Skip to contents

Source and models

Efzofitimod is an immunomodulatory Fc fusion protein (64.5 kDa) that binds neuropilin-2. Walker 2023 first fits a population PK model to a Phase 1 healthy-volunteer study and a Phase 1b/2a study in pulmonary sarcoidosis. It then uses post hoc individual exposures from that model in five separate exposure-response (E-R) regressions. The authors fitted the six models separately, so the package carries them as six files:

Model file What it is Source
Walker_2023_efzofitimod Three-compartment IV population PK Supplementary Table S.1, Results 3.2, Figure 2
Walker_2023_efzofitimod_ocs Linear: % change from baseline in post-taper mean daily OCS dose Supplementary Eq. A.1, Figure 4
Walker_2023_efzofitimod_fvcpp Linear: Week-24 ppFVC Supplementary Eq. A.2, Figure 5
Walker_2023_efzofitimod_ksqlung Linear: Week-24 KSQ-Lung score Supplementary Eq. A.3, Figure 6
Walker_2023_efzofitimod_fvcpp_mcid Logistic: P(ppFVC increase >= 2.5%) Supplementary Eq. A.4, Figures 7 and 9
Walker_2023_efzofitimod_ksqlung_mcid Logistic: P(KSQ-Lung increase >= 4 points) Supplementary Eq. A.5, Figures 8 and 9

The five E-R models have no PK layer. Exposure enters as the column CAV, the paper’s “time-averaged AUC”: an individual’s cumulative AUC over the analysis window divided by the window length in days. The paper prints its unit as ng*day/mL, but numerically it is an average concentration in ng/mL. The window is Day 51 to Week 24 (the post-taper period) for the OCS endpoint and Day 1 to Week 24 for the other four. Placebo subjects have CAV = 0. This vignette makes CAV from the PK model with PKNCA, the same way the source did.

Population

#> List of 13
#>  $ species       : chr "human"
#>  $ n_subjects    : int 49
#>  $ n_studies     : int 2
#>  $ n_observations: int 565
#>  $ age_range     : chr "Phase 1 mean 23.8 (SD 3.94) years; Phase 1b/2a mean 52.3 (SD 10.3) years"
#>  $ weight_range  : chr "55-157 kg"
#>  $ weight_median : chr "82 kg"
#>  $ sex_female_pct: num 53.1
#>  $ race_ethnicity: Named num [1:4] 61.2 22.4 10.2 6.1
#>   ..- attr(*, "names")= chr [1:4] "White" "Black" "Asian" "Other"
#>  $ disease_state : chr "Healthy volunteers (Phase 1, n = 24) and patients with pulmonary sarcoidosis on oral corticosteroids (Phase 1b/2a, n = 25)"
#>  $ dose_range    : chr "Phase 1: single 1-h IV infusion of 0.03, 0.1, 0.3, 1, 3 or 5 mg/kg. Phase 1b/2a: 1, 3 or 5 mg/kg IV every 4 wee"| __truncated__
#>  $ regions       : chr "Australia (Phase 1), United States (Phase 1b/2a)"
#>  $ notes         : chr "Demographics from Walker 2023 Table 3 (PK-evaluable population). Sex: 23 male (11 + 12), 26 female (13 + 13). R"| __truncated__

The PK analysis pooled 24 healthy volunteers (Phase 1, ATYR1923-C-001, Australia; single 1-h IV infusion of 0.03-5 mg/kg; rich sampling to 672 h) and 25 patients with pulmonary sarcoidosis (Phase 1b/2a, ATYR1923-C-002, NCT03824392, USA; 1, 3 or 5 mg/kg every 4 weeks for six doses; sparse sampling). That is 565 concentrations in total (Walker 2023 Table 1 and Results 3.1). Healthy volunteers were younger (mean 23.8 vs 52.3 years) and lighter (mean 74 vs 97.6 kg) than the patients (Table 3).

The E-R analysis used only the Phase 1b/2a efficacy-evaluable population: 32 participants who received at least four doses (11 placebo, 21 efzofitimod; Tables 1, 3 and 4). All were on oral corticosteroids (10-25 mg/day prednisone or equivalent), tapered to 5 mg/day between Day 15 and Day 50.

