Skip to contents

Rifampicin (Clewe 2018)

Source: vignettes/articles/Clewe_2018_rifampicin.Rmd
Clewe_2018_rifampicin.Rmd

Model and source

  • Citation: Clewe O., Wicha S. G., de Vogel C. P., de Steenwinkel J. E. M., Simonsson U. S. H. (2018). A model-informed preclinical approach for prediction of clinical pharmacodynamic interactions of anti-tuberculosis drug combinations. J Antimicrob Chemother 73(2):437-447. doi:10.1093/jac/dkx380.
  • DDMORE Foundation Model Repository: DDMODEL00000259 (Scenario = 4, the triple-combination MTP-GPDI fit).
  • Article: https://doi.org/10.1093/jac/dkx380.

The model is the Multistate Tuberculosis Pharmacometric (MTP) model coupled to the General Pharmacodynamic Interaction (GPDI) model for the triple combination of rifampicin (RIF), isoniazid (INH), and ethambutol (EMB) against an in vitro Mycobacterium tuberculosis B1585 culture. The MTP block carries three bacterial subpopulations – fast-multiplying (Fbugs), slow-multiplying (Sbugs), and non-replicating persisters (Nbugs) – that exchange via first-order rates and a time-dependent F-to-S transfer (KFS = kfslin * t). INH adaptive resistance is captured by a two-state ARON / AROFF system whose dynamic ARON value linearly inflates the INH EC50 on Fbugs and Sbugs. Each of the three drugs acts on each subpopulation through a Hill-type or hyperbolic exposure-response, combined across drugs via Bliss independence on Fbugs and linear summation on Sbugs / Nbugs; pairwise GPDI interaction parameters shift the Emax / EC50 of the affected drug-effect terms.

Population

The dataset is in vitro bacterial culture, not human subjects. The single strain is M. tuberculosis B1585 (Beijing genotype clinical isolate) cultured under static drug exposures of rifampicin, isoniazid, and ethambutol alone or in combination. Time is in days; bacterial counts are in CFU/mL. The full population description is available at readModelDb("Clewe_2018_rifampicin")$population.

The bundle’s simulated dataset spans seven experimental codes (EXPR codes 1, 2, 3, 7, 8, 11, 13) corresponding to control, single-drug, two-drug, and three-drug exposure arms; the present model fixes the MTP block and the single-drug exposure-response blocks at their published values and uses the GPDI parameters fitted on the combination data (DDMORE bundle annotation Scenario = 4).

Source trace

Every parameter line in inst/modeldb/ddmore/Clewe_2018_rifampicin.R carries a trailing comment pointing to the THETA index in the DDMORE bundle’s Executable_MTP-GPDI.mod. The block below reproduces the cross-walk for review. The bundle’s Output_real_MTP-GPDI.lst is a $TABLE export rather than a NONMEM listing file, so a direct MINIMIZATION SUCCESSFUL cross-check is not available; the Output_simulated_MTP-GPDI.lst re-prints the same THETA values in its FINAL PARAMETER ESTIMATE block (run with MAXEVAL = 0), confirming the .mod $THETA initial values are the operating final estimates for this scenario.

