Skip to contents

Telisotuzumab vedotin conjugate, MMAE payload and exposure-response (Babel 2026)

Source: vignettes/articles/Babel_2026_telisotuzumab.Rmd
Babel_2026_telisotuzumab.Rmd

Source and models

  • Citation: Babel H, Brunsdon P, Engelhardt B, Schmitt V, Ratajczak C, Mensing S, Menon RM, Parikh A. Population pharmacokinetics and exposure-response analyses for telisotuzumab vedotin in patients with c-Met protein overexpressing tumors. CPT Pharmacometrics Syst Pharmacol. 2026;15(1):e70219. doi:10.1002/psp4.70219. PMCID PMC12945708. All parameter values are from Data S1 (Supporting Information) Table S7, ‘Final Model Parameter Estimates and Variability of Teliso-V Conjugate and Unconjugated MMAE Payload Pharmacokinetics’, Teliso-V Conjugate block.
  • Article (open access): https://doi.org/10.1002/psp4.70219
  • PubMed Central: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12945708/
  • Supporting Information (Data S1; Tables S1-S7 and Figures S1-S5): distributed with the article and retrieved from the Europe PMC supplementary-files endpoint for PMC12945708.

Babel 2026 reports six models, and this package carries all six as separate files because the authors fitted them separately:

Model file What it is Source
Babel_2026_telisotuzumab Conjugate (Teliso-V) population PK, 2-compartment Table S7 conjugate block, Figure S1A
Babel_2026_telisotuzumab_mmae Unconjugated MMAE payload population PK, 1-compartment with first-order deconjugation Table S7 payload block, Figure S1B
Babel_2026_telisotuzumab_orr Exposure-efficacy logistic for ORR per Independent Central Review Figure 2
Babel_2026_telisotuzumab_neuropathy Exposure-safety logistic for grade >= 3 peripheral neuropathy Figure 4 left
Babel_2026_telisotuzumab_corneal Exposure-safety logistic for grade >= 2 corneal epitheliopathy Figure 4 right
Babel_2026_telisotuzumab_teae Exposure-safety logistic for any grade >= 3 TEAE, driven by payload Figure 5, CavgMMAE panel

The conjugate and payload PK models were “developed independently since independent models show similar performance compared to integrated models” (Methods). Figure S1 draws them as two separate panels, and the payload panel shows a single MMAE compartment with an inbound first-order arrow labelled Ka and an outbound clearance arrow – not a link from the conjugate compartments. That is why Babel_2026_telisotuzumab_mmae is a standalone depot-plus-central model rather than a coupled system, and why it can be solved without the conjugate model.

Population

Population of the Babel 2026 population PK analysis (Table S4).
Field Value
Species human
Subjects 304
Studies 2
Age median 65 years, range 30-87 (Babel 2026 Table S4, All Participants)
Body weight median 68.9 kg, range 36.0-144 (Babel 2026 Table S4, All Participants)
Female 37.8%
Race White 65%, Asian 32%, Black or African American 3
Disease Advanced solid tumours likely to express c-Met (phase 1, NCT02099058, n = 35, all tumour types, monotherapy cohorts receiving Process II drug material only) and locally advanced or metastatic c-Met protein overexpressing non-small cell lung cancer (LUMINOSITY phase 2, NCT03539536, n = 269).
Doses Phase 1: 0.15-3.3 mg/kg every 3 weeks and 1.6-2.2 mg/kg every 2 weeks. LUMINOSITY: 1.6 or 1.9 mg/kg every 2 weeks. Approved regimen 1.9 mg/kg Q2W as an intravenous infusion, capped at 190 mg for patients weighing at least 100 kg.
Regions Europe 26%, North America 26%, Asia 28%, rest of world 20% (Babel 2026 Table S4)

The population PK analysis pooled 304 patients: 35 from the phase 1 study NCT02099058 (monotherapy cohorts that received Process II drug material, all solid tumour types) and 269 from the LUMINOSITY phase 2 study NCT03539536 (c-Met protein overexpressing NSCLC). The exposure-efficacy analysis is a 193-patient subset of LUMINOSITY only – the phase 1 study was excluded because its c-Met overexpression criteria differed – and the exposure-safety analysis uses 284 patients pooled across both studies.

The three covariate medians that all six PK covariate terms are normalised to come from Table S4 (All Participants, N = 304) and are independently confirmed by Table S3, whose forest-plot reference strata are split at exactly those values: body weight 68.9 kg, baseline albumin 41.6 g/L and age 65 years.

Source trace

Every value in the six ini() blocks, with where it came from.