Source trace

Element Value Source location
Structure: 3-compartment, linear elimination from central – Results 3.2, Figure 2
lcl log(1.68 L/day) Table S.1 ‘CL’; Results 3.2
lvc log(3.94 L) Table S.1 ‘V1’; Results 3.2
lq log(2.4 L/day) Table S.1 ‘Q2’
lvp log(3.3 L) Table S.1 ‘V2’; Results 3.2
lq2 log(1.2 L/day) Table S.1 ‘Q3’
lvp2 log(7.43 L) Table S.1 ‘V3’; Results 3.2
e_crcl_cl 0.71 Table S.1 ‘CRCL_CL’
e_tbili_cl -0.17 Table S.1 ‘BILI_CL’
e_wt_vc 0.77 Table S.1 ‘BWT_V1’
e_wt_vp 1.52 Table S.1 ‘BWT_V2’
etalcl, etalvc, etalq, etalvp, etalq2, etalvp2 0.03, 0.01, 0.06, 0.1, 0 FIX, 0.01 Table S.1 omega^2 rows (variances)
expSdHv sqrt(0.01) = 0.1 Table S.1 ‘epsilon^2 (Phase 1)’, log-additive
expSdPat sqrt(1.63) = 1.277 Table S.1 ‘epsilon^2 (Phase 2)’, log-additive
WT reference 82 kg Results 3.2 (‘relative to the median value of 82 kg’)
CRCL reference 139 mL/min Not printed; geometric mean of the Results 3.2 10th/90th percentiles (103, 187)
TBILI reference 7.9 umol/L Not printed; geometric mean of the Results 3.2 10th/90th percentiles (3.2, 19.6)
ocsdosecfb_int, e_cav_ocsdosecfb -43.01, -0.001731 Supplementary Eq. A.1
fvcpp_int, e_cav_fvcpp, e_fvc_pctpred_fvcpp -17.15, 0.0005061, 1.182 Supplementary Eq. A.2
ksqlung_int, e_cav_ksqlung 50.1, 0.001935 Supplementary Eq. A.3
logite0, e_cav_logit, e_fvc_pctpred_logit (ppFVC MCID) -10.25, 0.0000953, 0.11 Supplementary Eq. A.4
logite0, e_cav_logit, e_ksq_lung_bl_logit (KSQ-Lung MCID) 6.38, 0.000219, -0.15 Supplementary Eq. A.5
Exposure metric AUC over window / window length Methods 2.3 equation; Results 3.3 (windows)

Population PK

Check 1 – terminal half-life

Table S.1 and Results 3.2 give a terminal half-life of 9 days. For the typical subject at the reference covariates, the half-lives follow from the eigenvalues of the three-compartment rate matrix.

th <- ui_pk$theta
cl <- exp(th[["lcl"]]); v1 <- exp(th[["lvc"]]); q <- exp(th[["lq"]])
v2 <- exp(th[["lvp"]]); q2 <- exp(th[["lq2"]]); v3 <- exp(th[["lvp2"]])
kmat <- matrix(c(
  -(cl + q + q2) / v1, q / v1, q2 / v1,
  q / v2, -q / v2, 0,
  q2 / v3, 0, -q2 / v3
), nrow = 3, byrow = TRUE)
thalf <- sort(log(2) / -eigen(kmat)$values)
thalf
#> [1] 0.3876676 1.9046924 9.0051138
# A deterministic function of the printed parameters: tight tolerance.
stopifnot(abs(thalf[3] - 9) < 0.05)

The terminal half-life is 9.005 days against the printed 9 days. The paper also quotes an “overall” half-life of 11.3 days “accounting for all phases of elimination” but does not say how it was computed. It is not log(2) * Vss / CL (6.05 days), so it is not checked here.

Virtual cohort

The cohort reproduces the Phase 1b/2a patients. Weight is drawn from Table 3 (mean 97.6 kg, SD 25.2), truncated to the 55-157 kg range of Results 3.2. Creatinine clearance and total bilirubin are drawn log-normally to match the pooled 10th-90th percentiles of Results 3.2 (103-187 mL/min and 3.2-19.6 umol/L). Each regimen gets 100 subjects. The six regimens are the three studied doses plus the untested regimens of Table 2: 7 mg/kg, a flat 450 mg, and a 900 mg loading dose followed by 4 mg/kg. Every dose is a 1-h IV infusion every 28 days, six doses in all.

