Skip to contents

Model and source

  • Citation: Wojciechowski J, Gaitonde P, Hughes JH, Ravva P. Population Modeling of Factor IX Activity Following Administration of Fidanacogene Elaparvovec Gene Therapy in Participants with Hemophilia B. Clin Pharmacokinet. 2025;64(10):1531-1548. doi:10.1007/s40262-025-01535-y. Parameter values from Table 2; model structure from the ESM NONMEM control stream (run57.mod). Extended half-life FIX disposition fixed from Diao L, et al. Clin Pharmacokinet. 2014;53(5):467-477 (doi:10.1007/s40262-013-0129-7), as hard-coded in run57.mod.
  • Description: Population model of factor IX (FIX) activity after the AAV gene therapy fidanacogene elaparvovec and/or FIX replacement therapy in hemophilia B (Wojciechowski 2025). A Hargrove-Schmidt gene/protein-expression model (productive transgene amount translated into FIX at a synthesis site, first-order transgene degradation, and for a transient-production mixture subpopulation a reversible transition to a non-productive transgene state) feeds a three-compartment FIX disposition model with an endogenous baseline. Standard half-life FIX products dose into the same central compartment; extended half-life products use the fixed three-compartment rFIXFc model of Diao 2014. Allometric weight scaling referenced to 70 kg, age effects on synthesis, transgene degradation and relative FIX activity, manufacturing-process effects on synthesis, Box-Cox-transformed IIV on the translation constant, study effects on baseline, relative activity and the log-scale residual error.
  • Article: https://doi.org/10.1007/s40262-025-01535-y (open access; the electronic supplementary material holds the final NONMEM control stream, run57.mod)

Fidanacogene elaparvovec is an adeno-associated virus (AAV) vector carrying a high-activity (Padua) factor IX (FIX) transgene. The model describes FIX activity from every source at once: endogenous FIX (baseline BASE), transgene-derived FIX, standard half-life (SHL) FIX replacement products such as nonacog alfa, and extended half-life (EHL) products. The gene-therapy part is the Hargrove-Schmidt gene/protein-expression model: the vector dose (vector genomes, vg) enters a productive-transgene state that is translated into FIX at a synthesis site at rate ktau * transgene without consuming the transgene, and the transgene itself is lost at kdeg. FIX leaves the synthesis site at ksyn into a three-compartment disposition model estimated from nonacog alfa data. A two-class mixture separates a constant-production phenotype (subpopulation 1) from a transient-production phenotype (subpopulation 2) in which the productive transgene moves reversibly into a non-productive state.

Population

The analysis pooled 337 participants with hemophilia B and 8931 FIX activity observations (one-stage clotting assays) from 11 trials: 274 participants from eight nonacog alfa studies (141 paediatric, 133 adult; median age 17.1 years, range < 0.1 to 64.7; median weight 60.0 kg) and 63 adult men from three fidanacogene elaparvovec studies (median age 29.0 years, range 18.0-62.0; median weight 86.1 kg, range 55.0-139). Most gene-therapy participants received 5e11 vg/kg as a single 60-minute infusion; 10, 8 and 45 received material from manufacturing processes 1, 2 and 3, and 31 of 63 received concomitant corticosteroids. Race overall was 82.8% White, 5.6% Asian, 5.0% Black and 5.3% other (Wojciechowski 2025 Table 1).

The same information is available programmatically via readModelDb("Wojciechowski_2025_fidanacogeneElaparvovec")()$population.

Source trace

Every ini() value carries an in-file comment naming its source. Structure follows the deposited control stream run57.mod (ESM pp. 2-6).

