Skip to contents

The paper

Yamada et al. (2025) used a multiple model-informed drug-development (MIDD) strategy to choose a first-in-human dosing regimen for an intratumorally injected oncolytic vaccinia virus carrying interleukin-7 and interleukin-12 genes (hIL-7/mIL-12-VV in mice, hIL-7/hIL-12-VV in humans) combined with an anti-PD-1 antibody (the murine surrogate RMP1-14, clinically pembrolizumab). The paper develops four models:

  1. a mouse preclinical QSP model (10 ODEs) calibrated to four CT26.WT tumour-growth experiments;
  2. a human clinical QSP model built by extending the Wang et al. 2020 immune-oncology platform with the oncolytic-virus mechanism;
  3. a mouse agent-based model (ABM); and
  4. a human ABM.

The two QSP models are ordinary-differential-equation systems and are packaged here:

  • Yamada_2025_oncolyticVirus_mouse_qsp
  • Yamada_2025_oncolyticVirus_human_qsp

The two ABMs are agent-based cell-cycle simulations (Physiomics’ Virtual Tumour) whose tumour-growth engine is an ensemble of cell-cycle agents with a proprietary volume algorithm rather than an ODE system; they are not packaged (see Assumptions and deviations).

mod_mouse <- rxode2::rxode2(readModelDb("Yamada_2025_oncolyticVirus_mouse_qsp"))
mod_human <- rxode2::rxode2(readModelDb("Yamada_2025_oncolyticVirus_human_qsp"))

Population

Both models are deterministic mechanism models: the authors generated virtual subjects by sampling parameter ranges (preclinical fit / Latin hypercube sampling), not by fitting inter-individual variability or residual error, so the packaged models carry no etas and no error model.

  • Mouse QSP was calibrated in NONMEM to four CT26.WT subcutaneous tumour-growth experiments in BALB/c mice (Exp1-Exp4, 19 conditions including 6 vehicle arms, 6-10 mice per arm; Supplementary Table S1). Tumour volumes were converted to cell counts assuming 1e6 cells/mm^3. Fourteen of the 28 parameters were estimated; the minimisation terminated before convergence and the authors used the terminated estimates (Supplementary Table S2).
  • Human QSP inherits its immune-network parameters from the Wang 2020 HER2-negative-breast-cancer model (Ref 4). 125 virtual-patient parameter sets were sampled from the Supplementary Table S6 ranges; 100 reached a preselected baseline tumour diameter (uniform 1.0-4.5 cm) within an 8000-day untreated run-in that started from 1e6 cancer cells. The viral infection rate constant was lowered relative to the mouse model to keep the single-agent objective response rate below 50%.

Source trace

Model element Source
Mouse QSP ODEs (virus, tumour Ni/I, nTeff, dTeff, PD-L1, anti-PD-1 depot/central, distant tumour) Supplement eqs 1-10
Mouse QSP PD-L1 equation numerator (dTeff) Kosinsky 2018 (Ref 3) eq 16; Table S2 K_pdl definition
Mouse QSP algebraic rules (TCD, DCm, Agsys, PD_L1free, ISC, IAR) Supplementary Table S2 (Rules)
Mouse QSP 28 parameters, per-experiment growth rate r Supplementary Table S2
Mouse QSP baseline tumour volume (100 mm^3) Supplementary Methods S1 (randomisation at > 100 mm^3)
Mouse dosing regimens Supplementary Table S1
Human QSP reactions R1-R158 Supplementary Table S3
Human QSP rules (incl. V_T.C1 = Cuit + Cit, IL7_act, IL12_act) Supplementary Table S4
Human QSP fixed parameters, incl. pembrolizumab PK/binding and the oncolytic-virus block Supplementary Table S5
Human QSP virtual-population parameters (typical values used) Supplementary Table S6
Human QSP parameters not tabulated by Yamada (durvalumab, ipilimumab PK, k_C1_therapy) and initial species amounts Wang 2020 (Ref 4) supplement / SBML deposit
Human run-in from 1e6 cells; dose volume by tumour size; regimens Methods, “Human QSP Model Simulations”

Mouse preclinical QSP

The mouse model describes an injected tumour (tumour_volume, the sum of noninfected and infected cells) and a distant noninjected tumour (tumour_volume_distant). Virus is injected intratumorally (into virus, pfu) and the anti-PD-1 antibody intraperitoneally (into mAbDep, nmol; 100 ug of a ~150 kDa antibody is 0.667 nmol). The tumour-growth rate constant r was fitted separately for each experiment, so each experiment is simulated with its own value.