Source trace for every ini() value in the six model files.
Model Parameter Value Source location
conjugate lcl 1.34 L/day Table S7, ‘CL (L/day)’
conjugate lvc 3.44 L Table S7, ‘Vc (L)’
conjugate lq 1.21 L/day Table S7, ‘Q (L/day)’
conjugate lvp 2.41 L Table S7, ‘Vp (L)’
conjugate e_wt_cl_q 0.405 Table S7, ‘Body Weight on CL and Q’
conjugate e_wt_vc_vp 0.590 Table S7, ‘Body Weight on Vc and Vp’
conjugate e_ada_cl 1.17 Table S7, ‘ADA on CL’
conjugate e_alb_cl -0.970 Table S7, ‘Albumin on CL’
conjugate e_black_cl 0.763 Table S7, ‘Black or African American vs. White on CL’
conjugate e_asian_cl 0.887 Table S7, ‘Asian vs. White on CL’
conjugate e_age_vc 0.223 Table S7, ‘Age on Vc’
conjugate e_alb_vc -0.464 Table S7, ‘Albumin on Vc’
conjugate e_sexf_vc 0.925 Table S7, ‘Sex on Vc’
conjugate etalcl, etalvc 0.0946, 0.0268 Table S7, ‘IIV on CL’ / ‘IIV on Vc’ (variances)
conjugate propSd, addSd sqrt(0.0528), sqrt(0.0447) Table S7, error rows (variances)
payload lka 0.154 /day Table S7, ‘Ka (1/day)’
payload lcl 76.3 L/day Table S7, ‘CL (L/day)’
payload lvc 96.0 L Table S7, ‘Vc (L)’
payload e_alb_cl 1.99 Table S7, ‘Albumin on CL’
payload e_renalmild_cl 0.842 Table S7, ‘Mild vs. Normal Renal Impairment on CL’
payload e_renalmodsev_cl 0.755 Table S7, ‘Moderate/Severe vs. Normal Renal Impairment on CL’
payload e_age_ka -0.454 Table S7, ‘Age on Ka’
payload e_alb_ka -1.01 Table S7, ‘Albumin on Ka’
payload e_black_ka 0.853 Table S7, ‘Black or African American vs. White on Ka’
payload e_asian_ka 0.874 Table S7, ‘Asian vs. White on Ka’
payload e_wt_vc 0.612 Table S7, ‘Body Weight on Vc’
payload etalcl, etalvc, etalka 0.240, 0.486, 0.0735 Table S7, IIV rows (variances)
payload propSd, addSd sqrt(0.0833), sqrt(7.53e-10) Table S7, error rows (variances)
ORR logit_ref, e_cav_orr -5.292, 2.450 DIGITISED, Figure 2 fitted line
neuropathy logit_ref, e_cav_pn -7.437, 2.793 DIGITISED, Figure 4 left fitted line
corneal logit_ref, e_cav_ce -9.786, 4.206 DIGITISED, Figure 4 right fitted line
TEAE logit_ref, e_cav_teae -1.0835, 0.7195 DIGITISED, Figure 5 CavgMMAE fitted line
all ER addSd_prob_* fixed(0)-like 0.001 NOT from source; placeholder so rxode2 accepts an observation

Every PK model equation in model() is either the standard two- or one-compartment mass-balance system drawn in Figure S1, or a covariate term of the form the Methods prescribes: “Continuous covariates were normalized to the median of the overall population and incorporated into the model using a power function. Categorical covariates were included multiplicatively to obtain the proportional change between the tested categorical groups.” The only hardcoded constants are the three normalisation medians (68.9 kg, 41.6 g/L, 65 years), all sourced above.

Check 1 – the IIV column of Table S7 is a variance, not a standard deviation

Table S7 prints a “Population Estimate” and a “%CV” for each random effect, and its footnote states the relationship: “%CV was calculated as SQRT(exp(omega2)-1)*100”. That footnote makes the table self-pinning: if the Population Estimate column is the variance omega^2, the formula must reproduce the printed %CV exactly. It does, for all five random effects, which is why the ini() eta lines carry the printed numbers unchanged rather than squared.

iiv <- tibble::tribble(
  ~Model,      ~Parameter, ~omega2,  ~cv_printed,
  "conjugate", "CL",       0.0946,   31.5,
  "conjugate", "Vc",       0.0268,   16.5,
  "payload",   "CL",       0.240,    52.1,
  "payload",   "Vc",       0.486,    79.1,
  "payload",   "Ka",       0.0735,   27.6
) |>
  dplyr::mutate(
    cv_from_variance = sqrt(exp(omega2) - 1) * 100,
    cv_if_it_were_sd = sqrt(exp(omega2^2) - 1) * 100
  )

knitr::kable(iiv, digits = 3,
             caption = "Table S7 footnote formula applied both ways.")
Table S7 footnote formula applied both ways.
Model Parameter omega2 cv_printed cv_from_variance cv_if_it_were_sd
conjugate CL 0.095 31.5 31.499 9.481
conjugate Vc 0.027 16.5 16.481 2.680
payload CL 0.240 52.1 52.082 24.350
payload Vc 0.486 79.1 79.108 51.616
payload Ka 0.074 27.6 27.617 7.360 

stopifnot(
  # The variance reading reproduces every printed %CV to the printed precision.
  all(abs(iiv$cv_from_variance - iiv$cv_printed) < 0.05),
  # The standard-deviation reading does not, for any of them.
  all(abs(iiv$cv_if_it_were_sd - iiv$cv_printed) > 1)
)

Check 2 – Table S7 point estimates are consistent with their own %RSE and 95% CI

A transcription check on the structural parameters: for a Wald interval the 95% CI half-width should be about 1.96 standard errors, and the standard error is %RSE / 100 times the estimate. Agreement confirms that the estimate, the %RSE and the interval were all read from the same row.

rse <- tibble::tribble(
  ~Model,      ~Parameter, ~est,   ~rse_pct, ~lo,    ~hi,
  "conjugate", "CL",       1.34,   2.98,     1.26,   1.42,
  "conjugate", "Vc",       3.44,   1.38,     3.35,   3.54,
  "conjugate", "Q",        1.21,   4.59,     1.11,   1.33,
  "conjugate", "Vp",       2.41,   2.68,     2.28,   2.54,
  "payload",   "CL",       76.3,   4.83,     69.4,   83.9,
  "payload",   "Vc",       96.0,   5.62,     86.0,   107,
  "payload",   "Ka",       0.154,  2.58,     0.147,  0.162
) |>
  dplyr::mutate(
    halfwidth_ci   = (hi - lo) / 2,
    halfwidth_wald = 1.96 * rse_pct / 100 * est,
    ratio          = halfwidth_ci / halfwidth_wald
  )

