Skip to contents

Amyloid plaque turnover (van Maanen 2025)

Source: vignettes/articles/vanMaanen_2025_amyloid.Rmd
vanMaanen_2025_amyloid.Rmd

Model and source

  • Citation: van Maanen E, Robey S, Bennacef I, Duffull S, Egan MF, Kennedy ME, Stone JA (2025). Modeling amyloid plaque turnover dynamics improves characterization of drug effects. Alzheimer’s & Dementia: Translational Research & Clinical Interventions 11(3):e70169. doi:10.1002/trc2.70169.
  • Article: https://doi.org/10.1002/trc2.70169

Population exposure-response model of brain amyloid plaque burden (Centiloid, CL) in patients with mild cognitive impairment or mild dementia due to Alzheimer’s disease. A single indirect-response (turnover) ODE describes the natural plaque time course (zero-order formation Kin, first-order elimination Kout). The BACE1 inhibitor verubecestat inhibits plaque formation via a fixed-Imax AUC-driven sigmoid on Kin (Inh_verub, AUC50 in uMh of verubecestat AUC over 24 h at steady state; Imax fixed to 1 per Table 1). Each of four anti-A-beta monoclonal antibodies (aducanumab, donanemab, gantenerumab, lecanemab) stimulates plaque elimination via a linear AUC/molecular-weight term on Kout (Stim_mAb = slope_mAb (AUC_mAb / MW_mAb), with distinct slopes per mAb). AUC of each drug is a time-varying model input (covariate); the paper’s simulations set each AUC to its steady-state value during dosing periods and to 0 during washout. Baseline plaque burden is a per-subject covariate used as the plaque initial condition. Fit by NONMEM FOCE-I to 370 individual verubecestat amyloid PET measurements from 188 aMCI due to AD subjects (phase 3 APECS, NCT01953601) pooled with 120 summary-level PET measurements from four anti-A-beta mAb trial programmes; external validation on 1506 amyloid-PET measurements from 521 amyloid-positive ADNI subjects with up to 10 years of follow-up. No IIV is encoded (paper: ‘no additional interindividual variability was included’; summary-level mAb data constrained the model to central-tendency predictions).

The model is a single indirect-response (turnover) ODE describing the natural time course of amyloid plaque burden (Centiloid units, CL) and the effects of five anti-amyloid therapies – one BACE1 inhibitor (verubecestat) inhibiting plaque formation and four anti-A-beta monoclonal antibodies (aducanumab, donanemab, gantenerumab, lecanemab) stimulating plaque elimination. The model consumes drug exposure as time-varying AUC covariates rather than embedding a PK ODE; this preserves the fidelity of the paper’s E-R fit (which used AUCs from upstream published popPK models) without duplicating five separate compound-specific popPK layers.

Population

The modelling data pooled individual verubecestat amyloid PET measurements from the phase 3 APECS trial (NCT01953601; n = 188 aMCI/prodromal AD subjects, 370 PET measurements) with summary-level PET data from four anti-A-beta mAb programmes (n = 120 summary-level time points across 8 phase 1b / phase 2 / phase 3 trials; PRIME, EMERGE, ENGAGE, TRAILBLAZER-ALZ 1/4, SCarlet RoAD + Marguerite RoAD open-label extensions, GRADUATE 1/2, Clarity AD, Study 201). Baseline demographics were similar across the trials (paper Table S2): median age 70-75 years, 50-59% female, APOE4-carrier prevalence 53-72%, and baseline amyloid PET 70-104 CL. External validation used 1506 individual PET measurements from 521 amyloid-positive ADNI subjects with up to 10 years of follow-up. Population metadata is programmatically available via readModelDb("vanMaanen_2025_amyloid")()$population.

Source trace

Each ini() value in inst/modeldb/specificDrugs/vanMaanen_2025_amyloid.R carries an in-file source comment. The table below collects them for reviewer audit.