mab_nmol <- 100e-6 / 150000 * 1e9 # 100 ug anti-PD-1 -> nmol
r_exp <- c(Exp1 = 0.36, Exp2 = 0.284, Exp3 = 0.348, Exp4 = 0.394)
mab_d1 <- c(1, 3, 7, 10, 15, 17, 20, 23, 28, 31)
mab_d7 <- c(7, 10, 15, 17)

design <- list(
  list(exp = "Exp1", arm = "vehicle", vd = 0, vdays = c(1, 3, 5), mdays = NULL),
  list(exp = "Exp1", arm = "2e4", vd = 2e4, vdays = c(1, 3, 5), mdays = NULL),
  list(exp = "Exp1", arm = "2e5", vd = 2e5, vdays = c(1, 3, 5), mdays = NULL),
  list(exp = "Exp1", arm = "2e6", vd = 2e6, vdays = c(1, 3, 5), mdays = NULL),
  list(exp = "Exp1", arm = "2e7", vd = 2e7, vdays = c(1, 3, 5), mdays = NULL),
  list(exp = "Exp2", arm = "vehicle", vd = 0, vdays = 1, mdays = NULL),
  list(exp = "Exp2", arm = "2e7 day1", vd = 2e7, vdays = 1, mdays = NULL),
  list(exp = "Exp2", arm = "2e7 day1, 8", vd = 2e7, vdays = c(1, 8), mdays = NULL),
  list(exp = "Exp2", arm = "2e7 day1, 15", vd = 2e7, vdays = c(1, 15), mdays = NULL),
  list(exp = "Exp3", arm = "vehicle", vd = 0, vdays = c(1, 3, 5), mdays = NULL),
  list(exp = "Exp3", arm = "2e6", vd = 2e6, vdays = c(1, 3, 5), mdays = NULL),
  list(exp = "Exp3", arm = "2e7", vd = 2e7, vdays = c(1, 3, 5), mdays = NULL),
  list(exp = "Exp4", arm = "vehicle", vd = 0, vdays = c(1, 3, 7), mdays = NULL),
  list(exp = "Exp4", arm = "2e7", vd = 2e7, vdays = c(1, 3, 7), mdays = NULL),
  list(exp = "Exp4", arm = "aPD1 (day1)", vd = 0, vdays = NULL, mdays = mab_d1),
  list(exp = "Exp4", arm = "aPD1 (day7)", vd = 0, vdays = NULL, mdays = mab_d7),
  list(exp = "Exp4", arm = "2e7 + aPD1 (day1)", vd = 2e7, vdays = c(1, 3, 7), mdays = mab_d1),
  list(exp = "Exp4", arm = "2e7 + aPD1 (day7)", vd = 2e7, vdays = c(1, 3, 7), mdays = mab_d7)
)

solve_mouse <- function(d, tmax = 21) {
  ev <- rxode2::et(seq(0, tmax, by = 0.25), cmt = "tumNi")
  if (d$vd > 0) {
    ev <- rxode2::et(ev, time = d$vdays, amt = d$vd, cmt = "virus")
  }
  if (!is.null(d$mdays)) {
    ev <- rxode2::et(ev, time = d$mdays[d$mdays <= tmax], amt = mab_nmol, cmt = "mAbDep")
  }
  s <- rxode2::rxSolve(mod_mouse, ev, params = c(r = r_exp[[d$exp]]),
                       method = "liblsoda", returnType = "data.frame")
  data.frame(exp = d$exp, arm = d$arm, time = s$time,
             injected = s$tumour_volume, distant = s$tumour_volume_distant)
}
mouse_sim <- dplyr::bind_rows(lapply(design, solve_mouse))
ggplot(mouse_sim, aes(time, injected, colour = arm)) +
  geom_line() +
  facet_wrap(~exp, scales = "free_y") +
  scale_y_log10() +
  labs(x = "Day after randomisation", y = "Injected tumour volume (mm^3)",
       colour = NULL,
       title = "Mouse QSP: injected tumour, Table S1 regimens (cf. Figure S2)")

mouse_sim |>
  dplyr::filter(exp %in% c("Exp3", "Exp4")) |>
  tidyr::pivot_longer(c(injected, distant), names_to = "site",
                      values_to = "volume") |>
  ggplot(aes(time, volume, colour = arm, linetype = site)) +
  geom_line() +
  facet_wrap(~exp) +
  scale_y_log10() +
  labs(x = "Day after randomisation", y = "Tumour volume (mm^3)",
       colour = NULL, linetype = NULL,
       title = "Mouse QSP: injected vs distant tumour (cf. Figure S2)")

