Skip to contents

Model and source

  • Citation: Oggianu L, Ottaviani M, Maddalena F, et al. PK/PD analysis of trazodone and gabapentin in neuropathic pain rodent models: Translational PK-PD modeling from nonclinical to clinical development. Clin Transl Sci 2023;16(4):606-617. doi:10.1111/cts.13472 (open access via PMC10087074).
  • Models:
    • Oggianu_2023_trazodone_gabapentin_mouse – Preclinical (mouse, male CD-1). Joint PK/PD model of the antinociceptive interaction of oral trazodone and intraperitoneal gabapentin in the acetic-acid writhing test. Each drug has zero-order input and two-compartment disposition parameterised by micro-constants. The intraperitoneal acetic-acid stimulus is a unit dose into a stimulus compartment scaled by the stimulus amplitude S; it passes through nine transit compartments into the observed writhing compartment, which empties first-order. Each drug removes stimulus in proportion to its plasma molar concentration, and a product term on the two concentrations carries the departure from additivity (synergy). Time is in minutes; the PK parameters keep their published hour units and are converted in model().
    • Oggianu_2023_trazodone_human – Two-compartment population PK model of oral trazodone (immediate-release tablet and 6% oral drops) with first-order absorption and a lag time in healthy adult volunteers, carrying the authors’ translational writhing layer: the human plasma concentration, rescaled to the mouse unbound fraction, drives the mouse acetic-acid writhing PD model (stimulus, nine transit compartments, writhing compartment) with the mouse trazodone potency, as used to predict the percent inhibition of a pain stimulus at low human doses. The writhing layer is a cross-species projection, not a fit to human PD data.

Angelini developed a low-dose fixed combination of trazodone and gabapentin for neuropathic pain. Oggianu and colleagues quantified the interaction of the two drugs in the mouse acetic-acid writhing test with a sequential PK/PD analysis, then projected the mouse PD onto human trazodone exposure to suggest clinical starting doses.

The writhing model (Supplementary Table S1 and Figure S1) treats the intraperitoneal acetic acid as a stimulus of amplitude S placed in a compartment a0. The stimulus passes through nine transit compartments with a common rate constant ktr into the observed writhing compartment W, which empties at rate k. A drug removes stimulus from a0 at a rate proportional to its plasma concentration, k2 * C(t). For the combination the loss rate is k2A * CA + k2B * CB + gamma * CA * CB, so gamma = 0 is pure additivity and gamma > 0 is synergy.

The paper reports several parameter sets for the drug-effect layer (Tables 3-5). The packaged mouse model uses the synergic combination fit of Table 4B, which is the model the authors carried into the human translation. The other sets are listed below and can be substituted with rxode2::ini().

Source Data k2 trazodone (L/umol/min) k2 gabapentin (L/umol/min) gamma omega2 k2 (TRZ / GBP) sigma2 add
Table 3 single agents 10.9 0.0459 – 1.42 / 1.97 3.20 / 2.17
Table 4A low-dose combinations, additive 12.7 0.0808 0 1.43 / 1.67 2.62
Table 4B (packaged) low-dose combinations, synergic 10.9 0.0490 16.0 1.42 / 1.78 2.64
Table 5A high-dose combinations, additive 10.2 0.040 0 1.40 / 1.71 2.71
Table 5B high-dose combinations, synergic 10.2 0.0401 0.0001 1.40 / 1.71 2.71

Population

The mouse data come from male CD-1 mice (20-25 g). Pharmacokinetics were measured in satellite groups of 12 fasted mice per dose with sparse sampling to 24 h: trazodone 1, 3 and 10 mg/kg by gavage and gabapentin 10, 30 and 100 mg/kg intraperitoneally, plus the combinations 1 + 3 and 10 + 30 mg/kg (trazodone + gabapentin) to test for a PK interaction, which was not found (Supplementary Tables S4-S5). In the writhing test the drugs were given 1 h before 0.7% acetic acid (16 uL/g i.p.) and writhings were counted in 5-min bins for 30 min: trazodone 0.075-4.5 mg/kg, gabapentin 1.5-15 mg/kg, and the combinations 0.075 + 1.5, 0.15 + 1.5, 0.5 + 5 and 1.5 + 15 mg/kg, with 6-9 mice per group.

Human trazodone PK came from a phase I four-way crossover in 23 fasted healthy volunteers aged 22-54 years given a 50 mg immediate-release tablet and 30, 60 and 90 mg as 6% oral drops (Supplementary Table S2). Human gabapentin PK was taken from Carlsson et al. (Ther Drug Monit 2009), whose parameters are not reprinted in this paper.

