Skip to contents

Model and source

Yang 2025 reports the exposure-response (E-R) analyses that support the approved axatilimab regimen of 0.3 mg/kg IV every 2 weeks (Q2W) in chronic graft-versus-host disease (cGVHD). Axatilimab is a humanized IgG4 monoclonal antibody against the colony-stimulating factor 1 receptor (CSF-1R). The paper fits twelve independent binomial logistic regressions with R’s glm(family = binomial):

  • two efficacy endpoints (overall response by Cycle 7 Day 1, and a >= 7-point improvement in the modified Lee Symptom Scale, mLSS) in 239 AGAVE-201 patients, driven by the single-dose AUC from time 0 to infinity (AUCSDINF);
  • ten safety endpoints in 278 patients pooled from SNDX-6352-0503 and AGAVE-201, driven by the steady-state AUC over a 4-week treatment cycle (AUCSS).

Duration of response (Cox model) and infections of unspecified etiology had no significant exposure association; the paper prints no coefficients for them, so they are not packaged.

  • Article: https://doi.org/10.1002/cpt.3731 (PMC12439004)
  • Coefficients: Supplementary Tables S5 (efficacy) and S7 (safety); the analysis scripts are Supplementary Data S2 (safety) and S3 (efficacy).
  • Upstream population PK/PD model (source of every exposure metric): https://doi.org/10.1002/cpt.3503, packaged as modellib("Yang_2024_axatilimab").

Each model has no PK layer: the exposure metric is a data column (AUC_AXATILIMAB_SD or AUC_AXATILIMAB_SS, in ug*day/mL). The probability output is prob_<endpoint> and the linear predictor logit_<endpoint>.

stems <- c(
  orr = "Yang_2025_axatilimab_orr",
  mlss_improve7 = "Yang_2025_axatilimab_mlss_improve7",
  teae_grade3 = "Yang_2025_axatilimab_teae_grade3",
  teae_dosemod = "Yang_2025_axatilimab_teae_dosemod",
  aesi_grade3 = "Yang_2025_axatilimab_aesi_grade3",
  sae = "Yang_2025_axatilimab_sae",
  teae_drug_related = "Yang_2025_axatilimab_teae_drug_related",
  amylase_lipase_increase = "Yang_2025_axatilimab_amylase_lipase_increase",
  cpk_increase = "Yang_2025_axatilimab_cpk_increase",
  liver_enzyme_increase = "Yang_2025_axatilimab_liver_enzyme_increase",
  periorbital_edema = "Yang_2025_axatilimab_periorbital_edema",
  irr = "Yang_2025_axatilimab_irr"
)
endpoint_label <- c(
  orr = "Overall response",
  mlss_improve7 = ">= 7-point improvement in mLSS",
  teae_grade3 = "Grade >= 3 TEAEs",
  teae_dosemod = "TEAEs leading to dose modification",
  aesi_grade3 = "Grade >= 3 AESI",
  sae = "Serious TEAEs",
  teae_drug_related = "Treatment-related TEAEs",
  amylase_lipase_increase = "Amylase and lipase increases",
  cpk_increase = "CPK elevations",
  liver_enzyme_increase = "Liver enzyme elevations",
  periorbital_edema = "Periorbital edema",
  irr = "Infusion-related reactions"
)
efficacy <- c("orr", "mlss_improve7")
mods <- lapply(stems, function(s) rxode2::rxode2(readModelDb(s)))

Population

The efficacy population is 239 AGAVE-201 patients (79, 81 and 79 on 0.3 mg/kg Q2W, 1 mg/kg Q2W and 3 mg/kg Q4W): median age 53 years (range 7-81), median weight 71.6 kg (18.1-151), 37.7% female, 82.8% White, 79.9% severe cGVHD, 32.2% with more than 4 involved organs (Table S2). The safety population adds 39 SNDX-6352-0503 patients for 278 in total (79, 107, 85, 6 and 2 on 0.3 mg/kg Q2W, 1 mg/kg Q2W, 3 mg/kg Q4W, 3 mg/kg Q2W and lower doses): median age 55 years, median weight 72.4 kg, 37.4% female, Karnofsky/Lansky score <= 70 in 35.7% (Table S3). The same information is in each model’s population metadata.

Source trace

Every coefficient is a Table S5 / S7 “Estimate”; main-text Tables 1 and 2 print the same fits as odds ratios (per 50 ug*day/mL for exposure). The model equation, logit(p) = alpha + sum(beta_k * X_k), is the Supplemental Methods logistic-regression equation; median-centring of the continuous covariates is in the Data S2 script (mutate(across(all_of(cont.predictors), ~ . - median(.)))).