Equation / parameter Value Source location
lrbase (BASE) 1.83 IU/dL Table 2
lcl, lvc (CL, V1) 0.617 L/h, 10.3 L Table 2
lq, lvp (Q2, V2) 1.46 L/h, 3.29 L Table 2
lq2, lvp2 (Q3, V3) 0.145 L/h, 14.0 L Table 2
lthalf_ksyn 34.9 day Table 2; run57.mod POPKSYN = LOG(2)/(24*THETA(8))
lktau 24.0 x 1e-13 IU/vg/h Table 2; run57.mod POPKT = THETA(9)/1E13
lthalf_kdeg 423 week Table 2; run57.mod LOG(2)/(168*THETA(10))
lthalf_ktrd, lthalf_ktrp 4.90, 62.6 week Table 2
boxcox_lktau -0.631 Table 2; run57.mod ETAKTT = (EXP(ETAKT)**BXCXK-1)/BXCXK
e_wt_cl, e_wt_vc, e_wt_krate 0.75, 1, -0.25 (fixed) Methods; Table 2 footnote; run57.mod
e_age_ksyn, e_age_kdeg, e_age_fixa -0.704, -1.57, 0.0395 Table 2 (age / 35)
e_form_p1_ksyn, e_form_p2_ksyn 0.356, 0.658 Table 2
e_wt_ktau 1.08 Table 2 (WT / 70)
e_study1002_rbase, e_study1034_rbase 0.788, 0.994 Table 2
e_study1010_fixa, e_study1036_fixa 0.252, 0.173 Table 2
e_study*_expSd (1002, 1010, 1034, 1036, C037) 1.32, 0.545, 0.843, -0.245, 0.305 Table 2
expSd 0.276 Table 2 (SD; $SIGMA 1 FIX)
etalrbase, etalktau 0.419, 0.657 Table 2 (variances)
*_ehl EHL product parameters 2.39 dL/h, 71.4 dL, 1.67 dL/h, 87.0 dL, 39.3 dL/h, 39.9 dL; WT/73 powers 0.436, 0.396 run57.mod POPCLA-POPV3A; Table S2 (Diao 2014)
Mixture probability (subpopulation 1) logit 0.945, i.e. 0.720 Table 2; run57.mod $MIX
d/dt(transgene), d/dt(transgene_np), d/dt(fix_cell) n/a Structural Model equations; run57.mod DADT(4), DADT(9), DADT(5)
d/dt(central), d/dt(peripheral1), d/dt(peripheral2), initial conditions n/a Structural Model equations; run57.mod DADT(1)-DADT(3), A_0
Cc <- fixa * central / vc + central_ehl / vc_ehl n/a Methods (IPRE = FIXA * CENT/V1 + EHL); run57.mod FIPRE
Cc ~ lnorm(expSd_study) n/a Methods (log-transformed both sides); run57.mod $ERROR

Simulation helpers

All simulations below are of the gene-therapy setting: process 3 material, STUDY_C037 = 1 (it only changes the residual SD) and the four nonacog alfa study indicators 0. Time is in hours; the plots convert to weeks or years. Simulated FIX activity is the individual prediction Cc (no residual error), as in the published simulations.

mod <- readModelDb("Wojciechowski_2025_fidanacogeneElaparvovec")
hy <- 24 * 365.25 # hours per year
hw <- 24 * 7 # hours per week
gt_dose_per_kg <- 5e11 # vg/kg, nominal dose

# Observation grid: weekly for two years, then every four weeks to 15 years
obs_times <- sort(unique(c(0, seq(hw, 104 * hw, by = hw), seq(108 * hw, 15 * hy, by = 4 * hw), 15 * hy)))

base_cov <- function(d) {
  d |>
    mutate(
      FORM_FIDANACOGENE_PROCESS1 = 0, FORM_FIDANACOGENE_PROCESS2 = 0,
      STUDY_B1821002 = 0, STUDY_B1821010 = 0, STUDY_B1821034 = 0,
      STUDY_B1821036 = 0, STUDY_C037 = 1
    )
}

# One gene-therapy infusion (60 min) at time 0 plus the observation grid.
# Subject-level covariates come from `subj` (one row per id).
make_gt_events <- function(subj, times = obs_times) {
  dose <- subj |>
    mutate(time = 0, amt = gt_dose_per_kg * WT, rate = amt, evid = 1, cmt = "transgene")
  obs <- tidyr::expand_grid(subj, time = times) |>
    mutate(amt = 0, rate = 0, evid = 0, cmt = "central")
  bind_rows(dose, obs) |>
    arrange(id, time, desc(evid)) |>
    base_cov()
}

Replicate Figure 1b: contributions of gene therapy and SHL replacement

Figure 1b shows a population-typical 86.1 kg individual given 5e11 vg/kg and five doses of SHL FIX replacement (40 IU/kg) at weeks 0, 24, 25, 26 and 52. The model is linear in its inputs, so each contribution is the difference between the full solve and a solve without that input.

typ <- rxode2::zeroRe(mod)
#> ℹ parameter labels from comments will be replaced by 'label()'
#> Warning: No sigma parameters in the model
subj_typ <- tibble(id = 1L, WT = 86.1, AGE = 35, MIX_TRANSIENT_EXPR = 0)
fig1b_times <- seq(0, 60 * hw, by = 2)

shl_doses <- subj_typ |>
  tidyr::expand_grid(week = c(0, 24, 25, 26, 52)) |>
  mutate(time = week * hw + 1 / 60, amt = 40 * WT, rate = amt / (10 / 60), evid = 1, cmt = "central") |>
  select(-week)

