Skip to contents

Models and source

Wu F, Zheng X, Burnett J, Masilamani M, Zhang W, Zhong X, Caia A, Cook M, Piasecki J, Kondic A, Lamba M, Zhou J. Characterizing cellular expansion of idecabtagene vicleucel and association with clinical efficacy and safety in patients with triple-class-exposed relapsed/refractory multiple myeloma. J Clin Pharmacol. 2025;65(11):1568-1576. doi:10.1002/jcph.70075.

Idecabtagene vicleucel (ide-cel) is a BCMA-directed CAR T-cell product given as one infusion. Wu 2025 summarised the in-vivo expansion of the cells in the ide-cel arm of the phase 3 KarMMa-3 trial by noncompartmental analysis of the peripheral-blood transgene level (copies per microgram of genomic DNA, droplet digital PCR), and then fitted a set of regressions:

Model Kind End point Exposure / driver Source
Wu_2025_idecabtageneVicleucel_auc linear regression log(AUC0-28D) actual dose Table S4
Wu_2025_idecabtageneVicleucel_pfs Cox PH progression-free survival log2(AUC0-28D) Table S7
Wu_2025_idecabtageneVicleucel_orr sigmoid Emax logistic overall response (IRC) log(AUC0-28D) Table S8
Wu_2025_idecabtageneVicleucel_crs_anti_il6 linear logistic CRS requiring tocilizumab or siltuximab (tCRS) log(Cmax/Tmax) Table S10
Wu_2025_idecabtageneVicleucel_crs_corticosteroid linear logistic CRS requiring corticosteroids (sCRS) log(Cmax/Tmax) Table S11
Wu_2025_idecabtageneVicleucel_dose_orr linear logistic overall response (IRC) actual dose Table S12
Wu_2025_idecabtageneVicleucel_dose_pfs Cox PH progression-free survival actual dose Table S13
Wu_2025_idecabtageneVicleucel_dose_crs_anti_il6 linear logistic tCRS actual dose Table S14
Wu_2025_idecabtageneVicleucel_dose_crs_corticosteroid linear logistic sCRS actual dose Table S15

None of the models has a cellular kinetic layer or an ODE. The exposure metrics are data columns: AUC_IDECEL_DDPCR (AUC0-28D, day*copies/ug) for efficacy and EXPANSION_RATE_IDECEL (Cmax/Tmax, copies/ug/day) for safety, which the authors chose as an early exposure measure because CRS starts within days. The dose models use the actual infused dose DOSE_IDECEL_MCELLS (10^6 CAR+ T cells) as a covariate column, not as a dosing event.

The library also carries the phase 2 KarMMa exposure-response models of Connarn 2023 (Connarn_2023_idecabtageneVicleucel_*). Wu 2025 notes that the two trials quantified the transgene with different bioanalytical methods and does not compare their exposure values directly. The two exposures therefore use separate columns: AUC_IDECEL holds the KarMMa quantitative-PCR AUC0-28 for the Connarn models, and AUC_IDECEL_DDPCR holds the KarMMa-3 droplet digital PCR AUC0-28D for the Wu models, so a value measured with one assay is not fed to a model fitted to the other.

mod_auc <- rxode2::rxode(readModelDb("Wu_2025_idecabtageneVicleucel_auc"))
mod_pfs <- rxode2::rxode(readModelDb("Wu_2025_idecabtageneVicleucel_pfs"))
mod_orr <- rxode2::rxode(readModelDb("Wu_2025_idecabtageneVicleucel_orr"))
mod_tcrs <- rxode2::rxode(readModelDb("Wu_2025_idecabtageneVicleucel_crs_anti_il6"))
mod_scrs <- rxode2::rxode(readModelDb("Wu_2025_idecabtageneVicleucel_crs_corticosteroid"))
mod_d_orr <- rxode2::rxode(readModelDb("Wu_2025_idecabtageneVicleucel_dose_orr"))
mod_d_pfs <- rxode2::rxode(readModelDb("Wu_2025_idecabtageneVicleucel_dose_pfs"))
mod_d_tcrs <- rxode2::rxode(readModelDb("Wu_2025_idecabtageneVicleucel_dose_crs_anti_il6"))
mod_d_scrs <- rxode2::rxode(readModelDb("Wu_2025_idecabtageneVicleucel_dose_crs_corticosteroid"))

