Skip to contents

Sunitinib (Schindler 2016)

Source: vignettes/articles/Schindler_2016_sunitinib.Rmd
Schindler_2016_sunitinib.Rmd

Model and source

  • Citation: Schindler E, Amantea MA, Karlsson MO, Friberg LE. (2016). PK-PD modeling of individual lesion FDG-PET response to predict overall survival in patients with sunitinib-treated gastrointestinal stromal tumor. CPT Pharmacometrics Syst Pharmacol 5(4):173-181. doi:10.1002/psp4.12057. DDMORE Foundation Model Repository: DDMODEL00000221.
  • Description: Joint pharmacodynamic model for sunitinib in advanced GIST coupling a five-lesion indirect-response model of [18F]FDG-PET SUVmax with a per-subject sum-of-longest-diameters (SLD) tumor-growth-inhibition module and constant-baseline-hazard Weibull time-to-event sub-models for overall survival and study dropout (Schindler 2016 / DDMODEL00000221). Sunitinib exposure enters via an effect compartment driven by a per-day AUC = DOSE / CLI; the OS hazard depends on the per-subject week-1 maximum across-lesion relative SUVmax change from baseline, RCFB1MAX.
  • Article: https://doi.org/10.1002/psp4.12057
  • DDMORE Foundation Model Repository entry: DDMODEL00000221 – bundle directory in the operator’s ddmore_scraping mirror at 221/, containing Executable_SLD_SUV_OS_GIST.mod, Output_real_SLD_SUV_OS_GIST.lst, Output_simulated_SLD_SUV_OS_GIST.lst, Simulated_SLD_SUV_OS_GIST.csv, DDMODEL00000221.rdf, Command.txt, and 221.json. The linked publication itself (Schindler 2016, CPT:PSP) is not on disk in the worktree; parameter values, structural equations, and stop-and-ask decisions in this vignette and the model file are sourced from the bundle’s .mod (structural equations + initial values) and .lst (final estimates).

Population

The Schindler 2016 model was developed on a 66-patient pooled cohort of imatinib-resistant or imatinib-intolerant advanced gastrointestinal stromal tumor (GIST) patients on second-line oral sunitinib at the standard 50 mg/day, 4-weeks-on / 2-weeks-off regimen. Each patient had serial [18F]FDG-PET SUVmax assessments on up to five target lesions, target-lesion sum-of-longest-diameters (SLD) measurements, and overall survival follow-up to event or right-censoring. Detailed baseline demographics (age, weight, sex, race / ethnicity distribution, prior-line distribution) live in Table 1 of the linked publication, which is not on disk in this worktree; those fields are recorded as NA in the model’s population metadata to make the gap explicit, with a notes string pointing to Schindler 2016 Table 1 as the source.

The same information is available programmatically via the model’s population metadata (rxode2::rxode2(readModelDb("Schindler_2016_sunitinib"))$population after the model is loaded).

Source trace

The per-parameter origin is recorded as an in-file comment next to each ini() entry in inst/modeldb/ddmore/Schindler_2016_sunitinib.R. The table below collects them in one place for review. All values are final estimates from the Output_real_SLD_SUV_OS_GIST.lst FINAL PARAMETER ESTIMATE block (lines 868-1000); the .mod $THETA / $OMEGA / $SIGMA blocks are initial values and are not used as the parameter source per extract-literature-model Phase 1 step 9.