trace <- bind_rows(lapply(names(mods), function(ep) {
  ini <- mods[[ep]]$iniDf
  ini <- ini[!is.na(ini$ntheta) & !grepl("^addSd", ini$name), ]
  data.frame(
    endpoint = endpoint_label[[ep]],
    parameter = ini$name,
    value = signif(ini$est, 4),
    source = ifelse(ep %in% efficacy, "Yang 2025 Table S5", "Yang 2025 Table S7")
  )
}))
trace |>
  dplyr::rename(
    "Endpoint" = endpoint, "Parameter" = parameter,
    "Value" = value, "Source" = source
  ) |>
  knitr::kable()
Endpoint Parameter Value Source
Overall response logit_ref 0.899000 Yang 2025 Table S5
Overall response e_auc_axatilimab_sd_logit -0.001570 Yang 2025 Table S5
>= 7-point improvement in mLSS logit_ref -0.114000 Yang 2025 Table S5
>= 7-point improvement in mLSS e_auc_axatilimab_sd_logit -0.001180 Yang 2025 Table S5
>= 7-point improvement in mLSS e_cgvhd_organs_gt4_logit 0.770000 Yang 2025 Table S5
Grade >= 3 TEAEs logit_ref -0.048200 Yang 2025 Table S7
Grade >= 3 TEAEs e_auc_axatilimab_ss_logit 0.001620 Yang 2025 Table S7
TEAEs leading to dose modification logit_ref -0.425000 Yang 2025 Table S7
TEAEs leading to dose modification e_auc_axatilimab_ss_logit 0.001090 Yang 2025 Table S7
Grade >= 3 AESI logit_ref -0.763000 Yang 2025 Table S7
Grade >= 3 AESI e_auc_axatilimab_ss_logit 0.001550 Yang 2025 Table S7
Serious TEAEs logit_ref -0.736000 Yang 2025 Table S7
Serious TEAEs e_auc_axatilimab_ss_logit 0.000546 Yang 2025 Table S7
Serious TEAEs e_alb_logit -0.096400 Yang 2025 Table S7
Serious TEAEs e_ecog_ge2_logit 0.687000 Yang 2025 Table S7
Treatment-related TEAEs logit_ref 0.927000 Yang 2025 Table S7
Treatment-related TEAEs e_auc_axatilimab_ss_logit 0.002120 Yang 2025 Table S7
Amylase and lipase increases logit_ref -1.790000 Yang 2025 Table S7
Amylase and lipase increases e_auc_axatilimab_ss_logit 0.002920 Yang 2025 Table S7
Amylase and lipase increases e_lipase_logit 0.013900 Yang 2025 Table S7
CPK elevations logit_ref -2.140000 Yang 2025 Table S7
CPK elevations e_auc_axatilimab_ss_logit 0.004200 Yang 2025 Table S7
CPK elevations e_cpk_logit 0.010200 Yang 2025 Table S7
Liver enzyme elevations logit_ref -0.985000 Yang 2025 Table S7
Liver enzyme elevations e_auc_axatilimab_ss_logit 0.002100 Yang 2025 Table S7
Periorbital edema logit_ref -2.010000 Yang 2025 Table S7
Periorbital edema e_auc_axatilimab_ss_logit 0.002360 Yang 2025 Table S7
Infusion-related reactions logit_ref -1.020000 Yang 2025 Table S7
Infusion-related reactions e_auc_axatilimab_ss_logit 0.001700 Yang 2025 Table S7
Infusion-related reactions e_alt_logit -0.017800 Yang 2025 Table S7

Structural verification against the published tables

These checks are deterministic: each evaluates the packaged model at published covariate values, so there is no cohort and no random draw.

# One observation row per grid point on a single subject; the covariates are
# columns of the event table (a static model cannot take a params grid).
er_eval <- function(ep, covs) {
  covs <- as.data.frame(covs)
  ev <- data.frame(id = 1L, time = seq_len(nrow(covs)), amt = 0, evid = 0L)
  ev <- cbind(ev, covs)
  s <- as.data.frame(rxode2::rxSolve(mods[[ep]], events = ev, returnType = "data.frame"))
  stopifnot(nrow(s) == nrow(covs))
  s[[paste0("prob_", ep)]]
}
# Every covariate at its reference (centring) value; exposure supplied separately.
ref_covs <- list(ALB = 40, LIPASE = 33, CPK = 64, ALT = 30, ECOG_GE2 = 0, CGVHD_ORGANS_GT4 = 0)
needs <- function(ep) mods[[ep]]$allCovs
make_covs <- function(ep, auc, ...) {
  out <- data.frame(row = seq_along(auc))
  exp_col <- if (ep %in% efficacy) "AUC_AXATILIMAB_SD" else "AUC_AXATILIMAB_SS"
  out[[exp_col]] <- auc
  extra <- list(...)
  for (cn in setdiff(needs(ep), exp_col)) {
    out[[cn]] <- if (!is.null(extra[[cn]])) extra[[cn]] else ref_covs[[cn]]
  }
  out$row <- NULL
  out
}

Check 1 – odds ratios per 50 ug*day/mL

The exposure odds ratio is exp(50 * slope); both columns are rounded in the paper, so agreement is to the third decimal.

pub_or <- c(
  orr = 0.924, mlss_improve7 = 0.942, teae_grade3 = 1.086, teae_dosemod = 1.055,
  aesi_grade3 = 1.08, sae = 1.028, teae_drug_related = 1.112,
  amylase_lipase_increase = 1.157, cpk_increase = 1.235,
  liver_enzyme_increase = 1.109, periorbital_edema = 1.125, irr = 1.088
)
or_tab <- bind_rows(lapply(names(mods), function(ep) {
  p <- er_eval(ep, make_covs(ep, c(0, 50)))
  odds <- p / (1 - p)
  data.frame(endpoint = endpoint_label[[ep]], model_or = odds[2] / odds[1], published_or = pub_or[[ep]])
}))
or_tab |>
  mutate(model_or = round(model_or, 4)) |>
  dplyr::rename("Endpoint" = endpoint, "Model OR per 50" = model_or, "Published OR" = published_or) |>
  knitr::kable()
Endpoint Model OR per 50 Published OR
Overall response 0.9245 0.924
>= 7-point improvement in mLSS 0.9427 0.942
Grade >= 3 TEAEs 1.0844 1.086
TEAEs leading to dose modification 1.0560 1.055
Grade >= 3 AESI 1.0806 1.080
Serious TEAEs 1.0277 1.028
Treatment-related TEAEs 1.1118 1.112
Amylase and lipase increases 1.1572 1.157
CPK elevations 1.2337 1.235
Liver enzyme elevations 1.1107 1.109
Periorbital edema 1.1252 1.125
Infusion-related reactions 1.0887 1.088
stopifnot(all(abs(or_tab$model_or - or_tab$published_or) < 0.003))

