Skip to contents

Model and source

  • Article: https://doi.org/10.1128/aac.00301-23
  • Supplementary material (aac.00301-23-s0001.docx): visual predictive checks for the three other isolates (Figs. S1-S3), observed-versus-predicted plots, and the NONMEM control streams of the expected-additivity (Code S1) and observed-interaction (Code S2) models for isolate NARA1295.
isolates <- c("NARA1295", "NARA1584", "NARA1182", "NARA864")
model_names <- paste0("Aubry_2023_ceftazidime_avibactam_colistin_", isolates)
mods <- setNames(lapply(model_names, function(nm) rxode2::rxode2(readModelDb(nm))), isolates)
cat(mods$NARA1295$reference, sep = "\n")
#> Aubry R, Buyck J, Prouvensier L, Decousser JW, Nordmann P, Wicha SG, Marchand S, Gregoire N. An improved PKPD modeling approach to characterize the pharmacodynamic interaction over time between ceftazidime/avibactam and colistin from in vitro time-kill experiments against multidrug-resistant Klebsiella pneumoniae isolates. Antimicrob Agents Chemother. 2023;67(10):e00301-23. doi:10.1128/aac.00301-23. Structural equations: main-text Eqs 1-6 and the NONMEM control stream of this isolate's observed interaction model (Supplementary Code S2). Parameter estimates: main-text Table 2.

Aubry 2023 fits one semi-mechanistic time-kill model per isolate to static time-kill curves of four KPC-3-producing Klebsiella pneumoniae clinical isolates exposed to ceftazidime/avibactam (CZA; avibactam always at 4 mg/L) and colistin (CST), alone and in combination. The structure is shared, but every isolate has its own parameter column in Table 2 and its own visual predictive check, so the library ships four models:

Model Isolate GPDI interaction (colistin as perpetrator)
Aubry_2023_ceftazidime_avibactam_colistin_NARA1295 NARA1295 (ST512) lowers the CZA EC50 by up to 51.4%
Aubry_2023_ceftazidime_avibactam_colistin_NARA1584 NARA1584 (ST101) lowers the CZA EC50 by up to 48.3%
Aubry_2023_ceftazidime_avibactam_colistin_NARA1182 NARA1182 (ST512) lowers the CZA EC50 by up to 48.6%
Aubry_2023_ceftazidime_avibactam_colistin_NARA864 NARA864 (ST348) lowers the CZA Emax by up to 65.8%

Each model has:

  • one homogeneous bacterial population cfu (CFU/mL) with logistic growth toward bmax;
  • a sigmoidal Emax kill by each drug, whose Emax falls in proportion to the fraction of bacteria in an “adapted” state (aron_caz, aron_col) – an empirical two-state adaptive-resistance model per drug, switched on only while that drug is present, with no reversal (koff = 0);
  • Bliss independence as the additivity criterion, with each kill rate normalised by the larger of the two mono-drug Emax values (Code S2);
  • a general pharmacodynamic interaction (GPDI) term through which colistin modifies one ceftazidime parameter;
  • the observation log_cfu = log10 CFU/mL with an additive residual error.

Drug concentrations are supplied as the covariates CONC_CAZ_MGL and CONC_CST_MGL. There is no PK and there are no dosing events: the solver integrates the bacterial and adaptation states from the inoculum. This is a pharmacodynamic model of an in vitro experiment, so PKNCA is not the right validation target; the checks below compare the model against the paper’s own model predictions instead.

Population (biological context)

