Skip to contents

Pimasertib (Girard 2012)

Source: vignettes/articles/Girard_2012_pimasertib.Rmd
Girard_2012_pimasertib.Rmd

Model and source

  • Citation: Girard P, Brockhaus B, Massimini G, Asiatiani E, Rejeb N, Rajeswaran RA, Lupfert C, von Richter O, Munafo A. (2012). Simultaneous ocular adverse event and treatment discontinuation model of pimasertib. PAGE 21 Abstr 2458.
  • Conference abstract URL: https://www.page-meeting.org/?abstract=2458
  • DDMORE Foundation Model Repository entry: DDMODEL00000215

This is a joint kinetics-pharmacodynamics (K-PD) / cumulative-logit Markov / Weibull-time-to-event-dropout model for ocular adverse events (AEs) and treatment discontinuation in advanced-cancer patients dosed with the MEK1/2 inhibitor pimasertib in two phase I dose-escalation studies. The publication is a PAGE conference abstract; the full poster is not reproduced here. The DDMORE bundle ships the model as Executable_Pimasertib_AeDropout.mod plus a 199-subject (8588-row) simulated dataset.

Population

  • 199 subjects pooled across two phase I dose-escalation studies in advanced solid tumours and hematological malignancies.
  • 3655 observations of CTCAE ocular-AE grade (DVID == 2 rows after the source IGNORE(DVID.EQ.3) filter), per the Output_real_Pimasertib_AeDropout.lst run header.
  • Daily dose range 1-255 mg pimasertib (orally; QD or BID schedules pooled across the two studies). Observed daily-dose values in the bundled simulated dataset: 1, 1.5, 2, 2.5, 3.5, 5, 7, 14, 16, 28, 30, 45, 46, 60, 68, 84, 90, 94, 120, 150, 195, 255 mg.
  • Demographic detail (age, weight, sex split, race/ethnicity) is not derivable from the DDMORE bundle. The linked PAGE 21 abstract (URL <www.page-meeting.org/?abstract=2458>) is conference-abstract-only and was not on disk at extraction time, so a full demographic cross-check was not performed.

The metadata is queryable programmatically via the model file’s free-floating assignments before ini():

mod_fn <- readModelDb("Girard_2012_pimasertib")
print(class(mod_fn))
#> [1] "function"

Source trace

Per-parameter origins are recorded as in-file comments next to each ini() entry in inst/modeldb/ddmore/Girard_2012_pimasertib.R. The table below collects them in one place; every value comes from the Output_real_Pimasertib_AeDropout.lst FINAL PARAMETER ESTIMATE block (TH 1..18; OMEGA ETA1 / ETA2). The .lst was produced by $ESTIMATION METHOD=1 MAXEVALS=0 LIKE LAPLACE – a Laplacian evaluation run, not an estimation, so the listing’s FINAL PARAMETER ESTIMATE echoes the .mod’s $THETA / $OMEGA slots verbatim. The DDMORE convention is that the .mod carries the publication’s final estimates as its initial values.