Check 2 – covariate centring is pinned by the percentile odds ratios

Tables 1 and 2 print the odds ratio at the 5th and 95th percentiles of each covariate relative to its median. Reproducing both percentiles with one reference value proves the centring value (40 g/L albumin, 33 U/L lipase, 64 U/L CPK, 30 U/L ALT). The exposure odds ratios at its percentiles are relative to zero exposure (the exposure term is not centred).

pct <- tribble(
  ~ep, ~cov, ~p05, ~p95, ~or05, ~or95,
  "sae", "ALB", 32.085, 47, 2.144, 0.509,
  "amylase_lipase_increase", "LIPASE", 11.85, 107.9, 0.748, 2.799,
  "cpk_increase", "CPK", 25.85, 224.45, 0.675, 5.21,
  "irr", "ALT", 11.85, 82.3, 1.368, 0.405,
  "orr", "AUC_AXATILIMAB_SD", 8.99, 714.525, 0.986, 0.325,
  "mlss_improve7", "AUC_AXATILIMAB_SD", 8.99, 715, 0.989, 0.427,
  "teae_grade3", "AUC_AXATILIMAB_SS", 18.876, 1442.416, 1.031, 10.023,
  "teae_dosemod", "AUC_AXATILIMAB_SS", 19.413, 1315.382, 1.022, 4.277,
  "aesi_grade3", "AUC_AXATILIMAB_SS", 18.197, 1333.464, 1.029, 8.015,
  "sae", "AUC_AXATILIMAB_SS", 18.197, 1112.353, 1.01, 1.817,
  "teae_drug_related", "AUC_AXATILIMAB_SS", 19.254, 1403.762, 1.042, 19.645,
  "amylase_lipase_increase", "AUC_AXATILIMAB_SS", 18.197, 1403.762, 1.055, 61.692,
  "cpk_increase", "AUC_AXATILIMAB_SS", 18.01, 1426.255, 1.079, 401.082,
  "liver_enzyme_increase", "AUC_AXATILIMAB_SS", 18.609, 1341.576, 1.04, 16.826,
  "periorbital_edema", "AUC_AXATILIMAB_SS", 18.197, 1233.407, 1.044, 18.541,
  "irr", "AUC_AXATILIMAB_SS", 18.272, 1339.036, 1.032, 9.749
)
pct_or <- function(ep, cov, x) {
  is_exp <- grepl("^AUC_", cov)
  base <- if (is_exp) 0 else ref_covs[[cov]]
  covs <- make_covs(ep, if (is_exp) c(base, x) else c(0, 0))
  if (!is_exp) covs[[cov]] <- c(base, x)
  p <- er_eval(ep, covs)
  (p[2] / (1 - p[2])) / (p[1] / (1 - p[1]))
}
pct$model05 <- mapply(pct_or, pct$ep, pct$cov, pct$p05)
pct$model95 <- mapply(pct_or, pct$ep, pct$cov, pct$p95)
pct |>
  mutate(endpoint = endpoint_label[ep], across(c(model05, model95), ~ signif(.x, 4))) |>
  select(endpoint, cov, or05, model05, or95, model95) |>
  dplyr::rename(
    "Endpoint" = endpoint, "Predictor" = cov, "Published OR 5th" = or05,
    "Model OR 5th" = model05, "Published OR 95th" = or95, "Model OR 95th" = model95
  ) |>
  knitr::kable()
Endpoint Predictor Published OR 5th Model OR 5th Published OR 95th Model OR 95th
Serious TEAEs ALB 2.144 2.1450 0.509 0.5093
Amylase and lipase increases LIPASE 0.748 0.7453 2.799 2.8320
CPK elevations CPK 0.675 0.6776 5.210 5.1380
Infusion-related reactions ALT 1.368 1.3810 0.405 0.3942
Overall response AUC_AXATILIMAB_SD 0.986 0.9860 0.325 0.3257
>= 7-point improvement in mLSS AUC_AXATILIMAB_SD 0.989 0.9894 0.427 0.4301
Grade >= 3 TEAEs AUC_AXATILIMAB_SS 1.031 1.0310 10.023 10.3500
TEAEs leading to dose modification AUC_AXATILIMAB_SS 1.022 1.0210 4.277 4.1940
Grade >= 3 AESI AUC_AXATILIMAB_SS 1.029 1.0290 8.015 7.9000
Serious TEAEs AUC_AXATILIMAB_SS 1.010 1.0100 1.817 1.8360
Treatment-related TEAEs AUC_AXATILIMAB_SS 1.042 1.0420 19.645 19.6100
Amylase and lipase increases AUC_AXATILIMAB_SS 1.055 1.0550 61.692 60.2800
CPK elevations AUC_AXATILIMAB_SS 1.079 1.0790 401.082 399.5000
Liver enzyme elevations AUC_AXATILIMAB_SS 1.040 1.0400 16.826 16.7300
Periorbital edema AUC_AXATILIMAB_SS 1.044 1.0440 18.541 18.3700
Infusion-related reactions AUC_AXATILIMAB_SS 1.032 1.0320 9.749 9.7410
# The published percentile ORs are medians of resampled fits, and the 95th
# percentile multiplies the rounding of the slope by ~1400, so a relative
# tolerance is used. A wrong centring value (e.g. 63 instead of 64 U/L CPK)
# or an uncentred covariate fails by far more.
stopifnot(
  all(abs(pct$model05 / pct$or05 - 1) < 0.02),
  all(abs(pct$model95 / pct$or95 - 1) < 0.05)
)