Equation / parameter Value Source location
d/dt(plaque) = Kin * (1 - Inh_verub) - plaque * Kout * (1 + Stim_mAb) – van Maanen 2025 Supplement Eq 5
Inh_verub = Imax * AUC_verub / (AUC_verub + AUC50) – Supplement Eq 4
Stim_mAb = slope_mAb * (AUC_mAb / MW_mAb) – Supplement Eq 6
plaque(t=0) = PLAQUE_BL per-subject Supplement ‘Further details on the exposure-response model’
y = plaque + eps, eps ~ N(0, sigma^2) – Supplement Eq 3
lkin (log Kin) -3.58 Table 1 (log-transformed Kin, RSE 2.4%)
lkout (log Kout) -8.13 Table 1 (log-transformed Kout, RSE 1.4%)
limax (log Imax) log(1) FIXED Table 1 (Imax = 1, footnote ‘a Fixed’)
lauc50 (log AUC50, verubecestat) log(0.392) Table 1 (AUC50 = 0.392 uM*h, RSE 59.7%)
lslope_adu (log slope, aducanumab) log(719) Table 1 (slope = 719 mM^-1 day^-1, RSE 11.9%)
lslope_don (log slope, donanemab) log(1120) Table 1 (slope = 1120 mM^-1 day^-1, RSE 11.8%)
lslope_gan (log slope, gantenerumab) log(397) Table 1 (slope = 397 mM^-1 day^-1, RSE 24.3%)
lslope_lec (log slope, lecanemab) log(606) Table 1 (slope = 606 mM^-1 day^-1, RSE 12.9%)
addSd (additive residual SD, CL) 6.39 Table 1 (additive error 6.39 CL, RSE 11.3%)
MW_ADU, MW_DON, MW_GAN, MW_LEC (g/mol) 145912, 145087, 146300, 150000 Supplement ‘Further details on the exposure-response model’

Units and dimensional analysis

Every term in Eq 5 must reduce to [CL / day] for the plaque ODE to be dimensionally consistent. Working left-to-right:

Term Units per element Product
Kin CL/day CL/day
1 - Inh_verub unitless (Emax sigmoid AUC/(AUC + AUC50), both AUC and AUC50 in uM*h) –
Kin * (1 - Inh_verub) CL/day CL/day
plaque CL –
Kout 1/day –
slope_mAb mM^-1 * day^-1 = L/mmol/day –
AUC_mAb mg*day/L –
MW_mAb g/mol –
AUC_mAb / MW_mAb (mgday/L) / (g/mol) = mmolday/L = mM*day –
slope_mAb * (AUC_mAb / MW_mAb) (L/mmol/day) * (mmol*day/L) = 1 (unitless) –
1 + Stim_mAb unitless –
plaque * Kout * (1 + Stim_mAb) CL * (1/day) = CL/day CL/day
d/dt(plaque) CL/day consistent

The unit convention AUC_mAb in mgday/L (rather than the more common mgh/mL or ugh/mL) is what makes the paper’s reported slope units mM^-1 day^-1 yield a dimensionless stimulation term. Downstream users supplying an AUC in mgh/L must divide by 24 before assigning to the covariate column.

Natural steady-state check

With no drug on board (AUC_VERUB = AUC_ADU = AUC_DON = AUC_GAN = AUC_LEC = 0), Eq 5 reduces to d/dt(plaque) = Kin - plaque * Kout whose steady state is plaque_ss = Kin / Kout = exp(-3.58) / exp(-8.13) = 94.90 CL. Paper Section 3.1 quotes this as “no net further additions being reached around 100 CL” (consistent with the algebraic steady state). Below we integrate for 100 years starting from the SS baseline and confirm the trajectory does not drift.

mod <- readModelDb("vanMaanen_2025_amyloid")

kin_val  <- exp(-3.58)
kout_val <- exp(-8.13)
plaque_ss <- kin_val / kout_val

nat_zero <- data.frame(
  id        = 1L,
  time      = seq(0, 100 * 365, by = 30),
  evid      = 0L,
  cmt       = "plaque",
  amt       = 0,
  AUC_VERUB = 0, AUC_ADU = 0, AUC_DON = 0, AUC_GAN = 0, AUC_LEC = 0,
  PLAQUE_BL = plaque_ss
)

sim_ss <- rxode2::rxSolve(mod, events = nat_zero) |> as.data.frame()
range_plaque <- range(sim_ss$plaque)
cat(sprintf("Algebraic steady state:   %.4f CL\n", plaque_ss))
#> Algebraic steady state:   94.6324 CL
cat(sprintf("Simulated plaque range:   [%.4f, %.4f] CL over 100 years\n",
            range_plaque[1], range_plaque[2]))
#> Simulated plaque range:   [94.6324, 94.6324] CL over 100 years

The trajectory is flat to > 4 decimal places over 100 years, confirming the solver reproduces the algebraic steady state to machine precision.