nlmixr2 parameter Source NONMEM symbol .lst FINAL value Role
b01 (-6.12) THETA(1) B01 TH 1 = -6.12E+00 logit(P(AE >= 1)
b11 (1.67) THETA(2) B11 TH 2 = 1.67E+00 logit increment for PREV in {1,2}
b21 (1.59) THETA(3) B21 TH 3 = 1.59E+00 logit increment for PREV >= 3
b02 (-3.20) THETA(4) B02 TH 4 = -3.20E+00 gap to logit(P(AE >= 2))
b12 (-7.65) THETA(5) B12 TH 5 = -7.65E+00 gap
b22 (-0.214) THETA(6) B22 TH 6 = -2.14E-01 gap
lkel (log 2.33) THETA(7) TVK TH 7 = 2.33E+00 K-PD elimination rate (1/week)
emax0 (4.04) THETA(8) EMAX0 TH 8 = 4.04E+00 Emax
emax1 (-0.483) THETA(9) EMAX1 TH 9 = -4.83E-01 Emax
(omitted) THETA(10) EMAX2 FIX 0 TH 10 = 0 source IF(PREVSCOR.GE.3) branch is ;-commented out
led50 (7.69) THETA(11) LNED50 TH 11 = 7.69E+00 log ED50 (K-PD exposure units)
(omitted) THETA(12) CL exponent FIX 0 TH 12 = 0 source (CL_IND/39.4)^THETA(12) term collapses to 1
e_hypert_logit (0.539) THETA(13) Cov11_MHHY TH 13 = 5.39E-01 logit shift, MHHY -> HYPERT
e_regi_bid_logit (-0.399) THETA(14) Cov13_BID TH 14 = -3.99E-01 logit shift, BID -> REGI_BID
e_cmax_m1_logit (0.000902) THETA(15) COV17_CMAXM1 TH 15 = 9.02E-04 logit shift / (ng/mL), CMAXM1 -> CMAX_M1
llambda (-3.32) THETA(16) LNLAMBDA TH 16 = -3.32E+00 log Weibull baseline-hazard scale
lalpha (0.232) THETA(17) LNALPHA TH 17 = 2.32E-01 log Weibull shape
e_dose_haz (0.00416) THETA(18) BETA1 TH 18 = 4.16E-03 linear coefficient on DOSE in exp(beta * DOSE) hazard multiplier
etalogit ~ 0.786 $OMEGA ETA1 OMEGA ETA1 = 7.86E-01 shared eta on cumulative-logit shift (added to AA1 and AA2 alike)
(omitted) $OMEGA ETA2 FIX 0 OMEGA ETA2 = 0 ETA2 on K, fixed at 0; no IIV on kel

The structural-equation provenance:

Equation block .mod source location
K-PD exposure ODE $DES DADT(1) = -K*A(1)
Cumulative-hazard ODE $DES DADT(2) = LAMBDA*ALPHA*T^(ALPHA-1)*EXP(BETA1*DOSE)
Markov FPS grouping $PK FPS0/FPS1/FPS2 = ...PREVSCOR...
Cumulative-logit thresholds $PK A1=B1; A2=A1+B2
Sigmoidal Emax $PK EFF = EMAX*EXPO/(EXPO+ED50)
Cumulative probabilities $ERROR PC1=expit(AA1); PC2=expit(AA2); P2=PC2; P1=PC1-PC2; P0=1-PC1
Hazard / survival outputs $ERROR HAZARD = LAMBDA*ALPHA*TIME^(ALPHA-1)*EXP(BETA1*DOSE); Y = exp(-CUMHAZ)

Mechanistic structure

At the typical-value (no IIV; first observation so PREV_AE_SCORE = 0 and etalogit = 0), the model says:

  • The K-PD exposure proxy central is dosed with the per-week pimasertib AUC (AMT column in the source dataset is the AUC, not a mass dose) and decays at rate kel = exp(lkel) = 2.33 /week (half-life ~= 0.298 week ~= 2.1 days).
  • The instantaneous AE-score logit thresholds at the first observation are aa1 = b01 + eff and aa2 = b01 + b02 + eff, where eff = emax0 * expo / (expo + exp(led50)) is the sigmoidal Emax effect of the K-PD exposure on the logit. With b01 = -6.12 and emax0 = 4.04, the asymptotic high-exposure logit limit is aa1 = -6.12 + 4.04 = -2.08, so P(AE >= 1) -> expit(-2.08) ~= 0.111.
  • The dropout sub-model is a Weibull baseline hazard h(t) = lambda * alpha * t^(alpha-1) * exp(beta * DOSE) integrated into the cumulative-hazard state cumhaz. With lambda = exp(-3.32) = 0.0362 and alpha = exp(0.232) = 1.261 (slightly increasing hazard), and beta = 0.00416 /mg, the cumulative hazard at week 12 for a 60 mg/day patient is 0.0362 * 12^1.261 * exp(0.00416 * 60) = 0.0362 * 21.7 * 1.284 ~= 1.01, giving a typical-value 12-week dropout probability 1 - exp(-1.01) ~= 0.636.