Source trace

Equation / parameter Value Source location
Mouse PK structure: zero-order input, two compartments, micro-constants – Results paragraph 1; Table 1; Supplementary Table S1B
Trazodone k, V1/F, k12, k21, D1 1.25 1/h, 5.95 L/kg, 0.982 1/h, 1.51 1/h, 0.0833 h Table 1a
Trazodone omega2 V, k, k12; sigma2 prop 0.0526, 0.00899, 0.577; 0.0752 Table 1a
Gabapentin k, V1/F, k12, k21, D1 0.112 1/h, 0.846 L/kg, 0.935 1/h, 0.110 1/h, 0.167 h Table 1b
Gabapentin omega2 V, k, k12; sigma2 prop 0.00287, 0.337, 1.43; 0.0598 Table 1b
Writhing chain: da0/dt = -(k2 C + ktr) a0, nine transits, dW/dt = ktr a9 - k W – Supplementary Table S1; Figure S1
Combination drug term k2A CA + k2B CB + gamma CA CB – Methods ‘Perturbed model’; Figure S1D
ktr, S, k, n 1.36 1/min, 26.10, 0.104 1/min, 9 Table 2
omega2 S, k, n; sigma2 add (control) 0.0231, 0.0349, 0.00848; 2.79 Table 2
k2 trazodone, k2 gabapentin, gamma 10.9, 0.0490 L/umol/min, 16.0 Table 4B
omega2 k2 trazodone, gabapentin; sigma2 add 1.42, 1.78; 2.64 Table 4B
Human CL/F, V2/F, Q/F, V3/F, ka, ALAG1 9.86 L/h, 50.8 L, 31.6 L/h, 76.9 L, 3.30 1/h, 0.0787 h Supplementary Table S2
Human omega2 CL/F, V2/F, Q/F, V3/F, ka, alag; sigma2 prop 0.0909, 0.0813, 0.0854, 0.0577, 0.454, 0.00148; 0.0634 Supplementary Table S2
Plasma protein binding, human / mouse 97% / 89% Methods ‘Prediction of human pharmacological effect’
Human translation design (stimulus without drug, reset, drug 1 h before stimulus; ratio of writhing AUCs) – Methods ‘Prediction of human pharmacological effect’
Molecular weights of trazodone / gabapentin free base 371.86 / 171.24 g/mol Not in the paper; standard values, used to convert ng/mL to umol/L

Mouse pharmacokinetics

Both drugs are given as a zero-order input over the fixed duration D1 (rate = -2 lets the model’s dur() apply). The model runs in minutes, so the simulated times are converted to hours for the NCA.

mod_mouse <- readModelDb("Oggianu_2023_trazodone_gabapentin_mouse")

pk_groups <- tibble::tribble(
  ~treatment,          ~cmt,                 ~amt,
  "trazodone 1 mg/kg",  "central",            1,
  "trazodone 10 mg/kg", "central",            10,
  "gabapentin 3 mg/kg", "central_gabapentin", 3,
  "gabapentin 30 mg/kg", "central_gabapentin", 30
)
n_pk <- 200
pk_times <- c(0, 2, 5, 10, 15, 20, 30, 45, 60, 90, 120, 180, 240, 360, 480, 720, 960, 1200, 1440)

ev_pk <- bind_rows(lapply(seq_len(nrow(pk_groups)), function(i) {
  ids <- (i - 1) * n_pk + seq_len(n_pk)
  bind_rows(
    data.frame(id = ids, time = 0, amt = pk_groups$amt[i], cmt = pk_groups$cmt[i],
               evid = 1, rate = -2),
    # Observations sit on the `writhing` state: with several declared
    # endpoints rxode2 needs an endpoint compartment on observation rows,
    # and every algebraic output (Cc, Cc_gabapentin) is returned anyway.
    tidyr::crossing(id = ids, time = pk_times) |>
      mutate(amt = 0, cmt = "writhing", evid = 0, rate = 0)
  ) |>
    mutate(treatment = pk_groups$treatment[i])
})) |>
  arrange(id, time, desc(evid))

rxode2::rxSetSeed(20230413)
sim_pk <- as.data.frame(rxode2::rxSolve(mod_mouse, ev_pk, keep = "treatment",
                                        useLinCmt = FALSE)) |>
  mutate(conc = ifelse(grepl("^trazodone", treatment), Cc, Cc_gabapentin),
         time_h = time / 60)
#> ℹ parameter labels from comments will be replaced by 'label()'
sim_pk |>
  filter(time_h > 0) |>
  group_by(treatment, time_h) |>
  summarise(med = median(conc), lo = quantile(conc, 0.05), hi = quantile(conc, 0.95),
            .groups = "drop") |>
  ggplot(aes(time_h, med)) +
  geom_ribbon(aes(ymin = lo, ymax = hi), alpha = 0.25) +
  geom_line() +
  scale_y_log10() +
  facet_wrap(~treatment, scales = "free_y") +
  labs(x = "Time after dose (h)", y = "Plasma concentration (ng/mL)")
Simulated plasma concentrations in mice (median and 5th-95th percentiles of 200 simulated mice per group).