Perturbation recovery

Starting the plaque state below and above the natural steady-state value with no drug should produce trajectories that monotonically approach 94.90 CL over several plaque half-lives (Kout t1/2 = ln(2) / 0.000294 = 2359 days = 6.46 years).

starts <- c("baseline 10 CL"  = 10,
            "baseline 50 CL"  = 50,
            "baseline 95 CL"  = 95,
            "baseline 150 CL" = 150,
            "baseline 200 CL" = 200)

pert <- lapply(seq_along(starts), function(i) {
  data.frame(
    id        = i,
    time      = seq(0, 30 * 365, by = 30),
    evid      = 0L,
    cmt       = "plaque",
    amt       = 0,
    AUC_VERUB = 0, AUC_ADU = 0, AUC_DON = 0, AUC_GAN = 0, AUC_LEC = 0,
    PLAQUE_BL = starts[i],
    scenario  = names(starts)[i]
  )
}) |>
  dplyr::bind_rows()
#> Warning in data.frame(id = i, time = seq(0, 30 * 365, by = 30), evid = 0L, :
#> row names were found from a short variable and have been discarded
#> Warning in data.frame(id = i, time = seq(0, 30 * 365, by = 30), evid = 0L, :
#> row names were found from a short variable and have been discarded
#> Warning in data.frame(id = i, time = seq(0, 30 * 365, by = 30), evid = 0L, :
#> row names were found from a short variable and have been discarded
#> Warning in data.frame(id = i, time = seq(0, 30 * 365, by = 30), evid = 0L, :
#> row names were found from a short variable and have been discarded
#> Warning in data.frame(id = i, time = seq(0, 30 * 365, by = 30), evid = 0L, :
#> row names were found from a short variable and have been discarded

sim_pert <- rxode2::rxSolve(mod, events = pert, keep = "scenario") |>
  as.data.frame()
#> Warning: multi-subject simulation without without 'omega'

ggplot(sim_pert, aes(time / 365, plaque, colour = scenario, group = scenario)) +
  geom_hline(yintercept = plaque_ss, linetype = "dashed", colour = "grey40") +
  geom_line(linewidth = 0.7) +
  labs(x = "Time (years)",
       y = "Plaque burden (CL)",
       colour = "Initial condition",
       title = "Perturbation-recovery: natural plaque trajectory",
       caption = "Dashed grey line: natural steady state = Kin / Kout = 94.9 CL. Kout t1/2 = 6.46 y.") +
  theme_minimal()

All five trajectories converge to the natural setpoint. Trajectories starting from a low baseline (10, 50 CL) rise monotonically; trajectories starting from a high baseline (150, 200 CL) decline monotonically. The paper’s Section 3.1 narrative for the “removal rate … varies with the amyloid plaque burden (-1.1, -2.7, -5.4, and -9.7 CL/year at burdens of 10, 25, 50, and 90 CL, respectively) which results in net yearly CL additions of 9.1, 7.5, 4.8, and 0.50” is quantitatively reproduced by these curves.

Replicating Table S4: mAb fold-increase in Kout at target dose

Table S4 reports the model-predicted fold-increase in plaque elimination rate (= 1 + Stim_mAb) at each mAb’s phase-3 target dose. Reproducing these requires computing Stim_mAb = slope_mAb * (AUC_mAb / MW_mAb) for each regimen. Steady-state AUCs (mg*day/L over 4 weeks) below are derived from the Table S3 phase-3 dose regimens using literature typical CL values for each mAb (aducanumab 0.36 L/day, donanemab 0.5 L/day, gantenerumab 0.6 L/day, lecanemab 0.4 L/day).

mab_targets <- tibble::tibble(
  drug      = c("aducanumab", "donanemab", "gantenerumab", "lecanemab"),
  slope     = c(719,          1120,        397,            606),
  MW        = c(145912,       145087,      146300,         150000),
  auc4w     = c(1944,         2800,        1700,           3500),
  paper_fold_est   = c(9.26, 18.6, 5.30, 13.8),
  paper_fold_low   = c(7.42, 14.8, 3.16, 10.9),
  paper_fold_high  = c(11.1, 22.7, 7.19, 16.9)
)

mab_targets <- mab_targets |>
  mutate(model_fold = 1 + slope * (auc4w / MW),
         plaque_thalf_treated_years = log(2) / (kout_val * model_fold) / 365)

