_ATP because Sang 2024 normalises the MMP and ATP baselines to 1.
  • kinj unit. Table 1 heads the column “kinj (uM) (x10^-2)”; because Einj is dimensionless in Eq. 2, kinj is a rate constant and is encoded as the printed value x 1e-2 in 1/h.
  • Imatinib H. Table S8 prints IC50,3D 33 uM and IC50,2D 78.2 uM for imatinib, which gives H = 2.37 and a Table S1 effective concentration of 2.68 uM. Transposing the two columns gives H = 0.42 and reproduces Table S1 (0.47 uM); it also matches the Discussion (IC50 is higher in 3D cultures, and accounting for heterogeneity improved the imatinib prediction). The model uses H = 0.422.
  • Absorption not fully specified in Table S2. Ibrutinib and nilotinib have only a lag time; their primary PK models (Al-Ghazawi 2021, Giles 2013) describe zero-order absorption whose duration Table S2 does not give, so these two TKIs are not packaged. Ponatinib uses the two parallel routes of Hanley 2022: a fraction 0.534 (the Table S2 “F”) through a depot and two transit compartments at ktr = 1.30 1/h, the rest by first-order absorption at ka = 4.41 1/h after a 3.93-h lag; total bioavailability 1 reproduces Table S1. Regorafenib uses the Keunecke 2020 chain of a depot and two transit compartments (ka = ktr = 0.48 1/h) with the printed 0.36-h lag; Table S2 does not give the number of transit compartments, so both are taken from the primary papers.
  • Printed values kept despite Table S1. Sorafenib, sunitinib and trametinib do not reproduce Table S1 (see above); the printed Table S2, S5, S7 and S8 values are kept. Vemurafenib fu (0.001%) and gefitinib fu (0.1%) are also kept as printed. Vemurafenib is dosed 960 mg once daily, as printed in Table S4.
  • Between-patient variability. The Table S2 CVs are applied as log-normal variability on the corresponding parameters (variance log(1 + CV^2)), including CV of F on fdepot; the Table 2 CVs of MAP0, LVEDV0 and Index_SV0 describe the covariate distribution of the virtual cohort rather than model random effects; CV_FB_LVESV is the random effect etalfb_lvesv. The TD parameters have no variability.
  • Cardiovascular QSP. The dissipation rate constants, the MAP feedback FB_MAP_0 with its baseline-MAP exponent, and the numerical guard are those of the Sang 2021 human model, as Section 4.3 states; FB_LVESV and the baselines are the Table 2 values. DIS_HYPERT and DIS_DILATED_CARDIOMYOPATHY select the population’s FB_LVESV; the baselines are supplied as covariates.
  • Population definition. Hypertensive MAP is sampled from the Table 2 log-normal and kept only at or above 106 mmHg; the incidence definition is read in percentage points of LVEF.
  • Solver. One-year daily-dosing solves can exceed the default step limit of the integrator; the vignette passes maxsteps = 5e5 to rxSolve(). Patients whose solve still fails are reported as unsolved rather than counted as non-events.
  • No residual error is reported; all residual SDs are fixed to 0.