nlmixr2 parameter NONMEM THETA Final value (.mod $THETA, scaled) Notes
kg THETA(1) FIX 0.796 1/day Fast-subpopulation exponential growth-rate
kfslin THETA(2)/100 FIX 0.00166 1/day^2 Time-dependent F-to-S transfer slope (KFS = kfslin * t)
kfn THETA(3)/1e6 FIX 8.97e-7 1/day F-to-N transfer rate
ksf THETA(4)/10 FIX 0.0145 1/day S-to-F transfer rate
ksn THETA(5) FIX 0.186 1/day S-to-N transfer rate
kns THETA(6)/100 FIX 0.00123 1/day N-to-S transfer rate
f0 THETA(7)*1000 FIX 209,000 CFU/mL Initial fast-subpopulation count
s0 THETA(8)*1000 FIX 324,000 CFU/mL Initial slow-subpopulation count
hfdemax THETA(9) FIX 22.2209 INH Emax on Fbugs (Bliss-scaled)
hfdec50 THETA(10) FIX 0.168 mg/L INH EC50 on Fbugs (reference)
hfdgam THETA(11) FIX 1.902 INH Hill exponent on Fbugs
hsdemax THETA(12) FIX 8.553 1/day INH Emax on Sbugs
hsdec50 THETA(13) FIX 0.0329 mg/L INH EC50 on Sbugs (reference)
hsdgam THETA(14) FIX 1.741 INH Hill exponent on Sbugs
kon THETA(15) FIX 0.0206 (mg/L * day)^-1 INH adaptive-resistance ON-rate
koff THETA(16) FIX 0 INH adaptive-resistance OFF-rate (irreversible)
arlinfd THETA(17) FIX 522.42 Linear adaptive-resistance slope on INH-FD EC50
arlinsd THETA(18) FIX 2352.28 Linear adaptive-resistance slope on INH-SD EC50
rfdemax THETA(19) FIX 1.969 (ratio) RIF Emax on Fbugs (relative to hfdemax)
rfdec50 THETA(20) FIX 0.00303 mg/L RIF EC50 on Fbugs
rfgemax THETA(21) FIX 1 RIF growth-inhibition Emax on Fbugs growth
rfgec50 THETA(22) FIX 0.388 mg/L RIF growth-inhibition EC50
rfggam THETA(23) FIX 2.802 RIF growth-inhibition Hill exponent
rsdemax THETA(24) FIX 1.792 1/day RIF Emax on Sbugs
rsdec50 THETA(25) FIX 0.0113 mg/L RIF EC50 on Sbugs
rndk THETA(26) FIX 3.286 (mg/L * day)^-1 RIF first-order kill on Nbugs
efdemax THETA(27) FIX 2.207 (ratio) EMB Emax on Fbugs (relative to hfdemax)
efdec50 THETA(28) FIX 0.860 mg/L EMB EC50 on Fbugs
efdgam THETA(29) FIX 2.458 EMB Hill exponent on Fbugs
esdk THETA(30) FIX 4.388 (mg/L * day)^-1 EMB first-order kill on Sbugs
fdirh THETA(31) -0.683 GPDI: RIF -> INH-FD EC50 shift
fdihr THETA(32) FIX 0 GPDI: INH -> RIF-FD EC50 shift
sdirh THETA(33) 1.529 GPDI: RIF -> INH-SD EC50 shift
sdihr THETA(34) 10.749 GPDI: INH -> RIF-SD EC50 shift
fdieh THETA(35) 1.809 GPDI: EMB -> INH-FD EC50 shift
fdihe THETA(36) FIX 0 GPDI: INH -> EMB-FD EC50 shift
sdieh THETA(37) 0.0855 (cond. EMB > 0) GPDI: EMB -> INH-SD EC50 shift
sdihe THETA(38) 91.422 (cond. INH > 0) GPDI: INH -> EMB-SD scalar shift
fdier THETA(39) -0.663 (cond. EMB > 0) GPDI: EMB -> RIF-FD EC50 shift
fdire THETA(40) FIX -0.99999 GPDI: RIF -> EMB-FD EC50 shift
sdier THETA(41) 1.706 (cond. EMB > 0) GPDI: EMB -> RIF-SD EC50 shift
sdire THETA(42) 479.458 GPDI: RIF -> EMB-SD scalar shift
sdierh THETA(43) -0.677 (cond. EMB > 0) GPDI: EMB modulating the RIF-INH SDIRH interaction
addSd SIGMA(1) FIX sqrt(0.937) ~= 0.968 (log scale) Additive residual SD on log(F + S)

Validation strategy

The DDMORE bundle ships a simulated event dataset (Simulated_Mtb-B1585_In-vitro-NATG-RIF-INH-EMB.csv) whose DV column is the NONMEM-generated stochastic prediction (IPRED + EPS(1)) for the scenario-4 model on a representative grid of single-, two-, and three-drug exposures. The associated publication (doi:10.1093/jac/dkx380) is not on disk in this worktree, so PKNCA-style comparison against published Cmax / AUC tables is not applicable (this is bacterial-count PD on log scale, not concentration-time PK). The validation strategy is therefore the F.2 self-consistency path of extract-literature-model references/verification-checklist.md augmented with the F.3 mechanistic-sanity path:

  1. Re-simulate selected representative experiments (control, single-drug monotherapy at three RIF / INH / EMB concentrations, and a three-drug combination) through rxode2::rxSolve() with typical-value parameters (no IIV, no residual error) and compare the trajectory against the bundle’s DV values.
  2. Demonstrate the mechanistic kill ordering – control (pure growth) < monotherapy (drug-specific kill) < combination (interaction-modulated strongest effect) – at standard pharmacologically-relevant concentrations spanning the bundle’s grid.