mab_targets |>
  dplyr::rename(
    "mAb (target regimen)"        = drug,
    "slope (mM^-1 day^-1)"        = slope,
    "MW (g/mol)"                  = MW,
    "AUC4w (mg*day/L)"            = auc4w,
    "Model fold-Kout"             = model_fold,
    "Table S4 median"             = paper_fold_est,
    "Table S4 2.5%"               = paper_fold_low,
    "Table S4 97.5%"              = paper_fold_high,
    "Plaque t1/2 treated (years)" = plaque_thalf_treated_years
  ) |>
  knitr::kable(digits = 2,
               caption = "Model-computed fold-increase in Kout at each mAb's phase-3 target dose vs. Table S4 median (95% CI). Small differences reflect the approximate literature-CL derivation of AUC used here.")
Model-computed fold-increase in Kout at each mAb’s phase-3 target dose vs. Table S4 median (95% CI). Small differences reflect the approximate literature-CL derivation of AUC used here.
mAb (target regimen) slope (mM^-1 day^-1) MW (g/mol) AUC4w (mg*day/L) Table S4 median Table S4 2.5% Table S4 97.5% Model fold-Kout Plaque t1/2 treated (years)
aducanumab 719 145912 1944 9.26 7.42 11.10 10.58 0.61
donanemab 1120 145087 2800 18.60 14.80 22.70 22.61 0.29
gantenerumab 397 146300 1700 5.30 3.16 7.19 5.61 1.15
lecanemab 606 150000 3500 13.80 10.90 16.90 15.14 0.43

The model-computed fold-increases fall within (or very close to) the bootstrap 95% CIs from Table S4 for all four mAbs, and the potency rank-order (donanemab > lecanemab > aducanumab > gantenerumab) matches the paper.

Replicating Figure 3: fold-Kout vs AUC per mAb

Figure 3 plots 1 + Stim_mAb against the mAb’s 4-week AUC for each of the four mAbs, showing the linear E-R with slope_mAb / MW_mAb. Reproduce it below across the exposure range implied by the phase-3 trials (roughly 0 to 4000 mgday/L for aducanumab / gantenerumab and 0 to 6000 mgday/L for donanemab / lecanemab).

fig3 <- tidyr::crossing(
  drug = mab_targets$drug,
  auc  = seq(0, 6000, by = 100)
) |>
  dplyr::left_join(mab_targets |> dplyr::select(drug, slope, MW), by = "drug") |>
  dplyr::mutate(fold_kout = 1 + slope * (auc / MW))

ggplot(fig3, aes(auc, fold_kout, colour = drug, group = drug)) +
  geom_line(linewidth = 0.8) +
  labs(x = "mAb 4-week AUC (mg*day/L)",
       y = "Fold increase in Kout",
       colour = "mAb",
       title = "Figure 3 -- fold increase in plaque elimination rate vs AUC",
       caption = "Replicates Figure 3 of van Maanen 2025 (median line; paper also shows 95% CIs).") +
  theme_minimal()

Donanemab has the steepest slope (highest potency per mmol*day of exposure) and gantenerumab the shallowest, matching the paper.

Replicating Figure 4A: two-year treatment effect and 5-year washout

Figure 4A simulates natural disease progression (baseline 10 CL, no drug) against BACE1 inhibition and each mAb treatment starting at a plaque burden of 85 CL for 2 years, followed by 5 years of washout. Below we reproduce the five arms.

sim_arm <- function(id, label, auc_col, auc_value,
                    treat_start_days = 365, treat_end_days = 3 * 365) {
  # Grid: 30-day steps across a 10-year horizon (0 to 3650 days).
  grid <- seq(0, 10 * 365, by = 30)
  on_treat <- grid >= treat_start_days & grid < treat_end_days
  df <- data.frame(
    id        = id,
    time      = grid,
    evid      = 0L,
    cmt       = "plaque",
    amt       = 0,
    AUC_VERUB = 0, AUC_ADU = 0, AUC_DON = 0, AUC_GAN = 0, AUC_LEC = 0,
    PLAQUE_BL = 10,   # start at 10 CL and grow toward 85 during t < treat_start
    arm       = label
  )
  if (!is.na(auc_col)) {
    df[[auc_col]][on_treat] <- auc_value
  }
  df
}

# Approximate steady-state AUCs used in the fig 4A phase-3-target arms:
# verubecestat 40 mg -> ~4.4 uM*h; aducanumab 10 mg/kg Q4W -> ~1944 mg*day/L;
# donanemab 1400 mg Q4W -> ~2800 mg*day/L; gantenerumab 1020 mg per 4w ->
# ~1700 mg*day/L; lecanemab 10 mg/kg Q2W -> ~3500 mg*day/L per 4-week window.

