Skip to contents

Glasdegib overall survival in AML (Lin 2020)

Source: vignettes/articles/Lin_2020_glasdegib_AML_overall_survival.Rmd
Lin_2020_glasdegib_AML_overall_survival.Rmd

Model and source

  • Citation: Lin S, Shaik N, Chan G, Cortes JE, Ruiz-Garcia A. An evaluation of overall survival in patients with newly diagnosed acute myeloid leukemia and the relationship with glasdegib treatment and exposure. Cancer Chemother Pharmacol. 2020;86(4):451-459.
  • Article: https://doi.org/10.1007/s00280-020-04132-x
  • Trial: BRIGHT AML 1003, ClinicalTrials.gov NCT01546038

This vignette packages and validates three parametric exponential time-to-event (TTE) overall-survival (OS) models from Lin 2020 of the BRIGHT AML 1003 Phase 1b/2 trial in older adults with newly diagnosed AML who were ineligible for intensive chemotherapy:

  • Lin_2020_glasdegib_treatment – Phase 2 treatment-response analysis (n = 116; glasdegib + LDAC vs LDAC alone). Treatment arm is the only retained covariate.
  • Lin_2020_glasdegib_exposure – Phase 2 exposure-response analysis (n = 75; glasdegib + LDAC subset with PK). Intercept-only exponential; no glasdegib exposure metric and no covariate was significant.
  • Lin_2020_glasdegib_decitabine – exploratory pooled Phase 1b + Phase 2 analysis (n = 162) including the small glasdegib + decitabine cohort (n = 7) alongside glasdegib + LDAC (n = 117) and LDAC alone (n = 38).

All three were fit in NONMEM 7.3.0 with the likelihood-based parametric TTE formulation (no estimated IIV, no residual error). For each, the hazard is constant in time (exponential distribution) and modulated multiplicatively by treatment arm where applicable. nlmixr2lib expresses these via the canonical TTE pattern d/dt(cumhaz) = hazard; sur = exp(-cumhaz).

Population

Phase 2 of BRIGHT AML 1003 randomised 116 older adults (>= 55 years) with newly diagnosed AML who were ineligible for intensive chemotherapy: 78 to glasdegib 100 mg orally once daily plus low-dose cytarabine (LDAC; 20 mg SC BID for 10 days per 28-day cycle), and 38 to LDAC alone. Lin 2020 Table 1 reports a median age of 76 years (range 58-92), median weight 78.2 kg (47.5-118.0), 71% male, 97% White (with 1% Black and 2% Asian); median baseline creatinine clearance (Cockcroft-Gault) of 62.7 mL/min, suggesting mild renal impairment in most patients; 52% had secondary AML and 40% had poor cytogenetic risk. Phase 1b added 39 patients to Arm A (glasdegib + LDAC) and 7 patients to Arm B (glasdegib + decitabine; 5 AML + 2 MDS), giving the n = 162 exploratory pooled cohort.

Source trace

The full per-parameter origin is recorded as in-file comments next to each ini() entry in inst/modeldb/therapeuticArea/oncology/Lin_2020_glasdegib_*.R. The tables below collect them.

Lin_2020_glasdegib_treatment (Phase 2 treatment-response, n = 116)

Equation / parameter Value Source location
S(t) = exp(-lambda * (1 + theta_ldac_alone * LDAC_alone) * t) n/a Lin 2020 Results page 4 typeset equation; Fig. 1b annotation
llam_haz (lambda for glasdegib + LDAC) log(0.00253) (1/day) Lin 2020 page 4: 0.00253, RSE 13.82%
e_ldac_alone (theta_ldac_alone) 1.376 Lin 2020 page 4: 1.376, RSE 37.74% (138% increase in base hazard for LDAC alone)
Hazard ratio glasdegib + LDAC vs LDAC alone 0.421 (computed) Lin 2020 abstract / page 4: published HR 0.42 (95% CI 0.28-0.66)
Estimated IIV on lambda none Source NONMEM TTE fit with no subject-level random effects
Residual error none Source uses NONMEM LIKE / F_FLAG parametric-TTE branch

Lin_2020_glasdegib_exposure (Phase 2 exposure-response subset, n = 75)