knitr::kable(rse, digits = 4,
             caption = "Printed 95% CI half-width against 1.96 x SE from the printed %RSE.")
Printed 95% CI half-width against 1.96 x SE from the printed %RSE.
Model Parameter est rse_pct lo hi halfwidth_ci halfwidth_wald ratio
conjugate CL 1.340 2.98 1.260 1.420 0.0800 0.0783 1.0221
conjugate Vc 3.440 1.38 3.350 3.540 0.0950 0.0930 1.0210
conjugate Q 1.210 4.59 1.110 1.330 0.1100 0.1089 1.0105
conjugate Vp 2.410 2.68 2.280 2.540 0.1300 0.1266 1.0269
payload CL 76.300 4.83 69.400 83.900 7.2500 7.2232 1.0037
payload Vc 96.000 5.62 86.000 107.000 10.5000 10.5746 0.9929
payload Ka 0.154 2.58 0.147 0.162 0.0075 0.0078 0.9631 

stopifnot(
  # Centre: a mis-transcribed estimate or %RSE would move the whole column.
  abs(median(rse$ratio) - 1) < 0.10,
  # Envelope: robust to one asymmetric or profile-likelihood interval.
  quantile(abs(rse$ratio - 1), 0.9) < 0.25
)

Virtual cohort

The cohort reproduces the marginal covariate distributions of Table S4 (All Participants). Continuous covariates are drawn from truncated normals with the printed mean, standard deviation and min/max; categorical covariates from the printed proportions. Two arms of 200 subjects each, at the two LUMINOSITY dose levels, with the same covariate draws in both arms so that the dose comparison is paired.

set.seed(20260212)
n_sub <- 200L

rtrunc_norm <- function(n, mean, sd, lo, hi) {
  x <- stats::rnorm(n, mean, sd)
  pmin(pmax(x, lo), hi)
}

race <- sample(c("White", "Asian", "Black"), n_sub, replace = TRUE,
               prob = c(0.65, 0.32, 0.03))
renal <- sample(c("normal", "mild", "moderate", "severe"), n_sub, replace = TRUE,
                prob = c(0.36, 0.43, 0.19, 0.01))

cohort <- tibble::tibble(
  id            = seq_len(n_sub),
  WT            = rtrunc_norm(n_sub, 70.7, 16.7, 36.0, 144),   # Table S4 mean (SD), min/max
  AGE           = rtrunc_norm(n_sub, 63,   10,   30,   87),    # Table S4
  ALB           = rtrunc_norm(n_sub, 40.9, 4.40, 29.0, 52.0),  # Table S4
  SEXF          = stats::rbinom(n_sub, 1, 0.378),              # Table S4, 115/304 female
  RACE_BLACK    = as.integer(race == "Black"),
  RACE_ASIAN    = as.integer(race == "Asian"),
  ADA_POS       = stats::rbinom(n_sub, 1, 61 / 304),           # Figure 1A, 61 positive of 304
  RENALIMP_MILD = as.integer(renal == "mild"),
  RENALIMP_MOD  = as.integer(renal == "moderate"),
  RENALIMP_SEV  = as.integer(renal == "severe")
)

# 1.9 mg/kg, capped at 190 mg for patients of at least 100 kg (Abstract).
teliso_dose <- function(wt, mgkg) pmin(mgkg * wt, 190)

knitr::kable(
  cohort |>
    dplyr::summarise(
      `WT median`  = median(WT), `AGE median` = median(AGE),
      `ALB median` = median(ALB), `Female %`  = 100 * mean(SEXF),
      `Asian %`    = 100 * mean(RACE_ASIAN), `ADA+ %` = 100 * mean(ADA_POS)
    ),
  digits = 1, caption = "Simulated cohort against the Table S4 medians 68.9 kg, 65 years, 41.6 g/L."
)
Simulated cohort against the Table S4 medians 68.9 kg, 65 years, 41.6 g/L.
WT median AGE median ALB median Female % Asian % ADA+ %
72.9 63.1 40.5 37 36 23

Conjugate concentration-time profiles

Eight every-2-week doses, an assumed 30-minute infusion (Babel 2026 states “intravenous infusion” and its PK sampling schedule references “the end of infusion” but never prints a duration – see Assumptions below; the conjugate has no absorption step, so the duration affects only the very first minutes of the peak).

tau      <- 14      # days
n_dose   <- 8L
inf_dur  <- 0.5 / 24  # 30 min, expressed in days

sim_arm <- function(mgkg, seed) {
  rxode2::rxSetSeed(seed)   # common random numbers per arm
  amt <- teliso_dose(cohort$WT, mgkg)
  dose <- data.frame(
    id   = rep(cohort$id, each = n_dose),
    time = rep(seq(0, by = tau, length.out = n_dose), times = n_sub),
    amt  = rep(amt, each = n_dose),
    rate = rep(amt / inf_dur, each = n_dose),
    evid = 1L,
    cmt  = "central"
  )
  # Dense observations over cycle 1 and over the eighth (near-steady-state)
  # interval; cmt is the ODE STATE, and rxode2 returns Cc alongside it.
  # The grid is deliberately fine over the first day of each interval: the
  # 30-minute infusion and the fast distribution phase make a 0.25-day
  # trapezoid underestimate the early AUC by several percent, which would
  # show up as a false failure of the mass-balance identity below.
  fine <- c(seq(0, 0.06, by = 0.005), seq(0.08, 1, by = 0.02))
  obs_t <- c(fine, seq(1.25, tau, by = 0.25),
             (n_dose - 1) * tau + fine,
             seq((n_dose - 1) * tau + 1.25, n_dose * tau, by = 0.25))
  obs <- data.frame(
    id   = rep(cohort$id, each = length(obs_t)),
    time = rep(obs_t, times = n_sub),
    amt  = 0, rate = 0, evid = 0L, cmt = "central"
  )
  ev <- dplyr::bind_rows(dose, obs) |>
    dplyr::arrange(id, time, dplyr::desc(evid)) |>
    dplyr::left_join(cohort, by = "id")
  out <- as.data.frame(rxode2::rxSolve(mod_adc, events = ev, returnType = "data.frame"))
  out$dose_mgkg <- mgkg
  out
}

sim19 <- sim_arm(1.9, 20260212L)
#> ℹ parameter labels from comments will be replaced by 'label()'
sim16 <- sim_arm(1.6, 20260212L)
sim_adc <- dplyr::bind_rows(sim19, sim16)
#> Warning in ggplot2::scale_y_log10(): log-10 transformation introduced infinite values.
#> log-10 transformation introduced infinite values.
#> log-10 transformation introduced infinite values.
#> log-10 transformation introduced infinite values.
Simulated Teliso-V conjugate concentrations, median and 5th-95th percentile band, both LUMINOSITY dose levels. Compare with the prediction-corrected VPC of Figure S4.