ev_full <- bind_rows(make_gt_events(subj_typ, fig1b_times), base_cov(shl_doses)) |>
  arrange(id, time, desc(evid))
ev_gt <- make_gt_events(subj_typ, fig1b_times)
ev_shl <- ev_full |> filter(!(evid == 1 & cmt == "transgene"))

solve_cc <- function(ev) as.data.frame(rxode2::rxSolve(typ, ev, returnType = "data.frame"))[, c("time", "Cc")]
s_full <- solve_cc(ev_full)
#> ℹ omega/sigma items treated as zero: 'etalrbase', 'etalktau'
s_gt <- solve_cc(ev_gt)
#> ℹ omega/sigma items treated as zero: 'etalrbase', 'etalktau'
s_shl <- solve_cc(ev_shl)
#> ℹ omega/sigma items treated as zero: 'etalrbase', 'etalktau'
s_base <- s_full$Cc[1]

fig1b <- tibble(
  time = s_full$time,
  total = s_full$Cc,
  gene_therapy = s_gt$Cc - s_base,
  shl = s_shl$Cc - s_base
)

ggplot(fig1b, aes(time / hw)) +
  geom_area(aes(y = gene_therapy, fill = "Fidanacogene elaparvovec"), alpha = 0.5) +
  geom_area(aes(y = shl, fill = "SHL FIX replacement"), alpha = 0.5) +
  geom_line(aes(y = total)) +
  scale_fill_manual(values = c("Fidanacogene elaparvovec" = "steelblue", "SHL FIX replacement" = "darkgreen")) +
  labs(
    x = "Time (weeks)", y = "FIX activity (IU/dL)", fill = NULL,
    caption = "Replicates Figure 1b of Wojciechowski 2025 (typical 86.1 kg individual, subpopulation 1)."
  ) +
  theme_bw() +
  theme(legend.position = "bottom")

The model is linear, so the full solve must equal baseline plus the two contributions. That identity is pure numerical error and is checked tightly.

stopifnot(max(abs(fig1b$total - (s_base + fig1b$gene_therapy + fig1b$shl))) < 1e-3 * max(fig1b$total))
# Without any dose the endogenous steady state is held at BASE = 1.83 IU/dL.
stopifnot(abs(s_base - 1.83) < 1e-6)

Replicate Figure 1c: the two production phenotypes

subj_ph <- tibble(id = 1:2, WT = 86.1, AGE = 35, MIX_TRANSIENT_EXPR = c(0, 1))
s_ph <- as.data.frame(rxode2::rxSolve(typ, make_gt_events(subj_ph), returnType = "data.frame")) |>
  mutate(phenotype = ifelse(id == 1, "Subpopulation 1 (constant)", "Subpopulation 2 (transient)"))
#> ℹ omega/sigma items treated as zero: 'etalrbase', 'etalktau'
#> Warning: multi-subject simulation without without 'omega'

ggplot(s_ph, aes(time / hy, Cc, colour = phenotype)) +
  geom_line(linewidth = 0.8) +
  scale_colour_manual(values = c("darkorange", "purple")) +
  labs(
    x = "Time (years)", y = "FIX activity (IU/dL)", colour = NULL,
    caption = "Replicates Figure 1c of Wojciechowski 2025 (86.1 kg, 35 years, process 3, 5e11 vg/kg)."
  ) +
  theme_bw() +
  theme(legend.position = "bottom")


typ_peak <- s_ph |>
  group_by(phenotype) |>
  summarise(peak = max(Cc), t_peak_wk = time[which.max(Cc)] / hw, at_15y = Cc[which.max(time)], .groups = "drop")
typ_peak |>
  rename(
    "Phenotype" = phenotype, "Peak (IU/dL)" = peak,
    "Time of peak (weeks)" = t_peak_wk, "FIX activity at 15 years (IU/dL)" = at_15y
  ) |>
  knitr::kable(digits = 2, caption = "Typical-value peak, time of peak and FIX activity at 15 years by phenotype.")
Typical-value peak, time of peak and FIX activity at 15 years by phenotype.
Phenotype Peak (IU/dL) Time of peak (weeks) FIX activity at 15 years (IU/dL)
Subpopulation 1 (constant) 18.83 34 7.20
Subpopulation 2 (transient) 8.43 8 2.22

The constant phenotype rises over about 30 weeks to a plateau near 19 IU/dL and then falls with the 423-week transgene half-life; the transient phenotype peaks within about two months and then drops to a slowly declining level, as in the published Figure 1c.