set.seed(20231011)
n_per <- 100
regimens <- c("1 mg/kg", "3 mg/kg", "5 mg/kg", "7 mg/kg", "450 mg fix dose", "900 L+4mg/kg")
q10 <- stats::qnorm(0.9)
make_subjects <- function(reg, offset) {
  tibble::tibble(
    id = offset + seq_len(n_per),
    regimen = reg,
    WT = pmin(pmax(stats::rnorm(n_per, 97.6, 25.2), 55), 157),
    CRCL = exp(stats::rnorm(n_per, log(sqrt(103 * 187)), log(187 / 103) / (2 * q10))),
    TBILI = exp(stats::rnorm(n_per, log(sqrt(3.2 * 19.6)), log(19.6 / 3.2) / (2 * q10))),
    DIS_HEALTHY = 0L
  )
}
subjects <- dplyr::bind_rows(lapply(seq_along(regimens), function(i) {
  make_subjects(regimens[i], (i - 1) * n_per)
}))

dose_amounts <- function(reg, wt) {
  switch(reg,
    "1 mg/kg" = rep(1 * wt, 6),
    "3 mg/kg" = rep(3 * wt, 6),
    "5 mg/kg" = rep(5 * wt, 6),
    "7 mg/kg" = rep(7 * wt, 6),
    "450 mg fix dose" = rep(450, 6),
    "900 L+4mg/kg" = c(900, rep(4 * wt, 5))
  )
}
dose_times <- (0:5) * 28
# Dense sampling around each infusion, sparse during the terminal phase.
cycle_grid <- c(0, 1 / 24, 2 / 24, 0.25, 0.5, 1, 2, 3, 5, 7, 10, 14, 21)
obs_times <- sort(unique(c(outer(cycle_grid, dose_times, "+"), 50, 168)))
obs_times <- obs_times[obs_times <= 168]

doses <- dplyr::bind_rows(lapply(seq_len(nrow(subjects)), function(i) {
  tibble::tibble(
    id = subjects$id[i], regimen = subjects$regimen[i], time = dose_times,
    amt = dose_amounts(subjects$regimen[i], subjects$WT[i])
  )
})) |>
  dplyr::mutate(evid = 1L, cmt = "central", rate = amt * 24)
obs <- tidyr::expand_grid(id = subjects$id, time = obs_times) |>
  dplyr::mutate(evid = 0L, cmt = "central", amt = 0, rate = 0)
events <- dplyr::bind_rows(doses |> dplyr::select(-regimen), obs) |>
  dplyr::left_join(subjects, by = "id") |>
  dplyr::arrange(id, time, dplyr::desc(evid))
rxode2::rxSetSeed(20231011)
sim <- rxode2::rxSolve(mod_pk, events = events, returnType = "data.frame") |>
  as.data.frame() |>
  dplyr::left_join(subjects |> dplyr::select(id, regimen), by = "id")
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalq2'

Figure 3 – concentration-time profiles

Walker 2023 Figure 3 shows the mean concentration-time profile, with three visible phases. The plot below shows the simulated median and 90% interval over 24 weeks for each regimen.

sim |>
  dplyr::filter(time > 0) |>
  dplyr::group_by(regimen, time) |>
  dplyr::summarise(
    med = median(Cc), lo = quantile(Cc, 0.05), hi = quantile(Cc, 0.95),
    .groups = "drop"
  ) |>
  ggplot(aes(time, med, colour = regimen, fill = regimen)) +
  geom_ribbon(aes(ymin = lo, ymax = hi), alpha = 0.12, colour = NA) +
  geom_line() +
  scale_y_log10() +
  labs(
    x = "Time (days)", y = "Efzofitimod (ng/mL)", colour = NULL, fill = NULL,
    caption = "Simulated median and 5th-95th percentiles; 100 subjects per regimen."
  ) +
  theme_bw()

Check 2 – PKNCA: exposure windows and terminal half-life

PKNCA computes each subject’s AUC over the two E-R windows: Day 1 to Week 24 (times 0-168 days) and Day 51 to Week 24 (times 50-168 days). Dividing by the window length gives CAV, exactly as Methods 2.3 defines it. A single-dose run at the reference covariates also gives a terminal half-life to compare with the printed 9 days.