Equation / parameter Value Source location
S(t) = exp(-lambda * t) n/a Lin 2020 Results page 5; Fig. 3 annotation
llam_haz (lambda) log(0.00246) (1/day) Lin 2020 page 5: 0.00246, RSE 14.19%
Covariates retained after SCM backward elimination none Lin 2020 page 5: seven glasdegib exposure metrics (raw and log-transformed) plus ECOG and cytogenetic risk were tested; none survived

Lin_2020_glasdegib_decitabine (exploratory pooled Phase 1b + Phase 2, n = 162)

Equation / parameter Value Source location
S(t) = exp(-lambda * (1 - theta_gl_ldac * GL_LDAC_arm) * (1 - theta_gl_dec * GL_DEC_arm) * t) n/a Lin 2020 Results page 6 typeset equation
llam_haz (lambda for LDAC alone) log(0.00540) (1/day) Lin 2020 page 6: 0.00540, RSE 14.1%
e_gl_ldac_haz (theta_gl_ldac) 0.480 Lin 2020 page 6 equation: 0.480 (= 48.0% hazard reduction for glasdegib + LDAC)
e_gl_dec_haz (theta_gl_dec) 0.618 Lin 2020 page 6 prose / equation: ~61.8% hazard reduction for glasdegib + decitabine (95% CI of change in base hazard -95.0% to -28.6%)

Virtual cohort

The published trial-level individual data are not redistributed. The virtual cohort below assigns each subject to one of the three trial arms in proportions matching the BRIGHT AML 1003 pooled cohort, and observes the survival probability sur on a daily grid out to 40 months (1216 days). The exponential-hazard structure produces a deterministic typical-value S(t) per arm; per the Lin 2020 source NONMEM fit, no subject-level IIV is added, so all subjects in an arm follow the same S(t) trajectory.

Per the vignette template’s cohort-size cap, each arm uses <= 200 subjects.

set.seed(20260624)

# One observation row per subject per visit day. Observation grid: every 14
# days out to 1216 days (~40 months), matching the Lin 2020 Fig. 1 / Fig. 4
# x-axis range; a time = 0 anchor row is included.
obs_times <- c(0, seq(14, 1216, by = 14))

make_arm <- function(arm_label, n, id_offset, CONMED_GLASDEGIB, CONMED_DECITABINE) {
  ids <- id_offset + seq_len(n)
  tidyr::crossing(
    id = ids,
    time = obs_times
  ) |>
    dplyr::mutate(
      evid  = 0L,
      amt   = 0,
      cmt   = NA_character_,
      arm   = arm_label,
      CONMED_GLASDEGIB  = CONMED_GLASDEGIB,
      CONMED_DECITABINE = CONMED_DECITABINE
    ) |>
    dplyr::select(id, time, evid, amt, cmt, arm,
                  CONMED_GLASDEGIB, CONMED_DECITABINE)
}

events <- dplyr::bind_rows(
  make_arm("LDAC alone",            n = 38,  id_offset =   0L,
           CONMED_GLASDEGIB = 0, CONMED_DECITABINE = 0),
  make_arm("Glasdegib + LDAC",      n = 117, id_offset = 100L,
           CONMED_GLASDEGIB = 1, CONMED_DECITABINE = 0),
  make_arm("Glasdegib + decitabine", n = 7,   id_offset = 300L,
           CONMED_GLASDEGIB = 1, CONMED_DECITABINE = 1)
)

stopifnot(!anyDuplicated(unique(events[, c("id", "time", "evid")])))
cat("Arms:\n"); print(events |> dplyr::distinct(id, arm) |> dplyr::count(arm))
#> Arms:
#> # A tibble: 3 × 2
#>   arm                        n
#>   <chr>                  <int>
#> 1 Glasdegib + LDAC         117
#> 2 Glasdegib + decitabine     7
#> 3 LDAC alone                38

Treatment-response model (Lin_2020_glasdegib_treatment, n = 116)

The treatment-response model uses only CONMED_GLASDEGIB (1 = glasdegib + LDAC arm, 0 = LDAC alone). Subjects in the glasdegib + decitabine arm of the exploratory cohort are not part of this analysis, so the simulation below is restricted to the LDAC alone and glasdegib + LDAC arms.