Representative experiments

The chunk below reproduces six representative IDs from the bundle’s simulated CSV. Each row is one observation; DV is the NONMEM-simulated stochastic prediction; RIF / INH / EMB are the time-fixed exposure concentrations (mg/L) supplied per replicate. The EXPR codes match the bundle’s experiment labels (1 = control, 2 = RIF mono, 7 = INH mono, 11 = RIF + EMB, 13 = RIF + INH + EMB).

bundle <- tibble::tribble(
  ~id, ~time, ~DV,   ~RIF,  ~INH,  ~EMB,  ~EXPR, ~scenario,
  # ID 1: control - exponential growth
  1L,  0,    13.19, 0.000, 0.000, 0.000, 1L,    "Control (no drug)",
  1L,  1,    13.51, 0.000, 0.000, 0.000, 1L,    "Control (no drug)",
  1L,  2,    14.05, 0.000, 0.000, 0.000, 1L,    "Control (no drug)",
  1L,  3,    14.73, 0.000, 0.000, 0.000, 1L,    "Control (no drug)",
  1L,  6,    17.03, 0.000, 0.000, 0.000, 1L,    "Control (no drug)",
  # ID 24: RIF 8 mg/L monotherapy - strong kill
  24L, 0,    13.19, 8.000, 0.000, 0.000, 2L,    "RIF 8 mg/L mono",
  24L, 1,    11.21, 8.000, 0.000, 0.000, 2L,    "RIF 8 mg/L mono",
  24L, 2,     9.24, 8.000, 0.000, 0.000, 2L,    "RIF 8 mg/L mono",
  24L, 3,     7.27, 8.000, 0.000, 0.000, 2L,    "RIF 8 mg/L mono",
  24L, 6,     1.35, 8.000, 0.000, 0.000, 2L,    "RIF 8 mg/L mono",
  # ID 111: INH 0.01 mg/L monotherapy - weak (sub-MIC); growth dominant
  111L, 0,   13.19, 0.000, 0.010, 0.000, 7L,    "INH 0.01 mg/L mono",
  111L, 1,   13.23, 0.000, 0.010, 0.000, 7L,    "INH 0.01 mg/L mono",
  111L, 2,   13.76, 0.000, 0.010, 0.000, 7L,    "INH 0.01 mg/L mono",
  111L, 3,   14.44, 0.000, 0.010, 0.000, 7L,    "INH 0.01 mg/L mono",
  111L, 6,   16.64, 0.000, 0.010, 0.000, 7L,    "INH 0.01 mg/L mono",
  # ID 115: RIF 0.125 mg/L monotherapy - moderate kill
  115L, 0,   13.19, 0.125, 0.000, 0.000, 7L,    "RIF 0.125 mg/L mono",
  115L, 1,   11.68, 0.125, 0.000, 0.000, 7L,    "RIF 0.125 mg/L mono",
  115L, 2,   10.28, 0.125, 0.000, 0.000, 7L,    "RIF 0.125 mg/L mono",
  115L, 3,    8.98, 0.125, 0.000, 0.000, 7L,    "RIF 0.125 mg/L mono",
  115L, 6,    5.35, 0.125, 0.000, 0.000, 7L,    "RIF 0.125 mg/L mono",
  # ID 247: RIF 0.002 + EMB 32 mg/L two-drug
  247L, 0,   13.19, 0.002, 0.000, 32.00, 11L,   "RIF 0.002 + EMB 32 mg/L",
  247L, 1,   11.05, 0.002, 0.000, 32.00, 11L,   "RIF 0.002 + EMB 32 mg/L",
  247L, 2,    8.98, 0.002, 0.000, 32.00, 11L,   "RIF 0.002 + EMB 32 mg/L",
  247L, 3,    6.98, 0.002, 0.000, 32.00, 11L,   "RIF 0.002 + EMB 32 mg/L",
  247L, 6,    3.09, 0.002, 0.000, 32.00, 11L,   "RIF 0.002 + EMB 32 mg/L",
  # ID 295: triple combination RIF 0.002 + INH 0.63 + EMB 0.5 mg/L
  295L, 0,   13.19, 0.002, 0.630, 0.500, 13L,   "RIF 0.002 + INH 0.63 + EMB 0.5",
  295L, 1,    7.73, 0.002, 0.630, 0.500, 13L,   "RIF 0.002 + INH 0.63 + EMB 0.5",
  295L, 2,    6.23, 0.002, 0.630, 0.500, 13L,   "RIF 0.002 + INH 0.63 + EMB 0.5",
  295L, 3,    5.46, 0.002, 0.630, 0.500, 13L,   "RIF 0.002 + INH 0.63 + EMB 0.5",
  295L, 6,    4.27, 0.002, 0.630, 0.500, 13L,   "RIF 0.002 + INH 0.63 + EMB 0.5"
)
bundle$scenario <- factor(bundle$scenario, levels = unique(bundle$scenario))
knitr::kable(
  bundle |> dplyr::group_by(id, scenario, RIF, INH, EMB) |>
    dplyr::summarise(t_grid = paste(time, collapse = ", "),
                     dv_grid = paste(sprintf("%.2f", DV), collapse = ", "),
                     .groups = "drop"),
  caption = "Six representative bundle experiments (subset of DDMODEL00000259's Simulated_*.csv)."
)
Six representative bundle experiments (subset of DDMODEL00000259’s Simulated_*.csv).
id scenario RIF INH EMB t_grid dv_grid
1 Control (no drug) 0.000 0.00 0.0 0, 1, 2, 3, 6 13.19, 13.51, 14.05, 14.73, 17.03
24 RIF 8 mg/L mono 8.000 0.00 0.0 0, 1, 2, 3, 6 13.19, 11.21, 9.24, 7.27, 1.35
111 INH 0.01 mg/L mono 0.000 0.01 0.0 0, 1, 2, 3, 6 13.19, 13.23, 13.76, 14.44, 16.64
115 RIF 0.125 mg/L mono 0.125 0.00 0.0 0, 1, 2, 3, 6 13.19, 11.68, 10.28, 8.98, 5.35
247 RIF 0.002 + EMB 32 mg/L 0.002 0.00 32.0 0, 1, 2, 3, 6 13.19, 11.05, 8.98, 6.98, 3.09
295 RIF 0.002 + INH 0.63 + EMB 0.5 0.002 0.63 0.5 0, 1, 2, 3, 6 13.19, 7.73, 6.23, 5.46, 4.27