conc_obj <- PKNCA::PKNCAconc(
  sim |> dplyr::filter(!is.na(Cc)) |> dplyr::select(id, regimen, time, Cc),
  Cc ~ time | regimen + id
)
dose_obj <- PKNCA::PKNCAdose(doses, amt ~ time | regimen + id)
intervals <- data.frame(
  start = c(0, 50), end = c(168, 168),
  aucint.last = TRUE
)
nca <- PKNCA::pk.nca(PKNCA::PKNCAdata(conc_obj, dose_obj, intervals = intervals))
cav <- as.data.frame(nca$result) |>
  dplyr::filter(PPTESTCD == "aucint.last") |>
  dplyr::mutate(window = ifelse(start == 0, "cav_d1_w24", "cav_d51_w24"),
                CAV = PPORRES / (end - start)) |>
  dplyr::select(id, regimen, window, CAV) |>
  tidyr::pivot_wider(names_from = window, values_from = CAV)
single_obs <- c(0, seq(0.5, 84, by = 0.5))
ev_single <- dplyr::bind_rows(
  tibble::tibble(time = 0, evid = 1L, amt = 5 * 82, rate = 5 * 82 * 24),
  tibble::tibble(time = single_obs, evid = 0L, amt = 0, rate = 0)
) |>
  dplyr::mutate(
    id = 1L, cmt = "central", WT = 82, CRCL = 139, TBILI = 7.9, DIS_HEALTHY = 1L,
    treatment = "5 mg/kg single dose (reference subject)"
  )
# zeroRe() removes the residual error as well, which rxode2 reports; expected.
sim_single <- withCallingHandlers(
  rxode2::rxSolve(rxode2::zeroRe(mod_pk), events = ev_single, returnType = "data.frame"),
  warning = function(w) {
    if (grepl("No sigma parameters", conditionMessage(w))) invokeRestart("muffleWarning")
  }
) |> as.data.frame()
#> ℹ parameter labels from comments will be replaced by 'label()'
#> ℹ omega/sigma items treated as zero: 'etalcl', 'etalvc', 'etalq', 'etalvp', 'etalq2', 'etalvp2'
if (is.null(sim_single$id)) sim_single$id <- 1L
sim_single$treatment <- ev_single$treatment[1]
nca_single <- PKNCA::pk.nca(PKNCA::PKNCAdata(
  PKNCA::PKNCAconc(sim_single |> dplyr::filter(!is.na(Cc)), Cc ~ time | treatment + id),
  PKNCA::PKNCAdose(ev_single |> dplyr::filter(evid == 1), amt ~ time | treatment + id),
  intervals = data.frame(start = 0, end = Inf, cmax = TRUE, aucinf.obs = TRUE, half.life = TRUE)
))
knitr::kable(
  nlmixr2lib::ncaComparisonTable(
    simulated = nca_single,
    reference = data.frame(treatment = "5 mg/kg single dose (reference subject)", half.life = 9),
    by = "treatment",
    params = "half.life",
    units = c(half.life = "day")
  ),
  caption = "Terminal half-life from PKNCA against Walker 2023 Table S.1 'Terminal HL'."
)
Terminal half-life from PKNCA against Walker 2023 Table S.1 ‘Terminal HL’.
NCA parameter treatment Reference Simulated % diff
t½ (day) 5 mg/kg single dose (reference subject) 9 8.96 -0.5%

Check 3 – exposure by dose group (Table 3 and Figure 5)

Tables 3 and 4 split the 21 efzofitimod participants into tertiles of Day 1 to Week 24 time-averaged AUC, with boundaries 805, 3,638, 6,198 and 14,860. Figure 5 plots each participant’s exposure by dose group, and the tertiles almost coincide with the dose groups. The per-dose medians below were read by the maintainers from the Figure 5 points (7 per dose group, or 8 at 5 mg/kg).

fig5_med <- tibble::tibble(
  regimen = c("1 mg/kg", "3 mg/kg", "5 mg/kg"),
  fig5_median = c(1850, 5500, 8500),
  lower = c(805, 3638, 6198), upper = c(3638, 6198, 14860)
)
tert_chk <- cav |>
  dplyr::filter(regimen %in% fig5_med$regimen) |>
  dplyr::group_by(regimen) |>
  dplyr::summarise(
    median_cav = median(cav_d1_w24),
    p10 = quantile(cav_d1_w24, 0.1), p90 = quantile(cav_d1_w24, 0.9),
    .groups = "drop"
  ) |>
  dplyr::left_join(fig5_med, by = "regimen") |>
  dplyr::mutate(ratio = median_cav / fig5_median)
tert_chk |>
  dplyr::select(regimen, median_cav, p10, p90, fig5_median, ratio, lower, upper) |>
  dplyr::rename(
    "Regimen" = regimen, "Simulated median CAV (ng/mL)" = median_cav,
    "Simulated P10" = p10, "Simulated P90" = p90,
    "Figure 5 median (read)" = fig5_median, "Simulated / Figure 5" = ratio,
    "Table 3 tertile lower" = lower, "Table 3 tertile upper" = upper
  ) |>
  knitr::kable(digits = c(0, 0, 0, 0, 0, 2, 0, 0))
Regimen Simulated median CAV (ng/mL) Simulated P10 Simulated P90 Figure 5 median (read) Simulated / Figure 5 Table 3 tertile lower Table 3 tertile upper
1 mg/kg 2092 1314 3435 1850 1.13 805 3638
3 mg/kg 6500 3740 9901 5500 1.18 3638 6198
5 mg/kg 10891 5946 16515 8500 1.28 6198 14860
# Structural gate on the centre: a unit slip (mg vs ng), a wrong clearance
# or a mis-scaled dose moves these ratios by a factor of 2 or 1000, not by tens of percent.
stopifnot(all(tert_chk$ratio > 0.75 & tert_chk$ratio < 1.5))