Nonacog alfa disposition check

A single 50 IU/kg nonacog alfa dose (10-minute infusion) in a typical 70 kg, 35-year-old adult. Above the constant endogenous baseline the disposition is linear, so dose / AUC(0-inf) of the baseline-corrected profile must recover the clearance, CL = 0.617 L/h = 6.17 dL/h. PKNCA computes the AUC.

subj_n <- tibble(id = 1L, WT = 70, AGE = 35, MIX_TRANSIENT_EXPR = 0, treatment = "nonacog alfa 50 IU/kg")
nonacog_times <- c(0, 0.25, 0.5, 1, 2, 4, 6, 8, 12, 18, 24, 36, 48, 72, 96, 144, 192, 288, 384, 576, 768, 1152)
ev_n <- bind_rows(
  subj_n |> mutate(time = 0, amt = 50 * WT, rate = amt / (10 / 60), evid = 1, cmt = "central"),
  tidyr::expand_grid(subj_n, time = nonacog_times) |> mutate(amt = 0, rate = 0, evid = 0, cmt = "central")
) |>
  arrange(id, time, desc(evid)) |>
  base_cov()
# A single-subject solve returns no id column; add it back with the group.
s_n <- as.data.frame(rxode2::rxSolve(typ, ev_n, returnType = "data.frame")) |>
  mutate(id = 1L, treatment = subj_n$treatment, Cinc = pmax(Cc - 1.83, 0))
#> ℹ omega/sigma items treated as zero: 'etalrbase', 'etalktau'

conc_n <- PKNCA::PKNCAconc(s_n |> filter(!is.na(Cinc)) |> select(id, time, Cinc, treatment), Cinc ~ time | treatment + id)
dose_n <- PKNCA::PKNCAdose(ev_n |> filter(evid == 1) |> select(id, time, amt, treatment), amt ~ time | treatment + id)
nca_n <- PKNCA::pk.nca(PKNCA::PKNCAdata(conc_n, dose_n,
  intervals = data.frame(start = 0, end = Inf, cmax = TRUE, aucinf.obs = TRUE, half.life = TRUE)
))
res_n <- as.data.frame(nca_n$result)
auc_n <- res_n$PPORRES[res_n$PPTESTCD == "aucinf.obs"]
cl_nca <- 50 * 70 / auc_n
knitr::kable(
  res_n |> filter(PPTESTCD %in% c("cmax", "aucinf.obs", "half.life")) |>
    select(treatment, PPTESTCD, PPORRES) |>
    rename("Group" = treatment, "NCA parameter" = PPTESTCD, "Value" = PPORRES),
  digits = 3, caption = "PKNCA of the baseline-corrected nonacog alfa profile (IU/dL, IU*h/dL, h)."
)
PKNCA of the baseline-corrected nonacog alfa profile (IU/dL, IU*h/dL, h).
Group NCA parameter Value
nonacog alfa 50 IU/kg cmax 32.813
nonacog alfa 50 IU/kg half.life 85.989
nonacog alfa 50 IU/kg aucinf.obs 569.109
cat(sprintf("Dose / AUC(0-inf) = %.3f dL/h (model CL = 6.17 dL/h)\n", cl_nca))
#> Dose / AUC(0-inf) = 6.150 dL/h (model CL = 6.17 dL/h)
# Deterministic identity: only trapezoid and extrapolation error separate them.
stopifnot(abs(cl_nca / 6.17 - 1) < 0.03)

Table 3 and Figure 3: covariate scenarios

Table 3 summarises 1000 simulated trials of 63 participants per covariate scenario as the median geometric mean of peak FIX activity, time to peak and the 15-year AUC, for a mixed population of 72% subpopulation 1 and 28% subpopulation 2. Here one cohort of 200 participants per scenario stands in for the trial distribution. As in the published summaries, exactly 28% of the participants (56 of 200) are assigned to subpopulation 2, the Table 2 mixture probability (1 - plogis(0.945) = 0.280); a random Bernoulli draw would add several percentage points of phenotype-mix noise to every median. The same participants and seed are used for every scenario so that between-scenario ratios (Figure 3) compare like with like.

scen <- tribble(
  ~scenario, ~WT, ~AGE, ~p1, ~p2,
  "Reference", 86.1, 35, 0, 0,
  "Younger age (18.8 y)", 86.1, 18.8, 0, 0,
  "Older age (56.5 y)", 86.1, 56.5, 0, 0,
  "Low body weight (56.5 kg)", 56.5, 35, 0, 0,
  "High body weight (108 kg)", 108, 35, 0, 0,
  "Manufacturing process 1", 86.1, 35, 1, 0,
  "Manufacturing process 2", 86.1, 35, 0, 1
)
n_per <- 200
p_sub2 <- 1 - plogis(0.945) # Table 2 log-odds of subpopulation 1
set.seed(2025)
mix_draw <- as.integer(seq_len(n_per) %in% sample(n_per, round(p_sub2 * n_per)))