Volumes on day 14, when most vehicle arms are still within the measured range:

d14 <- mouse_sim |>
  dplyr::filter(time == 14) |>
  dplyr::select(exp, arm, injected, distant)
d14 |>
  dplyr::rename("Experiment" = exp, "Arm" = arm,
                "Injected (mm^3)" = injected, "Distant (mm^3)" = distant) |>
  knitr::kable(digits = 0, caption = "Simulated tumour volume on day 14")
Simulated tumour volume on day 14
Experiment Arm Injected (mm^3) Distant (mm^3)
Exp1 vehicle 2298 2298
Exp1 2e4 2258 2258
Exp1 2e5 1987 1992
Exp1 2e6 1314 1358
Exp1 2e7 642 980
Exp2 vehicle 434 434
Exp2 2e7 day1 112 154
Exp2 2e7 day1, 8 81 142
Exp2 2e7 day1, 15 112 154
Exp3 vehicle 1802 1802
Exp3 2e6 1001 1037
Exp3 2e7 443 717
Exp4 vehicle 4453 4453
Exp4 2e7 1685 2232
Exp4 aPD1 (day1) 3753 3753
Exp4 aPD1 (day7) 4209 4209
Exp4 2e7 + aPD1 (day1) 1511 1741
Exp4 2e7 + aPD1 (day7) 1543 1957
v14 <- function(e, a, site = "injected") d14[[site]][d14$exp == e & d14$arm == a]

# Exp1: tumour-growth inhibition is ordered by virus dose.
exp1 <- vapply(c("vehicle", "2e4", "2e5", "2e6", "2e7"),
               function(a) v14("Exp1", a), numeric(1))
stopifnot(all(diff(exp1) < 0))

# Exp3: the systemic dTeff pool slows the distant tumour as well, but less
# than the injected one.
stopifnot(
  v14("Exp3", "2e7", "distant") < v14("Exp3", "vehicle", "distant"),
  v14("Exp3", "2e7", "distant") > v14("Exp3", "2e7", "injected")
)

# Exp4: anti-PD-1 adds to the virus effect in both schedules.
stopifnot(
  v14("Exp4", "2e7 + aPD1 (day1)") < v14("Exp4", "2e7"),
  v14("Exp4", "2e7 + aPD1 (day7)") < v14("Exp4", "2e7")
)

The packaged model reproduces the dose-ordered response at the injected tumour and a weaker abscopal effect at the distant tumour. With the published parameter values the anti-PD-1-enhanced virus clearance is a small effect (Imax = 0.139, a 13.9% increase), so in Exp4 the immune benefit of the extra anti-PD-1 doses in the simultaneous arm (10 doses from day 1, against 4 from day 7) outweighs it: on day 14 the simultaneous arm is slightly smaller than the sequential arm at both sites. The paper reports that its model favoured the sequential regimen (Results, “Mouse QSP Modeling”); Figure S2, which shows that fit, is not reproduced here, and the parameters were not tuned to recover it.

Anti-PD-1 antibody pharmacokinetics (PKNCA)

The anti-PD-1 arm adds a first-order depot-to-central PK model. The serum PK of a single 100 ug intraperitoneal dose is summarised with PKNCA.

ev_pk <- rxode2::et(seq(0, 30, by = 0.25), cmt = "mAbC") |>
  rxode2::et(time = 0, amt = mab_nmol, cmt = "mAbDep")
pk <- rxode2::rxSolve(mod_mouse, ev_pk, returnType = "data.frame")
pk_in <- pk |>
  dplyr::mutate(conc = mAbC / 0.003, id = 1L, # nmol / Vd (L) = nM
                treatment = "anti-PD-1 100 ug IP") |>
  dplyr::filter(!is.na(conc))