Check 3 – predicted efficacy at the exposure percentiles (Table 1, Results)

The Results print the predicted probability of each efficacy response at the median, 5th and 95th percentile of AUCSDINF (patients with 4 or fewer involved organs for mLSS).

eff_pts <- data.frame(
  ep = rep(efficacy, each = 3),
  auc = c(64.0, 8.99, 714.525, 64.0, 8.99, 715),
  published = c(0.69, 0.708, 0.444, 0.454, 0.468, 0.277)
)
eff_pts$model <- unlist(lapply(efficacy, function(ep) {
  er_eval(ep, make_covs(ep, eff_pts$auc[eff_pts$ep == ep]))
}))
eff_pts |>
  mutate(endpoint = endpoint_label[ep], model = round(model, 4)) |>
  select(endpoint, auc, published, model) |>
  dplyr::rename(
    "Endpoint" = endpoint, "AUCSDINF (ug*day/mL)" = auc,
    "Published probability" = published, "Model probability" = model
  ) |>
  knitr::kable()
Endpoint AUCSDINF (ug*day/mL) Published probability Model probability
Overall response 64.000 0.690 0.6897
Overall response 8.990 0.708 0.7078
Overall response 714.525 0.444 0.4445
>= 7-point improvement in mLSS 64.000 0.454 0.4528
>= 7-point improvement in mLSS 8.990 0.468 0.4689
>= 7-point improvement in mLSS 715.000 0.277 0.2773
stopifnot(all(abs(eff_pts$model - eff_pts$published) < 0.006))

Check 4 – body-weight quartiles at 0.3 mg/kg Q2W (Figure 5)

The paper evaluates the final models at the median exposure of the first and fourth body-weight quartiles (AUCSDINF 9.98 and 15.2 ugday/mL; AUCSS 18.3 and 29.8 ugday/mL) and reports a CPK-elevation probability of 0.113 versus 0.118 and a relative change below 0.4% for both efficacy endpoints.

cpk_q <- er_eval("cpk_increase", make_covs("cpk_increase", c(18.3, 29.8)))
orr_q <- er_eval("orr", make_covs("orr", c(9.98, 15.2)))
mlss_q <- er_eval("mlss_improve7", make_covs("mlss_improve7", c(9.98, 15.2)))
data.frame(
  endpoint = c("CPK elevations", "Overall response", ">= 7-point improvement in mLSS"),
  q1 = round(c(cpk_q[1], orr_q[1], mlss_q[1]), 4),
  q4 = round(c(cpk_q[2], orr_q[2], mlss_q[2]), 4)
) |>
  mutate(relative_change_pct = round(100 * (q4 / q1 - 1), 2)) |>
  dplyr::rename(
    "Endpoint" = endpoint, "Weight Q1" = q1, "Weight Q4" = q4,
    "Relative change (%)" = relative_change_pct
  ) |>
  knitr::kable()
Endpoint Weight Q1 Weight Q4 Relative change (%)
CPK elevations 0.1127 0.1177 4.44
Overall response 0.7075 0.7058 -0.24
>= 7-point improvement in mLSS 0.4686 0.4671 -0.32
stopifnot(
  abs(cpk_q[1] - 0.113) < 0.001, abs(cpk_q[2] - 0.118) < 0.001,
  abs(orr_q[2] / orr_q[1] - 1) < 0.004, abs(mlss_q[2] / mlss_q[1] - 1) < 0.004
)

Exposure driver: the Yang 2024 population PK/PD model

The paper derived every exposure metric from the companion population PK/PD model, which is packaged as Yang_2024_axatilimab. Two vignette-local additions turn it into an exposure calculator: an auc_c state that integrates Cc (mgh/L; divided by 24 for ugday/mL), and a floor of zero on the concentration entering the Hill term, because the solver can overshoot to a tiny negative concentration deep in the terminal phase and a negative base raised to the Hill coefficient 2.5 is NaN.

pk <- rxode2::rxode2(readModelDb("Yang_2024_axatilimab")) |>
  rxode2::model(Cc_nM <- max(Cc, 0) * 1000 / MW_PK) |>
  rxode2::model(d/dt(auc_c) <- Cc, append = Cp) |>
  rxode2::zeroRe()
#> ℹ parameter labels from comments will be replaced by 'label()'
pk_omega <- diag(rxode2::rxode2(readModelDb("Yang_2024_axatilimab"))$omega)
#> ℹ parameter labels from comments will be replaced by 'label()'
pk_covs <- c("WT", "CSF1", "CPK", "ADA_POS", "DIS_CANCER", "DIS_HEALTHY", names(pk_omega))