Simulated plasma concentrations in mice (median and 5th-95th percentiles of 200 simulated mice per group).

PKNCA

# Trazodone is ~30 half-lives below its peak by 24 h, where the solver
# can return values of order -1e-7 ng/mL; they are set to zero for the NCA.
conc_pk <- sim_pk |>
  filter(!is.na(conc)) |>
  mutate(conc = pmax(conc, 0)) |>
  select(id, time_h, conc, treatment)
dose_pk <- ev_pk |>
  filter(evid == 1) |>
  transmute(id, time_h = time / 60, amt, treatment)

nca_pk <- PKNCA::pk.nca(PKNCA::PKNCAdata(
  PKNCA::PKNCAconc(conc_pk, conc ~ time_h | treatment + id),
  PKNCA::PKNCAdose(dose_pk, amt ~ time_h | treatment + id),
  intervals = data.frame(start = 0, end = 24, cmax = TRUE, auclast = TRUE)
))

Supplementary Table S6 reports, for the same four dose groups, the NCA of the observed data and the mean of the individual (post-hoc) model predictions. The comparison uses the post-hoc means (Table S6 panels B and D).

published_pk <- tibble::tribble(
  ~treatment,            ~cmax, ~auclast,
  "trazodone 1 mg/kg",     282,     220,
  "trazodone 10 mg/kg",   2156,    1509,
  "gabapentin 3 mg/kg",   3345,    9887,
  "gabapentin 30 mg/kg", 33543,   86130
)
cmp_pk <- nlmixr2lib::ncaComparisonTable(
  simulated = nca_pk,
  reference = published_pk,
  by = "treatment",
  units = c(cmax = "ng/mL", auclast = "ng*h/mL"),
  tolerance_pct = 20
)
knitr::kable(cmp_pk, caption = attr(cmp_pk, "footnote"))
* differs from reference by more than ±20%.
NCA parameter treatment Reference Simulated % diff
Cmax (ng/mL) trazodone 1 mg/kg 282 149 -47.0%*
Cmax (ng/mL) trazodone 10 mg/kg 2160 1480 -31.2%*
Cmax (ng/mL) gabapentin 3 mg/kg 3340 3230 -3.4%
Cmax (ng/mL) gabapentin 30 mg/kg 33500 32400 -3.4%
AUClast (ng*h/mL) trazodone 1 mg/kg 220 131 -40.5%*
AUClast (ng*h/mL) trazodone 10 mg/kg 1510 1350 -10.8%
AUClast (ng*h/mL) gabapentin 3 mg/kg 9890 9460 -4.3%
AUClast (ng*h/mL) gabapentin 30 mg/kg 86100 92500 +7.4%

Gabapentin agrees with the post-hoc means within about 10%. Trazodone is lower than the post-hoc means, by about 40-45% at 1 mg/kg and 10-30% at 10 mg/kg. That gap is in the published estimates rather than in the simulation: for a linear model the typical AUC is dose / (k * V1/F), which is 1 / (1.25 x 5.95) = 0.134 mg h/L = 134 ng h/mL at 1 mg/kg against the reported 220. The observed NCA (Table S6A: 219 and 1406 ng h/mL) is itself less than dose-proportional, and the Table 1a values were fitted jointly to 1, 3 and 10 mg/kg, so the model sits between the low and high doses. The Table 1a values are kept as published.

auc_typ_trz_1 <- 1000 * 1 / (1.25 * 5.95)
stopifnot(abs(auc_typ_trz_1 - 134.45) < 0.01)
gab_rows <- grepl("gabapentin", cmp_pk$treatment)
gab_pct <- as.numeric(gsub("[%*+]", "", cmp_pk[["% diff"]][gab_rows]))
stopifnot(all(abs(gab_pct) < 15))

Mouse writhing model

Control animals

The acetic acid is given at 60 min, 1 h after the (here absent) drug, as in the experiments. It is dosed as amt = 1 into stimulus; the model’s bioavailability f(stimulus) scales that unit dose to the animal’s stimulus amplitude S.