The simulated exposures are dose-proportional, as the paper reports (“a good correlation between time-averaged AUC and dose”). They sit about 13-28% above the Figure 5 values. The 1 and 5 mg/kg medians fall inside their tertiles; the 3 mg/kg median lies just above the 6,198 boundary. The paper used post hoc individual clearances and the doses each participant actually received; four of the 21 received only four or five doses (Methods 2.1). It also does not give the exact length of its “study duration”. Any of these would lower the published values relative to a six-dose typical-cohort simulation. The offset is reported here and not adjusted.

Covariate effects on exposure (Results 3.2)

Body weight acts only on V1 and V2, so at steady state it does not change exposure for a fixed dose. Under mg/kg dosing, exposure is proportional to the dose, and so to weight. The printed changes of +30% at the 90th percentile of weight and -19% at the 10th percentile (relative to 82 kg) therefore imply weight percentiles of about 106.6 kg and 66.4 kg. Both lie inside the 55-157 kg range.

The printed CrCL and bilirubin ranges are not consistent with the printed exponents. Steady-state AUC is proportional to 1 / CL, so the change at each percentile follows from the power terms:

cov_tab <- tibble::tibble(
  Covariate = c("CrCL 10th pct (103 mL/min)", "CrCL 90th pct (187 mL/min)",
                "Bilirubin 10th pct (3.2 umol/L)", "Bilirubin 90th pct (19.6 umol/L)"),
  `Model AUC change (%)` = 100 * (c((103 / 139)^-0.71, (187 / 139)^-0.71,
                                    (3.2 / 7.9)^0.17, (19.6 / 7.9)^0.17) - 1),
  `Printed AUC change (%)` = c(10, -10, -8, 10)
)
knitr::kable(cov_tab, digits = 1)
Covariate Model AUC change (%) Printed AUC change (%)
CrCL 10th pct (103 mL/min) 23.7 10
CrCL 90th pct (187 mL/min) -19.0 -10
Bilirubin 10th pct (3.2 umol/L) -14.2 -8
Bilirubin 90th pct (19.6 umol/L) 16.7 10

No single reference value reconciles the printed changes with the exponents, because their ratio depends only on the 10th/90th-percentile ratio. The model keeps the Table S.1 exponents (see Assumptions and deviations).

Exposure-response

solve_er <- function(model, output, data) {
  ev <- dplyr::mutate(data, id = dplyr::row_number(), time = 0, amt = 0, evid = 0L)
  # The E-R models have no random effects by design, so rxode2's notice
  # about a multi-subject solve without omega is expected; muffle only it.
  out <- withCallingHandlers(
    as.data.frame(rxode2::rxSolve(model, events = ev, returnType = "data.frame")),
    warning = function(w) {
      if (grepl("without 'omega'", conditionMessage(w))) invokeRestart("muffleWarning")
    }
  )
  if (is.null(out$id)) out$id <- 1L
  out[[output]][match(ev$id, out$id)]
}
cav_grid <- seq(0, 15000, by = 250)

Figure 4 – OCS dose (Eq. A.1)

ocs_line <- tibble::tibble(CAV = cav_grid,
                           ocsdosecfb = solve_er(mod_ocs, "ocsdosecfb", tibble::tibble(CAV = cav_grid)))
ggplot(ocs_line, aes(CAV, ocsdosecfb)) +
  geom_line(colour = "blue") +
  coord_cartesian(ylim = c(-100, 25)) +
  labs(x = "Time-averaged efzofitimod AUC, Day 51-Week 24 (ng/mL)",
       y = "% change from baseline,\nmean daily OCS dose post-taper",
       caption = "Replicates the fitted line of Walker 2023 Figure 4.") +
  theme_bw()

ends <- ocs_line$ocsdosecfb[ocs_line$CAV %in% c(0, 15000)]
ends
#> [1] -43.010 -68.975
# Figure 4's fitted line runs from about -43% at zero exposure to about -69%
# at 15,000 (read from the figure).
stopifnot(abs(ends[1] - (-43)) < 2, abs(ends[2] - (-69)) < 3)