Typical-value re-simulation 

events <- bundle |>
  dplyr::transmute(
    id    = id,
    time  = time,
    evid  = 0L,
    cmt   = NA_integer_,
    RIF   = RIF,
    INH   = INH,
    EMB   = EMB,
    scenario = scenario,
    DV    = DV
  )

# Densify the time grid for smoother trajectories. Add 0.25-day-step
# observation rows for plotting; keep the bundle's discrete time points
# (with their DV values) for the side-by-side comparison.
plot_grid <- expand.grid(
  id   = unique(events$id),
  time = seq(0, 6, by = 0.25)
) |>
  dplyr::left_join(
    events |> dplyr::select(id, RIF, INH, EMB, scenario) |> dplyr::distinct(),
    by = "id"
  ) |>
  dplyr::mutate(evid = 0L, cmt = NA_integer_, DV = NA_real_)

events_plot <- dplyr::bind_rows(events, plot_grid) |>
  dplyr::distinct(id, time, .keep_all = TRUE) |>
  dplyr::arrange(id, time)
mod <- readModelDb("Clewe_2018_rifampicin")
mod_typical <- rxode2::zeroRe(mod)
#> Warning: No omega parameters in the model
sim <- rxode2::rxSolve(mod_typical, events = events_plot,
                       keep = c("scenario", "DV", "RIF", "INH", "EMB"),
                       returnType = "data.frame")
#> Warning: multi-subject simulation without without 'omega'
cat("Simulation rows:", nrow(sim),
    " unique IDs:", length(unique(sim$id)), "\n")
#> Simulation rows: 150  unique IDs: 6

Self-consistency check (F.2)

The DDMORE bundle’s DV column is a NONMEM-simulated stochastic observation (IPRED + EPS(1), EPS variance 0.937 on the natural-log scale). Our typical-value re-simulation produces logFSbugs = ln(Fbugs + Sbugs) which corresponds directly to NONMEM’s IPRED. For each observation we compare logFSbugs (rxode2 typical-value) against DV (NONMEM-simulated stochastic).