o_conc <- PKNCA::PKNCAconc(pk_in, conc ~ time | treatment + id)
o_dose <- PKNCA::PKNCAdose(
  data.frame(time = 0, dose = mab_nmol, id = 1L,
             treatment = "anti-PD-1 100 ug IP"),
  dose ~ time | treatment + id
)
res <- PKNCA::pk.nca(PKNCA::PKNCAdata(
  o_conc, o_dose,
  intervals = data.frame(start = 0, end = 30, cmax = TRUE, tmax = TRUE,
                         auclast = TRUE, half.life = TRUE)
))
nca <- as.data.frame(res$result)
nca |>
  dplyr::select(PPTESTCD, PPORRES) |>
  dplyr::rename("NCA parameter" = PPTESTCD, "Value" = PPORRES) |>
  knitr::kable(digits = 3, caption = "Anti-PD-1 serum NCA (mouse QSP, nM and days)")
Anti-PD-1 serum NCA (mouse QSP, nM and days)
NCA parameter Value
auclast 1455.866
cmax 205.957
tmax 0.500
tlast 30.000
lambda.z 0.150
r.squared 1.000
adj.r.squared 1.000
lambda.z.time.first 0.750
lambda.z.time.last 30.000
lambda.z.n.points 118.000
clast.pred 2.516
half.life 4.621
span.ratio 6.330

# Terminal half-life equals log(2) / kel_mAb (0.15/day) = 4.62 days.
hl <- nca$PPORRES[nca$PPTESTCD == "half.life"]
stopifnot(abs(hl / (log(2) / 0.15) - 1) < 0.02)

Human clinical QSP

The human model grows a tumour from 1e6 cancer cells; therapy starts once the tumour reaches its baseline diameter, with the node values at that moment as the patient’s initial conditions (as in the paper). One untreated run-in gives the initial conditions for four typical patients at baseline diameters of 1.5, 2.5, 3.5 and 4.4 cm, one in each of the paper’s dose-volume bands (1, 2, 4 and 6 mL of virus suspension). The virus dose in pfu is the concentration (pfu/mL) times that volume.

human_ctl <- list(method = "cvode", maxsteps = 5e5, atol = 1e-9, rtol = 1e-9)
solve_human <- function(ev, inits = NULL) {
  rxode2::rxSolve(mod_human, ev, inits = inits, method = human_ctl$method,
                  maxsteps = human_ctl$maxsteps, atol = human_ctl$atol,
                  rtol = human_ctl$rtol, returnType = "data.frame")
}
runin <- solve_human(rxode2::et(seq(0, 3000, by = 2), cmt = "q_V_T_Cit"))

patients <- data.frame(
  patient = c("1.5 cm", "2.5 cm", "3.5 cm", "4.4 cm"),
  base_cm = c(1.5, 2.5, 3.5, 4.4),
  dose_mL = c(1, 2, 4, 6)
)
patients$t_base <- vapply(patients$base_cm, function(b) {
  runin$time[which(runin$tumour_diameter >= b)[1]]
}, numeric(1))
knitr::kable(
  patients |>
    dplyr::rename("Patient" = patient, "Baseline diameter (cm)" = base_cm,
                  "Virus volume (mL)" = dose_mL,
                  "Run-in time to baseline (days)" = t_base),
  caption = "Typical patients taken from the untreated run-in"
)
Typical patients taken from the untreated run-in
Patient Baseline diameter (cm) Virus volume (mL) Run-in time to baseline (days)
1.5 cm 1.5 1 884
2.5 cm 2.5 2 1094
3.5 cm 3.5 4 1234
4.4 cm 4.4 6 1328

Arms: virus alone at 1e7, 1e8 or 5e8 pfu/mL every 2 weeks (Figure S3), and 5e8 pfu/mL plus pembrolizumab 400 mg every 6 weeks starting 0, 1, 2 or 4 weeks after the first virus dose (H-0W to H-4W, Figure 4).