sim_scen <- lapply(seq_len(nrow(scen)), function(i) {
  subj <- tibble(id = seq_len(n_per), WT = scen$WT[i], AGE = scen$AGE[i], MIX_TRANSIENT_EXPR = mix_draw)
  ev <- make_gt_events(subj) |>
    mutate(FORM_FIDANACOGENE_PROCESS1 = scen$p1[i], FORM_FIDANACOGENE_PROCESS2 = scen$p2[i])
  rxode2::rxSetSeed(2025)
  out <- as.data.frame(rxode2::rxSolve(mod, ev, returnType = "data.frame"))
  out$scenario <- scen$scenario[i]
  out
}) |>
  bind_rows()
#> ℹ parameter labels from comments will be replaced by 'label()'

PKNCA gives each participant’s peak, time to peak and AUC over 15 years (converted to IU x year/dL); the geometric means per scenario are compared with Table 3.

conc_s <- PKNCA::PKNCAconc(
  sim_scen |> filter(!is.na(Cc)) |> mutate(time_wk = time / hw) |> select(id, time_wk, Cc, scenario),
  Cc ~ time_wk | scenario + id
)
dose_s <- PKNCA::PKNCAdose(
  sim_scen |> distinct(scenario, id) |> mutate(time_wk = 0, amt = 1),
  amt ~ time_wk | scenario + id
)
nca_s <- PKNCA::pk.nca(PKNCA::PKNCAdata(conc_s, dose_s,
  intervals = data.frame(start = 0, end = 15 * hy / hw, cmax = TRUE, tmax = TRUE, auclast = TRUE)
))
ind <- as.data.frame(nca_s$result) |>
  select(scenario, id, PPTESTCD, PPORRES) |>
  tidyr::pivot_wider(names_from = PPTESTCD, values_from = PPORRES) |>
  mutate(auc_y = auclast / (365.25 / 7)) # IU*week/dL -> IU*year/dL

geo <- ind |>
  group_by(scenario) |>
  summarise(
    peak = exp(mean(log(cmax))), tpeak = exp(mean(log(tmax))), auc = exp(mean(log(auc_y))),
    .groups = "drop"
  )

published3 <- tribble(
  ~scenario, ~peak_pub, ~tpeak_pub, ~auc_pub,
  "Reference", 13.8, 22.8, 115,
  "Younger age (18.8 y)", 13.7, 14.4, 76.4,
  "Older age (56.5 y)", 13.7, 32.3, 140,
  "Low body weight (56.5 kg)", 9.02, 20.9, 80.2,
  "High body weight (108 kg)", 17.6, 24.1, 144,
  "Manufacturing process 1", 14.3, 18.7, 116,
  "Manufacturing process 2", 14.6, 15.9, 116
)
cmp3 <- published3 |>
  left_join(geo, by = "scenario") |>
  mutate(
    peak_diff = 100 * (peak / peak_pub - 1),
    tpeak_diff = 100 * (tpeak / tpeak_pub - 1),
    auc_diff = 100 * (auc / auc_pub - 1)
  )
stopifnot(nrow(cmp3) == 7, !anyNA(cmp3$peak))
cmp3 |>
  select(scenario, peak_pub, peak, peak_diff, tpeak_pub, tpeak, tpeak_diff, auc_pub, auc, auc_diff) |>
  rename(
    "Scenario" = scenario,
    "Peak, published (IU/dL)" = peak_pub, "Peak, simulated" = peak, "Peak diff (%)" = peak_diff,
    "Time to peak, published (wk)" = tpeak_pub, "Time to peak, simulated" = tpeak, "Tpeak diff (%)" = tpeak_diff,
    "AUC15y, published (IU*y/dL)" = auc_pub, "AUC15y, simulated" = auc, "AUC diff (%)" = auc_diff
  ) |>
  knitr::kable(digits = 1, caption = "Geometric means per covariate scenario vs Wojciechowski 2025 Table 3.")
Geometric means per covariate scenario vs Wojciechowski 2025 Table 3.
Scenario Peak, published (IU/dL) Peak, simulated Peak diff (%) Time to peak, published (wk) Time to peak, simulated Tpeak diff (%) AUC15y, published (IU*y/dL) AUC15y, simulated AUC diff (%)
Reference 13.8 15.0 8.8 22.8 22.7 -0.6 115.0 122.3 6.3
Younger age (18.8 y) 13.7 15.0 9.4 14.4 14.3 -0.9 76.4 79.4 4.0
Older age (56.5 y) 13.7 14.9 8.9 32.3 32.8 1.5 140.0 149.8 7.0
Low body weight (56.5 kg) 9.0 9.6 6.7 20.9 20.4 -2.2 80.2 82.9 3.3
High body weight (108 kg) 17.6 19.4 10.1 24.1 24.4 1.3 144.0 154.1 7.0
Manufacturing process 1 14.3 15.6 9.0 18.7 18.5 -1.1 116.0 122.4 5.5
Manufacturing process 2 14.6 15.9 9.1 15.9 15.8 -0.7 116.0 122.4 5.5