n_pd <- 100
pd_times <- seq(60, 120, by = 1)

make_pd_events <- function(groups, n) {
  bind_rows(lapply(seq_len(nrow(groups)), function(i) {
    g <- groups[i, ]
    ids <- (i - 1) * n + seq_len(n)
    bind_rows(
      if (g$trz > 0) data.frame(id = ids, time = 0, amt = g$trz, cmt = "central", evid = 1, rate = -2),
      if (g$gab > 0) data.frame(id = ids, time = 0, amt = g$gab, cmt = "central_gabapentin", evid = 1, rate = -2),
      data.frame(id = ids, time = 60, amt = 1, cmt = "stimulus", evid = 1, rate = 0),
      tidyr::crossing(id = ids, time = pd_times) |>
        mutate(amt = 0, cmt = "writhing", evid = 0, rate = 0)
    ) |>
      mutate(group = g$group)
  })) |>
    arrange(id, time, desc(evid))
}

pd_percentiles <- function(sim) {
  sim |>
    group_by(group, time) |>
    summarise(p05 = quantile(writhing, 0.05), p50 = median(writhing),
              p95 = quantile(writhing, 0.95), .groups = "drop")
}
ctrl <- tibble::tibble(group = "control", trz = 0, gab = 0)
rxode2::rxSetSeed(20230414)
sim_ctrl <- as.data.frame(rxode2::rxSolve(mod_mouse, make_pd_events(ctrl, n_pd),
                                          keep = "group", useLinCmt = FALSE))
pct_ctrl <- pd_percentiles(sim_ctrl)
ggplot(pct_ctrl, aes(time, p50)) +
  geom_ribbon(aes(ymin = p05, ymax = p95), alpha = 0.25) +
  geom_line() +
  labs(x = "Time (min; acetic acid at 60 min)", y = "Number of writhings")
Simulated writhing counts in control mice (median and 5th-95th percentiles of 100 simulated mice; model predictions without residual error).

Simulated writhing counts in control mice (median and 5th-95th percentiles of 100 simulated mice; model predictions without residual error).

With the typical values the control curve can be checked against the structure directly: the stimulus leaves a0 at ktr, and the writhing count peaks about ten minutes after the injection.

ev_typ <- make_pd_events(ctrl, 1)
sim_typ <- as.data.frame(rxode2::rxSolve(mod_mouse, ev_typ, omega = NA, sigma = NA,
                                         useLinCmt = FALSE))
t_peak <- sim_typ$time[which.max(sim_typ$writhing)] - 60
w_peak <- max(sim_typ$writhing)
c(t_peak_min = t_peak, peak_writhings = round(w_peak, 2))
#>     t_peak_min peak_writhings 
#>          10.00          16.42
# A total stimulus S passes through the chain, so the area under W over all
# time is S / k. The 60-minute grid captures nearly all of it.
auc_w <- sum(diff(sim_typ$time) * (head(sim_typ$writhing, -1) + tail(sim_typ$writhing, -1)) / 2)
stopifnot(t_peak >= 9, t_peak <= 12, abs(auc_w / (26.10 / 0.104) - 1) < 0.05)

Replication of Figure 1: combinations under additivity

Figure 1 simulates 100 mice for each combination group assuming additivity. The single-agent potencies of Table 3 are used here with gamma = 0.

mod_additive_t3 <- mod_mouse |>
  rxode2::ini(
    lslope = log(10.9), lslope_gabapentin = log(0.0459), synergy = 0,
    etalslope ~ 1.42, etalslope_gabapentin ~ 1.97
  )
combos <- tibble::tribble(
  ~group,                                  ~trz,  ~gab,
  "(a) 0.075 mg/kg TRZ + 1.5 mg/kg GBP",  0.075,  1.5,
  "(b) 0.15 mg/kg TRZ + 1.5 mg/kg GBP",   0.15,   1.5,
  "(c) 0.5 mg/kg TRZ + 5 mg/kg GBP",      0.5,    5,
  "(d) 1.5 mg/kg TRZ + 15 mg/kg GBP",     1.5,    15
)
rxode2::rxSetSeed(20230415)
sim_fig1 <- as.data.frame(rxode2::rxSolve(mod_additive_t3, make_pd_events(combos, n_pd),
                                          keep = "group", useLinCmt = FALSE))
pct_fig1 <- pd_percentiles(sim_fig1)
ggplot(pct_fig1, aes(time)) +
  geom_line(aes(y = p50), colour = "red", linewidth = 1) +
  geom_line(aes(y = p05), colour = "red") +
  geom_line(aes(y = p95), colour = "red") +
  facet_wrap(~group, scales = "free_y") +
  labs(x = "Time (min)", y = "Number of writhings")
Replicates Figure 1 of Oggianu 2023: simulated writhings (5th, 50th and 95th percentiles of 100 mice) for the four combinations under additivity, using the Table 3 single-agent potencies.