pop <- lapply(mods, function(m) m$population)
data.frame(
  isolate = isolates,
  disease_state = vapply(pop, function(p) p$disease_state, character(1)),
  inoculum = vapply(pop, function(p) p$initial_inoculum, character(1))
) |>
  knitr::kable()
isolate disease_state inoculum
NARA1295 NARA1295 Clinical K. pneumoniae isolate from the human strain collection (2019-2021) of the National Reference Center for Emerging Antibiotic Resistance (NARA, Fribourg, Switzerland). Sequence type ST512; resistance genes KPC-3, SHV-11 (35Q), TEM-1D, truncated OmpK35 (25%) and OmpK36 with a glycine-aspartate insertion (Table 1). MICs (broth microdilution, EUCAST): ceftazidime/avibactam 8 mg/L (avibactam fixed at 4 mg/L), colistin 0.125 mg/L; susceptible to both. Estimated 10^5.63 CFU/mL (Table 2)
NARA1584 NARA1584 Clinical K. pneumoniae isolate from the human strain collection (2019-2021) of the National Reference Center for Emerging Antibiotic Resistance (NARA, Fribourg, Switzerland). Sequence type ST101; resistance genes KPC-3, SHV-1, truncated OmpK35 (17%) and OmpK36 with a threonine-aspartate insertion (Table 1). MICs (broth microdilution, EUCAST): ceftazidime/avibactam 8 mg/L (avibactam fixed at 4 mg/L), colistin 0.125 mg/L; susceptible to both. Estimated 10^5.62 CFU/mL (Table 2)
NARA1182 NARA1182 Clinical K. pneumoniae isolate from the human strain collection (2019-2021) of the National Reference Center for Emerging Antibiotic Resistance (NARA, Fribourg, Switzerland). Sequence type ST512; resistance genes KPC-3, SHV-11 (35Q), truncated OmpK35 (25%) and OmpK36 with a glycine-aspartate insertion (Table 1). MICs (broth microdilution, EUCAST): ceftazidime/avibactam 4 mg/L (avibactam fixed at 4 mg/L), colistin 0.125 mg/L; susceptible to both. Estimated 10^5.53 CFU/mL (Table 2)
NARA864 NARA864 Clinical K. pneumoniae isolate from the human strain collection (2019-2021) of the National Reference Center for Emerging Antibiotic Resistance (NARA, Fribourg, Switzerland). Sequence type ST348; resistance genes KPC-3, SHV-11 (35Q), TEM-1D and OXA-9, with no porin alteration reported (Table 1). MICs (broth microdilution, EUCAST): ceftazidime/avibactam 2 mg/L (avibactam fixed at 4 mg/L), colistin 0.25 mg/L; susceptible to both. Estimated 10^5.68 CFU/mL (Table 2)

All four isolates come from the 2019-2021 human strain collection of the National Reference Center for Emerging Antibiotic Resistance (NARA, Fribourg, Switzerland) and are susceptible to both drugs (Table 1: CZA MICs 2-8 mg/L, CST MICs 0.125-0.25 mg/L). Static time-kill experiments ran for 30 h at 35 degrees C in 10 mL cation-adjusted Mueller-Hinton broth II from an inoculum of about 5 x 10^5 CFU/mL, with counts at 0, 2, 4, 8, 24 and 30 h and a lower limit of quantification of 200 CFU/mL (2.3 log10 CFU/mL).

Source trace

Element Source location Value(s) in the files
Sigmoidal Emax kill Eq 1 emax_*, ec50_*, hill_*
Two-state adaptive resistance Eqs 2-3; Code S2 DADT(4)-DADT(7) aron_* / aroff_*, kon_*, koff
Emax reduced by the adapted fraction Eq 4; Code S2 E_ADAPT_CZA = AR_CZA*A(4) armax_*
Adaptation only while the drug is present Code S2 IF(A(1).GT.0) ADAPT_CZA=KON_CZA kon_*_i <- kon_* * (CONC > 0)
Logistic growth on the linear CFU scale Eq 5; Code S2 DADT(3), A_0(3)=10**INOC kg, log10bmax, log10inoc
Bliss independence, normalised by max Emax Code S2 E_COMB emax_norm, kill_comb
GPDI interaction Eq 6; Code S2 INTERACTION int_col_caz, ec50int_col_caz, hillint_col_caz
Victim parameter (EC50 vs Emax) Table 2 rows INT_EC50 / INT_Emax; Results EC50 for NARA1295/1584/1182, Emax for NARA864
Residual error Table 2 RES_ADD; Code S2 W = SQRT(SIGMA(1)) addSd = sqrt(RES_ADD) (see below)
Observation floor Code S2 IF(BTOT_ERR<1e-6) BTOT_ERR=1e-6 log10(max(cfu, 1e-6))
All parameter estimates Table 2 (one column per isolate) ini() of each file

Parameter tables (paper vs. file)

Table 2 is transcribed below in the paper’s units (percentages for ARmax and INT, the residual entry as printed). The file values are read from each model’s ini() and converted to the same units, so the comparison checks the transcription and every back-transform in one pass.