arms <- dplyr::bind_rows(
  sim_arm(1, "Natural progression",           NA,         NA_real_),
  sim_arm(2, "Verubecestat 40 mg QD",         "AUC_VERUB", 4.4),
  sim_arm(3, "Aducanumab 10 mg/kg Q4W",       "AUC_ADU",   1944),
  sim_arm(4, "Donanemab 1400 mg Q4W",         "AUC_DON",   2800),
  sim_arm(5, "Gantenerumab 1020 mg per 4w",   "AUC_GAN",   1700),
  sim_arm(6, "Lecanemab 10 mg/kg Q2W",        "AUC_LEC",   3500)
)

# Plaque baseline is per-subject, not per-time-point; take the row-1 value.
sim_arms <- rxode2::rxSolve(mod, events = arms, keep = "arm") |> as.data.frame()
#> Warning: multi-subject simulation without without 'omega'

# Use the arm label from the events table so line colours are stable.
ggplot(sim_arms, aes(time / 365, plaque, colour = arm, group = id)) +
  geom_vline(xintercept = c(1, 3), linetype = "dashed", colour = "grey60") +
  geom_line(linewidth = 0.7) +
  labs(x = "Time (years)",
       y = "Plaque burden (CL)",
       colour = "Arm",
       title = "Figure 4A -- 2-year treatment starting at plaque burden 85 CL",
       caption = "Dashed lines at year 1 (treatment start) and year 3 (treatment end / washout begin). Replicates Figure 4A of van Maanen 2025.") +
  theme_minimal() +
  theme(legend.position = "bottom")

Qualitatively (per Figure 4A):

  • Natural progression rises from 10 CL toward the ~95 CL steady state.
  • Verubecestat produces a modest reduction of plaque over 2 years (BACE1 inhibition slows the growth rate; it does not accelerate elimination).
  • Donanemab and lecanemab produce dramatic reductions (fold-Kout 13-19, so plaque t1/2 drops from 6.4 years to about 0.35-0.47 years) with rapid regrowth after washout.
  • Aducanumab produces intermediate reduction; gantenerumab produces the smallest reduction of the four mAbs.
  • Post-treatment regrowth returns to the natural progression rate, as predicted by the paper.

Replicating Table S5: BACE1 inhibition levels to stabilise plaque

Table S5 reports the verubecestat dose (and equivalent BACE1 inhibition level from PET) required to hold plaque burden constant at each of six baselines. At steady state, d/dt(plaque) = 0 implies Kin * (1 - Inh_verub) = plaque * Kout * (1 + 0) (no mAb), so Inh_verub = 1 - plaque * Kout / Kin. Given Inh_verub = AUC / (AUC + AUC50) we can back-compute AUC for each stable-plaque target.

stable_plaque_cl <- c(10, 20, 25, 40, 50, 62)
paper_dose_mg    <- c(27.3, 11.9, 9.0, 4.4, 2.9, 1.7)
paper_inh_pct    <- c(89.4, 78.7, 73.6, 57.7, 47.2, 34.5)

auc50 <- 0.392

table_s5 <- tibble::tibble(
  `Baseline plaque (CL)`      = stable_plaque_cl,
  `Paper verubecestat (mg)`   = paper_dose_mg,
  `Paper BACE1 inhibition %`  = paper_inh_pct
) |>
  dplyr::mutate(
    `Model inhibition (fraction)` = 1 - `Baseline plaque (CL)` * kout_val / kin_val,
    `Model inhibition %`          = 100 * `Model inhibition (fraction)`,
    `Model AUC (uM*h)`            = `Model inhibition (fraction)` /
                                    (1 - `Model inhibition (fraction)`) * auc50
  ) |>
  dplyr::select(-`Model inhibition (fraction)`)

knitr::kable(table_s5, digits = 2,
             caption = "Model-computed BACE1 inhibition levels and verubecestat AUC to hold plaque stable at each baseline vs. Table S5. Model inhibition % is derived directly from Kin/Kout ratio and reproduces the paper's Table S5 percentages within rounding.")
Model-computed BACE1 inhibition levels and verubecestat AUC to hold plaque stable at each baseline vs. Table S5. Model inhibition % is derived directly from Kin/Kout ratio and reproduces the paper’s Table S5 percentages within rounding.
Baseline plaque (CL) Paper verubecestat (mg) Paper BACE1 inhibition % Model inhibition % Model AUC (uM*h)
10 27.3 89.4 89.43 3.32
20 11.9 78.7 78.87 1.46
25 9.0 73.6 73.58 1.09
40 4.4 57.7 57.73 0.54
50 2.9 47.2 47.16 0.35
62 1.7 34.5 34.48 0.21

The model-derived BACE1 inhibition percentages align with Table S5. The per-mg conversion is not embedded in this model (it requires the upstream Dockendorf 2022 verubecestat popPK model to translate mg dose to uM*h AUC); the AUC values in the last column are the model’s back-computed exposure requirement.