Population

KarMMa-3 (NCT03651128) randomised adults with triple-class-exposed relapsed/refractory multiple myeloma after two to four prior regimens 2:1 to ide-cel or a standard regimen. In the ide-cel arm, 225 patients were infused by the 1 March 2022 cellular-kinetic cutoff; 220 had evaluable exposure parameters and form the dose-exposure and exposure-response analysis set, while the exploratory dose-response models use all 225. Table S3 describes the 220: median age 63 years (range 30-81), median weight 83 kg (41.1-144), 37.7% female, 70% White, 31% / 37% / 31% with two / three / four prior regimens, 61.8% triple-class refractory, 23.6% with extramedullary plasmacytoma, 61.4% anti-CAR antibody positive after infusion, and 26% infused with Formulation B.

str(mod_orr$population, max.level = 1)
#> List of 12
#>  $ species       : chr "human"
#>  $ n_subjects    : int 220
#>  $ n_studies     : int 1
#>  $ n_observations: chr "220 binary outcomes (180 responders, 81.8%; Table S6)"
#>  $ age_range     : chr "30-81 years (median 63; 43% aged 65 or older)"
#>  $ weight_range  : chr "41.1-144 kg (median 83)"
#>  $ sex_female_pct: num 37.7
#>  $ race_ethnicity: Named num [1:5] 70 7.3 2.3 0.9 19.5
#>   ..- attr(*, "names")= chr [1:5] "White" "Black" "Asian" "Other" ...
#>  $ disease_state : chr "Triple-class-exposed relapsed/refractory multiple myeloma after two to four prior regimens with progression wit"| __truncated__
#>  $ dose_range    : chr "Single IV infusion; target 150-450 x 10^6 CAR+ T cells (up to 540 x 10^6 permitted)"
#>  $ regions       : chr "International (KarMMa-3, NCT03651128)"
#>  $ notes         : chr "Ide-cel arm of KarMMa-3, CK data cutoff 1 March 2022. Observed ORR by AUC0-28D quartile (Figure S7): about 56%,"| __truncated__

Source trace

Every coefficient is printed in a Supporting Information table. The trailing comment on each ini() line in the model files names the table row.

Model Parameters Source
_auc intercept 13.1186; dose 0.0022; log ferritin 0.0458; weight -0.0111; log sBCMA 0.1615; Formulation B -0.3939; 3 / 4 vs 2 prior regimens -0.1576 / -0.2029; log IL-15 0.5157; extramedullary plasmacytoma -0.5036; ADA +0.4621 Table S4
_auc centring: ferritin 163 ug/L, sBCMA 144.5 ng/mL, IL-15 43 pg/mL, weight 75 kg Table S5 note
_pfs 14 log hazard ratios (log2 AUC0-28D -0.43, …) Table S7
_orr E0 -5 (fixed), Emax 7.654, Hill 34.006, EC50 11.551, sex 0.121, Formulation B -0.19, prior regimens 0.03 / -0.005, extramedullary plasmacytoma -1.343 Table S8; E0 fixed per Results
_crs_anti_il6 intercept -4.5924, log(Cmax/Tmax) 0.6551, Formulation B 0.0972, prior regimens -0.1757 / 0.3399, ADA -0.8765 Table S10
_crs_corticosteroid intercept -6.1186, log(Cmax/Tmax) 0.506, formulation 0.277, prior regimens 0.4423 / 0.2028 Table S11
_dose_orr 9 coefficients Table S12
_dose_pfs 13 log hazard ratios Table S13
_dose_crs_anti_il6 9 coefficients Table S14
_dose_crs_corticosteroid 8 coefficients Table S15
logistic models addSd_prob_* 0.001 (fixed) not from the source (Bernoulli likelihood)
Equation Source
h(t) = h0(t) exp(b1 log2(AUC0-28D) + b2 COV) Methods, first equation
logit(P) = b0 + b1 log(X) + b2 COV, X = AUC0-28D (ORR) or Cmax/Tmax (tCRS, sCRS) Methods, second equation
sigmoid Emax with E0 fixed at -5 for ORR Results, ORR model
natural-log-transformed AUC0-28D in the dose-exposure model Results, CK characterisation