mod_trt <- readModelDb("Lin_2020_glasdegib_treatment")

events_trt <- events |> dplyr::filter(arm != "Glasdegib + decitabine")

sim_trt <- rxode2::rxSolve(mod_trt, events = events_trt, keep = c("arm")) |>
  as.data.frame()

# All subjects within an arm share the same typical-value S(t); pick one.
typ_trt <- sim_trt |>
  dplyr::group_by(arm, time) |>
  dplyr::summarise(sur = mean(sur), hazard = mean(hazard), .groups = "drop")
# Replicates Fig. 1b of Lin 2020: predicted S(t) for the final treatment-
# response model by treatment arm.
month_per_day <- 12 / 365.25

ggplot(typ_trt, aes(time * month_per_day, sur, colour = arm)) +
  geom_line(linewidth = 1) +
  scale_x_continuous(breaks = seq(0, 40, by = 10), limits = c(0, 40)) +
  scale_y_continuous(limits = c(0, 1), labels = scales::percent_format(accuracy = 1)) +
  labs(x = "Time (months)", y = "Probability of OS",
       title = "Replicates Fig. 1b of Lin 2020 -- final treatment-response model",
       caption = "Glasdegib + LDAC: S(t) = exp(-0.00253 t). LDAC alone: S(t) = exp(-0.00601 t).",
       colour = "Treatment arm")

Validation against the published survival functions and median OS

Both arms have an exponential survival function S(t) = exp(-h * t) with analytical median OS ln(2) / h. The table below compares the implementation against the published hazards and median OS.

trt_check <- tibble::tribble(
  ~arm,                ~h_published_per_day, ~h_simulated_per_day,
  "Glasdegib + LDAC",  0.00253,              typ_trt |> dplyr::filter(arm == "Glasdegib + LDAC", time == 0) |> dplyr::pull(hazard),
  "LDAC alone",        0.00253 * (1 + 1.376),
                                              typ_trt |> dplyr::filter(arm == "LDAC alone", time == 0) |> dplyr::pull(hazard)
) |>
  dplyr::mutate(
    median_OS_pub_months = log(2) / h_published_per_day * month_per_day,
    median_OS_sim_months = log(2) / h_simulated_per_day * month_per_day,
    abs_pct_diff_h       = abs(h_simulated_per_day - h_published_per_day) /
                           h_published_per_day * 100
  )
knitr::kable(trt_check, digits = 5,
             caption = "Treatment-response model: published vs simulated hazards and median OS.")
Treatment-response model: published vs simulated hazards and median OS.
arm h_published_per_day h_simulated_per_day median_OS_pub_months median_OS_sim_months abs_pct_diff_h
Glasdegib + LDAC 0.00253 0.00253 9.00111 9.00111 0
LDAC alone 0.00601 0.00601 3.78835 3.78835 0 

# Hard checks: simulated hazards within 0.1% of published, and median OS
# difference within 5 months of the paper's headline (~5 month gain).
stopifnot(all(trt_check$abs_pct_diff_h < 0.1))
stopifnot(abs(trt_check$median_OS_pub_months[1] - trt_check$median_OS_pub_months[2]) > 4)
stopifnot(abs(trt_check$median_OS_pub_months[1] - trt_check$median_OS_pub_months[2]) < 6)

# Hazard ratio glasdegib + LDAC vs LDAC alone should match the published 0.42.
hr_sim <- trt_check$h_simulated_per_day[1] / trt_check$h_simulated_per_day[2]
cat(sprintf("Hazard ratio glasdegib+LDAC vs LDAC alone: %.3f (paper: 0.42 [0.28-0.66])\n",
            hr_sim))
#> Hazard ratio glasdegib+LDAC vs LDAC alone: 0.421 (paper: 0.42 [0.28-0.66])
stopifnot(abs(hr_sim - 0.42) < 0.01)

Exposure-response model (Lin_2020_glasdegib_exposure, n = 75)

This is an intercept-only exponential model fit to the n = 75 glasdegib + LDAC subset with PK data. The hazard does not depend on any covariate.

mod_exp <- readModelDb("Lin_2020_glasdegib_exposure")