table2 <- read.csv(text = "
parameter,NARA1295,NARA1584,NARA1182,NARA864
INOC,5.63,5.62,5.53,5.68
KG,1.40,1.77,1.78,1.84
Bmax,8.80,9.13,8.43,9.01
Emax_CZA,4.59,6.27,7.51,9.17
EC50_CZA,13.39,13.47,2.51,0.46
gamma_CZA,2.13,1.84,1.53,1.79
KON_CZA,0.24,0.32,0.36,0.66
KOFF,0,0,0,0
ARmax_CZA,71.9,73.0,78.6,82.5
Emax_CST,13.12,8.78,14.31,8.81
EC50_CST,0.60,0.33,0.33,0.21
gamma_CST,0.96,1.13,1.21,1.16
KON_CST,0.65,0.52,1.18,0.75
ARmax_CST,91.8,83.1,87.0,80.3
INT,-51.4,-48.3,-48.6,-65.8
EC50_INT,0.06,0.06,0.07,0.39
gamma_INT,10,10,10,1
RES_ADD,1.74,1.01,1.44,2.25
")

file_values <- function(m) {
  th <- m$theta
  c(
    INOC = th[["log10inoc"]], KG = th[["kg"]], Bmax = th[["log10bmax"]],
    Emax_CZA = th[["emax_caz"]], EC50_CZA = th[["ec50_caz"]],
    gamma_CZA = th[["hill_caz"]], KON_CZA = th[["kon_caz"]], KOFF = th[["koff"]],
    ARmax_CZA = 100 * th[["armax_caz"]],
    Emax_CST = th[["emax_col"]], EC50_CST = th[["ec50_col"]],
    gamma_CST = th[["hill_col"]], KON_CST = th[["kon_col"]],
    ARmax_CST = 100 * th[["armax_col"]],
    INT = 100 * th[["int_col_caz"]], EC50_INT = th[["ec50int_col_caz"]],
    gamma_INT = th[["hillint_col_caz"]],
    # Table 2 prints the $SIGMA variance; the file carries its square root.
    RES_ADD = th[["addSd"]]^2
  )
}
cmp <- table2 |>
  pivot_longer(-parameter, names_to = "isolate", values_to = "paper") |>
  mutate(file = mapply(function(p, i) file_values(mods[[i]])[[p]], parameter, isolate))
stopifnot(all(abs(cmp$file - cmp$paper) < 1e-8))
cmp |>
  pivot_wider(names_from = isolate, values_from = c(paper, file)) |>
  select(parameter, ends_with("NARA1295"), ends_with("NARA1584"),
         ends_with("NARA1182"), ends_with("NARA864")) |>
  knitr::kable(digits = 3)
parameter paper_NARA1295 file_NARA1295 paper_NARA1584 file_NARA1584 paper_NARA1182 file_NARA1182 paper_NARA864 file_NARA864
INOC 5.63 5.63 5.62 5.62 5.53 5.53 5.68 5.68
KG 1.40 1.40 1.77 1.77 1.78 1.78 1.84 1.84
Bmax 8.80 8.80 9.13 9.13 8.43 8.43 9.01 9.01
Emax_CZA 4.59 4.59 6.27 6.27 7.51 7.51 9.17 9.17
EC50_CZA 13.39 13.39 13.47 13.47 2.51 2.51 0.46 0.46
gamma_CZA 2.13 2.13 1.84 1.84 1.53 1.53 1.79 1.79
KON_CZA 0.24 0.24 0.32 0.32 0.36 0.36 0.66 0.66
KOFF 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
ARmax_CZA 71.90 71.90 73.00 73.00 78.60 78.60 82.50 82.50
Emax_CST 13.12 13.12 8.78 8.78 14.31 14.31 8.81 8.81
EC50_CST 0.60 0.60 0.33 0.33 0.33 0.33 0.21 0.21
gamma_CST 0.96 0.96 1.13 1.13 1.21 1.21 1.16 1.16
KON_CST 0.65 0.65 0.52 0.52 1.18 1.18 0.75 0.75
ARmax_CST 91.80 91.80 83.10 83.10 87.00 87.00 80.30 80.30
INT -51.40 -51.40 -48.30 -48.30 -48.60 -48.60 -65.80 -65.80
EC50_INT 0.06 0.06 0.06 0.06 0.07 0.07 0.39 0.39
gamma_INT 10.00 10.00 10.00 10.00 10.00 10.00 1.00 1.00
RES_ADD 1.74 1.74 1.01 1.01 1.44 1.44 2.25 2.25

For NARA864 the interaction row is INT_Emax / EC50_INT_Emax / gamma_INT_Emax; for the other three it is INT_EC50 / EC50_INT_EC50 / gamma_INT_EC50. The NARA864 growth rate, all interaction Hill coefficients, the NARA864 interaction EC50 and KOFF were not estimated (Table 2 “fixed”) and are wrapped in fixed().