Helper

A static model has no states and no doses, so every covariate row can be an observation record of a single subject; time only orders the rows.

solve_static <- function(model, covs) {
  ev <- covs |> mutate(id = 1L, time = seq_len(n()), evid = 0L, amt = 0)
  out <- as.data.frame(rxode2::rxSolve(model, events = ev, returnType = "data.frame"))
  stopifnot(nrow(out) == nrow(covs))
  out
}

# Reference subject of every model: female, Formulation A, two prior
# regimens, no extramedullary plasmacytoma, ADA-negative, not triple-class
# refractory, no steroid or anti-CD38 in the last regimen, continuous
# covariates at their reference values.
ref_covs <- data.frame(
  SEXF = 1, FORM_IDECEL_B = 0, LINE_4L = 0, LINE_5L = 0, DIS_EMD = 0,
  ADA_POS = 0, REFRACTORY_TRIPLE_CLASS = 0, LAST_PRIOR_CORTICOSTEROID = 0,
  LAST_PRIOR_ANTICD38 = 0, WT = 75, FERRITIN_BL = 163, SBCMA = 144.5,
  IL15 = 43, TNFA = 27, SCD137 = 0.105, DOSE_IDECEL_MCELLS = 450,
  AUC_IDECEL_DDPCR = 1.3e6, EXPANSION_RATE_IDECEL = 1.35e4
)
with_ref <- function(...) {
  x <- data.frame(...)
  cbind(x, ref_covs[rep(1, nrow(x)), setdiff(names(ref_covs), names(x)), drop = FALSE])
}

# Expected value over the categorical covariates of the analysis set
# (Table S3), treated as independent: sex, formulation (57 / 216 with a
# known formulation were B), prior regimens 69 / 82 / 69, extramedullary
# plasmacytoma 52 / 220, post-infusion ADA 135 / 220.
cat_grid <- expand.grid(
  SEXF = 0:1, FORM_IDECEL_B = 0:1, nreg = 2:4, DIS_EMD = 0:1, ADA_POS = 0:1
) |>
  mutate(
    LINE_4L = as.integer(nreg == 3), LINE_5L = as.integer(nreg == 4),
    w = ifelse(SEXF == 1, 83 / 220, 137 / 220) *
      ifelse(FORM_IDECEL_B == 1, 57 / 216, 159 / 216) *
      c(69, 82, 69)[nreg - 1] / 220 *
      ifelse(DIS_EMD == 1, 52 / 220, 168 / 220) *
      ifelse(ADA_POS == 1, 135 / 220, 85 / 220)
  ) |>
  select(-nreg)
stopifnot(abs(sum(cat_grid$w) - 1) < 1e-12)

expected_prob <- function(model, output, exposure_col, exposure) {
  sapply(exposure, function(x) {
    covs <- cat_grid |> select(-w)
    covs[[exposure_col]] <- x
    covs <- cbind(covs, ref_covs[rep(1, nrow(covs)), setdiff(names(ref_covs), names(covs)), drop = FALSE])
    sum(solve_static(model, covs)[[output]] * cat_grid$w)
  })
}

Gate 1: the dose-exposure model reproduces Table S5 and Figure 1

Table S5 prints, for each covariate of the dose-log(AUC0-28D) model, the percent change in exposure at the 5th and 95th percentile of that covariate relative to its reference value. Those percent changes depend on the centring and on which logarithm is used, so reproducing them confirms both.