diff_df <- sim |>
  dplyr::filter(!is.na(DV)) |>
  dplyr::mutate(
    residual_log = DV - logFSbugs,
    abs_residual = abs(residual_log)
  )

summary_df <- diff_df |>
  dplyr::group_by(scenario) |>
  dplyr::summarise(
    n              = dplyr::n(),
    median_dv      = stats::median(DV),
    median_logfs   = stats::median(logFSbugs),
    rmse           = sqrt(mean(residual_log^2)),
    median_abs_res = stats::median(abs_residual),
    .groups = "drop"
  )
knitr::kable(summary_df, digits = 3,
             caption = paste("Per-scenario self-consistency summary.",
                             "RMSE / median |DV - logFSbugs| are on natural-log scale."))
Per-scenario self-consistency summary. RMSE / median |DV - logFSbugs| are on natural-log scale.
scenario n median_dv median_logfs rmse median_abs_res
Control (no drug) 5 14.05 14.049 0.002 0.001
RIF 8 mg/L mono 5 9.24 9.241 0.003 0.002
INH 0.01 mg/L mono 5 13.76 13.762 0.004 0.004
RIF 0.125 mg/L mono 5 10.28 10.282 0.003 0.002
RIF 0.002 + EMB 32 mg/L 5 8.98 8.983 0.004 0.004
RIF 0.002 + INH 0.63 + EMB 0.5 5 6.23 6.231 0.003 0.004

The expected per-observation residual under perfect translation is the NONMEM stochastic noise term, with theoretical SD ~= sqrt(0.937) ~= 0.968 on the natural-log scale. RMSE values comparable to that magnitude indicate the trajectory is consistent with the source model; values much larger flag a translation defect.

ggplot(sim, aes(time, logFSbugs)) +
  geom_line(color = "steelblue", linewidth = 0.7) +
  geom_point(data = diff_df, aes(y = DV), alpha = 0.7, size = 1.6) +
  facet_wrap(~ scenario, scales = "free_y") +
  labs(x = "Time (days)", y = "ln(Fbugs + Sbugs)  (CFU/mL on log scale)",
       title = "Self-consistency: rxode2 typical value vs DDMORE-simulated DV",
       caption = paste("Points: bundle DV (NONMEM-simulated stochastic);",
                       "line: rxode2::rxSolve typical-value `logFSbugs`."))
Typical-value rxode2 trajectory (line) vs DDMORE-bundle simulated DV (points), one panel per scenario.

Typical-value rxode2 trajectory (line) vs DDMORE-bundle simulated DV (points), one panel per scenario.

Mechanistic-sanity check (F.3)

Below we plot all three bacterial subpopulations (Fbugs, Sbugs, Nbugs) to confirm the model exhibits the published mechanistic behavior:

  • Control (no drug): exponential growth of Fbugs dominates; Sbugs and Nbugs remain small relative to Fbugs.
  • Single-drug monotherapy (RIF 8 mg/L): strong kill of Fbugs and Sbugs; Nbugs initially seeded by F-to-N transfer is itself killed by RIF (only RIF acts on Nbugs in this model).
  • Triple combination (RIF + INH + EMB): combined kill substantially faster than any monotherapy at the same low concentrations.
sub_long <- sim |>
  dplyr::select(id, time, scenario, Fbugs, Sbugs, Nbugs) |>
  tidyr::pivot_longer(c(Fbugs, Sbugs, Nbugs),
                      names_to = "subpop", values_to = "count") |>
  dplyr::mutate(subpop = factor(subpop, levels = c("Fbugs", "Sbugs", "Nbugs")),
                count_safe = pmax(count, 1e-3))   # avoid log10(0) artefacts on the plot

ggplot(sub_long, aes(time, count_safe, color = subpop)) +
  geom_line(linewidth = 0.7) +
  facet_wrap(~ scenario, scales = "free_y") +
  scale_y_log10(labels = function(x) format(x, scientific = TRUE, digits = 2)) +
  labs(x = "Time (days)", y = "Bacterial count (CFU/mL, log10)",
       color = "Subpopulation",
       title = "Mechanistic-sanity: F / S / N trajectories per scenario")
Bacterial subpopulation trajectories on log10 scale across the six representative scenarios. F = fast-multiplying; S = slow-multiplying; N = non-replicating persisters.

Bacterial subpopulation trajectories on log10 scale across the six representative scenarios. F = fast-multiplying; S = slow-multiplying; N = non-replicating persisters.