Simulated Teliso-V conjugate concentrations, median and 5th-95th percentile band, both LUMINOSITY dose levels. Compare with the prediction-corrected VPC of Figure S4.

Check 3 – mass balance: AUC(0, inf) x CL equals the dose exactly

A closed-form identity that holds for any linear compartmental system at any time: the amount eliminated up to time T equals the dose in minus the amount still in the body. Evaluated per subject on the same solve that produced the profiles, so the two sides are numerically independent only through the ODE solver – which is exactly what this checks. A tight bound is correct here because the only source of disagreement is integration error.

mb <- sim19 |>
  dplyr::filter(time <= tau) |>
  dplyr::group_by(id) |>
  dplyr::summarise(
    auc     = sum(diff(time) * (head(Cc, -1) + tail(Cc, -1)) / 2),
    remain  = dplyr::last(central) + dplyr::last(peripheral1),
    cl_i    = dplyr::first(cl),        # the individual's own clearance
    .groups = "drop"
  ) |>
  dplyr::left_join(cohort, by = "id") |>
  dplyr::mutate(
    dose    = teliso_dose(WT, 1.9),
    balance = (auc * cl_i + remain) / dose
  )

stopifnot(
  abs(median(mb$balance) - 1) < 0.01,
  quantile(abs(mb$balance - 1), 0.9) < 0.02
)
knitr::kable(
  tibble::tibble(
    Quantity = c("median (eliminated + remaining) / dose", "90th pct |deviation|"),
    Value    = c(median(mb$balance), quantile(abs(mb$balance - 1), 0.9))
  ),
  digits = 4,
  caption = "Mass balance of the conjugate model over cycle 1, per subject."
)
Mass balance of the conjugate model over cycle 1, per subject.
Quantity Value
median (eliminated + remaining) / dose 1.0006
90th pct |deviation| 0.0010

This identity uses each subject’s own realised clearance, so it tests the ODE system and the observation equation Cc = central / vc but not the covariate algebra. The next check does that separately, and deterministically.

Check 3b – covariate algebra reproduces the Table S7 formulas exactly

Solving with the random effects zeroed makes each subject’s cl, vc, q and vp deterministic functions of the covariates, so they can be compared against the Table S7 numbers recomputed independently here. A tight bound is correct: the two sides should agree to machine precision or the model file has a wrong exponent, a wrong reference value, or a covariate attached to the wrong parameter.

ev_typ <- dplyr::bind_rows(
  data.frame(id = cohort$id, time = 0, amt = 1, rate = 1 / inf_dur,
             evid = 1L, cmt = "central"),
  data.frame(id = cohort$id, time = 1, amt = 0, rate = 0, evid = 0L, cmt = "central")
) |>
  dplyr::arrange(id, time, dplyr::desc(evid)) |>
  dplyr::left_join(cohort, by = "id")

typ <- as.data.frame(
  rxode2::rxSolve(rxode2::zeroRe(mod_adc), events = ev_typ, returnType = "data.frame")
) |>
  dplyr::distinct(WT, AGE, ALB, SEXF, RACE_BLACK, RACE_ASIAN, ADA_POS,
                  cl, vc, q, vp) |>
  dplyr::mutate(
    cl_ref = 1.34 * (WT / 68.9)^0.405 * (ALB / 41.6)^-0.970 *
             1.17^ADA_POS * 0.763^RACE_BLACK * 0.887^RACE_ASIAN,
    vc_ref = 3.44 * (WT / 68.9)^0.590 * (AGE / 65)^0.223 *
             (ALB / 41.6)^-0.464 * 0.925^SEXF,
    q_ref  = 1.21 * (WT / 68.9)^0.405,
    vp_ref = 2.41 * (WT / 68.9)^0.590
  )
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc'
#> Warning: multi-subject simulation without without 'omega'

stopifnot(
  max(abs(typ$cl / typ$cl_ref - 1)) < 1e-8,
  max(abs(typ$vc / typ$vc_ref - 1)) < 1e-8,
  max(abs(typ$q  / typ$q_ref  - 1)) < 1e-8,
  max(abs(typ$vp / typ$vp_ref - 1)) < 1e-8
)

knitr::kable(
  tibble::tibble(
    Parameter = c("CL", "Vc", "Q", "Vp"),
    `Max relative deviation` = c(
      max(abs(typ$cl / typ$cl_ref - 1)), max(abs(typ$vc / typ$vc_ref - 1)),
      max(abs(typ$q  / typ$q_ref  - 1)), max(abs(typ$vp / typ$vp_ref - 1))
    ),
    `Typical-covariate value` = c(1.34, 3.44, 1.21, 2.41),
    `Model at the medians` = c(
      1.34 * 1, 3.44 * 1, 1.21 * 1, 2.41 * 1
    )
  ),
  digits = 12,
  caption = "Model covariate algebra against the Table S7 formulas, over all 200 simulated covariate vectors."
)
Model covariate algebra against the Table S7 formulas, over all 200 simulated covariate vectors.
Parameter Max relative deviation Typical-covariate value Model at the medians
CL 0 1.34 1.34
Vc 0 3.44 3.44
Q 0 1.21 1.21
Vp 0 2.41 2.41