Replicates Figure 1 of Oggianu 2023: simulated writhings (5th, 50th and 95th percentiles of 100 mice) for the four combinations under additivity, using the Table 3 single-agent potencies.

The published panels peak near 70 min with a median of roughly 12, 11, 6.5 and 3 writhings for panels (a)-(d) and a 95th percentile of roughly 17, 18, 14 and 10 (read from the figure by the maintainers). The simulated peak percentiles are:

fig1_peak <- pct_fig1 |>
  group_by(group) |>
  slice_max(p50, n = 1, with_ties = FALSE) |>
  ungroup() |>
  mutate(figure_p50 = c(12, 11, 6.5, 3), figure_p95 = c(17, 18, 14, 10))
fig1_peak |>
  transmute(
    "Group" = group,
    "Peak time (min)" = time,
    "Simulated 5th" = round(p05, 1),
    "Simulated median" = round(p50, 1),
    "Figure 1 median (approx.)" = figure_p50,
    "Simulated 95th" = round(p95, 1),
    "Figure 1 95th (approx.)" = figure_p95
  ) |>
  knitr::kable()
Group Peak time (min) Simulated 5th Simulated median Figure 1 median (approx.) Simulated 95th Figure 1 95th (approx.)
(a) 0.075 mg/kg TRZ + 1.5 mg/kg GBP 70 6.9 13.0 12.0 18.4 17
(b) 0.15 mg/kg TRZ + 1.5 mg/kg GBP 70 5.2 11.9 11.0 17.2 18
(c) 0.5 mg/kg TRZ + 5 mg/kg GBP 70 2.0 7.9 6.5 15.1 14
(d) 1.5 mg/kg TRZ + 15 mg/kg GBP 70 1.0 4.6 3.0 10.6 10
# Structural check: the higher-dose combinations suppress writhing more.
stopifnot(
  fig1_peak$p50[3] < fig1_peak$p50[1],
  fig1_peak$p50[4] < fig1_peak$p50[3]
)

The simulated peak medians and 95th percentiles fall within about 1.5 writhings of the published panels.

Additive versus synergic fits (Tables 4 and 5)

For the two low-dose combinations the synergic model (Table 4B, the packaged parameters) lowered the objective function by 17 points relative to the additive model (Table 4A). The typical-value writhing areas over the 30-min counting window show what gamma adds:

mod_additive_t4a <- mod_mouse |>
  rxode2::ini(lslope = log(12.7), lslope_gabapentin = log(0.0808), synergy = 0)
mod_additive_t5a <- mod_mouse |>
  rxode2::ini(lslope = log(10.2), lslope_gabapentin = log(0.040), synergy = 0)
low_high <- tibble::tribble(
  ~group,                                ~trz,   ~gab,
  "control",                             0,      0,
  "0.075 mg/kg TRZ + 1.5 mg/kg GBP",    0.075,  1.5,
  "0.15 mg/kg TRZ + 1.5 mg/kg GBP",     0.15,   1.5,
  "0.5 mg/kg TRZ + 5 mg/kg GBP",        0.5,    5,
  "1.5 mg/kg TRZ + 15 mg/kg GBP",       1.5,    15
)
ev_lh <- make_pd_events(low_high, 1) |> filter(time <= 90)
w_area <- function(m) {
  as.data.frame(rxode2::rxSolve(m, ev_lh, omega = NA, sigma = NA, keep = "group",
                                useLinCmt = FALSE)) |>
    group_by(group) |>
    summarise(area = sum(diff(time) * (head(writhing, -1) + tail(writhing, -1)) / 2),
              .groups = "drop")
}
syn_tab <- w_area(mod_mouse) |>
  rename(synergic_4B = area) |>
  left_join(rename(w_area(mod_additive_t4a), additive_4A = area), by = "group") |>
  left_join(rename(w_area(mod_additive_t5a), additive_5A = area), by = "group")
ctrl_area <- syn_tab$synergic_4B[syn_tab$group == "control"]
syn_tab |>
  mutate(across(-group, ~ round(100 * (1 - .x / ctrl_area), 1))) |>
  filter(group != "control") |>
  rename(
    "Group" = group,
    "% inhibition, synergic (Table 4B)" = synergic_4B,
    "% inhibition, additive (Table 4A)" = additive_4A,
    "% inhibition, additive (Table 5A)" = additive_5A
  ) |>
  knitr::kable()
Group % inhibition, synergic (Table 4B) % inhibition, additive (Table 4A) % inhibition, additive (Table 5A)
0.075 mg/kg TRZ + 1.5 mg/kg GBP 37.2 24.1 15.2
0.15 mg/kg TRZ + 1.5 mg/kg GBP 51.0 28.1 19.2
0.5 mg/kg TRZ + 5 mg/kg GBP 90.5 56.7 44.3
1.5 mg/kg TRZ + 15 mg/kg GBP 98.8 79.7 70.5