The OCS equation matches the Figure 4 line exactly. Typical post-taper exposures in the cohort above put the predicted OCS reduction at:

cav |>
  dplyr::mutate(ocsdosecfb = solve_er(mod_ocs, "ocsdosecfb", dplyr::select(cav, CAV = cav_d51_w24))) |>
  dplyr::group_by(regimen) |>
  dplyr::summarise(`Median CAV Day 51-Wk 24` = median(cav_d51_w24),
                   `Median % change in OCS dose` = median(ocsdosecfb), .groups = "drop") |>
  dplyr::rename("Regimen" = regimen) |>
  knitr::kable(digits = 1)
Regimen Median CAV Day 51-Wk 24 Median % change in OCS dose
1 mg/kg 2044.8 -46.5
3 mg/kg 6374.4 -54.0
450 mg fix dose 9547.8 -59.5
5 mg/kg 10672.4 -61.5
7 mg/kg 14477.8 -68.1
900 L+4mg/kg 7858.7 -56.6

Figures 5 and 6 – ppFVC and KSQ-Lung (Eqs. A.2 and A.3)

Equations A.2 and A.3 are labelled “percent change from baseline”, but their values are the absolute Week-24 score. Read literally, A.2 at a baseline ppFVC of 78 would predict a change of about +75%, and A.3 a change of at least +50% at any exposure, while Figures 5 and 6 centre near -5% and +15% at zero exposure. The models therefore emit the absolute scores (fvcpp, ksqlung). The percent change is formed below at the efficacy-population baselines (Table 4: ppFVC 78.1; KSQ-Lung 42.4 on placebo and 50.9 on efzofitimod).

fvc_line <- tibble::tibble(CAV = cav_grid, FVC_PCTPRED = 78.1) |>
  dplyr::mutate(fvcpp = solve_er(mod_fvc, "fvcpp", dplyr::pick(CAV, FVC_PCTPRED)),
                pct = 100 * (fvcpp - FVC_PCTPRED) / FVC_PCTPRED,
                Endpoint = "ppFVC (Figure 5), baseline 78.1")
ksq_line <- tibble::tibble(CAV = cav_grid) |>
  dplyr::mutate(ksqlung = solve_er(mod_ksq, "ksqlung", dplyr::pick(CAV)),
                bl = ifelse(CAV == 0, 42.4, 50.9),
                pct = 100 * (ksqlung - bl) / bl,
                Endpoint = "KSQ-Lung (Figure 6), baseline 42.4 placebo / 50.9 active")
dplyr::bind_rows(fvc_line, ksq_line) |>
  ggplot(aes(CAV, pct)) +
  geom_line(colour = "blue") +
  facet_wrap(~Endpoint, scales = "free_y", ncol = 1) +
  labs(x = "Time-averaged efzofitimod AUC, Day 1-Week 24 (ng/mL)",
       y = "% change from baseline at Week 24") +
  theme_bw()

fig56 <- tibble::tibble(
  Endpoint = c("ppFVC", "ppFVC", "KSQ-Lung", "KSQ-Lung"),
  CAV = c(0, 15000, 0, 15000),
  `Model (% change)` = c(fvc_line$pct[fvc_line$CAV %in% c(0, 15000)],
                         ksq_line$pct[ksq_line$CAV %in% c(0, 15000)]),
  `Figure line (% change, read)` = c(-5, 8, 15, 62)
)
knitr::kable(fig56, digits = 1)
Endpoint CAV Model (% change) Figure line (% change, read)
ppFVC 0 -3.8 -5
ppFVC 15000 6.0 8
KSQ-Lung 0 18.2 15
KSQ-Lung 15000 55.5 62
# Direction of the exposure effect; the figure lines are unadjusted fits of
# the percent change and need not coincide with the baseline-adjusted
# equations, so only the sign and the zero-exposure placement are asserted.
stopifnot(
  diff(fig56$`Model (% change)`[1:2]) > 0,
  diff(fig56$`Model (% change)`[3:4]) > 0,
  abs(fig56$`Model (% change)`[1] - (-5)) < 5,
  abs(fig56$`Model (% change)`[3] - 15) < 10
)

Figures 7 and 8 – responder probabilities (Eqs. A.4 and A.5)