PKNCA analysis of the simulated conjugate profiles 

conc_c1 <- sim19 |>
  dplyr::filter(time <= tau, !is.na(Cc)) |>
  dplyr::select(id, time, Cc) |>
  dplyr::distinct(id, time, .keep_all = TRUE)

# Defensive time-zero record so PKNCA never asks for an AUC starting before the
# first measurement.
conc_c1 <- conc_c1 |>
  dplyr::bind_rows(
    tibble::tibble(id = unique(conc_c1$id), time = 0, Cc = 0)
  ) |>
  dplyr::distinct(id, time, .keep_all = TRUE) |>
  dplyr::arrange(id, time)

dose_c1 <- cohort |>
  dplyr::transmute(id, time = 0, amt = teliso_dose(WT, 1.9), treatment = "1.9 mg/kg Q2W")

conc_c1$treatment <- "1.9 mg/kg Q2W"

o_conc <- PKNCA::PKNCAconc(conc_c1, Cc ~ time | id / treatment)
o_dose <- PKNCA::PKNCAdose(dose_c1, amt ~ time | id + treatment)
o_data <- PKNCA::PKNCAdata(
  o_conc, o_dose,
  intervals = data.frame(
    start = 0, end = tau,
    cmax = TRUE, tmax = TRUE, auclast = TRUE, cav = TRUE, half.life = TRUE
  )
)
res <- suppressWarnings(PKNCA::pk.nca(o_data))
nca <- as.data.frame(res)
Non-compartmental analysis of the simulated cycle-1 conjugate profiles at 1.9 mg/kg Q2W.
NCA parameter Median P25 P75
adj.r.squared 1.000 1.000 1.000
auclast 90.085 74.239 111.751
cav 6.435 5.303 7.982
clast.pred 0.892 0.481 1.778
cmax 39.378 34.401 45.559
half.life 3.563 3.085 4.368
lambda.z 0.195 0.159 0.225
lambda.z.n.points 38.000 37.000 40.000
lambda.z.time.first 4.750 4.250 5.000
lambda.z.time.last 14.000 14.000 14.000
r.squared 1.000 1.000 1.000
span.ratio 2.549 2.122 3.077
tlast 14.000 14.000 14.000
tmax 0.025 0.025 0.025

Comparison against the published exposure distribution

Babel 2026 tabulates no conjugate NCA parameters, but the legend of Figure 3 prints the cycle-1 average conjugate concentration (CavgC1) quartile boundaries of the 193-patient exposure-efficacy analysis set, in ug/mL: Q1 [2.20, 4.71], Q2 [4.71, 6.17], Q3 [6.17, 7.29], Q4 [7.29, 10.8]. PKNCA’s cav over the interval [0, 14] days is the same quantity, so the two can be compared directly.

cav_sim <- nca |>
  dplyr::filter(PPTESTCD == "cav") |>
  dplyr::pull(PPORRES)

cav_cmp <- tibble::tibble(
  Quantile  = c("25th (Q1/Q2 boundary)", "50th (Q2/Q3 boundary)", "75th (Q3/Q4 boundary)"),
  Simulated = as.numeric(quantile(cav_sim, c(0.25, 0.5, 0.75))),
  Published = c(4.71, 6.17, 7.29)
) |>
  dplyr::mutate(`Ratio` = Simulated / Published)

knitr::kable(cav_cmp, digits = 3,
             caption = "Simulated cycle-1 Cavg quartile boundaries against the Figure 3 legend.")
Simulated cycle-1 Cavg quartile boundaries against the Figure 3 legend.
Quantile Simulated Published Ratio
25th (Q1/Q2 boundary) 5.303 4.71 1.126
50th (Q2/Q3 boundary) 6.435 6.17 1.043
75th (Q3/Q4 boundary) 7.982 7.29 1.095 

stopifnot(
  # Centre: a mis-transcribed CL, dose or unit would move the whole
  # distribution by tens of percent.
  abs(cav_cmp$Ratio[2] - 1) < 0.20,
  # Envelope, robust to which subjects land in the tails.
  all(abs(cav_cmp$Ratio - 1) < 0.30)
)

The simulated quartiles sit slightly above the published ones, and in a direction the paper explains: its exposure metrics are computed “utilizing actual doses received by patients” over the whole treatment period, whereas the simulation gives every subject the full nominal dose at every cycle. The published analysis set also mixes in 25 patients dosed at 1.6 mg/kg.

Median simulated cycle-1 Cavg against the Figure 3 published median (Q2/Q3 boundary).
NCA parameter Reference Simulated % diff
Cavg 6.17 6.43 +4.3%

Check 4 – covariate effects reproduce the Figure 1 forest plot

For the conjugate, AUCtau depends on clearance only, so a purely categorical covariate on CL implies an exposure ratio of exactly 1 / factor when nothing else changes. Figure 1A prints the simulated geometric-mean ratios, which do include the covariate correlations of the real cohort, so exact agreement is not expected – but the sign and the order of magnitude are a real check.

Univariate implication of the Table S7 categorical CL factors against the Figure 1A simulated ratios.
Comparison Table S7 factor on CL Implied AUCtau ratio Figure 1A AUCtau ratio
ADA positive vs negative 1.170 0.855 0.846
Asian vs White 0.887 1.127 1.100
Black or African American vs White 0.763 1.311 1.560

The ADA and Asian rows agree closely (0.855 against 0.846; 1.127 against 1.10). The Black or African American row does not (1.31 against 1.56), which is expected: that stratum has only 8 patients, whose weights and albumin values enter the simulated ratio alongside the race factor itself. Babel 2026 draws the same conclusion, noting that those patients’ exposures “were within the variability of exposures in White patients” and that “the small number of Black or African American patients (n = 8) limits conclusions on relevance”.