s5 <- tribble(
  ~row, ~cov, ~value, ~published,
  "Ferritin 5th percentile", "FERRITIN_BL", 25.5, -8.144,
  "Ferritin 95th percentile", "FERRITIN_BL", 1836.25, 11.728,
  "Body weight 5th percentile", "WT", 54.97, 24.834,
  "Body weight 95th percentile", "WT", 115.845, -36.385,
  "Soluble BCMA 5th percentile", "SBCMA", 13.3, -31.967,
  "Soluble BCMA 95th percentile", "SBCMA", 487.4, 21.69,
  "IL-15 5th percentile", "IL15", 14, -43.936,
  "IL-15 95th percentile", "IL15", 100.4, 54.849,
  "Formulation B vs A", "FORM_IDECEL_B", 1, -32.557,
  "3 vs 2 prior regimens", "LINE_4L", 1, -14.58,
  "4 vs 2 prior regimens", "LINE_5L", 1, -18.366,
  "Extramedullary plasmacytoma", "DIS_EMD", 1, -39.567,
  "Post-infusion ADA positive", "ADA_POS", 1, 58.747
)
ref_auc <- solve_static(mod_auc, ref_covs)$auc0_28
s5$simulated <- sapply(seq_len(nrow(s5)), function(i) {
  covs <- ref_covs
  covs[[s5$cov[i]]] <- s5$value[i]
  100 * (solve_static(mod_auc, covs)$auc0_28 / ref_auc - 1)
})
s5 |>
  mutate(difference = simulated - published) |>
  select(-cov) |>
  rename(
    "Covariate level" = row, "Value" = value,
    "Published change (%)" = published, "Model change (%)" = simulated,
    "Difference (points)" = difference
  ) |>
  knitr::kable(digits = 3)
Covariate level Value Published change (%) Model change (%) Difference (points)
Ferritin 5th percentile 25.500 -8.144 -8.145 -0.001
Ferritin 95th percentile 1836.250 11.728 11.730 0.002
Body weight 5th percentile 54.970 24.834 24.899 0.065
Body weight 95th percentile 115.845 -36.385 -36.452 -0.067
Soluble BCMA 5th percentile 13.300 -31.967 -31.973 -0.006
Soluble BCMA 95th percentile 487.400 21.690 21.696 0.006
IL-15 5th percentile 14.000 -43.936 -43.937 -0.001
IL-15 95th percentile 100.400 54.849 54.851 0.002
Formulation B vs A 1.000 -32.557 -32.558 -0.001
3 vs 2 prior regimens 1.000 -14.580 -14.581 -0.001
4 vs 2 prior regimens 1.000 -18.366 -18.364 0.002
Extramedullary plasmacytoma 1.000 -39.567 -39.565 0.002
Post-infusion ADA positive 1.000 58.747 58.740 -0.007
# The coefficients are printed to four decimals, so the largest gap (body
# weight, -0.0111 per kg) is a rounding effect of under 0.1 point.
stopifnot(all(abs(s5$simulated - s5$published) < 0.1))

Figure 1 draws the dose-log(AUC0-28D) line for the reference patient; it runs from about 13.5 at 170 x 10^6 cells to about 14.3 at 530 x 10^6 cells. The dose therefore enters uncentred, while every other covariate is centred.

dose_grid <- seq(150, 550, by = 10)
fig1 <- solve_static(mod_auc, with_ref(DOSE_IDECEL_MCELLS = dose_grid))
fig1_ends <- fig1$lauc0_28[dose_grid %in% c(170, 530)]
fig1_ends
#> [1] 13.4926 14.2846
stopifnot(abs(fig1_ends - c(13.5, 14.3)) < 0.1)

ggplot(data.frame(dose = dose_grid, lauc = fig1$lauc0_28), aes(dose, lauc)) +
  geom_line() +
  coord_cartesian(ylim = c(9.5, 17)) +
  labs(
    x = "Actual dose (x10^6 CAR+ T cells)", y = "Predicted log(AUC0-28D)",
    title = "Replicates Figure 1 of Wu 2025 (reference patient, typical line)"
  )

The dose slope is 0.0022 per 10^6 cells, so the whole approved range of 300-510 x 10^6 cells moves the typical log(AUC0-28D) by 0.46, a 59% change, which is smaller than the effect of a single categorical covariate such as extramedullary plasmacytoma. This is the flat dose-exposure relationship the paper uses to support the extended dose range.