Simulation helpers

grid_events <- function(panels, times) {
  panels$id <- seq_len(nrow(panels))
  ev <- merge(panels, data.frame(time = times))
  ev$CONC_CAZ_MGL <- ev$caz
  ev$CONC_CST_MGL <- ev$col
  # Observations on the bacterial ODE state; log_cfu is returned alongside.
  ev$evid <- 0
  ev$cmt <- "cfu"
  dplyr::arrange(ev, id, time)
}
# The models have no between-tube variability, and each `id` is a
# drug-concentration scenario, not a subject. rxode2 warns about a
# multi-subject solve without omega; only that message is muffled.
muffle_no_omega <- function(w) {
  if (grepl("multi-subject simulation without", conditionMessage(w), fixed = TRUE)) {
    invokeRestart("muffleWarning")
  }
}
solve_typical <- function(mod, panels, times = seq(0, 30, by = 0.25)) {
  ev <- grid_events(panels, times)
  withCallingHandlers(
    rxode2::rxSolve(rxode2::zeroRe(mod, "sigma"), ev, omega = NA,
                    keep = c("caz", "col"), returnType = "data.frame"),
    warning = muffle_no_omega
  )
}

Carrying capacity and inoculum

With no drug the model is a pure logistic growth curve: it starts at 10^INOC and settles at 10^Bmax. A mis-transcribed log10bmax or a missed 10^ back-transform shows up here immediately.

gc <- bind_rows(lapply(isolates, function(i) {
  solve_typical(mods[[i]], data.frame(caz = 0, col = 0), times = c(0, 30)) |>
    mutate(isolate = i)
}))
gc_chk <- gc |>
  select(isolate, time, log_cfu) |>
  pivot_wider(names_from = time, values_from = log_cfu, names_prefix = "t") |>
  mutate(
    inoc = vapply(isolate, function(i) mods[[i]]$theta[["log10inoc"]], numeric(1)),
    bmax = vapply(isolate, function(i) mods[[i]]$theta[["log10bmax"]], numeric(1))
  )
knitr::kable(gc_chk, digits = 3)
isolate t0 t30 inoc bmax
NARA1295 5.63 8.80 5.63 8.80
NARA1584 5.62 9.13 5.62 9.13
NARA1182 5.53 8.43 5.53 8.43
NARA864 5.68 9.01 5.68 9.01
stopifnot(
  all(abs(gc_chk$t0 - gc_chk$inoc) < 1e-6),
  all(abs(gc_chk$t30 - gc_chk$bmax) < 0.01)
)

Mass balance of the adaptation states

aron + aroff is a fraction of the population and must stay at 1 for each drug. The adaptation switch must also stay off for a drug that is absent.

ab <- solve_typical(mods$NARA1295, data.frame(caz = c(8, 0), col = c(0, 0.25)))
stopifnot(
  all(abs(ab$aron_caz + ab$aroff_caz - 1) < 1e-6),
  all(abs(ab$aron_col + ab$aroff_col - 1) < 1e-6),
  # colistin-free arm: no colistin adaptation; CZA-free arm: no CZA adaptation
  max(ab$aron_col[ab$col == 0]) < 1e-9,
  max(ab$aron_caz[ab$caz == 0]) < 1e-9
)
# kon_caz = 0.24 1/h, so 1 - exp(-0.24 * 30) of the bacteria are adapted by 30 h
ab |>
  filter(time == 30) |>
  select(caz, col, aron_caz, aron_col) |>
  mutate(aron_caz_closed_form = ifelse(caz > 0, 1 - exp(-0.24 * 30), 0)) |>
  knitr::kable(digits = 4)
caz col aron_caz aron_col aron_caz_closed_form
8 0.00 0.9993 0 0.9993
0 0.25 0.0000 1 0.0000

Replicate the visual predictive checks (Figure 3, Figures S1-S3)

The models have no random effect except the residual error, and that error is additive and symmetric on the log10 scale. The median of the paper’s VPC simulations is therefore the typical prediction itself, so the red median lines of the VPCs are a direct, quantitative target for zeroRe() simulations.