Unconjugated MMAE payload

The payload model is flip-flop: the deconjugation rate is much slower than MMAE elimination, so the apparent terminal half-life of the payload is set by ka.

ka_typ  <- 0.154
cl_typ  <- 76.3
vc_typ  <- 96.0
kel_typ <- cl_typ / vc_typ

payload_halflives <- tibble::tibble(
  Quantity = c("deconjugation half-life = log(2)/ka",
               "MMAE elimination half-life = log(2)/(CL/Vc)"),
  Days     = c(log(2) / ka_typ, log(2) / kel_typ)
)
knitr::kable(payload_halflives, digits = 3,
             caption = "The payload system is flip-flop: input is slower than elimination.")
The payload system is flip-flop: input is slower than elimination.
Quantity Days
deconjugation half-life = log(2)/ka 4.501
MMAE elimination half-life = log(2)/(CL/Vc) 0.872 

stopifnot(ka_typ < kel_typ)

Because Babel 2026 does not report the drug-antibody ratio, the molecular weights, or the systemically available payload fraction, the depot of Babel_2026_telisotuzumab_mmae must be dosed in MMAE-equivalent mass, not in milligrams of conjugate. The vignette therefore solves the payload model per unit MMAE-equivalent dose for the structural checks, and only afterwards applies an illustrative conversion for plotting.

rxode2::rxSetSeed(20260213L)
obs_t_m <- c(seq(0, 1, by = 0.02), seq(1.25, tau, by = 0.25))
ev_m <- dplyr::bind_rows(
  data.frame(id = cohort$id, time = 0, amt = 1, evid = 1L, cmt = "depot"),
  data.frame(
    id   = rep(cohort$id, each = length(obs_t_m)),
    time = rep(obs_t_m, times = n_sub),
    amt  = 0, evid = 0L, cmt = "central"
  )
) |>
  dplyr::arrange(id, time, dplyr::desc(evid)) |>
  dplyr::left_join(cohort, by = "id")

sim_mmae <- as.data.frame(
  rxode2::rxSolve(mod_mmae, events = ev_m, returnType = "data.frame")
)
#> ℹ parameter labels from comments will be replaced by 'label()'

Check 5 – payload mass balance 

mb_m <- sim_mmae |>
  dplyr::group_by(id) |>
  dplyr::summarise(
    auc    = sum(diff(time) * (head(Cc, -1) + tail(Cc, -1)) / 2),
    remain = dplyr::last(central) + dplyr::last(depot),
    cl_i   = dplyr::first(cl),
    .groups = "drop"
  ) |>
  dplyr::mutate(balance = (auc * cl_i + remain) / 1)

stopifnot(
  abs(median(mb_m$balance) - 1) < 0.01,
  quantile(abs(mb_m$balance - 1), 0.9) < 0.03
)
knitr::kable(
  tibble::tibble(
    Quantity = c("median (eliminated + remaining) / dose", "90th pct |deviation|"),
    Value    = c(median(mb_m$balance), quantile(abs(mb_m$balance - 1), 0.9))
  ),
  digits = 4,
  caption = "Mass balance of the payload model over 14 days, per unit MMAE-equivalent dose."
)
Mass balance of the payload model over 14 days, per unit MMAE-equivalent dose.
Quantity Value
median (eliminated + remaining) / dose 0.9998
90th pct |deviation| 0.0003

Exposure-response

How the four logistic models were recovered

Babel 2026 reports a nominal p value for each of the four relationships but tabulates no regression coefficients anywhere – not in the article, not in Data S1. The only representation of the fitted models is the solid line in each figure panel. The library admits a figure-derived parameter exactly when there is no competing printed value, which is the case here.

Two properties of this paper make the digitisation sound rather than a guess:

  1. The plotted curve is the final model. The usual objection to digitising an exposure-response panel is that the plotted line is the unadjusted, exposure-only fit while the tabulated model is covariate-adjusted. Babel 2026 states that “no covariates were found to have a significant effect on efficacy or safety”, so adjusted and unadjusted coincide for every endpoint.
  2. The functional form is decidable from the figure itself. The Methods evaluated “linear and logarithmic logistic regression analyses”; fitting both to each digitised curve separates them decisively, and the winner differs between the conjugate endpoints and the payload endpoint.
Root-mean-square residual of each candidate form against the digitised curve, in percentage points of probability. The selected form is the smaller of the two in every panel.
Panel Selected form RMS residual, log form (pp) RMS residual, linear form (pp)
Figure 2 (ORR) logarithmic 0.186 0.703
Figure 4 left (PN) logarithmic 0.083 0.247
Figure 4 right (CE) logarithmic 0.078 0.441
Figure 5 (TEAE) linear 7.598 0.053

The payload panel is additionally decidable by eye: its curve meets the left axis at a finite 25.3%, whereas any logarithmic form is pinned to zero there.

Replicating the published curves 

#> Warning: multi-subject simulation without without 'omega'
#> Warning: multi-subject simulation without without 'omega'
#> Warning: multi-subject simulation without without 'omega'
Replicates Figure 2 and Figure 4 of Babel 2026: fitted exposure-response curves against conjugate CavgADC, with the binned observed proportions read off the same panels.

Replicates Figure 2 and Figure 4 of Babel 2026: fitted exposure-response curves against conjugate CavgADC, with the binned observed proportions read off the same panels. 

#> Warning: multi-subject simulation without without 'omega'
Replicates the CavgMMAE panel of Figure 5 of Babel 2026.

Replicates the CavgMMAE panel of Figure 5 of Babel 2026.