Figures 7 and 8 draw an unadjusted logistic fit (exposure only). The equations add the baseline score. The curves below evaluate the equations at the mean efficacy-population baselines (ppFVC 78.1; KSQ-Lung 50.9).

mcid_curves <- dplyr::bind_rows(
  tibble::tibble(CAV = cav_grid, FVC_PCTPRED = 78.1) |>
    dplyr::mutate(p = solve_er(mod_fvc_mcid, "prob_fvcpp_incr2p5", dplyr::pick(CAV, FVC_PCTPRED)),
                  Endpoint = "ppFVC increase >= 2.5% (Figure 7)"),
  tibble::tibble(CAV = cav_grid, KSQ_LUNG_BL = 50.9) |>
    dplyr::mutate(p = solve_er(mod_ksq_mcid, "prob_ksqlung_mcid", dplyr::pick(CAV, KSQ_LUNG_BL)),
                  Endpoint = "KSQ-Lung increase >= 4 points (Figure 8)")
)
ggplot(mcid_curves, aes(CAV, p)) +
  geom_line(colour = "blue") +
  facet_wrap(~Endpoint) +
  coord_cartesian(ylim = c(0, 1)) +
  labs(x = "Time-averaged efzofitimod AUC, Day 1-Week 24 (ng/mL)",
       y = "Probability of response") +
  theme_bw()

Figure 9 – simulated response by regimen

Figure 9 is the paper’s dose-selection forest plot: the predicted Week-24 probability of each responder endpoint for placebo and the six regimens of Table 2. Here every simulated subject from the PK cohort receives its own CAV. Baseline ppFVC and KSQ-Lung are drawn from the Table 4 efficacy-population distributions, and the mean probability is taken per regimen.

set.seed(9)
fig9_cohort <- dplyr::bind_rows(
  cav |> dplyr::select(id, regimen, CAV = cav_d1_w24),
  tibble::tibble(id = -seq_len(n_per), regimen = "Placebo", CAV = 0)
) |>
  dplyr::mutate(
    FVC_PCTPRED = stats::rnorm(dplyr::n(), 78.1, 14.9),
    KSQ_LUNG_BL = pmin(pmax(stats::rnorm(dplyr::n(), 50.9, 13), 0), 100)
  )
fig9_cohort$p_fvc <- solve_er(mod_fvc_mcid, "prob_fvcpp_incr2p5",
                              dplyr::select(fig9_cohort, CAV, FVC_PCTPRED))
fig9_cohort$p_ksq <- solve_er(mod_ksq_mcid, "prob_ksqlung_mcid",
                              dplyr::select(fig9_cohort, CAV, KSQ_LUNG_BL))
fig9_pub <- tibble::tibble(
  regimen = c("Placebo", "3 mg/kg", "5 mg/kg", "450 mg fix dose", "900 L+4mg/kg", "7 mg/kg"),
  pub_fvc = c(0.25, 0.33, 0.39, 0.38, 0.39, 0.45),
  pub_ksq = c(0.60, 0.73, 0.79, 0.78, 0.79, 0.84)
)
fig9 <- fig9_cohort |>
  dplyr::group_by(regimen) |>
  dplyr::summarise(sim_fvc = mean(p_fvc), sim_ksq = mean(p_ksq), .groups = "drop") |>
  dplyr::inner_join(fig9_pub, by = "regimen") |>
  dplyr::arrange(match(regimen, fig9_pub$regimen))
fig9 |>
  dplyr::rename(
    "Regimen" = regimen,
    "ppFVC >= 2.5%: simulated" = sim_fvc, "ppFVC >= 2.5%: Figure 9" = pub_fvc,
    "KSQ-Lung MCID: simulated" = sim_ksq, "KSQ-Lung MCID: Figure 9" = pub_ksq
  ) |>
  knitr::kable(digits = 2)
Regimen ppFVC >= 2.5%: simulated KSQ-Lung MCID: simulated ppFVC >= 2.5%: Figure 9 KSQ-Lung MCID: Figure 9
Placebo 0.24 0.32 0.25 0.60
3 mg/kg 0.31 0.52 0.33 0.73
5 mg/kg 0.43 0.61 0.39 0.79
450 mg fix dose 0.39 0.67 0.38 0.78
900 L+4mg/kg 0.42 0.65 0.39 0.79
7 mg/kg 0.47 0.80 0.45 0.84
# ppFVC: the whole chain (PK model -> PKNCA CAV -> Eq. A.4) reproduces the
# published probabilities. Centre and robust envelope only: the cohort mean
# moves by about 0.02 between random-number streams.
fvc_err <- abs(fig9$sim_fvc - fig9$pub_fvc)
stopifnot(median(fvc_err) < 0.03, quantile(fvc_err, 0.8) < 0.05)

The ppFVC column agrees with Figure 9 in every regimen, with a median absolute difference of 0.018 and a largest difference of 0.041. Exposures here run slightly above the published ones (Check 3), which accounts for most of that. This agreement tests the PK model, the exposure definition (Day 1 to Week 24 AUC divided by 168 days), the ng/mL reading of the exposure unit and Eq. A.4 together.