The ratios to the reference scenario (Figure 3) remove most of the cohort noise, because every scenario shares the same random effects and phenotypes.

ref <- geo |> filter(scenario == "Reference")
ratios <- geo |>
  filter(scenario != "Reference") |>
  transmute(
    scenario,
    peak_ratio = peak / ref$peak, tpeak_ratio = tpeak / ref$tpeak, auc_ratio = auc / ref$auc
  )
pub_ratios <- published3 |>
  filter(scenario != "Reference") |>
  transmute(
    scenario,
    peak_ratio_pub = peak_pub / 13.8, tpeak_ratio_pub = tpeak_pub / 22.8, auc_ratio_pub = auc_pub / 115
  )
cmp_ratio <- left_join(pub_ratios, ratios, by = "scenario")
cmp_ratio |>
  select(scenario, peak_ratio_pub, peak_ratio, tpeak_ratio_pub, tpeak_ratio, auc_ratio_pub, auc_ratio) |>
  rename(
    "Scenario" = scenario,
    "Peak ratio, published" = peak_ratio_pub, "Peak ratio, simulated" = peak_ratio,
    "Tpeak ratio, published" = tpeak_ratio_pub, "Tpeak ratio, simulated" = tpeak_ratio,
    "AUC ratio, published" = auc_ratio_pub, "AUC ratio, simulated" = auc_ratio
  ) |>
  knitr::kable(digits = 2, caption = "Geometric mean ratios to the reference scenario (Figure 3 of Wojciechowski 2025, from Table 3).")
Geometric mean ratios to the reference scenario (Figure 3 of Wojciechowski 2025, from Table 3).
Scenario Peak ratio, published Peak ratio, simulated Tpeak ratio, published Tpeak ratio, simulated AUC ratio, published AUC ratio, simulated
Younger age (18.8 y) 0.99 1.00 0.63 0.63 0.66 0.65
Older age (56.5 y) 0.99 0.99 1.42 1.45 1.22 1.23
Low body weight (56.5 kg) 0.65 0.64 0.92 0.90 0.70 0.68
High body weight (108 kg) 1.28 1.29 1.06 1.08 1.25 1.26
Manufacturing process 1 1.04 1.04 0.82 0.82 1.01 1.00
Manufacturing process 2 1.06 1.06 0.70 0.70 1.01 1.00

ratios |>
  tidyr::pivot_longer(-scenario, names_to = "metric", values_to = "ratio") |>
  mutate(metric = recode(metric, peak_ratio = "Peak", tpeak_ratio = "Time to peak", auc_ratio = "AUC15y")) |>
  ggplot(aes(ratio, scenario)) +
  annotate("rect", xmin = 0.8, xmax = 1.25, ymin = -Inf, ymax = Inf, fill = "lightblue", alpha = 0.4) +
  geom_vline(xintercept = 1, linetype = 2) +
  geom_point() +
  facet_wrap(~metric) +
  labs(x = "Geometric mean ratio to reference", y = NULL, caption = "Replicates Figure 3 of Wojciechowski 2025.") +
  theme_bw()

The geometric means reproduce Table 3 to within about 10% for every scenario and metric, and the covariate ratios of Figure 3 to within a few percent: age moves mainly the time to peak, weight mainly the peak, and manufacturing processes 1 and 2 shorten the time to peak without changing the 15-year AUC.

# Structural gates. A mis-transcribed rate constant, translation constant or
# covariate exponent moves these by tens of percent; the cohort-to-cohort
# spread of a 200-participant geometric mean is a few percent.
stopifnot(
  abs(median(cmp3$peak_diff)) < 15,
  abs(median(cmp3$tpeak_diff)) < 15,
  abs(median(cmp3$auc_diff)) < 15,
  max(abs(c(cmp3$peak_diff, cmp3$tpeak_diff, cmp3$auc_diff))) < 30,
  # Ratios share random effects across scenarios, so they are tighter.
  max(abs(cmp_ratio$peak_ratio / cmp_ratio$peak_ratio_pub - 1)) < 0.10,
  max(abs(cmp_ratio$tpeak_ratio / cmp_ratio$tpeak_ratio_pub - 1)) < 0.10,
  max(abs(cmp_ratio$auc_ratio / cmp_ratio$auc_ratio_pub - 1)) < 0.10
)