# Yang 2024 is a five-endpoint model, so observation rows must name the
# endpoint they belong to; cmt = "Cc" here selects the axatilimab-concentration
# endpoint (it is declared by the model's error lines, so no compartment slot
# is added). Doses go to the ODE state `central`.
# One dose list per subject; AUCSDINF = auc_c at 5 years after one dose,
# AUCSS = auc_c over the last 28 days of 104 weeks of dosing.
pk_events <- function(subj, single) {
  bind_rows(lapply(seq_len(nrow(subj)), function(i) {
    s <- subj[i, ]
    dose <- s$mgkg * s$WT
    if (single) {
      d_t <- 0
      o_t <- c(0, 5 * 365 * 24)
    } else {
      d_t <- seq(0, 728 * 24 - 1, by = s$ii * 24)
      o_t <- c(700, 728) * 24
    }
    ev <- data.frame(
      id = s$id, time = c(d_t, o_t), amt = c(rep(dose, length(d_t)), rep(0, length(o_t))),
      evid = c(rep(1L, length(d_t)), rep(0L, length(o_t))),
      cmt = c(rep("central", length(d_t)), rep("Cc", length(o_t)))
    )
    cbind(ev, s[rep(1, nrow(ev)), pk_covs, drop = FALSE])
  })) |>
    arrange(id, time, desc(evid))
}
pk_auc <- function(subj) {
  # A zeroRe() model solved for many subjects warns that no omega is present;
  # the etas are supplied as data columns, which is the intent.
  solve_one <- function(single) {
    suppressWarnings(rxode2::rxSolve(pk, pk_events(subj, single), returnType = "data.frame")) |>
      group_by(id) |>
      summarise(auc = (max(auc_c) - min(auc_c)) / 24, .groups = "drop")
  }
  subj |>
    left_join(solve_one(TRUE) |> rename(AUC_AXATILIMAB_SD = auc), by = "id") |>
    left_join(solve_one(FALSE) |> rename(AUC_AXATILIMAB_SS = auc), by = "id")
}

Typical-patient exposure by body-weight quartile

The paper reports median AUCSDINF of 9.98 and 15.2 ugday/mL (body-weight quartile medians 53.6 and 94.8 kg) and median AUCSS of 18.3 and 29.8 ugday/mL (53.8 and 97.6 kg) at 0.3 mg/kg Q2W. The typical patient here is ADA-negative with the efficacy-population median baseline CSF-1 of 584 pg/mL (Table S2) and all etas at zero.

typ <- data.frame(id = 1:4, WT = c(53.6, 94.8, 53.8, 97.6), mgkg = 0.3, ii = 14,
                  CSF1 = 584, CPK = 64, ADA_POS = 0, DIS_CANCER = 0, DIS_HEALTHY = 0)
for (nm in names(pk_omega)) typ[[nm]] <- 0
typ <- pk_auc(typ)
#> ℹ omega/sigma items treated as zero: 'etalvc', 'etalcl', 'etalvmax', 'etalbl_csf1', 'etalbl_ncmc', 'etalbl_ast', 'etalbl_cpk'
#> ℹ omega/sigma items treated as zero: 'etalvc', 'etalcl', 'etalvmax', 'etalbl_csf1', 'etalbl_ncmc', 'etalbl_ast', 'etalbl_cpk'
typ_cmp <- data.frame(
  metric = c("AUCSDINF Q1", "AUCSDINF Q4", "AUCSS Q1", "AUCSS Q4"),
  weight = typ$WT,
  published = c(9.98, 15.2, 18.3, 29.8),
  model = c(typ$AUC_AXATILIMAB_SD[1:2], typ$AUC_AXATILIMAB_SS[3:4])
) |>
  mutate(pct_diff = round(100 * (model / published - 1), 1), model = round(model, 2))
typ_cmp |>
  dplyr::rename(
    "Metric" = metric, "Weight (kg)" = weight, "Published median" = published,
    "Typical patient (model)" = model, "Difference (%)" = pct_diff
  ) |>
  knitr::kable()
Metric Weight (kg) Published median Typical patient (model) Difference (%)
AUCSDINF Q1 53.6 9.98 10.46 4.8
AUCSDINF Q4 94.8 15.20 15.04 -1.0
AUCSS Q1 53.8 18.30 19.93 8.9
AUCSS Q4 97.6 29.80 29.79 0.0
# The published values are medians over individual empirical Bayes estimates
# (a third of patients ADA-positive), the model value a single typical
# patient; a mis-scaled dose or unit would miss by a factor, not 30%.
stopifnot(all(abs(typ_cmp$pct_diff) < 30))

PKNCA cross-check of the AUC accumulator

The auc_c state is an exact integral of the model concentration. As an independent check, PKNCA computes AUC0-inf from a dense sampled profile of typical patients after a single dose.