The median lines of Figs. S1-S3 (high-resolution images in the supplement) were digitised by the maintainers from the image pixels: the y axis was calibrated on the printed 0.0 and 10.0 log10 CFU/mL ticks of each panel row, the x axis on the 0-30 h tick marks of each panel column, and the median was read at the six sampling times. Fig. 3 (NARA1295) is published only as a low-resolution image in which the median line could not be traced reliably, so NARA1295 is shown without digitised points; its grid of concentrations is the same as that of NARA1584.

# Digitised VPC medians (log10 CFU/mL) at 2, 4, 8, 24 and 30 h; caz and col
# are the ceftazidime and colistin concentrations (mg/L) of each panel.
dig <- read.csv(text = "
isolate,caz,col,t2,t4,t8,t24,t30
NARA1584,0,0,7.11,8.36,9.01,9.01,9.08
NARA1584,2,0,6.9,8.22,9.04,9.04,9.01
NARA1584,4,0,6.6,7.79,8.99,9.01,9.01
NARA1584,8,0,5.87,6.55,8.36,8.92,9.01
NARA1584,16,0,4.58,4.37,5.0,8.64,8.73
NARA1584,32,0,3.62,2.44,2.11,4.08,4.93
NARA1584,64,0,3.03,1.78,1.13,1.5,1.9
NARA1584,0,0.03,6.83,8.09,9.05,9.08,9.08
NARA1584,2,0.03,6.69,7.78,9.01,9.08,9.08
NARA1584,4,0.03,6.36,7.45,8.96,9.03,9.01
NARA1584,8,0.03,5.67,6.26,7.87,8.96,9.01
NARA1584,16,0.03,4.35,4.04,4.54,8.49,8.58
NARA1584,32,0.03,3.4,2.2,1.91,3.14,3.83
NARA1584,0,0.06,6.48,7.6,9.0,9.14,9.14
NARA1584,2,0.06,6.29,7.26,8.9,9.0,9.0
NARA1584,4,0.06,5.88,6.6,8.55,9.0,9.0
NARA1584,8,0.06,4.95,5.12,6.21,8.88,8.93
NARA1584,16,0.06,3.79,2.98,2.79,5.45,6.5
NARA1584,32,0.06,3.02,1.55,0.98,0.79,1.0
NARA1584,0,0.125,5.91,6.67,8.63,9.01,9.01
NARA1584,2,0.125,5.53,6.1,7.97,8.94,8.94
NARA1584,4,0.125,4.92,5.04,6.22,8.82,8.87
NARA1584,8,0.125,3.88,3.12,3.14,6.08,7.16
NARA1584,16,0.125,2.96,1.58,0.5,0.43,0.5
NARA1584,32,0.125,2.7,0.78,NA,NA,NA
NARA1584,0,0.25,4.95,5.38,7.02,8.92,8.94
NARA1584,2,0.25,4.69,4.88,6.17,8.83,8.87
NARA1584,4,0.25,4.18,3.94,4.58,8.59,8.66
NARA1584,8,0.25,3.29,2.18,1.78,3.1,3.66
NARA1584,16,0.25,2.51,0.82,NA,NA,NA
NARA1584,32,0.25,2.3,0.28,NA,NA,NA
NARA1584,0,0.5,3.97,3.8,4.88,8.78,8.8
NARA1584,0,1,3.05,2.56,3.12,7.16,8.31
NARA1584,0,2,2.46,1.74,1.95,4.95,5.99
NARA1584,0,4,2.18,1.22,1.24,3.64,4.51
NARA1182,0,0,6.85,8.17,8.33,8.38,8.38
NARA1182,1,0,5.96,6.83,8.24,8.31,8.31
NARA1182,2,0,4.98,5.09,6.5,8.15,8.17
NARA1182,4,0,3.64,3.12,3.45,7.58,7.89
NARA1182,8,0,2.79,1.69,1.29,3.52,4.61
NARA1182,16,0,2.32,0.94,0.28,1.48,2.11
NARA1182,0,0.03,6.74,7.99,8.32,8.42,8.37
NARA1182,1,0.03,5.72,6.57,8.06,8.23,8.23
NARA1182,2,0.03,4.78,4.78,6.08,8.16,8.23
NARA1182,4,0.03,3.5,2.84,2.98,6.81,7.77
NARA1182,8,0.03,2.6,1.39,0.78,2.65,3.4
NARA1182,0,0.06,6.33,7.64,8.4,8.43,8.43
NARA1182,1,0.06,5.43,6.07,7.69,8.26,8.29
NARA1182,2,0.06,4.36,4.29,5.19,8.14,8.21
NARA1182,4,0.06,3.17,2.36,2.29,5.14,6.29
NARA1182,8,0.06,2.36,0.95,0.14,1.1,1.57
NARA1182,0,0.125,5.74,6.83,8.18,8.3,8.3
NARA1182,1,0.125,4.0,3.83,4.49,7.99,8.01
NARA1182,2,0.125,2.88,1.91,1.63,3.52,4.26
NARA1182,4,0.125,1.99,0.57,NA,NA,NA
NARA1182,8,0.125,1.58,0.21,NA,NA,NA
NARA1182,0,0.25,4.81,5.38,7.02,8.15,8.17
NARA1182,1,0.25,3.12,2.56,2.82,5.33,6.09
NARA1182,2,0.25,2.18,1.01,NA,NA,NA
NARA1182,4,0.25,1.41,0.28,NA,NA,NA
NARA1182,8,0.25,1.03,NA,NA,NA,NA
NARA1182,0,0.5,3.62,3.83,4.77,7.93,7.89
NARA1182,0,1,2.68,2.54,2.91,5.02,5.63
NARA1182,0,2,2.11,1.62,1.74,2.37,2.61