Gate 2: the ORR sigmoid model reproduces the Figure S7 plateau

Table S8 prints back-transformed estimates (Emax 7.654, Hill 34.006, EC50 11.551) next to 95% confidence intervals on the log scale (log(7.654) = 2.035 lies in 1.9-2.17). The exposure is log(AUC0-28D), as in the KarMMa model of Connarn 2023, so EC50 corresponds to an AUC0-28D of about 10^{5} day*copies/ug, and a Hill coefficient of 34 on that scale makes the switch close to a step.

The model reads Emax as the increment on the logit, logit(P) = E0 + Emax * f, so the reference-subject plateau is expit(-5 + 7.654) = 93%. The alternative reading, Emax as the logit at full effect, would put the plateau at expit(7.654) = 99.95%. The Figure S7 VPC shows observed ORR of about 89%, 91% and 94% in AUC0-28D quartiles 2-4, all far above EC50, with model-predicted medians near 91%; only the increment reading reproduces that.

# Representative log(AUC0-28D) of quartiles 2-4: log-midpoints of the
# Figure 2a / S7 quartile bounds (0.4-1.3, 1.3-2.7, 2.7-17.8 x 10^6).
q_lnauc <- c(Q2 = mean(log(c(0.4e6, 1.3e6))), Q3 = mean(log(c(1.3e6, 2.7e6))), Q4 = mean(log(c(2.7e6, 17.8e6))))
orr_pred <- 100 * expected_prob(mod_orr, "prob_orr_central", "AUC_IDECEL_DDPCR", exp(q_lnauc))
orr_obs <- c(Q2 = 89, Q3 = 91, Q4 = 94.5) # Figure S7 observed points, digitised
orr_alt <- 100 * sum(cat_grid$w * plogis(7.654 + 0.121 * (1 - cat_grid$SEXF) -
  0.19 * cat_grid$FORM_IDECEL_B + 0.03 * cat_grid$LINE_4L -
  0.005 * cat_grid$LINE_5L - 1.343 * cat_grid$DIS_EMD))
data.frame(
  quartile = names(q_lnauc), observed = orr_obs, model = orr_pred,
  emax_as_full_effect_logit = orr_alt, row.names = NULL
) |>
  rename(
    "AUC0-28D quartile" = quartile, "Observed ORR (%)" = observed,
    "Model, increment Emax (%)" = model,
    "Alternative, Emax as full-effect logit (%)" = emax_as_full_effect_logit
  ) |>
  knitr::kable(digits = 1)
AUC0-28D quartile Observed ORR (%) Model, increment Emax (%) Alternative, Emax as full-effect logit (%)
Q2 89.0 89.9 99.9
Q3 91.0 90.2 99.9
Q4 94.5 90.2 99.9
stopifnot(
  abs(mean(orr_pred) - mean(orr_obs)) < 4,
  orr_alt > 99
)

The lowest quartile (AUC0-28D below 0.4 x 10^6, observed ORR about 56%) straddles EC50, so its predicted rate depends on how many of its patients fall below 1.0 x 10^5, which the paper does not tabulate; it is not gated here.

lnauc_grid <- seq(10, 17, by = 0.02)
orr_curve <- solve_static(mod_orr, with_ref(AUC_IDECEL_DDPCR = exp(lnauc_grid)))
orr_curve_emd <- solve_static(mod_orr, with_ref(AUC_IDECEL_DDPCR = exp(lnauc_grid), DIS_EMD = 1))
bind_rows(
  data.frame(lnauc = lnauc_grid, p = orr_curve$prob_orr_central, who = "reference"),
  data.frame(lnauc = lnauc_grid, p = orr_curve_emd$prob_orr_central, who = "extramedullary plasmacytoma")
) |>
  ggplot(aes(lnauc, 100 * p, colour = who)) +
  geom_line() +
  labs(
    x = "log(AUC0-28D) (day*copies/ug)", y = "Predicted ORR (%)", colour = NULL,
    title = "ORR exposure-response (Table S8)"
  )

Gate 3: the CRS models reproduce the Figure 3 quartile rates

Figure 3a/3b and Figures S8-S9 report tCRS and sCRS by quartile of the expansion rate Cmax/Tmax (quartile bounds 0-0.4, 0.4-1.3, 1.4-3.2 and 3.4-11.6 x 10^4 copies/ug/day; 55 patients each). Each quartile is represented by the log-midpoint of its bounds, except quartile 1, whose lower bound is not printed; its median log(Cmax/Tmax), 7.8, is read off Figure 3. The categorical covariates are averaged over their Table S3 distribution.