nca_subj <- data.frame(id = 1:3, WT = 72.4, mgkg = c(0.3, 1, 3), ii = 14, CSF1 = 584,
                       CPK = 64, ADA_POS = 0, DIS_CANCER = 0, DIS_HEALTHY = 0)
for (nm in names(pk_omega)) nca_subj[[nm]] <- 0
nca_subj$treatment <- paste(nca_subj$mgkg, "mg/kg single dose")
grid_h <- sort(unique(c(seq(0, 48, by = 2), seq(48, 24 * 120, by = 12), 24 * seq(120, 720, by = 10))))
nca_ev <- bind_rows(lapply(seq_len(nrow(nca_subj)), function(i) {
  s <- nca_subj[i, ]
  ev <- data.frame(
    id = s$id, time = c(0, grid_h), amt = c(s$mgkg * s$WT, rep(0, length(grid_h))),
    evid = c(1L, rep(0L, length(grid_h))), cmt = c("central", rep("Cc", length(grid_h)))
  )
  cbind(ev, s[rep(1, nrow(ev)), pk_covs, drop = FALSE])
}))
nca_sim <- suppressWarnings(rxode2::rxSolve(pk, nca_ev, returnType = "data.frame")) |>
  left_join(nca_subj[, c("id", "treatment")], by = "id") |>
  mutate(time_d = time / 24)
#> ℹ omega/sigma items treated as zero: 'etalvc', 'etalcl', 'etalvmax', 'etalbl_csf1', 'etalbl_ncmc', 'etalbl_ast', 'etalbl_cpk'
conc <- nca_sim |>
  dplyr::filter(!is.na(Cc)) |>
  select(id, treatment, time_d, Cc)
dose_df <- nca_subj |>
  transmute(id, treatment, time_d = 0, amt = mgkg * WT)
o_conc <- PKNCAconc(conc, Cc ~ time_d | treatment + id)
o_dose <- PKNCAdose(dose_df, amt ~ time_d | treatment + id)
o_data <- PKNCAdata(o_conc, o_dose,
  intervals = data.frame(start = 0, end = Inf, cmax = TRUE, aucinf.obs = TRUE, half.life = TRUE)
)
nca_res <- as.data.frame(pk.nca(o_data))
nca_auc <- nca_res |>
  dplyr::filter(PPTESTCD == "aucinf.obs") |>
  select(id, treatment, pknca = PPORRES) |>
  left_join(nca_sim |> group_by(id) |> summarise(accumulator = max(auc_c) / 24), by = "id") |>
  mutate(pct_diff = round(100 * (pknca / accumulator - 1), 2))
nca_auc |>
  select(treatment, pknca, accumulator, pct_diff) |>
  mutate(across(c(pknca, accumulator), ~ round(.x, 2))) |>
  dplyr::rename(
    "Treatment" = treatment, "PKNCA AUC0-inf (ug*day/mL)" = pknca,
    "Accumulator AUC (ug*day/mL)" = accumulator, "Difference (%)" = pct_diff
  ) |>
  knitr::kable()
Treatment PKNCA AUC0-inf (ug*day/mL) Accumulator AUC (ug*day/mL) Difference (%)
0.3 mg/kg single dose 12.64 12.66 -0.16
1 mg/kg single dose 84.30 84.32 -0.03
3 mg/kg single dose 457.83 457.81 0.00
# Same parameters on both sides; only trapezoidal error separates them.
stopifnot(all(abs(nca_auc$pct_diff) < 2))

Virtual cohort and model-based regimen predictions (Figure 4)

The paper projected each regimen with 100 replicate populations of 500 patients whose covariates were bootstrapped from the study population. Here 200 virtual patients per regimen are drawn from distributions matched to Tables S2, S3 and 2. Their individual etas are drawn from the Yang 2024 omega matrix with R’s random number generator and supplied as data, so the cohort does not depend on the number of solver threads.

set.seed(20250619)
regimens <- data.frame(
  regimen = c("0.15 mg/kg Q2W", "0.3 mg/kg Q2W", "0.5 mg/kg Q2W", "1 mg/kg Q2W", "3 mg/kg Q2W", "3 mg/kg Q4W"),
  mgkg = c(0.15, 0.3, 0.5, 1, 3, 3),
  ii = c(14, 14, 14, 14, 14, 28)
)
n_per_arm <- 200
cohort <- regimens[rep(seq_len(nrow(regimens)), each = n_per_arm), ]
n <- nrow(cohort)
cohort$id <- seq_len(n)
cohort$WT <- pmin(pmax(exp(rnorm(n, log(72.4), 0.27)), 18.1), 151)
cohort$CSF1 <- exp(rnorm(n, log(584), 0.5))
cohort$CPK <- exp(rnorm(n, log(64), 0.65))
cohort$ALB <- pmin(pmax(rnorm(n, 40, 4.55), 22), 53)
cohort$LIPASE <- exp(rnorm(n, log(33), 0.67))
cohort$ALT <- exp(rnorm(n, log(30), 0.6))
cohort$ADA_POS <- rbinom(n, 1, 0.34)
cohort$ECOG_GE2 <- rbinom(n, 1, 0.357)
cohort$CGVHD_ORGANS_GT4 <- rbinom(n, 1, 0.322)
cohort$DIS_CANCER <- 0
cohort$DIS_HEALTHY <- 0
for (nm in names(pk_omega)) cohort[[nm]] <- rnorm(n, 0, sqrt(pk_omega[[nm]]))
cohort <- pk_auc(cohort)
#> ℹ omega/sigma items treated as zero: 'etalvc', 'etalcl', 'etalvmax', 'etalbl_csf1', 'etalbl_ncmc', 'etalbl_ast', 'etalbl_cpk'
#> ℹ omega/sigma items treated as zero: 'etalvc', 'etalcl', 'etalvmax', 'etalbl_csf1', 'etalbl_ncmc', 'etalbl_ast', 'etalbl_cpk'
stopifnot(!anyNA(cohort$AUC_AXATILIMAB_SD), !anyNA(cohort$AUC_AXATILIMAB_SS))