NARA1182,0,4,1.81,1.2,1.2,1.15,1.13
NARA864,0,0,7.22,8.47,8.92,8.97,8.97
NARA864,0.25,0,5.95,6.98,8.73,8.86,8.89
NARA864,0.5,0,4.6,4.87,6.35,8.65,8.65
NARA864,1,0,3.28,2.86,3.52,7.35,8.25
NARA864,2,0,2.67,1.93,2.17,4.47,5.36
NARA864,4,0,2.3,1.53,1.67,3.54,4.21
NARA864,8,0,2.49,1.46,1.51,3.2,3.81
NARA864,16,0,2.25,1.35,1.56,3.1,3.65
NARA864,0,0.03,6.72,8.04,8.81,8.94,8.97
NARA864,0.25,0.03,5.58,6.4,8.28,8.81,8.81
NARA864,0.5,0.03,4.29,4.55,5.77,8.62,8.65
NARA864,1,0.03,3.28,2.78,3.25,6.53,7.54
NARA864,2,0.03,2.67,1.83,2.01,3.76,4.48
NARA864,4,0.03,2.33,1.61,1.59,2.94,3.41
NARA864,0,0.06,6.32,7.35,8.78,8.89,8.89
NARA864,0.25,0.06,5.34,5.95,7.67,8.73,8.73
NARA864,0.5,0.06,4.15,4.18,5.24,8.54,8.57
NARA864,1,0.06,3.17,2.62,2.96,5.45,6.35
NARA864,2,0.06,2.67,1.72,1.72,2.96,3.37
NARA864,4,0.06,2.3,1.48,1.35,2.06,2.38
NARA864,0,0.125,5.71,6.38,8.23,8.81,8.81
NARA864,0.25,0.125,4.79,5.11,6.48,8.73,8.73
NARA864,0.5,0.125,3.76,3.65,4.31,7.88,8.41
NARA864,1,0.125,2.91,2.25,2.25,3.81,4.13
NARA864,2,0.125,2.65,1.61,1.35,1.59,1.67
NARA864,4,0.125,2.35,1.32,1.01,0.79,0.71
NARA864,0,0.25,4.77,5.04,6.4,8.72,8.72
NARA864,0.25,0.25,4.19,4.03,4.85,8.51,8.56
NARA864,0.5,0.25,3.28,2.91,3.12,5.15,5.88
NARA864,1,0.25,2.77,1.87,1.6,1.6,1.52
NARA864,2,0.25,2.48,1.28,0.69,NA,NA
NARA864,4,0.25,2.24,1.07,0.45,NA,NA
NARA864,0,0.5,3.95,3.84,4.69,8.21,8.48
NARA864,0.25,0.5,3.49,3.04,3.39,5.92,6.72
NARA864,0.5,0.5,2.83,2.19,2.08,2.56,2.8
NARA864,1,0.5,2.45,1.36,0.91,NA,NA
NARA864,2,0.5,2.4,0.96,0.32,NA,NA
NARA864,4,0.5,2.05,0.83,NA,NA,NA
NARA864,0,1,3.33,3.02,3.33,5.71,6.67
NARA864,0,2,2.99,2.45,2.6,4.14,4.69
NARA864,0,4,2.8,2.22,2.17,3.25,3.65
NARA864,0,8,2.64,2.08,2.08,2.93,3.12
")
dig_long <- dig |>
  pivot_longer(starts_with("t"), names_to = "time", values_to = "log_cfu_paper",
               names_prefix = "t", names_transform = as.numeric) |>
  filter(!is.na(log_cfu_paper))
nrow(dig_long)
#> [1] 499
grid_1584 <- dig |> filter(isolate == "NARA1584") |> select(caz, col)
panels <- list(
  NARA1295 = grid_1584,
  NARA1584 = grid_1584,
  NARA1182 = dig |> filter(isolate == "NARA1182") |> select(caz, col),
  NARA864  = dig |> filter(isolate == "NARA864")  |> select(caz, col)
)
typ <- bind_rows(lapply(isolates, function(i) {
  solve_typical(mods[[i]], panels[[i]]) |> mutate(isolate = i)
}))
plot_isolate <- function(i, fig) {
  d <- typ |> filter(isolate == i)
  p <- ggplot(d, aes(time, log_cfu)) +
    geom_hline(yintercept = log10(200), linetype = "dashed", colour = "grey50") +
    geom_line(colour = "firebrick") +
    facet_grid(col ~ caz, labeller = label_both) +
    coord_cartesian(ylim = c(0, 10.5)) +
    scale_x_continuous(breaks = c(0, 8, 24)) +
    labs(x = "Time (h)", y = "log10 CFU/mL",
         title = paste0(i, ": typical prediction (line)",
                        if (i != "NARA1295") " vs digitised VPC median (points)"),
         caption = paste0("Replicates ", fig, " of Aubry 2023.\nColumns: ceftazidime (mg/L, ",
                          "with 4 mg/L avibactam); rows: colistin (mg/L). Dashed: LLOQ."))
  if (i != "NARA1295") {
    p <- p + geom_point(data = dig_long |> filter(isolate == i),
                        aes(time, log_cfu_paper), size = 0.8)
  }
  p
}
plot_isolate("NARA1295", "Figure 3")