q_ler <- c(Q1 = 7.8, Q2 = mean(log(c(4e3, 1.3e4))), Q3 = mean(log(c(1.4e4, 3.2e4))), Q4 = mean(log(c(3.4e4, 1.16e5))))
crs <- data.frame(
  quartile = names(q_ler),
  tcrs_obs = 100 * c(28, 37, 43, 48) / 55,
  tcrs_model = 100 * expected_prob(mod_tcrs, "prob_crs_anti_il6", "EXPANSION_RATE_IDECEL", exp(q_ler)),
  scrs_obs = 100 * c(7, 11, 16, 26) / 55,
  scrs_model = 100 * expected_prob(mod_scrs, "prob_crs_corticosteroid", "EXPANSION_RATE_IDECEL", exp(q_ler)),
  row.names = NULL
)
crs |>
  rename(
    "Cmax/Tmax quartile" = quartile, "tCRS observed (%)" = tcrs_obs,
    "tCRS model (%)" = tcrs_model, "sCRS observed (%)" = scrs_obs,
    "sCRS model (%)" = scrs_model
  ) |>
  knitr::kable(digits = 1)
Cmax/Tmax quartile tCRS observed (%) tCRS model (%) sCRS observed (%) sCRS model (%)
Q1 50.9 51.0 12.7 13.6
Q2 67.3 67.2 20.0 21.3
Q3 78.2 80.1 29.1 31.7
Q4 87.3 89.0 47.3 44.4
# Observed quartile counts carry a binomial SE of 5-7 points with n = 55.
stopifnot(
  all(abs(crs$tcrs_model - crs$tcrs_obs) < 6),
  all(abs(crs$scrs_model - crs$scrs_obs) < 6)
)
ler_grid <- seq(5, 12, by = 0.05)
bind_rows(
  data.frame(ler = ler_grid, p = expected_prob(mod_tcrs, "prob_crs_anti_il6", "EXPANSION_RATE_IDECEL", exp(ler_grid)), endpoint = "tCRS"),
  data.frame(ler = ler_grid, p = expected_prob(mod_scrs, "prob_crs_corticosteroid", "EXPANSION_RATE_IDECEL", exp(ler_grid)), endpoint = "sCRS")
) |>
  ggplot(aes(ler, 100 * p, colour = endpoint)) +
  geom_line() +
  geom_point(
    data = crs |> select(quartile, tcrs_obs, scrs_obs) |>
      pivot_longer(-quartile, names_to = "endpoint", values_to = "pct") |>
      mutate(endpoint = ifelse(endpoint == "tcrs_obs", "tCRS", "sCRS"), ler = q_ler[quartile]),
    aes(ler, pct, colour = endpoint)
  ) +
  labs(
    x = "log(Cmax/Tmax) (copies/ug/day)", y = "Rate (%)", colour = NULL,
    title = "Replicates Figure 3a-b of Wu 2025 (population-averaged model, observed quartiles)"
  )

Gate 4: the Cox models return the published hazard ratios

A Cox regression has no baseline hazard, so these models return the relative hazard against their reference subject. For the AUC0-28D term the reference is 1.3 x 10^6 day*copies/ug, the median quartile boundary in Figure 2a; moving it changes which subject has hr = 1 but no hazard ratio between subjects.