# A single typical subject is sufficient because no covariate enters the
# hazard.
events_exp <- tibble::tibble(
  id   = 1L,
  time = obs_times,
  evid = 0L,
  amt  = 0,
  cmt  = NA_character_
)
sim_exp <- rxode2::rxSolve(mod_exp, events = events_exp) |> as.data.frame()
# Replicates Fig. 3 of Lin 2020: predicted S(t) for the final exposure-
# response model (intercept-only).
ggplot(sim_exp, aes(time * month_per_day, sur)) +
  geom_line(linewidth = 1, colour = "firebrick") +
  scale_x_continuous(breaks = seq(0, 40, by = 10), limits = c(0, 40)) +
  scale_y_continuous(limits = c(0, 1), labels = scales::percent_format(accuracy = 1)) +
  labs(x = "Time (months)", y = "Probability of OS",
       title = "Replicates Fig. 3 of Lin 2020 -- final exposure-response model",
       caption = "S(t) = exp(-0.00246 t).")

h_exp_pub <- 0.00246
h_exp_sim <- sim_exp$hazard[1]
exp_check <- tibble::tibble(
  source              = c("published", "simulated"),
  h_per_day           = c(h_exp_pub, h_exp_sim),
  median_OS_months    = log(2) / c(h_exp_pub, h_exp_sim) * month_per_day
)
knitr::kable(exp_check, digits = 5,
             caption = "Exposure-response model: published vs simulated hazard and median OS.")
Exposure-response model: published vs simulated hazard and median OS.
source h_per_day median_OS_months
published 0.00246 9.25724
simulated 0.00246 9.25724 
stopifnot(abs(h_exp_sim - h_exp_pub) / h_exp_pub * 100 < 0.1)

Exploratory pooled model (Lin_2020_glasdegib_decitabine, n = 162)

The exploratory model includes all three trial arms. The published per-arm median OS is approximately 11.1 months for glasdegib + decitabine, 9.1 months for glasdegib + LDAC (computed from log(2) / (0.00540 * (1 - 0.480))), and 4.2 months for LDAC alone (computed from log(2) / 0.00540).

mod_dec <- readModelDb("Lin_2020_glasdegib_decitabine")

sim_dec <- rxode2::rxSolve(mod_dec, events = events, keep = c("arm")) |>
  as.data.frame()

typ_dec <- sim_dec |>
  dplyr::group_by(arm, time) |>
  dplyr::summarise(sur = mean(sur), hazard = mean(hazard), .groups = "drop")
# Replicates Fig. 4 of Lin 2020: predicted S(t) for the exploratory three-arm
# model.
arm_levels <- c("LDAC alone", "Glasdegib + LDAC", "Glasdegib + decitabine")
ggplot(typ_dec |> dplyr::mutate(arm = factor(arm, levels = arm_levels)),
       aes(time * month_per_day, sur, colour = arm)) +
  geom_line(linewidth = 1) +
  scale_x_continuous(breaks = seq(0, 40, by = 10), limits = c(0, 40)) +
  scale_y_continuous(limits = c(0, 1), labels = scales::percent_format(accuracy = 1)) +
  labs(x = "Time (months)", y = "Probability of OS",
       title = "Replicates Fig. 4 of Lin 2020 -- exploratory three-arm model",
       caption = paste("LDAC alone: S(t) = exp(-0.00540 t).",
                       "Glasdegib + LDAC: S(t) = exp(-0.00540*0.520 t).",
                       "Glasdegib + decitabine: S(t) = exp(-0.00540*0.382 t)."),
       colour = "Treatment arm")

h_ldac     <- 0.00540
h_gl_ldac  <- 0.00540 * (1 - 0.480)
h_gl_dec   <- 0.00540 * (1 - 0.618)

dec_check <- tibble::tribble(
  ~arm,                     ~h_published_per_day, ~h_simulated_per_day,
  "LDAC alone",             h_ldac,
                            typ_dec |> dplyr::filter(arm == "LDAC alone", time == 0) |> dplyr::pull(hazard),
  "Glasdegib + LDAC",       h_gl_ldac,
                            typ_dec |> dplyr::filter(arm == "Glasdegib + LDAC", time == 0) |> dplyr::pull(hazard),
  "Glasdegib + decitabine", h_gl_dec,
                            typ_dec |> dplyr::filter(arm == "Glasdegib + decitabine", time == 0) |> dplyr::pull(hazard)
) |>
  dplyr::mutate(
    median_OS_pub_months = log(2) / h_published_per_day * month_per_day,
    median_OS_sim_months = log(2) / h_simulated_per_day * month_per_day,
    abs_pct_diff_h       = abs(h_simulated_per_day - h_published_per_day) /
                           h_published_per_day * 100
  )