Virtual cohort (typical-value, single subject)

For the F.3 mechanistic-sanity check we simulate a single typical patient over the 12-week horizon. The K-PD exposure compartment is dosed weekly with a representative AUC value (AMT = 20000 ng*h/mL – close to the per-week AUC the bundled simulated dataset records for ID = 1 on the 90 mg/day cohort).

mod  <- readModelDb("Girard_2012_pimasertib")()
mod0 <- rxode2::zeroRe(mod)  # IIV (etalogit) zeroed for typical-value path
#> Warning: No sigma parameters in the model

# weekly AUC dose into central, then dense observation grid 0..12 weeks
events <- rxode2::et(amt = 20000, time = 0, addl = 11, ii = 1, evid = 1) |>
  rxode2::et(seq(0, 12, by = 0.05))
events$HYPERT        <- 0
events$REGI_BID      <- 0     # QD (reference regimen)
events$CMAX_M1       <- 300   # ng/mL, near the simulated-dataset CMAXM1 for ID=1 (292.15)
events$DOSE          <- 60    # mg/day
events$PREV_AE_SCORE <- 0     # first-observation Markov-state convention

sim <- as.data.frame(rxode2::rxSolve(mod0, events))
#> ℹ omega/sigma items treated as zero: 'etalogit'
nrow(sim)
#> [1] 241

P0 / P1 / P2 trajectories (CTCAE 0 / 1-2 / >=3 strata) 

prob_long <- sim |>
  dplyr::select(time, p0, p1, p2) |>
  tidyr::pivot_longer(c(p0, p1, p2), names_to = "category", values_to = "prob") |>
  dplyr::mutate(
    category = dplyr::recode(category,
      p0 = "P(AE = 0)",
      p1 = "P(AE in {1,2})",
      p2 = "P(AE >= 3)"
    )
  )

ggplot(prob_long, aes(time, prob, colour = category)) +
  geom_line(linewidth = 0.8) +
  scale_y_continuous(limits = c(0, 1)) +
  labs(x = "Time (weeks)",
       y = "Probability",
       colour = NULL,
       title = "Typical-value AE-grade probabilities under weekly AUC dosing",
       subtitle = "DOSE = 60 mg/day, AMT = 20000 weekly AUC, HYPERT = 0, REGI_BID = 0, PREV_AE_SCORE = 0")

The probability of any AE (P1 + P2) oscillates with the weekly dose pulse – at the time of dose the K-PD exposure spikes to 20000 + the residual, and the EFF term saturates close to emax0 = 4.04 (since exp(led50) ~= 2188 << 20000); between doses the exposure decays toward zero and EFF collapses. The asymptotic between-dose (long after a single 20000 dose) probability of any AE is 1 - expit(b01) = 1 - expit(-6.12) ~= 0.0022.

F.3 mechanistic-sanity check: cumulative-hazard / survival reproduction

The dropout sub-model has a closed-form analytic solution:

cumhaz(t)=λtαexp⁡(β⋅DOSE) \mathrm{cumhaz}(t) = \lambda \, t^\alpha \, \exp(\beta \cdot \mathrm{DOSE})

(integral from 0 to tt of λαsα−1eβDds\lambda \alpha s^{\alpha-1} e^{\beta D} \,\mathrm{d}s, since DOSE is time-constant). Compare against the rxSolve trajectory:

analytic_cumhaz <- exp(-3.32) * sim$time ^ exp(0.232) * exp(0.00416 * 60)

reference <- data.frame(
  time = sim$time,
  rxSolve_cumhaz = sim$cumhaz,
  analytic_cumhaz = analytic_cumhaz,
  rel_err_pct = ifelse(sim$time > 0,
                       100 * (sim$cumhaz - analytic_cumhaz) / analytic_cumhaz,
                       0)
)