pembro_nmol <- 400 / 149000 * 1e6 # 400 mg pembrolizumab, MW ~149 kDa -> nmol
arms <- data.frame(
  arm = c("untreated", "1e7", "1e8", "5e8", "H-0W", "H-1W", "H-2W", "H-4W"),
  conc = c(0, 1e7, 1e8, 5e8, 5e8, 5e8, 5e8, 5e8),
  pembro_start = c(NA, NA, NA, NA, 0, 7, 14, 28)
)
horizon <- 400
run_arm <- function(p, a) {
  inits <- unlist(runin[runin$time == patients$t_base[p], mod_human$state])
  ev <- rxode2::et(seq(0, horizon, by = 5), cmt = "q_V_T_Cit")
  if (arms$conc[a] > 0) {
    ev <- rxode2::et(ev, time = seq(0, horizon, by = 14),
                     amt = arms$conc[a] * patients$dose_mL[p],
                     cmt = "q_V_T_Virus_T")
  }
  if (!is.na(arms$pembro_start[a])) {
    ev <- rxode2::et(ev, time = seq(arms$pembro_start[a], horizon, by = 42),
                     amt = pembro_nmol, cmt = "q_V_C_pembro")
  }
  s <- solve_human(ev, inits = inits)
  data.frame(patient = patients$patient[p], arm = arms$arm[a], time = s$time,
             diameter = s$tumour_diameter,
             pct_change = 100 * (s$tumour_diameter / s$tumour_diameter[1] - 1))
}
human_sim <- dplyr::bind_rows(lapply(seq_len(nrow(patients)), function(p) {
  dplyr::bind_rows(lapply(seq_len(nrow(arms)), function(a) run_arm(p, a)))
}))
human_sim$arm <- factor(human_sim$arm, levels = arms$arm)
ggplot(human_sim, aes(time, pct_change, colour = arm)) +
  geom_line() +
  geom_hline(yintercept = -30, linetype = "dashed") +
  facet_wrap(~patient) +
  coord_cartesian(ylim = c(-100, 150)) +
  labs(x = "Day after treatment start",
       y = "Change in tumour diameter from baseline (%)", colour = NULL,
       title = "Human QSP: virus alone and with pembrolizumab (cf. Figures 4 and S3)")

resp <- human_sim |>
  dplyr::filter(time %in% c(100, 365)) |>
  dplyr::select(patient, arm, time, pct_change) |>
  tidyr::pivot_wider(names_from = time, values_from = pct_change,
                     names_prefix = "day_")
resp |>
  dplyr::rename("Patient" = patient, "Arm" = arm,
                "Day 100 change (%)" = day_100,
                "Day 365 change (%)" = day_365) |>
  knitr::kable(digits = 1,
               caption = "Change in tumour diameter from baseline (RECIST response at <= -30%)")
Change in tumour diameter from baseline (RECIST response at <= -30%)
Patient Arm Day 100 change (%) Day 365 change (%)
1.5 cm untreated 27.7 142.5
1.5 cm 1e7 26.7 138.8
1.5 cm 1e8 18.4 105.7
1.5 cm 5e8 -18.4 -99.1
1.5 cm H-0W -8.3 -96.7
1.5 cm H-1W -10.7 -96.8
1.5 cm H-2W -10.5 -96.8
1.5 cm H-4W -12.1 -96.9
2.5 cm untreated 27.3 138.6
2.5 cm 1e7 26.9 136.8
2.5 cm 1e8 23.3 123.4
2.5 cm 5e8 7.4 49.6
2.5 cm H-0W 8.6 3.0
2.5 cm H-1W 7.9 1.1
2.5 cm H-2W 8.4 3.0
2.5 cm H-4W 8.3 3.8
3.5 cm untreated 27.0 132.2
3.5 cm 1e7 26.7 130.8
3.5 cm 1e8 24.0 121.6
3.5 cm 5e8 12.5 75.8
3.5 cm H-0W 11.6 18.6
3.5 cm H-1W 11.3 17.9
3.5 cm H-2W 11.8 19.5
3.5 cm H-4W 12.1 20.8
4.4 cm untreated 26.5 123.9
4.4 cm 1e7 26.3 122.7
4.4 cm 1e8 24.3 116.3
4.4 cm 5e8 15.5 85.9
4.4 cm H-0W 13.6 28.5
4.4 cm H-1W 13.4 28.1
4.4 cm H-2W 13.9 29.4
4.4 cm H-4W 14.4 30.8
chg <- function(p, a, day = "day_365") resp[[day]][resp$patient == p & resp$arm == a]

for (p in patients$patient) {
  # virus monotherapy is ordered by dose at every baseline size
  mono <- vapply(c("untreated", "1e7", "1e8", "5e8"), function(a) chg(p, a),
                 numeric(1))
  stopifnot(all(diff(mono) < 0))
  # the four pembrolizumab schedules end within 5 percentage points of each
  # other by day 365 (paper: no obvious difference between schedules)
  combo <- vapply(c("H-0W", "H-1W", "H-2W", "H-4W"), function(a) chg(p, a),
                  numeric(1))
  stopifnot(diff(range(combo)) < 5)
}