plot_isolate("NARA1584", "Figure S1")

plot_isolate("NARA1182", "Figure S2")

plot_isolate("NARA864", "Figure S3")

Quantitative agreement

Below the limit of quantification the paper’s plotted medians are clipped at the 0.0 axis line, so the comparison is restricted to points where both the digitised median and the model are above 2.5 log10 CFU/mL.

vpc_cmp <- typ |>
  select(isolate, caz, col, time, log_cfu) |>
  inner_join(dig_long, by = c("isolate", "caz", "col", "time")) |>
  filter(log_cfu_paper > 2.5, log_cfu > 2.5) |>
  mutate(diff = log_cfu - log_cfu_paper)
vpc_sum <- vpc_cmp |>
  group_by(isolate) |>
  summarise(
    n = n(),
    median_diff = median(diff),
    median_abs_diff = median(abs(diff)),
    q90_abs_diff = unname(quantile(abs(diff), 0.9)),
    max_abs_diff = max(abs(diff)),
    .groups = "drop"
  )
knitr::kable(vpc_sum, digits = 3)
isolate n median_diff median_abs_diff q90_abs_diff max_abs_diff
NARA1182 97 0.014 0.051 0.137 0.238
NARA1584 136 0.028 0.050 0.141 0.316
NARA864 140 0.068 0.080 0.203 0.421
# Digitising error is about 0.1-0.2 log10 (one image pixel is 0.14-0.16 log10
# CFU/mL). A mis-transcribed EC50, Emax or adaptation rate moves whole panels
# by one log10 or more.
stopifnot(
  all(abs(vpc_sum$median_diff) < 0.25),
  all(vpc_sum$q90_abs_diff < 0.5)
)

Residual error: variance, not standard deviation

Table 2 prints RES_ADD as an “additive residual error (log10CFU/mL)”. The control streams (Codes S1-S2) estimate it as $SIGMA and use W = SQRT(SIGMA(1)) as the standard deviation, which means NONMEM reports a variance. The VPC bands settle the reading. In the growth-control panel the band is the 10th-90th percentile of an additive normal error around a flat median, so its width is 2 * qnorm(0.9) * SD. Measured on the image pixels of the growth-control panels (the outer edge lines of the band, at the plateau): NARA1182 (Fig. S2) 3.14 log10 wide, NARA864 (Fig. S3) 3.81 log10 wide.