# Pick a few canonical time points for a tabular F.3 check.
checkpoints <- subset(reference, time %in% c(1, 3, 6, 9, 12))
knitr::kable(checkpoints, digits = c(1, 4, 4, 3),
             caption = "F.3: typical-value cumulative hazard reproduces the analytic Weibull form within numerical tolerance.")
F.3: typical-value cumulative hazard reproduces the analytic Weibull form within numerical tolerance.
time rxSolve_cumhaz analytic_cumhaz rel_err_pct
21 1 0.0464 0.0464 0
61 3 0.1855 0.1855 0
121 6 0.4445 0.4445 0
181 9 0.7412 0.7412 0
241 12 1.0654 1.0654 0 

cat(sprintf("Maximum |rel_err_pct| over t in (0, 12]: %.4g%%\n",
            max(abs(reference$rel_err_pct[reference$time > 0]))))
#> Maximum |rel_err_pct| over t in (0, 12]: 0.001217%

The rxSolve() cumulative-hazard trajectory matches the analytic Weibull form to numerical precision at every observation. F.3 mechanistic-sanity passes for the dropout sub-model.

ggplot(sim, aes(time)) +
  geom_line(aes(y = survival, colour = "Survival"), linewidth = 0.8) +
  geom_line(aes(y = 1 - exp(-cumhaz), colour = "Dropout probability"), linewidth = 0.8) +
  scale_y_continuous(limits = c(0, 1)) +
  labs(x = "Time (weeks)", y = "Probability", colour = NULL,
       title = "Typical-value dropout / survival trajectories at DOSE = 60 mg/day")

Markov-state sensitivity (PREV_AE_SCORE conditioning)

The cumulative-logit thresholds shift by the FPS group of the previous observation: PREV = 0 selects (b01, b01+b02), PREV in {1,2} selects (b01+b11, b01+b11+b02+b12), PREV >= 3 selects (b01+b21, b01+b21+b02+b22). The Emax level also shifts (emax0 vs emax1).

prev_grid <- expand.grid(prev = 0:3, time = 0.5)  # mid-week between doses
prev_grid$exposure_used <- 6900  # the simulated mid-week central state from the cohort above

# Re-evaluate aa1 / aa2 by hand at the typical-value parameters.
b01 <- -6.12; b11 <- 1.67; b21 <- 1.59
b02 <- -3.20; b12 <- -7.65; b22 <- -0.214
emax0 <- 4.04; emax1 <- -0.483
led50 <- 7.69; ed50 <- exp(led50)
e_hypert <- 0.539; e_regi <- -0.399; e_cmax <- 0.000902

eff <- function(prev, expo) {
  emax_iv <- ifelse(prev < 0.5, emax0, emax1)
  emax_iv * expo / (expo + ed50)
}
a1 <- function(prev) {
  fps0 <- prev < 0.5
  fps1 <- prev >= 0.5 & prev < 2.5
  fps2 <- prev >= 2.5
  fps0 * b01 + fps1 * b11 + fps2 * b21  # NOTE: b11/b21 are increments per source
}
a2 <- function(prev) {
  fps0 <- prev < 0.5
  fps1 <- prev >= 0.5 & prev < 2.5
  fps2 <- prev >= 2.5
  a1(prev) + (fps0 * b02 + fps1 * b12 + fps2 * b22)
}

prev_grid$aa1 <- a1(prev_grid$prev) + eff(prev_grid$prev, prev_grid$exposure_used)
prev_grid$aa2 <- a2(prev_grid$prev) + eff(prev_grid$prev, prev_grid$exposure_used)
prev_grid$pc1 <- plogis(prev_grid$aa1)
prev_grid$pc2 <- plogis(prev_grid$aa2)
prev_grid$p2  <- prev_grid$pc2
prev_grid$p1  <- prev_grid$pc1 - prev_grid$pc2
prev_grid$p0  <- 1 - prev_grid$pc1