pfs_hr <- function(model, ...) solve_static(model, with_ref(...))$hr
hr_checks <- data.frame(
  contrast = c(
    "AUC0-28D doubled", "Post-infusion ADA positive", "Formulation B",
    "Triple-class refractory", "Soluble BCMA doubled",
    "Dose +100 x10^6 cells (dose model)", "TNF-alpha doubled (dose model)"
  ),
  model = c(
    pfs_hr(mod_pfs, AUC_IDECEL_DDPCR = 2.6e6), pfs_hr(mod_pfs, ADA_POS = 1),
    pfs_hr(mod_pfs, FORM_IDECEL_B = 1), pfs_hr(mod_pfs, REFRACTORY_TRIPLE_CLASS = 1),
    pfs_hr(mod_pfs, SBCMA = 289), pfs_hr(mod_d_pfs, DOSE_IDECEL_MCELLS = 550),
    pfs_hr(mod_d_pfs, TNFA = 54)
  ),
  published = exp(c(-0.43, -0.89, -0.67, 0.61, 0.28, -0.003 * 100, 0.406))
)
hr_checks |>
  rename("Contrast vs reference" = contrast, "Model HR" = model, "exp(Table S7 / S13 coefficient)" = published) |>
  knitr::kable(digits = 3)
Contrast vs reference Model HR exp(Table S7 / S13 coefficient)
AUC0-28D doubled 0.651 0.651
Post-infusion ADA positive 0.411 0.411
Formulation B 0.512 0.512
Triple-class refractory 1.840 1.840
Soluble BCMA doubled 1.323 1.323
Dose +100 x10^6 cells (dose model) 0.741 0.741
TNF-alpha doubled (dose model) 1.501 1.501
stopifnot(all(abs(hr_checks$model / hr_checks$published - 1) < 1e-8))

Each doubling of AUC0-28D lowers the hazard of progression or death by 35%; over the interquartile range of AUC0-28D (0.4 to 2.7 x 10^6, Figure 2a) that is a hazard ratio of 0.31, consistent with the separation of the lowest-quartile Kaplan-Meier curve in Figure 2a.

Gate 5: the dose-response models are flat and reproduce the observed rates

The exploratory dose-response logistic models carry no exposure term, so their population-averaged prediction at the KarMMa-3 doses must match the observed event rates (for a fitted logistic regression with an intercept, the mean fitted probability equals the observed rate). A virtual cohort of 200 patients is drawn from the Table S3 distributions (weight normal 83.1 +/- 18.3 kg truncated to 41-144 kg; soluble BCMA log-normal with median 143 ng/mL and mean 184 ng/mL; doses uniform over 350-530 x 10^6 cells, the span of the Figure 1 scatter) and compared with the analysis-set rates (ORR 180 / 220, tCRS 156 / 220, sCRS 60 / 220).

set.seed(20251106)
n <- 200
cohort <- data.frame(
  SEXF = rbinom(n, 1, 83 / 220),
  WT = pmin(pmax(rnorm(n, 83.1, 18.3), 41.1), 144),
  SBCMA = pmin(pmax(exp(rnorm(n, log(143), sqrt(2 * log(184 / 143)))), 1.8), 789),
  FORM_IDECEL_B = rbinom(n, 1, 57 / 216),
  nreg = sample(2:4, n, replace = TRUE, prob = c(69, 82, 69)),
  DIS_EMD = rbinom(n, 1, 52 / 220),
  ADA_POS = rbinom(n, 1, 135 / 220),
  DOSE_IDECEL_MCELLS = runif(n, 350, 530)
) |>
  mutate(LINE_4L = as.integer(nreg == 3), LINE_5L = as.integer(nreg == 4)) |>
  select(-nreg)
cohort_full <- cbind(cohort, ref_covs[rep(1, n), setdiff(names(ref_covs), names(cohort)), drop = FALSE])
dose_rates <- data.frame(
  endpoint = c("ORR", "tCRS", "sCRS"),
  observed = 100 * c(180, 156, 60) / 220,
  model = 100 * c(
    mean(solve_static(mod_d_orr, cohort_full)$prob_orr_central),
    mean(solve_static(mod_d_tcrs, cohort_full)$prob_crs_anti_il6),
    mean(solve_static(mod_d_scrs, cohort_full)$prob_crs_corticosteroid)
  )
)
dose_rates |>
  rename("End point" = endpoint, "Observed rate (%)" = observed, "Virtual-cohort mean (%)" = model) |>
  knitr::kable(digits = 1)
End point Observed rate (%) Virtual-cohort mean (%)
ORR 81.8 79.3
tCRS 70.9 70.9
sCRS 27.3 27.3
stopifnot(all(abs(dose_rates$model - dose_rates$observed) < 8))
dgrid <- seq(300, 510, by = 10)
bind_rows(
  data.frame(dose = dgrid, p = solve_static(mod_d_orr, with_ref(DOSE_IDECEL_MCELLS = dgrid))$prob_orr_central, endpoint = "ORR"),
  data.frame(dose = dgrid, p = solve_static(mod_d_tcrs, with_ref(DOSE_IDECEL_MCELLS = dgrid))$prob_crs_anti_il6, endpoint = "tCRS"),
  data.frame(dose = dgrid, p = solve_static(mod_d_scrs, with_ref(DOSE_IDECEL_MCELLS = dgrid))$prob_crs_corticosteroid, endpoint = "sCRS")
) |>
  ggplot(aes(dose, 100 * p, colour = endpoint)) +
  geom_line() +
  labs(
    x = "Actual dose (x10^6 CAR+ T cells)", y = "Predicted rate, reference subject (%)",
    colour = NULL, title = "Dose-response over the approved 300-510 x10^6 range (Tables S12, S14, S15)"
  )