Replicate Figure 4: 15-year FIX activity in a virtual population

The published simulation drew 100,000 individuals with demographics representative of the adult analysis population. Here 200 participants are drawn with age and weight from truncated normal distributions matching the gene-therapy studies in Table 1 (age mean 34.0, SD 12.0, range 18-62 years; weight mean 85.5, SD 16.0, range 55-139 kg), process 3, 5e11 vg/kg, and the same exact 28% allocation to subpopulation 2.

rtnorm <- function(n, mean, sd, lo, hi) {
  x <- rnorm(n, mean, sd)
  while (any(bad <- x < lo | x > hi)) x[bad] <- rnorm(sum(bad), mean, sd)
  x
}
set.seed(4)
n_pop <- 200
subj_pop <- tibble(
  id = seq_len(n_pop),
  AGE = rtnorm(n_pop, 34.0, 12.0, 18, 62),
  WT = rtnorm(n_pop, 85.5, 16.0, 55, 139),
  MIX_TRANSIENT_EXPR = as.integer(seq_len(n_pop) %in% sample(n_pop, round(p_sub2 * n_pop))),
  treatment = "5e11 vg/kg"
)
rxode2::rxSetSeed(44)
sim_pop <- as.data.frame(rxode2::rxSolve(mod, make_gt_events(subj_pop), returnType = "data.frame", keep = "treatment"))

pop_q <- sim_pop |>
  group_by(time) |>
  summarise(
    Q05 = quantile(Cc, 0.05), Q50 = median(Cc), Q95 = quantile(Cc, 0.95),
    .groups = "drop"
  )
ggplot(pop_q, aes(time / hy, Q50)) +
  geom_ribbon(aes(ymin = Q05, ymax = Q95), fill = "steelblue", alpha = 0.3) +
  geom_line(colour = "steelblue4") +
  geom_vline(xintercept = 6, linetype = 2) +
  scale_y_log10() +
  labs(
    x = "Time (years)", y = "FIX activity (IU/dL)",
    caption = "Replicates Figure 4b of Wojciechowski 2025: median and 90% prediction interval."
  ) +
  theme_bw()

at_year <- function(y) {
  x <- sim_pop |>
    group_by(id) |>
    summarise(Cc = Cc[which.min(abs(time - y * hy))], .groups = "drop")
  x$Cc
}
conc_p <- PKNCA::PKNCAconc(
  sim_pop |> filter(!is.na(Cc)) |> mutate(time_wk = time / hw) |> select(id, time_wk, Cc, treatment),
  Cc ~ time_wk | treatment + id
)
dose_p <- PKNCA::PKNCAdose(subj_pop |> mutate(time_wk = 0, amt = 1) |> select(id, time_wk, amt, treatment), amt ~ time_wk | treatment + id)
nca_p <- PKNCA::pk.nca(PKNCA::PKNCAdata(conc_p, dose_p,
  intervals = data.frame(start = 0, end = 15 * hy / hw, cmax = TRUE, tmax = TRUE, auclast = TRUE)
))
ind_p <- as.data.frame(nca_p$result) |>
  select(id, PPTESTCD, PPORRES) |>
  tidyr::pivot_wider(names_from = PPTESTCD, values_from = PPORRES)