knitr::kable(dec_check, digits = 5,
             caption = "Exploratory three-arm model: published vs simulated hazards and median OS.")
Exploratory three-arm model: published vs simulated hazards and median OS.
arm h_published_per_day h_simulated_per_day median_OS_pub_months median_OS_sim_months abs_pct_diff_h
LDAC alone 0.00540 0.00540 4.21719 4.21719 0
Glasdegib + LDAC 0.00281 0.00281 8.10997 8.10997 0
Glasdegib + decitabine 0.00206 0.00206 11.03975 11.03975 0 

stopifnot(all(dec_check$abs_pct_diff_h < 0.1))

# Paper reports glasdegib + decitabine median OS = 11.1 months. Computed
# value should agree within 0.5 months.
dec_median <- dec_check$median_OS_pub_months[dec_check$arm == "Glasdegib + decitabine"]
cat(sprintf("Glasdegib + decitabine median OS: %.2f months (paper: 11.1 months)\n",
            dec_median))
#> Glasdegib + decitabine median OS: 11.04 months (paper: 11.1 months)
stopifnot(abs(dec_median - 11.1) < 0.5)

Assumptions and deviations

  • Treatment arm encoded via canonical drug indicators. Lin 2020 expresses the treatment-response model with the binary covariate LDAC_alone (1 = LDAC alone, 0 = glasdegib + LDAC), and the exploratory three-arm model with two binary covariates glasdegib+LDAC and glasdegib+decitabine (reference = LDAC alone). The nlmixr2lib model files use the canonical covariates CONMED_GLASDEGIB and CONMED_DECITABINE (1 if the named drug is present in the subject’s treatment regimen) and derive the paper’s arm indicators inside model() – ldac_alone = 1 - CONMED_GLASDEGIB for the treatment-response model, and gl_ldac_arm = CONMED_GLASDEGIB * (1 - CONMED_DECITABINE), gl_dec_arm = CONMED_GLASDEGIB * CONMED_DECITABINE for the exploratory model. Published parameter values (lambda, theta_ldac_alone, theta_gl_ldac, theta_gl_dec) are preserved verbatim; the substitution is purely a column-naming choice that aligns with the canonical register.

  • No estimated IIV. All three source NONMEM runs were fit without subject-level random effects (parametric population TTE). The packaged models therefore produce a single typical-value S(t) per covariate setting; per-subject survival trajectories within an arm are identical.

  • No residual error. The source NONMEM $ERROR / $EST framework for parametric TTE uses the LIKE / F_FLAG branch (the likelihood is the survival / event-density itself). The packaged models expose sur and hazard as derived outputs intended for forward simulation; downstream users who want stochastic event-time sampling should layer a survival- sampling step on top (e.g. inverse-CDF sampling of an exponential with rate hazard).

  • Documented-but-unused covariates. All baseline demographics, baseline safety laboratory values, and disease characteristics that Lin 2020 screened in SCM forward inclusion but did not retain after backward elimination are documented in each model file’s covariatesDataExcluded list (with the cohort-level summary statistics from Table 1). The covariatesDataExcluded mechanism preserves the screening provenance without flagging the entries as model-required.

  • Exposure-response cohort size. The exposure-response analysis (n = 75) is a strict subset of the treatment-response analysis glasdegib + LDAC arm (n = 78); the difference of 3 patients is those without available glasdegib PK information.

  • Phase 1b Arm A subjects. Phase 1b Arm A evaluated glasdegib at 100 or 200 mg QD plus LDAC; Phase 2 used the 100 mg QD dose. The exploratory pooled analysis pools both phases into the glasdegib + LDAC arm because the exposure-response analysis showed no significant glasdegib dose / exposure effect on OS. The packaged exploratory model encodes a single CONMED_GLASDEGIB = 1 for all glasdegib-containing subjects regardless of dose level.

  • No errata. A search of the journal landing page, publisher corrections feed, and PubMed at extraction time (2026-06-24) found no errata or corrigenda for Lin 2020.