# Adding pembrolizumab improves day-365 control over the high virus dose alone
# for the patients who do not respond to the virus alone.
for (p in c("2.5 cm", "3.5 cm", "4.4 cm")) {
  stopifnot(chg(p, "H-0W") < chg(p, "5e8") - 20)
}

# The infected-cell / virus cascade engages after a single dose: infected
# cells appear (lysed within hours, Delta_I = 57.4/day, so the infected
# fraction stays small) and the cancer-cell count falls over the first days.
tx <- solve_human(
  rxode2::et(seq(0, 28, by = 1), cmt = "q_V_T_Cit") |>
    rxode2::et(time = 0, amt = 5e8, cmt = "q_V_T_Virus_T"),
  inits = unlist(runin[runin$time == patients$t_base[1], mod_human$state])
)
stopifnot(
  max(tx$infected_fraction) > 1e-4,
  min(tx$cancer_cells[tx$time <= 7]) < 0.95 * tx$cancer_cells[1]
)

The paper’s virtual population gave objective response rates of 0%, 2% and 22% for virus alone at 1e7, 1e8 and 5e8 pfu/mL and 48% for every combination schedule at day 365. The four typical patients here follow the same pattern: virus alone is dose-ordered and shrinks only the smallest tumour, while pembrolizumab adds a large benefit by day 365 for the larger tumours and the four schedules end within a few percentage points of each other. At day 100 the paper ranked the schedules 1 week >= 0 weeks > 2 weeks > 4 weeks by response rate; for the typical 1.5 cm patient here the early differences are within a few percentage points and the later starts are marginally better, because the pembrolizumab-enhanced virus clearance acts before the immune benefit builds up. A response rate needs the Table S6 virtual population, which is not reproduced (see below).

Pembrolizumab pharmacokinetics (PKNCA)

base1 <- unlist(runin[runin$time == patients$t_base[1], mod_human$state])
pk_h <- solve_human(
  rxode2::et(seq(0, 42, by = 0.5), cmt = "q_V_C_pembro") |>
    rxode2::et(time = 0, amt = pembro_nmol, cmt = "q_V_C_pembro"),
  inits = base1
)
pk_hin <- pk_h |>
  dplyr::mutate(conc = Cc_pembrolizumab, id = 1L,
                treatment = "pembrolizumab 400 mg IV") |>
  dplyr::filter(!is.na(conc))
o_conc_h <- PKNCA::PKNCAconc(pk_hin, conc ~ time | treatment + id)
o_dose_h <- PKNCA::PKNCAdose(
  data.frame(time = 0, dose = pembro_nmol, id = 1L,
             treatment = "pembrolizumab 400 mg IV"),
  dose ~ time | treatment + id
)
res_h <- PKNCA::pk.nca(PKNCA::PKNCAdata(
  o_conc_h, o_dose_h,
  intervals = data.frame(start = 0, end = 42, cmax = TRUE, tmax = TRUE,
                         auclast = TRUE, half.life = TRUE)
))
nca_h <- as.data.frame(res_h$result)
nca_h |>
  dplyr::select(PPTESTCD, PPORRES) |>
  dplyr::rename("NCA parameter" = PPTESTCD, "Value" = PPORRES) |>
  knitr::kable(digits = 3, caption = "Pembrolizumab plasma NCA (human QSP, nM and days)")
Pembrolizumab plasma NCA (human QSP, nM and days)
NCA parameter Value
auclast 6305.782
cmax 447.427
tmax 0.000
tlast 42.000
lambda.z 0.016
r.squared 1.000
adj.r.squared 1.000
lambda.z.time.first 12.000
lambda.z.time.last 42.000
lambda.z.n.points 61.000
clast.pred 95.107
half.life 42.639
span.ratio 0.704

# Cmax after the bolus is the dose over the 6 L central compartment.
cmax_h <- nca_h$PPORRES[nca_h$PPTESTCD == "cmax"]
stopifnot(abs(cmax_h / (pembro_nmol / 6) - 1) < 0.01)