q3 <- function(x) quantile(x, c(0.5, 0.05, 0.95), names = FALSE)
pop_tab <- tribble(
  ~metric, ~published,
  "Peak FIX activity (IU/dL)", c(13.5, 3.12, 41.3),
  "Time to peak (weeks)", c(25.3, 6.70, 48.7),
  "FIX activity at year 1 (IU/dL)", c(11.0, 1.84, 40.1),
  "FIX activity at year 6 (IU/dL)", c(7.16, 1.50, 27.9),
  "FIX activity at year 15 (IU/dL)", c(4.11, 1.15, 17.6),
  "AUC over 15 years (IU*year/dL)", c(105, 23.4, 394)
)
pop_tab$simulated <- list(
  q3(ind_p$cmax), q3(ind_p$tmax), q3(at_year(1)), q3(at_year(6)), q3(at_year(15)),
  q3(ind_p$auclast / (365.25 / 7))
)
pop_cmp <- pop_tab |>
  mutate(
    pub_median = vapply(published, `[`, numeric(1), 1),
    sim_median = vapply(simulated, `[`, numeric(1), 1),
    pub_pi = vapply(published, function(x) sprintf("%.3g-%.3g", x[2], x[3]), character(1)),
    sim_pi = vapply(simulated, function(x) sprintf("%.3g-%.3g", x[2], x[3]), character(1)),
    diff_pct = 100 * (sim_median / pub_median - 1)
  )
pop_cmp |>
  select(metric, pub_median, pub_pi, sim_median, sim_pi, diff_pct) |>
  rename(
    "Metric" = metric, "Published median" = pub_median, "Published 90% PI" = pub_pi,
    "Simulated median" = sim_median, "Simulated 90% PI" = sim_pi, "Median diff (%)" = diff_pct
  ) |>
  knitr::kable(digits = 2, caption = "Virtual-population summaries vs Wojciechowski 2025 Results and Figure 4.")
Virtual-population summaries vs Wojciechowski 2025 Results and Figure 4.
Metric Published median Published 90% PI Simulated median Simulated 90% PI Median diff (%)
Peak FIX activity (IU/dL) 13.50 3.12-41.3 13.96 3.48-44.6 3.41
Time to peak (weeks) 25.30 6.7-48.7 29.00 7-49 14.62
FIX activity at year 1 (IU/dL) 11.00 1.84-40.1 12.45 2.42-41.2 13.20
FIX activity at year 6 (IU/dL) 7.16 1.5-27.9 8.34 1.99-29.2 16.53
FIX activity at year 15 (IU/dL) 4.11 1.15-17.6 4.81 1.73-19.2 17.12
AUC over 15 years (IU*year/dL) 105.00 23.4-394 120.02 30.7-414 14.30

# Centre of the distribution only; the tails of a 200-person cohort are not
# asserted on. The published sampling distributions of age and weight are not
# given, and this cohort sits about 15% above the published medians (the
# covariate scenarios above, where demographics are fixed, agree within 10%).
# A mis-transcribed rate constant, dose or unit moves these by far more.
stopifnot(
  nrow(pop_cmp) == 6,
  abs(median(pop_cmp$diff_pct)) < 25,
  max(abs(pop_cmp$diff_pct)) < 40
)

In this cohort the simulated medians sit about 3-17% above the published ones, most for the late time points and the 15-year AUC. The covariate scenarios, whose demographics are fixed and stated in Table 3, agree to within 10%, so the gap points to the unstated demographic sampling of the published 100,000-participant simulation (an older cohort has slower transgene loss, kdeg falling with age to the power -1.57) rather than to the model.

Assumptions and deviations

  • Non-productive transgene degradation. The printed structural equations in the article omit transgene loss from the non-productive state (dPERIG/dt = ktr,d * DEPOT - ktr,p * PERIG). The deposited final control stream degrades it at kdeg (DADT(9) = KTRD*DEPOT-KTRP*PERIG-KDEG*PERIG), and Table S3 run 56 (“assume the transgene degradation still occurs while in non-productive state … final structural model”) confirms it. The model follows the control stream.
  • Allometric scaling of the transition rates. The article lists allometric scaling on ksyn and kdeg; the control stream also scales ktr,d and ktr,p by (WT/70)^-0.25. The model follows the control stream.
  • Mixture model. rxode2 cannot estimate a $MIX block, so the latent class is supplied as the covariate MIX_TRANSIENT_EXPR (1 = transient production, subpopulation 2). For population simulation it is drawn as Bernoulli(0.280), from the Table 2 log-odds 0.945 of subpopulation 1. The log-odds is documented in the covariate notes rather than carried as an unused ini() parameter.
  • Extended half-life products. The authors accounted for every EHL product (eftrenonacog alfa, nonacog beta pegol, rFIX-FP) with the eftrenonacog alfa model of Diao 2014, hard-coded without variability in the control stream (Table S2). Those values are carried as fixed() parameters with the Diao 2014 reference weight of 73 kg. SHL products (nonacog alfa and plasma-derived FIX) dose into central.
  • Study effects. Study-level effects on BASE, relative FIX activity and the residual SD are carried as study indicators so the model can describe each source study; FIXA is 1 with no variability for the reference studies, including the gene-therapy studies, apart from the age effect.
  • Simulation demographics. The published simulations do not give the sampling distributions of age and weight. Figure 4 here uses truncated normal distributions matching the gene-therapy study summary in Table 1, which yields an older and heavier population than the pooled analysis.
  • Cohort size. Table 3 summarises 1000 trials of 63 participants and Figure 4 100,000 participants; this article uses 200 per scenario, so the 90% intervals of single participants are noisier than published.
  • Screened covariates not retained. Concomitant corticosteroids (significant on ktr,d but judged descriptive) and BMI are documented in covariatesDataExcluded.
  • No erratum or correction notice was found on Europe PMC as of 2026-11-06.