The synergy estimate comes from the two low-dose combinations only. At the high-dose combinations the authors found gamma indistinguishable from zero (Table 5B, gamma = 0.0001), so the packaged gamma = 16 overstates the interaction there; use rxode2::ini(mod, synergy = 0) with the Table 5 potencies for doses at or above 0.5 + 5 mg/kg.

Human trazodone

Pharmacokinetics (Supplementary Table S3)

Table S3 reports the simulated AUC from time zero to infinity after single oral trazodone doses of 0.5-50 mg, and the same AUC rescaled for protein binding (AUCeq).

mod_human <- readModelDb("Oggianu_2023_trazodone_human")
hum_doses <- c(0.5, 1, 5, 10, 25, 50)
n_hum <- 200
hum_times <- c(0, 0.083, 0.167, 0.25, 0.5, 0.75, 1, 1.5, 2, 3, 4, 6, 8, 12, 16, 24, 36, 48, 72)
ev_hum <- bind_rows(lapply(seq_along(hum_doses), function(i) {
  ids <- (i - 1) * n_hum + seq_len(n_hum)
  bind_rows(
    data.frame(id = ids, time = 0, amt = hum_doses[i], cmt = "depot", evid = 1),
    tidyr::crossing(id = ids, time = hum_times) |>
      mutate(amt = 0, cmt = "central", evid = 0)
  ) |>
    mutate(treatment = paste(hum_doses[i], "mg"))
})) |>
  arrange(id, time, desc(evid))

rxode2::rxSetSeed(20230416)
sim_hum <- as.data.frame(rxode2::rxSolve(mod_human, ev_hum, keep = "treatment"))
#> ℹ parameter labels from comments will be replaced by 'label()'

nca_hum <- PKNCA::pk.nca(PKNCA::PKNCAdata(
  PKNCA::PKNCAconc(filter(sim_hum, !is.na(Cc)), Cc ~ time | treatment + id),
  PKNCA::PKNCAdose(filter(ev_hum, evid == 1), amt ~ time | treatment + id),
  intervals = data.frame(start = 0, end = Inf, aucinf.obs = TRUE)
))
published_hum <- tibble::tibble(
  treatment = paste(hum_doses, "mg"),
  aucinf.obs = c(46.1, 92.7, 458, 931, 2323, 4643)
)
cmp_hum <- nlmixr2lib::ncaComparisonTable(
  simulated = nca_hum,
  reference = published_hum,
  by = "treatment",
  units = c(aucinf.obs = "ng*h/mL"),
  tolerance_pct = 20
)
knitr::kable(cmp_hum, caption = attr(cmp_hum, "footnote"))
NCA parameter treatment Reference Simulated % diff
AUC0-∞ (obs) (ng*h/mL) 0.5 mg 46.1 50.5 +9.6%
AUC0-∞ (obs) (ng*h/mL) 1 mg 92.7 96.1 +3.6%
AUC0-∞ (obs) (ng*h/mL) 5 mg 458 518 +13.1%
AUC0-∞ (obs) (ng*h/mL) 10 mg 931 1030 +10.9%
AUC0-∞ (obs) (ng*h/mL) 25 mg 2320 2550 +9.9%
AUC0-∞ (obs) (ng*h/mL) 50 mg 4640 4960 +6.7%