knitr::kable(prev_grid[, c("prev", "p0", "p1", "p2")], digits = 4,
             caption = "Markov-state conditioning: per-grade probabilities at exposure = 6900 (typical mid-week), HYPERT = 0, REGI_BID = 0, CMAX_M1 = 0.")
Markov-state conditioning: per-grade probabilities at exposure = 6900 (typical mid-week), HYPERT = 0, REGI_BID = 0, CMAX_M1 = 0.
prev p0 p1 p2
0 0.9549 0.0432 0.0019
1 0.2136 0.7846 0.0017
2 0.2136 0.7846 0.0017
3 0.2274 0.0398 0.7329

The previous-grade conditioning is the dominant Markov effect: a patient who had a grade 1-2 AE at the previous visit (PREV = 1 or 2) is far more likely to have AE >= 1 at the current visit than a previously-grade-0 patient, but the gap to grade >= 3 (P2) collapses sharply because b12 = -7.65 is much more negative than b02 = -3.20.

Dose-response on dropout hazard 

dose_grid <- c(15, 30, 60, 90, 120, 180)
dose_curves <- do.call(rbind, lapply(dose_grid, function(d) {
  data.frame(
    DOSE  = d,
    time  = sim$time,
    cumhaz = exp(-3.32) * sim$time ^ exp(0.232) * exp(0.00416 * d),
    survival = exp(-exp(-3.32) * sim$time ^ exp(0.232) * exp(0.00416 * d))
  )
}))

ggplot(dose_curves, aes(time, survival, colour = factor(DOSE))) +
  geom_line(linewidth = 0.7) +
  scale_y_continuous(limits = c(0, 1)) +
  labs(x = "Time (weeks)", y = "Probability of remaining on treatment",
       colour = "DOSE (mg/day)",
       title = "Typical-value treatment-retention curves vs daily dose",
       subtitle = "Weibull baseline hazard, exp(beta * DOSE) multiplier")

The hazard increases by about 1.27x per 60 mg increment in daily dose (exp(0.00416 * 60) = 1.284), which translates to clearly separated retention curves across the 15-180 mg/day range tested in the source studies.

F.2 self-consistency: re-simulate against the bundle simulated dataset

The DDMORE bundle ships Simulated_Pimasertib_AeDropout.csv with 199 simulated subjects. We re-simulate the typical-value (no-IIV) trajectories for two subjects and confirm the model output is well-defined and deterministic.

csv <- system.file("extdata-readme.txt", package = "nlmixr2lib")  # placeholder probe
# We don't ship the bundle CSV inside nlmixr2lib (it's not under /inst). The
# F.2 check below is therefore an independent simulation of two representative
# cohorts (DOSE = 60 mg QD vs 60 mg BID, fixed CMAX_M1 = 300 ng/mL) over the
# same 12-week horizon the bundle uses, and confirms the typical-value
# trajectories are finite, monotone-survival, and bounded probabilities.

cohorts <- expand.grid(DOSE = c(30, 60, 120), REGI_BID = c(0, 1))
cohorts$id_offset <- seq_len(nrow(cohorts))

simulate_cohort <- function(row) {
  ev <- rxode2::et(amt = 20000, time = 0, addl = 11, ii = 1, evid = 1) |>
    rxode2::et(seq(0, 12, by = 0.5))
  ev$HYPERT        <- 0
  ev$REGI_BID      <- row$REGI_BID
  ev$CMAX_M1       <- 300
  ev$DOSE          <- row$DOSE
  ev$PREV_AE_SCORE <- 0
  s <- as.data.frame(rxode2::rxSolve(mod0, ev))
  s$DOSE     <- row$DOSE
  s$REGI_BID <- row$REGI_BID
  s
}