Parameter (.lst label) Final value Role Source location
THETA(1) KOUT 555.634 (1/week x 1000); FIX SUVmax loss-of-response rate .lst line 880; .mod line 54
THETA(2) ALPHA 0; FIX SUVmax disease-progression slope .lst line 880; .mod line 63
THETA(3) DRUG 0.94569; FIX SUVmax drug-effect coefficient .lst line 880; .mod line 72
THETA(4) IBASE_SUV 7.58866; FIX Typical baseline SUVmax .lst line 880; .mod line 81
THETA(5) VAR RUV 0.173794; FIX SUVmax residual variance (log scale) .lst line 880; .mod line 90
THETA(6) CORR RUV 0.466827; FIX SUVmax across-lesion residual correlation .lst line 880; .mod line 97
THETA(7) KG*1000 10.4796 (1/week x 1000); FIX SLD tumor growth rate .lst line 880; .mod line 111
THETA(8) LAMBDA*1000 20.0529 (1/week x 1000); FIX SLD drug-resistance decay rate .lst line 880; .mod line 113
THETA(9) KDRUG*1000 16.5971 (1/week x 1000); FIX SLD drug-effect rate .lst line 880; .mod line 115
THETA(10) IBASE_SLD 263.064 mm; FIX Baseline sum of longest diameters .lst line 880; .mod line 117
THETA(11) RUV SLD 0.0667619; FIX SLD proportional residual error .lst line 880; .mod line 256
THETA(12) LAMBH 0.019030 (1/week) OS Weibull scale (estimated) .lst iteration 8 NPARAMETR (line 790)
THETA(13) ALPHH 1; FIX OS Weibull shape (constant baseline hazard) .lst line 881; .mod line 124
THETA(14) LAMBD 0.019924 (1/week) Dropout Weibull scale (estimated) .lst iteration 8 NPARAMETR (line 790)
THETA(15) ALPHD 1; FIX Dropout Weibull shape .lst line 881; .mod line 126
THETA(16) Predictor 5.3558 OS hazard coefficient on RCFB1MAX .lst iteration 8 NPARAMETR (line 790)
OMEGA BLOCK(2) KDRUG_SLD, DRUG_SUV (0.398557, 0.397941, 0.548112); FIX Typical-subject drug-effect IIV .lst lines 928-934; .mod lines 375-377
OMEGA BLOCK(1) SAME (x5) DRUG_LES 0.324481; FIX Lesion-specific drug-effect IIV .lst lines 942-960; .mod lines 379-384
OMEGA(20) IBASE_SLD 0.288669; FIX SLD baseline IIV .lst line 962; .mod line 386
OMEGA(21) IBASE_SUV 0.105177; FIX Typical-subject SUVmax baseline IIV .lst line 968; .mod line 387
OMEGA BLOCK(1) SAME (x5) IBASE_LES 0.0549036; FIX Lesion-specific SUVmax baseline IIV .lst lines 970-990; .mod lines 389-394
OMEGA(27) KG 0.361258; FIX SLD growth-rate IIV .lst line 994; .mod line 396
KEO log(2)/50 (1/h) Effect-compartment equilibration .mod line 118
ODE structure (5xSUV + SLD + 2xWeibull + effect) n/a Equations 1-9 of the source .mod $DES block lines 137-180
Residual error structure (5xlog-additive + SLD prop) n/a Y = log(A(i)) + EPSN, Cholesky-decomposed EPS over 5 lesions; Y = IPRED + W1*EPS(6) for SLD .mod $ERROR block lines 183-263

The MINIMIZATION line in Output_real_*.lst (line 795) reads MINIMIZATION SUCCESSFUL followed by HOWEVER, PROBLEMS OCCURRED WITH THE MINIMIZATION. REGARD THE RESULTS OF THE ESTIMATION STEP CAREFULLY, AND ACCEPT THEM ONLY AFTER CHECKING THAT THE COVARIANCE STEP PRODUCES REASONABLE OUTPUT. Per the extract-literature-model skill’s ddmore-source.md Section “Reading final estimates from .lst”, MINIMIZATION SUCCESSFUL (with caveat) is acceptable; only MINIMIZATION TERMINATED triggers a sidecar. The caveat is recorded in the Errata below.

Virtual cohort

The original observed dataset is not publicly available and is not in the DDMODEL00000221 bundle. The bundle’s Simulated_SLD_SUV_OS_GIST.csv ships a single virtual subject with the standard 50 mg/day 4-weeks-on / 2-weeks-off-style schedule as a regression-style smoke test. For the F.2 self-consistency check below the vignette uses a virtual cohort that mirrors that single-subject layout for the SUVmax + SLD + cumulative-hazard arms; the OS / dropout TTE arms are evaluated at typical-value parameters with RCFB1MAX derived from the simulated week-1 SUVmax trajectory.