Assumptions and deviations

  • No IIV encoded. The paper explicitly states that no additional interindividual variability was estimated (Supplement ‘Further details on the exposure-response model’). The summary-level nature of the mAb data constrained the fit to typical-value predictions; per-subject variability was inherited only from the individual baseline plaque burden. This is faithfully reproduced in the model file with no eta* parameters.
  • Verubecestat AUC50 reported twice. Table 1 gives AUC50 = 0.392 uMh (RSE 59.7%); Section 3.1 mentions AUC50 = 0.402 uMh in the derived “reduce plaque formation by 91.8%” calculation. The Table 1 point estimate (0.392) is used as the primary value in the model file; the difference is a rounding artifact within the parameter’s confidence interval (Table 1 bootstrap 95% CI: 0.0609 to 0.920 uM*h).
  • plaque observation is a convention warning. checkModelConventions() flags the observation variable plaque as non-canonical (the canonical concentration-output name is Cc). Because the observation here is a brain amyloid-plaque burden measured by PET on the Centiloid scale – neither a drug concentration nor a hematological / oncology endpoint that would fit an existing canonical – the model retains the paper-native plaque name (matching the model’s compartment). This deviation is intentional; the warning is informational only.
  • Time-varying AUC covariates. AUC_VERUB, AUC_ADU, AUC_DON, AUC_GAN, AUC_LEC are supplied as time-varying step-wise covariates. The paper’s E-R model was fit against AUCs from upstream compound-specific popPK models (Dockendorf 2022 for verubecestat; Sevigny 2016 for aducanumab; Lowe 2021 for donanemab; Portron 2020 for gantenerumab; Logovinsky 2016 for lecanemab; see Table S1). This model does not embed those PK layers; downstream users compute or look up per-subject AUCs from the appropriate upstream popPK (or in typical-subject simulations, use the steady-state values quoted in the covariate description tables of the model file).
  • **Slope units require AUC in mg*day/L.** The paper’s slope units mM^-1 day^-1 make Stim dimensionless only when AUC / MW evaluates in mmolday/L (= mMday). This forces AUC_mAb to be supplied in mgday/L (not mgh/mL / ugh/mL / mgh/L). Users converting from a typical clinical-PK unit mg*h/L must divide by 24 before assigning to the covariate. This is documented in each covariate’s notes field.
  • Molecular weights hard-coded in model(). The four mAb MWs are fixed physical constants (aducanumab 145912; donanemab 145087; gantenerumab 146300; lecanemab 150000) from Supplement ‘Further details on the exposure-response model’; they appear as unnamed constants at the top of model() rather than as ini() parameters (mirroring the Bosch 2024 cotadutide QSP convention for structural physical constants).
  • AUC_<DRUG> and PLAQUE_BL canonicals ratified. Six new covariate canonicals were registered in inst/references/covariate-columns.md alongside this extraction: AUC_VERUB, AUC_ADU, AUC_DON, AUC_GAN, AUC_LEC (AUC_<DRUG> family; siblings of AUC_CARBO, AUC_GEM, AUC_GCV, AUC_LCM, AUC_CBZ, AUC_PAZO, AUC_RTV, AUC_EMPA), and PLAQUE_BL (per-subject baseline plaque as initial condition; sibling of HGB_BL).
  • Coincident mAb use. The paper simulates a single mAb at a time (Figures 4A, 4B). The model’s `Stim_mAb = Stim_adu + Stim_don + Stim_gan
    • Stim_lec` sums linearly across the four mAbs; simultaneous use of multiple mAbs is a linear extrapolation beyond the fit and should be treated with caution (competitive-binding and dose-limiting-toxicity considerations are outside the scope of the paper’s E-R model).
  • Additive-error scale. The paper’s residual error is additive on the raw Centiloid scale (Supplement Eq 3). The model uses plaque ~ add(addSd) which encodes the same additive form.