Assumptions and deviations

  • The two agent-based models are not packaged. The mouse and human ABMs are Physiomics’ Virtual Tumour, an agent-based cell-cycle simulation whose tumour volume comes from a proprietary algorithm; it is not an ODE system and cannot be expressed in rxode2. Only the immune-system component of the ABM is an ODE set (Supplement eqs 1-12, Table S7), and it is driven by the cell count from the agent engine, so it is not a standalone model either.
  • Mouse QSP PD-L1 equation. Supplement eq 6 is printed with PDL1 in the numerator of the saturable term. The source model (Kosinsky 2018, Ref 3, eq 16) and the definition of K_pdl (“sensitivity of PD-L1 expression up-regulation to dTeff count”, in cells) show that the numerator is the differentiated-effector-T-cell count; the packaged model uses dTeff / (dTeff + K_pdl). Eq 2 likewise prints T_eff for the killing term, which is dT_eff in eqs 3 and 10 and in the source model.
  • Mouse QSP eradication threshold. The infection flux V_beta * Virus * Tum_ni / Tum_tot / 1e6 does not shrink with tumour size while noninfected cells dominate, so the injected tumour reaches zero in finite time and the Tum_ni / Tum_tot and Tum_i / Tum_tot ratios become infinitely stiff (the solver fails, for example, in the Exp2 day 1 + 8 arm). Below one tumour cell (1e-6 mm^3) the packaged model freezes the two injected-tumour volumes and floors the ratio denominators at one cell. This changes nothing while the tumour is larger than one cell.
  • Mouse QSP growth rate. r was fitted per experiment (0.36, 0.284, 0.348 and 0.394 1/day for Exp1-Exp4). The model default is the Exp1 value; this vignette overrides r for each experiment.
  • Mouse QSP control virus. The unmodified control virus (Exp1) has no IL-7 or IL-12 genes, but the model ties IL-7 and IL-12 stimulation to infected cells, so it cannot distinguish the two viruses; the control-virus arm is not simulated.
  • Mouse QSP baseline and antibody dose. Mice were randomised when the tumour exceeded 100 mm^3; the injected and distant tumours start at 100 mm^3 with the immune and virus states at rest. The 100 ug anti-PD-1 dose is converted to nmol with an assumed molecular weight of 150 kDa.
  • Mouse QSP fit did not converge. The NONMEM minimisation terminated before convergence and the authors used the terminated estimates; the packaged values are those estimates.
  • Human QSP is an extension of Wang 2020. The immune network, its parameters and the initial species amounts are inherited from the Wang 2020 model (Ref 4), packaged separately as Wang_2020_entinostat_nivolumab_ipilimumab_qsp. The nivolumab arm is replaced by pembrolizumab (Table S5 parameters); the anti-PD-L1 (durvalumab), anti-CTLA-4 (ipilimumab) and entinostat modules are carried over but not dosed in this paper. Values are converted to the model’s counts/litre/dm^2/day system as in the Wang 2020 translation; each ini() line’s comment records the tabulated value and unit. Where Yamada’s tables differ from Wang 2020, Yamada’s value is used (k_C1_growth = 0.0075 1/day from Table S6; IC50_ENT_ArgI = 3.74e-7 M from Table S5). Parameters Yamada does not tabulate (durvalumab and ipilimumab PK, k_C1_therapy) and the initial species amounts come from Wang 2020.
  • Human QSP run-in. The untreated run-in starts from 1e6 cancer cells (Methods), not from the single cell of the Wang 2020 platform. Yamada’s Table S5 still lists the Wang pre-treatment diameter of 0.1 cm, which is not used.
  • Human QSP virus units. Virus_T is held in molecule-equivalents, the unit of the Table S5 virus parameters (Alpha in molecule/cell, Beta_infect in cell/day/molecule); one pfu is taken as one molecule-equivalent, as in the mouse model, where the same parameters are in pfu. The dose in pfu is the concentration times the tumour-size-dependent volume (Methods).
  • Human QSP C2 tracker. Table S3 reaction 155 defines a cancer-cell tracker V_T.C2 without infection, while Table S4 also lists the rule V_T.C2 = V_T.Cuit. C2 feeds back on nothing; the packaged model integrates reaction 155 as printed.
  • Human QSP virtual population. The paper sampled 15 parameters from the Table S6 distributions by Latin hypercube; some entries are ambiguous as printed (the lognormal k_C1_growth has “mean 0.0075, SD 1”; the k_cln_ENT range is given as [6e-4, 1.22e-4]). The packaged model carries the typical values, and this vignette uses four typical patients at different baseline diameters rather than the virtual population, so it checks the pattern of response, not the published response rates.
  • No NCA was reported in the paper for either antibody; the PKNCA sections check the antibody PK sub-models against their own parameters.