PKNCA

Not applicable. These are static regressions on per-subject exposure summaries that the authors computed by noncompartmental analysis of observed transgene levels; no model in this paper predicts a transgene time course, so there is nothing to submit to PKNCA. The library’s mechanistic ide-cel model, Singh_2021_idecabtageneVicleucel_human, simulates CAR T cells in cells rather than transgene copies per microgram of DNA and cannot supply these exposure columns directly.

Assumptions and deviations

  • PFS exposure-response model is the full model. Table S7 prints the full multivariable Cox model. The Results state that backward elimination then removed last prior anti-CD38 therapy and baseline soluble CD137, but the refitted final coefficients are not printed anywhere, so the library carries the full model as published, including those two terms.
  • Emax is an increment on the logit. The paper prints no sigmoid Emax equation. Gate 2 shows that only logit(P) = E0 + Emax * f reproduces the Figure S7 plateau.
  • Back-transformed Table S8 estimates. Emax, Hill and EC50 are encoded as the natural logs of the printed estimates, the scale on which their printed 95% confidence intervals lie.
  • Cox reference values. Continuous covariates of the Cox models are centred on the medians the paper reports (ferritin 163 ug/L and sBCMA 144.5 ng/mL from Table S5; TNF-alpha 27 pg/mL and sCD137 0.105 ng/mL from Table S3). The AUC0-28D median is not printed; 1.3 x 10^6 day*copies/ug, the Q2/Q3 boundary of Figure 2a, is used. The dose Cox model is referenced to 450 x 10^6 cells and 75 kg. None of these choices changes a hazard ratio.
  • TNF-alpha in the dose PFS model is on the log2 scale. Table S13 labels the row ‘Baseline TNF-alpha (pg/mL)’ without a transform; per pg/mL the coefficient 0.406 would imply hazard ratios above 10^26 across the observed range, and the same covariate is log2 in Table S7.
  • Centring of the dose-response logistic models. The paper does not state it. The dose is uncentred (the intercept-plus-dose term reproduces the observed rates only that way), and body weight and log sBCMA are centred on 75 kg and 144.5 ng/mL, the reference values Table S5 gives for the dose-exposure model. Centring weight on the 83 kg median instead would move the logits by 0.15-0.20; Gate 5 cannot distinguish the two.
  • Formulation in the sCRS model. Table S11 labels the row ‘Formulation (sLVV vs aLVV)’, while every other table and the Results compare Formulation B with A. The model reads it as the same Formulation B indicator.
  • Post-infusion ADA status is a post-baseline covariate (positive at any visit after the infusion, mostly from month 3). The authors caution that it may mark longer follow-up rather than cause higher exposure, longer PFS or less tCRS; a simulation that draws it independently of outcome inherits that confounding.
  • Bernoulli likelihood. The logistic models carry a fixed additive residual of 0.001 on the probability only so that rxode2 can attach an error model; it is not a published quantity. The dose-exposure regression’s residual standard deviation is not reported, so that model declares no endpoint.
  • Not extracted. No regression was fitted for duration of response, MRD-negative complete response, all-grade or grade 3+ CRS, neurotoxicity or cytopenia (graphical analyses only, Figures 2-3).