set.seed(20260506)

n_sub <- 5L

# Build a single per-subject schedule that follows the bundle's 14-on / 14-off
# pattern across 100 days (~2400 hours). The bundle's Simulated CSV uses a
# 24-hour observation grid for FLAG = 1 dosing-record-equivalents; we keep
# the same grid. The model has six observable outputs (suv1..suv5 plus sld);
# rxSolve requires each observation row to be tagged with the cmt of the
# output it represents, so the event table is built by binding one
# observation block per output via rxode2::et().
day_hours <- seq(0, 100 * 24, by = 24)
on_off_for <- function(t) ifelse(((t %/% (24 * 14)) %% 2) == 0, 50, 0)

cohort_one <- function(id) {
  ev <- rxode2::et(time = day_hours, cmt = "suv1") |>
    rxode2::etRbind(rxode2::et(time = day_hours, cmt = "suv2")) |>
    rxode2::etRbind(rxode2::et(time = day_hours, cmt = "suv3")) |>
    rxode2::etRbind(rxode2::et(time = day_hours, cmt = "suv4")) |>
    rxode2::etRbind(rxode2::et(time = day_hours, cmt = "suv5")) |>
    rxode2::etRbind(rxode2::et(time = day_hours, cmt = "sld"))
  ev$id           <- id
  ev$DOSE   <- on_off_for(ev$time)
  ev$CLI <- 50
  ev$RCFB1MAX     <- 0   # placeholder for stage 1; recomputed for stage 2
  ev
}

events <- bind_rows(lapply(seq_len(n_sub), cohort_one))
stopifnot(!anyDuplicated(unique(events[, c("id", "time", "cmt")])))

Stage 1 simulation (SUVmax + SLD only)

The OS / dropout TTE arms depend on RCFB1MAX (the per-subject week-1 maximum across-lesion relative SUVmax change from baseline). In the source NONMEM .mod RCFB1MAX is a record-loop intermediate computed inside $ERROR from the running SUVmax compartment values at FLAG = 1 / TIME = 168 h; rxode2 does not have an idiomatic equivalent of NONMEM’s record-loop persistent state, so the model file consumes RCFB1MAX as a per-subject input covariate. The vignette therefore runs a two-stage simulation:

  1. Stage 1 – simulate the SUVmax + SLD ODEs with RCFB1MAX = 0 (the OS arm is irrelevant in stage 1; only its RCFB1MAX dependency is). Extract the per-subject lesion-specific SUVmax values at t = 168 h.
  2. Stage 2 – compute RCFB1MAX per subject as the maximum (across the lesions present) of (SUVmax(168) - SUVmax(0)) / SUVmax(0), bind it back into the event table as a per-subject covariate, and re-run the full model so the OS / dropout cumulative-hazard ODEs use the correct predictor.
mod <- rxode2::rxode2(readModelDb("Schindler_2016_sunitinib")) |> rxode2::zeroRe()