Check 6 – the exposure ratio implied by Table 1 equals the dose ratio

This is the sharpest available gate on the digitised slopes, and it does not involve the intercepts at all. Table 1 prints the simulated median probability of each endpoint at 1.6 and at 1.9 mg/kg Q2W. For a logistic in log(Cavg),

logit(p_1.9) - logit(p_1.6) = b * log(Cavg_1.9 / Cavg_1.6)

so the printed probability pair and the digitised b together imply an exposure ratio, which must equal the dose ratio 1.9/1.6 = 1.1875 because conjugate clearance is linear.

logit <- function(p) log(p / (1 - p))

ratio_chk <- tibble::tribble(
  ~Endpoint,                          ~p16,    ~p19,    ~b,
  "ORR per ICR",                       0.222,   0.314,   2.450,
  "Grade >= 3 peripheral neuropathy",  0.0610,  0.0960,  2.793,
  "Grade >= 2 corneal epitheliopathy", 0.0780,  0.150,   4.206
) |>
  dplyr::mutate(
    `Implied exposure ratio` = exp((logit(p19) - logit(p16)) / b),
    `Dose ratio`             = 1.9 / 1.6,
    `Relative error`         = `Implied exposure ratio` / `Dose ratio` - 1
  )

knitr::kable(ratio_chk, digits = 4,
             caption = "Table 1 printed probabilities plus the digitised slope imply the dose ratio.")
Table 1 printed probabilities plus the digitised slope imply the dose ratio.
Endpoint p16 p19 b Implied exposure ratio Dose ratio Relative error
ORR per ICR 0.222 0.314 2.450 1.2127 1.1875 0.0213
Grade >= 3 peripheral neuropathy 0.061 0.096 2.793 1.1924 1.1875 0.0041
Grade >= 2 corneal epitheliopathy 0.078 0.150 4.206 1.1910 1.1875 0.0030 

stopifnot(all(abs(ratio_chk$`Relative error`) < 0.05))

All three land within 2.2% of the nominal dose ratio – 0.6% and 0.2% for the two safety endpoints – on a quantity that was not used to fit the curves.

Check 7 – absolute probabilities against Table 1 at the published exposure

Check 6 gates the slopes without touching the intercepts; this one gates the intercepts. Both sides are printed quantities: the exposure is the Figure 3 legend’s median cycle-1 conjugate Cavg of 6.17 ug/mL at the LUMINOSITY dose levels, scaled dose-proportionally to the 1.6 mg/kg arm (conjugate clearance is linear, and non-linear clearance was tested and rejected), and the target is Table 1.

cav19_pub <- 6.17                      # Figure 3 legend, CavgC1 Q2/Q3 boundary
cav16_pub <- cav19_pub * 1.6 / 1.9     # linear clearance

table1 <- tibble::tribble(
  ~Endpoint,                           ~Model,  ~Output,                               ~p16_pub, ~p19_pub,
  "ORR per ICR",                       "orr",   "prob_orr_central",                     22.2,     31.4,
  "Grade >= 3 peripheral neuropathy",  "pn",    "prob_peripheral_neuropathy_grade3",     6.10,     9.60,
  "Grade >= 2 corneal epitheliopathy", "ce",    "prob_corneal_epitheliopathy_grade2",    7.80,    15.0
)
er_mods <- list(orr = mod_orr, pn = mod_pn, ce = mod_ce)

table1 <- table1 |>
  dplyr::rowwise() |>
  dplyr::mutate(
    p16_mod = 100 * solve_prob(er_mods[[Model]], Output, cav16_pub),
    p19_mod = 100 * solve_prob(er_mods[[Model]], Output, cav19_pub)
  ) |>
  dplyr::ungroup() |>
  dplyr::mutate(d16 = p16_mod - p16_pub, d19 = p19_mod - p19_pub)

knitr::kable(
  table1 |>
    dplyr::select(Endpoint, p16_pub, p16_mod, d16, p19_pub, p19_mod, d19) |>
    dplyr::rename(
      "Published 1.6 (%)" = p16_pub, "Model 1.6 (%)" = p16_mod,
      "Difference 1.6 (pp)" = d16,
      "Published 1.9 (%)" = p19_pub, "Model 1.9 (%)" = p19_mod,
      "Difference 1.9 (pp)" = d19
    ),
  digits = 2,
  caption = "Table 1 of Babel 2026 against the digitised logistic models evaluated at the Figure 3 published median exposure."
)
Table 1 of Babel 2026 against the digitised logistic models evaluated at the Figure 3 published median exposure.
Endpoint Published 1.6 (%) Model 1.6 (%) Difference 1.6 (pp) Published 1.9 (%) Model 1.9 (%) Difference 1.9 (pp)
ORR per ICR 22.2 22.19 -0.01 31.4 30.29 -1.11
Grade >= 3 peripheral neuropathy 6.1 5.55 -0.55 9.6 8.67 -0.93
Grade >= 2 corneal epitheliopathy 7.8 5.44 -2.36 15.0 10.60 -4.40 

stopifnot(
  # Centre: a wrong intercept or a wrong exposure scale moves every row.
  abs(median(c(table1$d16, table1$d19))) < 3,
  # Envelope: the digitised intercepts carry a few percentage points of error.
  max(abs(c(table1$d16, table1$d19))) < 6
)

Two caveats keep this from being a tighter gate than it is. Table 1’s exposure metric is the whole-treatment Cavg, not the cycle-1 Cavg whose quartiles Figure 3 prints, and the two differ by the accumulation ratio; and Table 1’s simulation resamples 500 patients from both studies while Figure 3’s quartiles describe the 193-patient efficacy set. The corneal row carries the largest residual, consistent with the internal spread of the digitisation: solving each endpoint’s own intercept for the exposure that would reproduce its Table 1 value gives 6.42 ug/mL from the neuropathy model and 6.78 ug/mL from the corneal model, a 5.5% spread on two endpoints fitted to the same analysis set. That spread is the practical accuracy of a figure-derived intercept.