Assumptions and deviations

The verbatim translation deliberately preserves several non-standard features of the source model. Each is flagged here for downstream review.

  • Drug-concentration compartments collapsed to scalar covariates. The source .mod declares INH (compartment 4), RIF (compartment 7), and EMB (compartment 8) as ODE compartments with DADT = 0 and initial values seeded from the dataset’s INH, RIF, EMB columns. Because the compartments are constant in time per replicate, the nlmixr2 translation drops them and uses scalar covariates of the same name; the bacterial-state predictions are unchanged.
  • Source .mod typo: RIFEFG vs RIFFG. The $DES block contains IF(RIFEFG.LT.0) RIFEFG=0 immediately after defining RIFFG = 1 - .... RIFEFG is never declared; with RFGEMAX = 1 (FIXED) the surrounding expression is mathematically bounded to [0, 1] and the guard is unnecessary. The guard is omitted in the nlmixr2 translation; the predictions match the source behavior because the guard never fired.
  • Source .mod parenthesisation in EMBSD. The line EMBSD = A(8) * (ESDK / ((1 + SDIHE * A(4)) * (1 + SDIRE * A(7) / RSDEC50 + A(7)))) parses the trailing + A(7) as part of the sum (1 + SDIRE*RIF/RSDEC50 + RIF) rather than the canonical Hill-shift RIF/(RSDEC50 + RIF). The parenthesisation is reproduced verbatim because (a) it is what the scenario-4 fit was run against and (b) editing it would change the predictions. Flagged here as a likely upstream typo; downstream users who want the canonical Hill-shift form should edit the model file.
  • EMB / INH presence guards reproduced via covariate factors. The source .mod gates five GPDI parameters (sdieh, sdihe, fdier, sdier, sdierh) behind IF(A(8) > 0) (or IF(A(4) > 0) for sdihe) blocks. Since the drug compartments are collapsed to scalar covariates in this implementation, the guard is reproduced as a multiplicative factor (EMB > 0) (resp. (INH > 0)) on the parameter inside model(). The numeric behavior is identical because the source compartments are constant in time at the covariate value.
  • Output_real_*.lst is a $TABLE export, not a NONMEM listing. The bundle’s Output_real_MTP-GPDI.lst is a 1,021-row text table starting TABLE NO. 1. ID ... (a NONMEM $TABLE output) rather than a NONMEM listing with a MINIMIZATION SUCCESSFUL block. Final estimates are taken from the .mod $THETA block (which holds the published scenario-4 fit values; the Output_simulated_*.lst re-prints these in its FINAL PARAMETER ESTIMATE block under MAXEVAL = 0 evaluation, confirming the cross-walk).
  • Linked publication not on disk. The associated paper (doi:10.1093/jac/dkx380) is not available locally in this worktree (/home/bill/github/mab_human_consensus/literature/). PKNCA-style comparison against published tables / figures from the publication is therefore not possible. Validation reduces to the F.2 self-consistency check against the bundle-shipped DV and the F.3 mechanistic-sanity check on bacterial-subpopulation trajectories above.
  • koff = 0 (irreversible adaptive resistance). The source fixes KOFF = 0, so the ARON / AROFF system is monotone (ARON increases with INH exposure; never returns to AROFF). Reproduced verbatim.
  • No IIV. The source $OMEGA = 0 FIX. The model is a population-typical mechanism; downstream users adding IIV must augment the model file with their own eta* parameters.
  • M3 LOQ handling deferred to data. The source .mod uses NONMEM’s M3 method (F_FLAG = 1, Y = PHI((LLOQ - IPRED)/SD)) with LLOQ = ln(5) = 1.6094. nlmixr2 handles below-LOQ data via the cens event-table column at fit time; the model file declares only the standard additive residual on logFSbugs. Users fitting against censored data should populate the dataset’s cens column accordingly.
  • Non-canonical observation, compartments, and covariates flagged as 10 checkModelConventions() warnings – Fbugs / Sbugs / Nbugs / aron / aroff are non-canonical compartments, logFSbugs is the paper-named observation (the canonical Cc is misleading for a log bacterial-density observation), RIF / INH / EMB are dataset-tied in vitro exposure columns rather than reusable population covariates (matching the precedent in Mohamed_2016_colistin_meropenem.R), and units$dosing (CFU/mL inoculum) and units$concentration (mg/L drug) are different physical quantities by design.