The KSQ-Lung column does not match. Figure 9 gives 0.60 on placebo. Equation A.5 gives 0.22 at the efficacy-population mean baseline of 50.9 and 0.5 at the placebo-arm mean of 42.4. A placebo probability of 0.60 for a typical subject needs a baseline of 39.8, lower than any arm mean in Table 4. The step from placebo to 7 mg/kg in Figure 9 is also flatter than the A.5 slope implies. The paper does not report the baseline distribution behind Figure 9, so the gap cannot be traced. Equation A.5 is kept as printed, because it is broadly consistent with the observed responder rates in Figure 8 (7 of 11 placebo participants, and a rise from about 14% to 83% across the exposure tertiles, whose mean baselines are 50.6, 53.8 and 48.3).

Assumptions and deviations

  • CrCL and bilirubin reference values. Walker 2023 does not print the reference values of the two power terms on CL. The model uses 139 mL/min and 7.9 umol/L, the geometric means of the printed 10th and 90th percentiles: an estimate of the median for a log-symmetric distribution. The same method recovers about 84 kg for weight, against the printed median of 82 kg. Results 3.2 does print 82 kg, so weight uses it.
  • CrCL is raw mL/min. Results 3.2 reports CrCL in mL/min, not BSA-normalised, and does not name the estimating equation. Supply creatinine clearance in mL/min in the CRCL column.
  • Printed covariate impacts are inconsistent with the printed exponents (table above). The Table S.1 exponents are kept, because the table is the primary source of the estimates.
  • Residual error. Table S.1 reports log-additive residual variances (0.01 for Phase 1 and 1.63 for Phase 1b/2a). The model uses their square roots as log-scale SDs, selected by DIS_HEALTHY (1 for the Phase 1 healthy-volunteer study, 0 for patients).
  • omega^2 for Q3 is 0 FIX in Table S.1 and is encoded as etalq2 ~ fixed(0).
  • Equations A.2 and A.3 are absolute Week-24 scores. They are labelled “percent change from baseline”, but their magnitudes are only consistent with the absolute score (Figures 5 and 6 section above). The outputs are fvcpp and ksqlung; compute the percent change from the baseline yourself.
  • Exposure unit. The source prints time-averaged AUC in ng*day/mL. Cumulative AUC divided by days is a concentration. The ng/mL reading puts each simulated dose group within about 30% of the Figure 5 exposures and reproduces the Figure 9 ppFVC probabilities.
  • Simulated exposures are about 10-30% above the published ones (Check 3). The reasons the paper allows for (actual doses received, post hoc clearances, an unstated study duration) are listed there; nothing is adjusted.
  • Exposure window lengths. Day 1 to Week 24 is taken as 168 days (times 0-168) and Day 51 to Week 24 as 118 days (times 50-168). The paper notes that some participants received only four or five doses. Their real exposure windows are not reported, so the simulation gives everyone six doses.
  • No residual error is reported for the E-R regressions. The linear models (A.1-A.3) are least-squares fits whose residual standard error is not printed. The logistic models (A.4, A.5) have a Bernoulli likelihood with no residual. All five carry a tiny fixed additive SD (0.001) only so rxode2 accepts an observation model. These are typical-value predictions: they do not describe between-patient scatter.
  • ppFVC responder threshold. The paper calls it “a 2.5% increase” and does not say whether that is a relative change or 2.5 percentage points. The model returns the probability as fitted, so the reading does not affect it.
  • Virtual-cohort covariates. CrCL and bilirubin are drawn from the pooled PK-population percentiles (Results 3.2), since no patient-only distribution is reported. Baseline ppFVC and KSQ-Lung are drawn as normal distributions with the Table 4 efzofitimod-arm mean and SD.
  • Figure 9 KSQ-Lung probabilities are not reproduced (see above). The equation is kept as printed.
  • Overall half-life of 11.3 days is not checked, because the paper does not define how it was computed.
  • No erratum or correction to Walker 2023 was found (checked 2026-11-07).