The simulated cohort’s own exposure, for reference

Steady-state conjugate Cavg of the simulated cohort, at nominal dosing.
Dose (mg/kg Q2W) Median steady-state Cavg (ug/mL) P25 P75
1.6 5.72 4.58 7.46
1.9 6.78 5.44 8.86

These sit above the published exposure used in Check 7, for the reason already given under Check 4: the paper’s metrics use the doses patients actually received across the whole treatment period, and Babel 2026’s own Discussion notes that the positive exposure-safety relationships “are managed in the clinical setting through dose modifications and/or dose reductions”. A simulation at full nominal dosing has no such attrition, so its exposure is an upper bound on the observed one and must not be substituted into a published-probability gate.

Payload exposure and the TEAE model 

# NOT a model parameter and NOT a published value. Babel 2026 never reports the
# drug-antibody ratio, the molecular weights, or the systemically available
# payload fraction, so this factor was back-solved from the CavgMMAE axis of
# Figure 5 purely so that the payload model can be plotted on a clinical dose.
# It is deliberately kept in the vignette and out of the model file.
payload_fraction_illustrative <- 0.0131

cav_mmae_ngml <-
  1000 * payload_fraction_illustrative * teliso_dose(68.9, 1.9) / (76.3 * tau)

teae_p <- 100 * solve_prob(mod_teae, "prob_teae_grade3", cav_mmae_ngml)

At the median weight of 68.9 kg and the illustrative payload fraction, the 1.9 mg/kg Q2W regimen gives a steady-state payload Cavg of 1.61 ng/mL, inside the 0-4 ng/mL range of Figure 5, and a predicted grade >= 3 TEAE probability of 51.8%. Because the fraction was calibrated to that same figure, the payload concentration level is not an independent check; only the shape of the profile and the mass-balance identity above are.

Assumptions and deviations

  • The payload depot is dosed in MMAE-equivalent mass, not in milligrams of conjugate. Babel 2026’s Table S7 payload clearance (76.3 L/day) and volume (96.0 L) are physiologically sized for MMAE itself, but the paper reports neither the drug-antibody ratio nor the molecular weights nor the systemically available payload fraction, so the conversion from an administered conjugate dose cannot be reconstructed from anything on disk. The model file states this in its description, its units$dosing and its compartmentData. The vignette’s illustrative fraction of 1.31% was back-solved from the Figure 5 CavgMMAE axis and is deliberately not part of the packaged model, following the precedent of Wang_2024_omega3PUFA.Rmd.
  • The payload CL-Vc random-effect correlation is not reproduced. Babel 2026 Results states that the payload model included “the inclusion of correlation between CL and Vc”, but Table S7 tabulates only the three diagonal variances and no covariance or correlation coefficient. The omega block is therefore left diagonal rather than invented. Simulated payload variability is consequently slightly mis-shaped, although the marginal variances are correct.
  • The four exposure-response models carry digitised, not printed, coefficients. See the Exposure-response section for the method, the form-selection residuals and the two independent cross-checks against Table 1. Treat the slopes as accurate to a few percent and the intercepts to a few percentage points of probability. Every ini() line and the model description carries this provenance.
  • Moderate and severe renal impairment share one estimated factor. Babel 2026 pooled the two strata (only 2 of 304 patients had severe impairment) and printed a single “Moderate/Severe vs. Normal” value. The model keeps RENALIMP_MOD and RENALIMP_SEV as separate columns so user data can retain the clinical distinction, and applies the one published factor to both.
  • Infusion duration is assumed to be 30 minutes. Babel 2026 describes the route as “an intravenous infusion” and its sampling schedule references “the end of infusion” but never prints a duration. The conjugate model has no absorption step, so the assumption affects only the shape of the first few minutes of the peak and not AUC.
  • Treatment-emergent ADA status is carried as a subject-level indicator. Nominally the covariate is time-varying, since a patient becomes ADA positive at seroconversion. Babel 2026 does not state whether it was implemented as a time-varying flag or an ever-positive subject-level flag, so the simpler reading is used.
  • The conjugate’s additive residual row is printed as “Additional Error (Variance)”. The payload block of the same table prints “Additive Error (Variance)” for the identical quantity; the conjugate wording is read as a typographical slip and the value is used as an additive error.
  • Grade >= 2 peripheral neuropathy, all-grade and grade >= 2 ILD / pneumonitis, dose interruption or discontinuation due to an adverse event, DCR, and the DoR / PFS / OS Cox models are all reported as significant but are not plotted with a fitted curve, so no coefficients are recoverable for them and they are not packaged. Figure 5’s three additional payload metrics (CavgC1MMAE, CmingmMMAE, CmaxgmMMAE) are plotted, but they describe the same endpoint through a different exposure metric, so only the Cavg panel is packaged in order to keep the exposure metric aligned with the three conjugate-driven models.
  • The virtual cohort uses independent marginal covariate draws. Table S4 reports marginal distributions only; the real correlations between body weight, sex, albumin, age and renal function are not published, so the cohort cannot reproduce them. This matters for the Figure 1 forest comparison, which is why that section is presented descriptively rather than as a gate.
  • New canonical names introduced by this extraction. prob_peripheral_neuropathy_grade3 and prob_corneal_epitheliopathy_grade2 are registered in inst/references/compartment-names.md, following the prob_<endpoint> shape founded by prob_roc. The four exposure-response models reuse the existing CAV covariate canonical, and their entries were added to that column’s example-model list; note that this paper is the first in the register where a single CAV name carries two different analytes and two different units within one publication.
  • Residual error on the exposure-response models. The source fits binomial logistic regressions with an exact Bernoulli likelihood and estimates no residual error. Each model emits its probability with a tiny fixed additive placeholder so that rxode2 accepts an observation declaration; this does not perturb the predicted probability. To simulate binary outcomes, apply rbinom(n, 1, prob_*) to the rxSolve() output.