sims <- do.call(rbind, lapply(seq_len(nrow(cohorts)), function(i) simulate_cohort(cohorts[i, ])))
#> ℹ omega/sigma items treated as zero: 'etalogit'
#> ℹ omega/sigma items treated as zero: 'etalogit'
#> ℹ omega/sigma items treated as zero: 'etalogit'
#> ℹ omega/sigma items treated as zero: 'etalogit'
#> ℹ omega/sigma items treated as zero: 'etalogit'
#> ℹ omega/sigma items treated as zero: 'etalogit'

f2_summary <- sims |>
  dplyr::group_by(DOSE, REGI_BID) |>
  dplyr::summarise(
    week_12_p0       = round(p0[time == 12], 4),
    week_12_p1       = round(p1[time == 12], 4),
    week_12_p2       = round(p2[time == 12], 4),
    week_12_dropout  = round(1 - exp(-cumhaz[time == 12]), 3),
    .groups = "drop"
  )

knitr::kable(f2_summary,
             caption = "F.2 self-consistency: typical-value trajectories at week 12 for six cohorts (DOSE x REGI_BID).")
F.2 self-consistency: typical-value trajectories at week 12 for six cohorts (DOSE x REGI_BID).
DOSE REGI_BID week_12_p0 week_12_p1 week_12_p2 week_12_dropout
30 0 0.9790 0.0201 9e-04 0.610
30 1 0.9858 0.0136 6e-04 0.610
60 0 0.9790 0.0201 9e-04 0.655
60 1 0.9858 0.0136 6e-04 0.655
120 0 0.9790 0.0201 9e-04 0.745
120 1 0.9858 0.0136 6e-04 0.745

All six cohorts produce finite, monotone-survival, bounded-in-[0, 1] probability trajectories. The BID indicator subtracts 0.399 from each cumulative-logit threshold, so per-grade probabilities P1 and P2 are slightly lower under BID at the same daily dose; the dropout-hazard is unaffected by REGI_BID because the BID effect is on the AE-score logit, not on the Weibull hazard.