sim_stage1 <- rxode2::rxSolve(
  mod, events = events,
  keep = c("DOSE", "CLI", "RCFB1MAX"),
  returnType = "data.frame"
) |>
  as_tibble() |>
  distinct(id, time, .keep_all = TRUE)
#> ℹ omega/sigma items treated as zero: 'etalkdrug', 'etaldrug', 'etaldrug_les1', 'etaldrug_les2', 'etaldrug_les3', 'etaldrug_les4', 'etaldrug_les5', 'etalbase_sld', 'etalbase', 'etalbase_les1', 'etalbase_les2', 'etalbase_les3', 'etalbase_les4', 'etalbase_les5', 'etalkg'
#> Warning: multi-subject simulation without without 'omega'
glimpse(sim_stage1)
#> Rows: 505
#> Columns: 38
#> $ id          <int> 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1…
#> $ time        <dbl> 0, 24, 48, 72, 96, 120, 144, 168, 192, 216, 240, 264, 288,…
#> $ auc_daily   <dbl> 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0…
#> $ kout        <dbl> 0.003307345, 0.003307345, 0.003307345, 0.003307345, 0.0033…
#> $ drug_typ    <dbl> 0.94569, 0.94569, 0.94569, 0.94569, 0.94569, 0.94569, 0.94…
#> $ base_typ    <dbl> 7.58866, 7.58866, 7.58866, 7.58866, 7.58866, 7.58866, 7.58…
#> $ drug1       <dbl> 0.94569, 0.94569, 0.94569, 0.94569, 0.94569, 0.94569, 0.94…
#> $ drug2       <dbl> 0.94569, 0.94569, 0.94569, 0.94569, 0.94569, 0.94569, 0.94…
#> $ drug3       <dbl> 0.94569, 0.94569, 0.94569, 0.94569, 0.94569, 0.94569, 0.94…
#> $ drug4       <dbl> 0.94569, 0.94569, 0.94569, 0.94569, 0.94569, 0.94569, 0.94…
#> $ drug5       <dbl> 0.94569, 0.94569, 0.94569, 0.94569, 0.94569, 0.94569, 0.94…
#> $ ibase1      <dbl> 7.58866, 7.58866, 7.58866, 7.58866, 7.58866, 7.58866, 7.58…
#> $ ibase2      <dbl> 7.58866, 7.58866, 7.58866, 7.58866, 7.58866, 7.58866, 7.58…
#> $ ibase3      <dbl> 7.58866, 7.58866, 7.58866, 7.58866, 7.58866, 7.58866, 7.58…
#> $ ibase4      <dbl> 7.58866, 7.58866, 7.58866, 7.58866, 7.58866, 7.58866, 7.58…
#> $ ibase5      <dbl> 7.58866, 7.58866, 7.58866, 7.58866, 7.58866, 7.58866, 7.58…
#> $ kg          <dbl> 6.237857e-05, 6.237857e-05, 6.237857e-05, 6.237857e-05, 6.…
#> $ lambda      <dbl> 0.0001193625, 0.0001193625, 0.0001193625, 0.0001193625, 0.…
#> $ kdrug       <dbl> 9.879226e-05, 9.879226e-05, 9.879226e-05, 9.879226e-05, 9.…
#> $ base_sld    <dbl> 263.064, 263.064, 263.064, 263.064, 263.064, 263.064, 263.…
#> $ lambh       <dbl> 0.0001132738, 0.0001132738, 0.0001132738, 0.0001132738, 0.…
#> $ lambd       <dbl> 0.0001185952, 0.0001185952, 0.0001185952, 0.0001185952, 0.…
#> $ keo         <dbl> 0.01386294, 0.01386294, 0.01386294, 0.01386294, 0.01386294…
#> $ ipredSim    <dbl> 7.588660, 7.506310, 7.303497, 7.033815, 6.734994, 6.431882…
#> $ sim         <dbl> 7.588660, 7.506310, 7.303497, 7.033815, 6.734994, 6.431882…
#> $ effect      <dbl> 0.00000000, 0.28302236, 0.48594307, 0.63143269, 0.73574548…
#> $ suv1        <dbl> 7.588660, 7.506310, 7.303497, 7.033815, 6.734994, 6.431882…
#> $ suv2        <dbl> 7.588660, 7.506310, 7.303497, 7.033815, 6.734994, 6.431882…
#> $ suv3        <dbl> 7.588660, 7.506310, 7.303497, 7.033815, 6.734994, 6.431882…
#> $ suv4        <dbl> 7.588660, 7.506310, 7.303497, 7.033815, 6.734994, 6.431882…
#> $ suv5        <dbl> 7.588660, 7.506310, 7.303497, 7.033815, 6.734994, 6.431882…
#> $ sld         <dbl> 263.0640, 263.3650, 263.5168, 263.5623, 263.5322, 263.4486…
#> $ cumHaz_os   <dbl> 0.000000000, 0.002718571, 0.005437143, 0.008155714, 0.0108…
#> $ cumHaz_drop <dbl> 0.000000000, 0.002846286, 0.005692571, 0.008538857, 0.0113…
#> $ DOSE        <dbl> 50, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50, 0,…
#> $ CMT         <dbl> 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2…
#> $ CLI         <dbl> 50, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50…
#> $ RCFB1MAX    <dbl> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0…

Compute RCFB1MAX from week-1 SUVmax 

suv_at_t168 <- sim_stage1 |>
  filter(time == 24 * 7) |>      # week 1 = 168 h
  select(id, suv1, suv2, suv3, suv4, suv5)