The reference values are the Table S3 medians. The typical-value AUC is dose / (CL/F) = 50.7 ng h/mL at 0.5 mg, about 10% above the published median of 46.1; the published simulation’s median-to-dose ratio is constant across doses (92.2-92.9 ng h/mL per mg), so the gap is a fixed scale offset rather than nonlinearity. The AUCeq column of Table S3 is AUC x (1 - 0.97) / (1 - 0.89), the same factor the model applies before the concentration drives the writhing layer:

fu_ratio <- (1 - 0.97) / (1 - 0.89)
table_s3 <- tibble::tibble(
  dose = hum_doses,
  auc_tot_mean = c(48.1, 96.3, 481, 967, 2438, 4784),
  auc_eq_mean = c(13.1, 26.3, 131, 264, 665, 1305)
)
table_s3 |>
  mutate(ratio = round(auc_eq_mean / auc_tot_mean, 4)) |>
  knitr::kable()
dose auc_tot_mean auc_eq_mean ratio
0.5 48.1 13.1 0.2723
1.0 96.3 26.3 0.2731
5.0 481.0 131.0 0.2723
10.0 967.0 264.0 0.2730
25.0 2438.0 665.0 0.2728
50.0 4784.0 1305.0 0.2728
stopifnot(all(abs(table_s3$auc_eq_mean / table_s3$auc_tot_mean - fu_ratio) < 0.003))
hum_pct <- as.numeric(gsub("[%*+]", "", cmp_hum[["% diff"]]))
stopifnot(abs(median(hum_pct)) < 15)

Translation of the writhing effect (Table 6, trazodone alone)

The paper’s design is replicated for each virtual subject: the stimulus is first applied without drug, the compartments are reset (evid = 3), and the stimulus is applied again 1 h after a trazodone dose. Percent inhibition is one minus the ratio of the two writhing areas over the 30-min counting window. Both phases share the subject’s parameters, so the comparison is within subject, as in the paper.

grid_30 <- seq(0, 0.5, by = 1 / 120)
ev_tr <- bind_rows(lapply(seq_along(hum_doses), function(i) {
  ids <- (i - 1) * n_hum + seq_len(n_hum)
  bind_rows(
    data.frame(id = ids, time = 0, amt = 1, cmt = "stimulus", evid = 1),
    tidyr::crossing(id = ids, time = grid_30) |> mutate(amt = 0, cmt = "writhing", evid = 0),
    data.frame(id = ids, time = 24, amt = 0, cmt = "writhing", evid = 3),
    data.frame(id = ids, time = 24, amt = hum_doses[i], cmt = "depot", evid = 1),
    data.frame(id = ids, time = 25, amt = 1, cmt = "stimulus", evid = 1),
    tidyr::crossing(id = ids, time = 25 + grid_30) |> mutate(amt = 0, cmt = "writhing", evid = 0)
  ) |>
    mutate(dose = hum_doses[i])
})) |>
  arrange(id, time, desc(evid))

rxode2::rxSetSeed(20230417)
sim_tr <- as.data.frame(rxode2::rxSolve(mod_human, ev_tr, keep = "dose"))
inh <- sim_tr |>
  mutate(phase = ifelse(time < 24, "no_drug", "drug")) |>
  group_by(id, dose, phase) |>
  summarise(area = sum(diff(time) * (head(writhing, -1) + tail(writhing, -1)) / 2),
            .groups = "drop") |>
  tidyr::pivot_wider(names_from = phase, values_from = area) |>
  mutate(pct_inh = 100 * (1 - drug / no_drug))