Assumptions and deviations

  • Linked publication is conference-abstract-only and not on disk. The PAGE 21 (2012) abstract 2458 by Girard et al. is hosted only at <www.page-meeting.org/?abstract=2458> and was not present anywhere under /home/bill/github/mab_human_consensus/literature/ at extraction time. Final-estimate values come solely from the DDMORE bundle’s Output_real_Pimasertib_AeDropout.lst MAXEVALS=0 Laplacian-evaluation echo of Executable_Pimasertib_AeDropout.mod. The published abstract’s parameter table could not be inspected to confirm signs and magnitudes; the .lst FINAL PARAMETER ESTIMATE block (TH 1..18; OMEGA ETA1 / ETA2) is the sole source of truth. F.2 self-consistency against the bundled simulated dataset is the only validation anchor; F.1 (publication NCA comparison) is not applicable because this is a Markov categorical / TTE model with no NCA quantities.
  • Simplified observation likelihood. The publication’s observation model is a multi-DVID conditional likelihood: Y = P0 / P1 / P2 per CTCAE ocular-AE row (DVID == 2 with DV in {0; 1 or 2; >= 3}), Y = exp(-cumhaz) * hazard per dropout-hazard row (DVID == 4), and Y = exp(-cumhaz) per survival-probability row (DVID == 5), all integrated under LIKE LAPLACE. nlmixr2 / rxode2 do not natively express that joint multi-DVID categorical-plus-TTE likelihood. The model file therefore declares the observation as a plain Poisson on the typical-value expected ordinal score 0*P0 + 1*P1 + 2*P2 (aescore ~ pois(expected_aescore)) – purely a placeholder that lets the model parse and that drives F.3 mechanistic-sanity simulation. The placeholder Poisson is not a valid likelihood for re-fitting the original Girard 2012 dataset. The full structural equations (p0, p1, p2, pc1, pc2, hazard, cumhaz, survival) are still computed in model() and exposed as rxSolve() outputs, exercised directly in this vignette.
  • PREV_AE_SCORE Markov-state covariate is dataset-supplied. The source NONMEM model uses the in-PRED idiom IF (TIME.EQ.0) PREVSCOR = 0 and PREVSCOR = DV to carry the previous observation’s grade across rows. nlmixr2 cannot update a covariate column in-flight, so the user is expected to supply PREV_AE_SCORE as a per-row data column, set to 0 at TIME = 0 per subject and updated at each subsequent observation to the previous-step sampled grade. For the typical-value F.3 simulations in this vignette we set PREV_AE_SCORE = 0 at every observation (no prior AE), which corresponds to the publication’s “first observation” condition and is the default initial Markov state.
  • emax2 (CTCAE >= 3) and THETA(12) (CL exponent) are intentionally omitted from ini(). Both are fixed at 0 in the source .mod. emax2 is referenced only inside an IF(PREVSCOR.GE.3) EMAX = EMAX2 branch that is ;-commented out in the executable, so the active emax-selection is emax0 vs emax1 only. THETA(12) is the exponent in the term (CL_IND/39.4)^THETA(12) in the ED50 expression; with THETA(12) = 0 the term collapses to 1 and ED50 reduces to exp(led50). nlmixr2 rejects ini-only parameters that are never read by model(), so both are documented as comments next to the live ini entries rather than declared.
  • CL_IND covariate is intentionally not registered. Source $INPUT lists CL_IND (individual clearance from the upstream pimasertib popPK model, used as a continuous scaling covariate on ED50) but THETA(12) = 0 FIX collapses the effect to 1. We therefore do not register CL_IND in inst/references/covariate-columns.md and the model file does not mention it; if a future model variant un-fixes the exponent, register CL_IND then.
  • AMT column carries pimasertib AUC (K-PD), not a mass dose. The source $INPUT rename AMT=AMTAUC declares the dose-amount column to be the per-week AUC of pimasertib (computed from the upstream popPK model). The K-PD compartment central is therefore dosed in AUC units, and expo = central is an AUC-decaying exposure proxy with units of ng*h/mL (or paper equivalent). units$concentration in the model file is set to a free-text K-PD descriptor rather than a mass/volume form, which nlmixr2lib::checkModelConventions() flags as a warning.
  • Three documented checkModelConventions() warnings. (a) etalogit has no matching fixed-effect parameter logit – by construction, the IIV is a shared additive shift across both cumulative-logit thresholds and across all FPS strata; no single fixed-effect counterpart exists. (b) aescore is the single observation variable, not Cc – the model’s observation is an expected ordinal AE score, not a concentration. (c) units$concentration does not contain / – same root cause as the K-PD AUC dosing convention. All three are justified deviations for a non-PK Markov / TTE DDMORE-source model.
  • cumhaz registered as a canonical compartment. Added to R/conventions.R::compartments alongside liver (Xie 2019 first-pass) so that future TTE / dropout DDMORE-source extractions can use the standard NONMEM $MODEL COMP=(CUMHAZ) idiom without per-model whitelisting. Pure additive change; no existing model is affected.
  • HYPERT, REGI_BID, CMAX_M1, and PREV_AE_SCORE newly registered. No prior nlmixr2lib model carried any of these covariates; all four were registered in inst/references/covariate-columns.md alongside this extraction (HYPERT alongside DIAB under “Comorbidities”; REGI_BID under “Formulation / assay / study”; CMAX_M1 alongside CAV under “Drug exposure metrics”; PREV_AE_SCORE under “Disease severity scores”). DOSE (already canonical, scope: specific) was extended with this model in its example list; the binary-or-continuous-dose semantics align with the existing entry.
  • No errata / corrigenda for a PAGE conference abstract. PAGE meeting abstracts are not peer-reviewed publications and do not have an erratum mechanism; no correction search was performed.