The efficacy regimens of AGAVE-201 give an AUCSDINF distribution to set against the Table 1 median (5th, 95th percentile) of 64.0 (8.99, 714.5).

eff_reg <- c("0.3 mg/kg Q2W", "1 mg/kg Q2W", "3 mg/kg Q4W")
q_sd <- quantile(cohort$AUC_AXATILIMAB_SD[cohort$regimen %in% eff_reg], c(0.05, 0.5, 0.95))
data.frame(percentile = c("5th", "50th", "95th"), published = c(8.99, 64.0, 714.525), cohort = unname(round(q_sd, 1))) |>
  dplyr::rename("Percentile" = percentile, "Published AUCSDINF" = published, "Virtual cohort AUCSDINF" = cohort) |>
  knitr::kable()
Percentile Published AUCSDINF Virtual cohort AUCSDINF
5th 8.990 6.8
50th 64.000 68.5
95th 714.525 726.1
stopifnot(abs(q_sd[["50%"]] / 64.0 - 1) < 0.3)
pred_cols <- lapply(names(mods), function(ep) {
  cov_df <- cohort[, needs(ep), drop = FALSE]
  er_eval(ep, cov_df)
})
names(pred_cols) <- names(mods)
pred <- bind_cols(cohort["regimen"], as.data.frame(pred_cols)) |>
  pivot_longer(-regimen, names_to = "ep", values_to = "prob") |>
  dplyr::filter(!(ep %in% efficacy) | regimen %in% eff_reg) |>
  group_by(regimen, ep) |>
  summarise(mean_prob = mean(prob), .groups = "drop")

pred |>
  mutate(endpoint = endpoint_label[ep], mean_prob = round(100 * mean_prob, 1)) |>
  select(endpoint, regimen, mean_prob) |>
  pivot_wider(names_from = regimen, values_from = mean_prob) |>
  dplyr::rename("Endpoint" = endpoint) |>
  knitr::kable(caption = "Mean predicted probability (%) by regimen")
Mean predicted probability (%) by regimen
Endpoint 0.15 mg/kg Q2W 0.3 mg/kg Q2W 0.5 mg/kg Q2W 1 mg/kg Q2W 3 mg/kg Q2W 3 mg/kg Q4W
Grade >= 3 AESI 32.1 32.6 33.6 42.5 74.4 54.7
Amylase and lipase increases 16.3 17.0 20.0 31.3 73.9 50.0
CPK elevations 13.6 15.0 16.8 30.9 79.3 53.8
Infusion-related reactions 25.2 26.2 27.2 36.1 72.2 50.2
Liver enzyme elevations 27.5 28.2 29.5 40.4 76.3 55.5
Periorbital edema 12.0 12.4 13.2 23.4 66.5 39.1
Serious TEAEs 36.9 37.4 38.6 42.0 61.3 48.5
TEAEs leading to dose modification 39.7 40.2 40.9 47.7 73.9 57.0
Treatment-related TEAEs 72.0 72.7 73.9 80.2 93.6 87.1
Grade >= 3 TEAEs 49.1 49.8 51.0 59.3 83.7 69.9
>= 7-point improvement in mLSS NA 51.8 NA 50.3 NA 41.4
Overall response NA 70.7 NA 68.4 NA 54.9
ggplot(pred |> mutate(endpoint = endpoint_label[ep]),
       aes(x = factor(regimen, levels = regimens$regimen), y = 100 * mean_prob)) +
  geom_col(fill = "steelblue") +
  facet_wrap(~endpoint, ncol = 3) +
  labs(x = NULL, y = "Mean predicted probability (%)",
       caption = "Replicates Figure 4 of Yang 2025 (point predictions only).") +
  theme_bw() +
  theme(axis.text.x = element_text(angle = 45, hjust = 1, size = 7))

The paper quotes these regimen-level predictions in the text:

getp <- function(ep, reg) 100 * pred$mean_prob[pred$ep == ep & pred$regimen == reg]
reg_cmp <- data.frame(
  quantity = c(
    "Overall response, 0.3 mg/kg Q2W", ">= 7-point mLSS improvement, 0.3 mg/kg Q2W",
    "Overall response, 3 mg/kg Q4W minus 0.3 mg/kg Q2W",
    ">= 7-point mLSS improvement, 3 mg/kg Q4W minus 0.3 mg/kg Q2W",
    "CPK elevations, 0.15 mg/kg Q2W", "CPK elevations, 3 mg/kg Q2W"
  ),
  published = c(70.8, 53.2, -14.1, -9.9, 13.9, 81.7),
  model = c(
    getp("orr", "0.3 mg/kg Q2W"), getp("mlss_improve7", "0.3 mg/kg Q2W"),
    getp("orr", "3 mg/kg Q4W") - getp("orr", "0.3 mg/kg Q2W"),
    getp("mlss_improve7", "3 mg/kg Q4W") - getp("mlss_improve7", "0.3 mg/kg Q2W"),
    getp("cpk_increase", "0.15 mg/kg Q2W"), getp("cpk_increase", "3 mg/kg Q2W")
  )
) |>
  mutate(model = round(model, 1), diff = model - published)
reg_cmp |>
  dplyr::rename("Quantity (percentage points)" = quantity, "Published" = published,
                "Model" = model, "Difference" = diff) |>
  knitr::kable()
Quantity (percentage points) Published Model Difference
Overall response, 0.3 mg/kg Q2W 70.8 70.7 -0.1
>= 7-point mLSS improvement, 0.3 mg/kg Q2W 53.2 51.8 -1.4
Overall response, 3 mg/kg Q4W minus 0.3 mg/kg Q2W -14.1 -15.8 -1.7
>= 7-point mLSS improvement, 3 mg/kg Q4W minus 0.3 mg/kg Q2W -9.9 -10.4 -0.5
CPK elevations, 0.15 mg/kg Q2W 13.9 13.6 -0.3
CPK elevations, 3 mg/kg Q2W 81.7 79.3 -2.4
# Cohort means, not extremes: the efficacy probabilities are flat at low
# exposure so the 0.3 mg/kg values hardly depend on the cohort; the 3 mg/kg
# CPK value depends on the right tail of AUCSS and gets a wider band.
stopifnot(
  abs(reg_cmp$diff[1]) < 3, abs(reg_cmp$diff[2]) < 4,
  abs(reg_cmp$diff[3]) < 4, abs(reg_cmp$diff[4]) < 4,
  abs(reg_cmp$diff[5]) < 4, abs(reg_cmp$diff[6]) < 12
)

Exposure-response curves (Figures 1, 2 and S3)

Typical-patient curves (continuous covariates at their reference values, binary covariates 0) over the exposure range the paper plots.

curve_df <- bind_rows(lapply(names(mods), function(ep) {
  grid <- if (ep %in% efficacy) seq(0, 1000, by = 10) else seq(0, 2000, by = 20)
  data.frame(endpoint = endpoint_label[[ep]], auc = grid, prob = er_eval(ep, make_covs(ep, grid)),
             metric = if (ep %in% efficacy) "AUCSDINF" else "AUCSS")
}))
ggplot(curve_df, aes(auc, prob)) +
  geom_line(linewidth = 0.8) +
  facet_wrap(~endpoint, ncol = 3, scales = "free_x") +
  scale_y_continuous(limits = c(0, 1)) +
  labs(x = "Axatilimab AUCSDINF (efficacy) or AUCSS (safety), ug*day/mL", y = "Probability",
       caption = "Replicates the prediction curves of Figures 1a-b, 2a-d and S3 of Yang 2025.") +
  theme_bw()

Assumptions and deviations

  • No PK layer, no Bernoulli likelihood. Each model is a typical-value probability. The addSd_prob_* residual is a fixed 0.001 placeholder so rxode2 can attach an error model; it is not a published quantity. The paper’s glm has no random effects.
  • Exposure units. Coefficients are per 1 ugday/mL (Table S5 / S7 print the per-unit estimate under a “50 ug/mLd” row label; the odds ratio column is per 50 ug*day/mL, as Check 1 confirms).
  • Covariate centring. Albumin, lipase, CPK and ALT enter as deviations from the analysis-population median (40 g/L, 33 U/L, 64 U/L, 30 U/L), as the analysis script does; the values are pinned by Check 2. The exposure term is not centred, so each intercept is the logit at zero exposure.
  • Karnofsky/Lansky score <= 70 is carried as ECOG_GE2. Under the Oken 1982 crosswalk a Karnofsky score of 60-70 is ECOG 2 (“unable to work”) and <= 50 is ECOG >= 3, matching the paper’s label “unable to work”.
  • Endpoint name mapping. “TEAEs leading to dose modifications” in this paper includes dose interruptions as well as discontinuations and reductions; it is carried as prob_teae_dosemod.
  • Virtual cohort. Covariate marginals are matched to Tables S2, S3 and 2 (log-normal CSF-1, CPK, lipase and ALT; normal albumin; Bernoulli ADA, performance-status and organ-involvement indicators) and drawn independently, without the correlations of the real population. All regimens share the safety-population covariate distributions.
  • Exposure derivation. The paper simulated AUCSS with each patient’s cycle-average dose up to the time of first event; the projections here and in the paper’s Figure 4 use the nominal regimen. AUCSS is the last 28-day cycle after 104 weeks of dosing, which is close to, but for the slowest clearers not fully at, steady state. ADA status is time-invariant (1 if ever positive), as in the paper’s exposure derivation.

Errata and source notes

  • Table 2 labels the albumin row “ALB, U/L”; Table S7 and the Discussion give g/L, which is the only unit consistent with the printed median of 40.
  • Table S7 marks “Amylase and lipase elevations” with the dagger footnote (“Including hypersensitivity reactions”); the main-text Table 2 marks it as a grouped adverse-event term, which is the intended footnote.
  • The Discussion gives the CPK-elevation odds-ratio increase per 50 ugday/mL as 23.5%; Table 2 prints 1.235 (23.5%) and the Results 23.8%. The model uses the Table S7 slope 0.0042 (exp(50 0.0042) = 1.234).
  • Table S3 prints a median baseline CPK of 63.0 U/L over the 272 patients with a value, while the Discussion and Table 2 use 64 U/L; both Table 2 percentile odds ratios are consistent with 64 only, so 64 is the centring value (the analysis dataset imputed the missing values).