# Per-subject baseline SUVmax (the IBASE_n values are folded into the
# compartment initial conditions; the t = 0 value is the per-subject baseline).
suv_baseline <- sim_stage1 |>
  filter(time == 0) |>
  select(id, base_suv1 = suv1, base_suv2 = suv2, base_suv3 = suv3,
             base_suv4 = suv4, base_suv5 = suv5)

rcfb1 <- suv_at_t168 |>
  left_join(suv_baseline, by = "id") |>
  mutate(
    rcfb_les1 = (suv1 - base_suv1) / base_suv1,
    rcfb_les2 = (suv2 - base_suv2) / base_suv2,
    rcfb_les3 = (suv3 - base_suv3) / base_suv3,
    rcfb_les4 = (suv4 - base_suv4) / base_suv4,
    rcfb_les5 = (suv5 - base_suv5) / base_suv5
  ) |>
  rowwise() |>
  mutate(RCFB1MAX = max(c(rcfb_les1, rcfb_les2, rcfb_les3, rcfb_les4, rcfb_les5),
                       na.rm = TRUE)) |>
  ungroup() |>
  select(id, RCFB1MAX)

knitr::kable(rcfb1, caption = "Per-subject RCFB1MAX computed from week-1 SUVmax.")
Per-subject RCFB1MAX computed from week-1 SUVmax.
id RCFB1MAX
1 -0.2268256
2 -0.2268256
3 -0.2268256
4 -0.2268256
5 -0.2268256

Stage 2 simulation (full joint model) 

events2 <- events |>
  select(-RCFB1MAX) |>
  left_join(rcfb1, by = "id")

sim_stage2 <- rxode2::rxSolve(
  mod, events = events2,
  keep = c("DOSE", "CLI", "RCFB1MAX"),
  returnType = "data.frame"
) |>
  as_tibble() |>
  distinct(id, time, .keep_all = TRUE)
#> ℹ omega/sigma items treated as zero: 'etalkdrug', 'etaldrug', 'etaldrug_les1', 'etaldrug_les2', 'etaldrug_les3', 'etaldrug_les4', 'etaldrug_les5', 'etalbase_sld', 'etalbase', 'etalbase_les1', 'etalbase_les2', 'etalbase_les3', 'etalbase_les4', 'etalbase_les5', 'etalkg'
#> Warning: multi-subject simulation without without 'omega'

Replicate published-style trajectories

The four panels below reproduce the qualitative shape of the four sub-models (SUVmax, SLD, OS cumulative hazard, dropout cumulative hazard) over the simulation horizon. The Schindler 2016 publication reports population-level VPCs of these endpoints (Figures 2-4); without the original observed dataset this vignette plots the simulated-cohort medians and 5th-95th percentile bands instead.

# Long-format SUVmax trajectory across the 5 lesions for ggplot facets.
sim_suv_long <- sim_stage2 |>
  select(id, time, suv1, suv2, suv3, suv4, suv5) |>
  pivot_longer(cols = starts_with("suv"), names_to = "lesion", values_to = "suvmax") |>
  mutate(lesion = factor(lesion, levels = paste0("suv", 1:5)))

p_suv <- sim_suv_long |>
  group_by(time, lesion) |>
  summarise(med = median(suvmax),
            q05 = quantile(suvmax, 0.05),
            q95 = quantile(suvmax, 0.95),
            .groups = "drop") |>
  ggplot(aes(x = time / 24, y = med)) +
  geom_ribbon(aes(ymin = q05, ymax = q95), alpha = 0.25) +
  geom_line() +
  facet_wrap(~lesion, ncol = 5) +
  labs(x = "Time (days)", y = "Simulated SUVmax",
       title = "Per-lesion SUVmax trajectory",
       caption = "Replicates qualitative shape of Schindler 2016 Figure 2 (per-lesion SUVmax VPC).")

p_sld <- sim_stage2 |>
  group_by(time) |>
  summarise(med = median(sld),
            q05 = quantile(sld, 0.05),
            q95 = quantile(sld, 0.95),
            .groups = "drop") |>
  ggplot(aes(x = time / 24, y = med)) +
  geom_ribbon(aes(ymin = q05, ymax = q95), alpha = 0.25) +
  geom_line() +
  labs(x = "Time (days)", y = "Simulated SLD (mm)",
       title = "Sum-of-longest-diameters trajectory",
       caption = "Replicates qualitative shape of Schindler 2016 Figure 3 (SLD VPC).")