table6 <- tibble::tibble(
  dose = hum_doses,
  paper_mean = c(4.93, 9.20, 27.6, 37.5, 56.2, 69.41),
  paper_median = c(2.76, 5.67, 21.7, 34.1, 58.7, 73.96),
  paper_q1 = c(1.18, 2.50, 10.5, 18.6, 38.4, 55.09),
  paper_q3 = c(6.03, 11.75, 39.6, 53.3, 75.0, 85.97)
)
inh_sum <- inh |>
  group_by(dose) |>
  summarise(sim_mean = mean(pct_inh), sim_median = median(pct_inh),
            sim_q1 = quantile(pct_inh, 0.25), sim_q3 = quantile(pct_inh, 0.75)) |>
  left_join(table6, by = "dose")
inh_sum |>
  transmute(
    "Trazodone dose (mg)" = dose,
    "Mean, paper" = paper_mean, "Mean, simulated" = round(sim_mean, 1),
    "Median, paper" = paper_median, "Median, simulated" = round(sim_median, 1),
    "Q1, paper" = paper_q1, "Q1, simulated" = round(sim_q1, 1),
    "Q3, paper" = paper_q3, "Q3, simulated" = round(sim_q3, 1)
  ) |>
  knitr::kable(caption = "Percent inhibition of the pain stimulus after single oral trazodone doses (Table 6, gabapentin 0 mg column).")
Percent inhibition of the pain stimulus after single oral trazodone doses (Table 6, gabapentin 0 mg column).
Trazodone dose (mg) Mean, paper Mean, simulated Median, paper Median, simulated Q1, paper Q1, simulated Q3, paper Q3, simulated
0.5 4.93 5.0 2.76 3.2 1.18 1.5 6.03 6.6
1.0 9.20 10.5 5.67 6.5 2.50 2.8 11.75 14.5
5.0 27.60 27.4 21.70 22.3 10.50 11.6 39.60 38.4
10.0 37.50 43.9 34.10 41.0 18.60 25.2 53.30 61.4
25.0 56.20 59.4 58.70 59.6 38.40 42.2 75.00 78.3
50.0 69.41 72.4 73.96 76.1 55.09 62.2 85.97 88.3
# Centre-of-distribution checks across the six doses. With a 119% CV on the
# potency, a 200-subject mean moves by several points between cohorts, so
# the check is on the median difference across doses, not on any one dose.
stopifnot(
  abs(median(inh_sum$sim_median - inh_sum$paper_median)) < 4,
  abs(median(inh_sum$sim_mean - inh_sum$paper_mean)) < 4
)
ggplot(inh_sum, aes(dose)) +
  geom_ribbon(aes(ymin = sim_q1, ymax = sim_q3), alpha = 0.25) +
  geom_line(aes(y = sim_median)) +
  geom_pointrange(aes(y = paper_median, ymin = paper_q1, ymax = paper_q3), colour = "red") +
  scale_x_log10() +
  labs(x = "Trazodone dose (mg)", y = "Inhibition of writhing stimulus (%)")
Percent inhibition of the writhing stimulus versus human trazodone dose: simulated median and interquartile range (line and band) against the Table 6 values (points and error bars).

Percent inhibition of the writhing stimulus versus human trazodone dose: simulated median and interquartile range (line and band) against the Table 6 values (points and error bars).

The gabapentin columns of Table 6 (gabapentin alone and with trazodone) need the human gabapentin model of Carlsson et al. (2009), which this paper uses but does not reprint, so they are not reproduced here.

Assumptions and deviations

  • Time units. The mouse model runs in minutes because the writhing parameters are per minute; the mouse PK parameters keep their published hour units in ini() and are divided by 60 in model(). The human model runs in hours and multiplies the per-minute writhing parameters by 60.
  • Molar conversion. The potencies k2 are per umol/L, so the plasma concentration is divided by the free-base molecular weight (trazodone 371.86, gabapentin 171.24 g/mol). The molecular weights are not printed in the paper. The Table 6 reproduction above depends on this conversion and agrees with the published values.
  • Stimulus as a dose. The paper gives the stimulus as the initial condition a0(0) = S. Because the drugs are given 1 h earlier, the model doses the stimulus as amt = 1 into stimulus at the time of the acetic acid, with f(stimulus) = S.
  • Number of transit compartments. Table 2 reports n = 9 with an interindividual variance (omega2 n = 0.00848). The PK/PD equations of Supplementary Table S1B write the chain as nine explicit compartments, and an explicit chain cannot carry a random number of compartments, so the model follows the printed equations with n = 9 for every animal and omits omega2 n.
  • Packaged drug-effect parameters. The mouse model carries the Table 4B synergic fit. It was estimated on the two low-dose combinations; at the high-dose combinations the authors found additivity (Table 5). The Table 3 single-agent fits had separate additive residual variances (3.20 and 2.17); the model carries the single Table 4B value (2.64).
  • Sequential estimation. The PK parameters and the control-group PD parameters were estimated first and held fixed while the potencies and gamma were estimated. They are not wrapped in fixed(), because they are estimates from this paper. The zero-order input durations D1 (5 and 10 min) have no reported standard error and are encoded as fixed.
  • Human translation. The writhing parameters in the human model are the mouse estimates and are wrapped in fixed(); the trazodone potency is assumed equal across species, as the paper states. Human concentrations are scaled by the unbound-fraction ratio (1 - 0.97) / (1 - 0.89) before they act. The per-subject percent inhibition uses the 30-min counting window of the mouse experiment; the paper does not state the window used for the human areas, and integrating the full writhing curve instead changes the percentages by a fraction of a percentage point.
  • Trazodone mouse exposure. The typical-value trazodone AUC and Cmax are below the post-hoc means of Table S6 (see the PKNCA section); the Table 1a values are used as published.
  • Human gabapentin. The human gabapentin PK model (Carlsson et al., Ther Drug Monit 2009;31:86-94) is cited but not reprinted, so the human model covers trazodone only.
  • Animal numbers. The total number of mice in the writhing experiments is not reported (6-9 per group), so n_subjects is NA in the mouse model.