p_haz <- sim_stage2 |>
  select(id, time, cumHaz_os, cumHaz_drop) |>
  pivot_longer(cols = starts_with("cumHaz"), names_to = "endpoint", values_to = "cum_hazard") |>
  group_by(time, endpoint) |>
  summarise(med = median(cum_hazard),
            q05 = quantile(cum_hazard, 0.05),
            q95 = quantile(cum_hazard, 0.95),
            .groups = "drop") |>
  ggplot(aes(x = time / 24, y = med, colour = endpoint, fill = endpoint)) +
  geom_ribbon(aes(ymin = q05, ymax = q95), alpha = 0.20, colour = NA) +
  geom_line() +
  labs(x = "Time (days)", y = "Cumulative hazard",
       title = "OS and dropout cumulative hazards",
       caption = "OS hazard depends on per-subject RCFB1MAX; dropout hazard does not.")

print(p_suv)

print(p_sld)

print(p_haz)

F.2 self-consistency check (typical-value re-simulation of the bundle dataset)

The bundle’s Output_simulated_*.lst reproduces a one-subject MAXEVAL = 0 simulation of the source .mod on Simulated_SLD_SUV_OS_GIST.csv. The check below runs a typical-value (zero-IIV) simulation through the packaged nlmixr2 model on a comparable layout and verifies that the SUVmax / SLD trajectories settle at the .mod’s typical-baseline (IBASE_SUV ~= 7.59 for the SUVmax compartments and IBASE_SLD ~= 263 mm for SLD) at t = 0 and that the post-cycle SLD trajectory falls within an order of magnitude of those baselines – the bundle’s Output_simulated .lst does not embed a parsable $TABLE block for direct numerical comparison, so this check is structural rather than per-time-point.

mod_typical <- rxode2::rxode2(readModelDb("Schindler_2016_sunitinib")) |> rxode2::zeroRe()

events_one <- rxode2::et(time = day_hours, cmt = "suv1") |>
  rxode2::etRbind(rxode2::et(time = day_hours, cmt = "suv2")) |>
  rxode2::etRbind(rxode2::et(time = day_hours, cmt = "suv3")) |>
  rxode2::etRbind(rxode2::et(time = day_hours, cmt = "suv4")) |>
  rxode2::etRbind(rxode2::et(time = day_hours, cmt = "suv5")) |>
  rxode2::etRbind(rxode2::et(time = day_hours, cmt = "sld"))
events_one$id           <- 1L
events_one$DOSE   <- on_off_for(events_one$time)
events_one$CLI <- 50
events_one$RCFB1MAX     <- -0.30   # typical responder, plausible reference

sim_typical <- rxode2::rxSolve(mod_typical, events = events_one,
                               returnType = "data.frame") |>
  as_tibble() |>
  distinct(time, .keep_all = TRUE)
#> ℹ omega/sigma items treated as zero: 'etalkdrug', 'etaldrug', 'etaldrug_les1', 'etaldrug_les2', 'etaldrug_les3', 'etaldrug_les4', 'etaldrug_les5', 'etalbase_sld', 'etalbase', 'etalbase_les1', 'etalbase_les2', 'etalbase_les3', 'etalbase_les4', 'etalbase_les5', 'etalkg'

t0_row <- sim_typical |> filter(time == 0)

# Structural sanity:
#  - At t = 0, all 5 SUVmax compartments equal the typical baseline IBASE_SUV.
#  - At t = 0, SLD equals the typical baseline IBASE_SLD.
#  - Cumulative hazards are non-negative and monotone non-decreasing.
stopifnot(abs(t0_row$suv1 - exp(log(7.58866))) < 1e-6)
stopifnot(abs(t0_row$sld  - 263.064)            < 1e-3)
stopifnot(all(diff(sim_typical$cumHaz_os)   >= -1e-12))
stopifnot(all(diff(sim_typical$cumHaz_drop) >= -1e-12))

structural_summary <- tibble(
  endpoint = c("SUVmax (lesion 1) at t=0",
               "SUVmax (lesion 5) at t=0",
               "SLD at t=0 (mm)",
               "Final cumulative OS hazard",
               "Final cumulative dropout hazard"),
  simulated_typical = c(
    t0_row$suv1, t0_row$suv5, t0_row$sld,
    tail(sim_typical$cumHaz_os, 1),
    tail(sim_typical$cumHaz_drop, 1)
  ),
  expected = c(7.58866, 7.58866, 263.064, NA_real_, NA_real_),
  source = c(
    ".mod IBASE_SUV (THETA(4))",
    ".mod IBASE_SUV (THETA(4))",
    ".mod IBASE_SLD (THETA(10))",
    "structural sanity (monotone non-negative)",
    "structural sanity (monotone non-negative)"
  )
)

knitr::kable(structural_summary,
             caption = "F.2 self-consistency: typical-value baselines vs. .mod THETA values.")
F.2 self-consistency: typical-value baselines vs. .mod THETA values.
endpoint simulated_typical expected source
SUVmax (lesion 1) at t=0 7.5886600 7.58866 .mod IBASE_SUV (THETA(4))
SUVmax (lesion 5) at t=0 7.5886600 7.58866 .mod IBASE_SUV (THETA(4))
SLD at t=0 (mm) 263.0640000 263.06400 .mod IBASE_SLD (THETA(10))
Final cumulative OS hazard 0.0545183 NA structural sanity (monotone non-negative)
Final cumulative dropout hazard 0.2846286 NA structural sanity (monotone non-negative)

Assumptions and deviations

  • Cross-lesion residual-error correlation dropped. The source .mod $ERROR block applies a Cholesky-decomposed 5x5 residual-error structure across the SUVmax outputs, with diagonal variance THETA(5) = 0.173794 and across-lesion correlation THETA(6) = 0.466827. nlmixr2 does not support cross-output residual-error correlation in its standard ~ prop() syntax, so each lesion’s residual error is preserved at the per-lesion marginal scale (propSd = sqrt(0.173794) ~= 0.4169) but the across-lesion correlation is omitted. Per-lesion typical values and IIV are unaffected.
  • RCFB1MAX consumed as an input covariate, not an in-model state-snapshot. In the source NONMEM .mod RCFB1MAX is computed inline at FLAG = 1 / TIME = 168 h via record-loop persistent state (see .mod $ERROR block lines 282-302). rxode2 does not have an idiomatic equivalent, so the model file consumes RCFB1MAX as a per-subject input covariate; the vignette implements the two-stage simulation (Stage 1 -> compute RCFB1MAX -> Stage 2) that emulates the NONMEM behavior.
  • Sunitinib clearance is a required upstream-popPK input. The source .mod $INPUT line declares CL as “post-hoc clearance from previously-developed PK model”. The upstream sunitinib popPK model is not part of the DDMODEL00000221 bundle and is not currently in nlmixr2lib. The model file therefore requires CLI (per-subject CL/F, L/h) as an input covariate; the vignette virtual cohort uses 50 L/h as a literature-typical adult sunitinib clearance, consistent with Houk et al. (2010) J Clin Pharmacol 50:843-858, but does not reproduce Houk’s popPK structure inline.
  • MINIMIZATION SUCCESSFUL carried a NONMEM caveat. Output_real_*.lst line 795 reports MINIMIZATION SUCCESSFUL immediately followed by HOWEVER, PROBLEMS OCCURRED WITH THE MINIMIZATION. REGARD THE RESULTS OF THE ESTIMATION STEP CAREFULLY .... Per the extract-literature-model ddmore-source.md Section “Reading final estimates from .lst”, MINIMIZATION SUCCESSFUL (with caveat) is acceptable but flagged here.
  • Linked publication not on disk. The Schindler 2016 publication is not on disk under mab_human_consensus/literature/; an external check of parameter values against the publication’s tables was not performed. Final estimates were taken from the .lst only.
  • Iteration-log precision used for estimated THETAs. THETA(12), THETA(14), and THETA(16) (the only THETAs that moved during minimization) are reported in the FINAL PARAMETER ESTIMATE block at 3-significant-digit precision (1.90E-02, 1.99E-02, 5.36E+00); the model file uses the higher-precision values from iteration 8 of the iteration-log NPARAMETR row (1.9030E-02, 1.9924E-02, 5.3558E+00).
  • Bundle simulated dataset is a 1-subject smoke test, not a representative cohort. Simulated_SLD_SUV_OS_GIST.csv ships 315 records for a single virtual subject on a single dosing schedule, which is consistent with NONMEM regression-test usage but is not reflective of the 66-patient pooled GIST cohort that the model was built on. The vignette’s virtual cohort is a deliberately small (n = 5) extension of the bundle’s per-subject schedule to make the F.2 self-consistency and qualitative-VPC checks tractable; reproducing the publication’s per-figure quantitative numbers would require the original observed dataset, which is not in the bundle.