res_chk <- data.frame(
  isolate = c("NARA1182", "NARA864"),
  RES_ADD = c(1.44, 2.25),
  band_width_figure = c(3.14, 3.81)
) |>
  mutate(
    sd_from_band = band_width_figure / (2 * qnorm(0.9)),
    sd_if_variance = sqrt(RES_ADD),
    sd_if_sd = RES_ADD,
    addSd_in_file = vapply(isolate, function(i) mods[[i]]$theta[["addSd"]], numeric(1))
  )
knitr::kable(res_chk, digits = 3)
isolate RES_ADD band_width_figure sd_from_band sd_if_variance sd_if_sd addSd_in_file
NARA1182 1.44 3.14 1.225 1.2 1.44 1.2
NARA864 2.25 3.81 1.486 1.5 2.25 1.5
stopifnot(
  all(abs(res_chk$sd_from_band - res_chk$sd_if_variance) <
        abs(res_chk$sd_from_band - res_chk$sd_if_sd)),
  all(abs(res_chk$addSd_in_file - res_chk$sd_if_variance) < 1e-8)
)

The variance reading reproduces the band width of both isolates within the pixel resolution; the standard-deviation reading would make the NARA864 band 5.8 log10 wide, half as wide again as printed.

Results-text checks

The Results section reports that each drug alone at its MIC let the bacteria regrow, and that the combination at 2x the MIC of each drug prevented regrowth for every isolate. With the Table 1 MICs:

mic <- data.frame(
  isolate = isolates,
  mic_caz = c(8, 8, 4, 2),
  mic_cst = c(0.125, 0.125, 0.125, 0.25)
)
claims <- bind_rows(lapply(seq_len(nrow(mic)), function(k) {
  i <- mic$isolate[k]
  p <- data.frame(
    scenario = c("CZA 1x MIC alone", "CST 1x MIC alone", "CZA 2x + CST 2x MIC"),
    caz = c(1, 0, 2) * mic$mic_caz[k],
    col = c(0, 1, 2) * mic$mic_cst[k]
  )
  solve_typical(mods[[i]], p) |>
    left_join(p, by = c("caz", "col")) |>
    group_by(scenario) |>
    summarise(nadir = min(log_cfu), at_30h = log_cfu[time == 30], .groups = "drop") |>
    mutate(isolate = i, regrowth = at_30h - nadir)
}))
claims |>
  select(isolate, scenario, nadir, at_30h, regrowth) |>
  knitr::kable(digits = 2)
isolate scenario nadir at_30h regrowth
NARA1295 CST 1x MIC alone 5.45 8.73 3.28
NARA1295 CZA 1x MIC alone 5.63 8.69 3.06
NARA1295 CZA 2x + CST 2x MIC -1.54 -1.54 0.00
NARA1584 CST 1x MIC alone 5.57 9.03 3.45
NARA1584 CZA 1x MIC alone 5.62 9.00 3.38
NARA1584 CZA 2x + CST 2x MIC -2.54 -2.54 0.00
NARA1182 CST 1x MIC alone 5.33 8.31 2.98
NARA1182 CZA 1x MIC alone 3.06 8.02 4.96
NARA1182 CZA 2x + CST 2x MIC -6.00 -6.00 0.00
NARA864 CST 1x MIC alone 4.70 8.69 4.00
NARA864 CZA 1x MIC alone 1.95 5.51 3.56
NARA864 CZA 2x + CST 2x MIC -2.54 -2.54 0.00
alone <- claims |> filter(grepl("alone", scenario))
combo <- claims |> filter(grepl("\\+", scenario))
stopifnot(
  all(alone$regrowth > 2),
  all(combo$at_30h < log10(200))
)

Direction of the interaction (Figures 4 and 5)

The paper judged synergy and antagonism by comparing bootstrap 95% CIs of the observed-interaction model with those of a separately fitted Bliss expected-additivity model (Code S1), whose parameter estimates are not tabulated. That comparison cannot be reproduced here, but its direction can: setting int_col_caz = 0 switches the GPDI term off while keeping every other parameter. For NARA1295 colistin lowers the CZA EC50, so the combination kills more than additivity predicts – the paper finds significant synergy late in the experiment (Fig. 4); the case below uses both drugs at their MIC (CZA 8 mg/L, CST 0.125 mg/L), where the EC50 shift matters most, since well below the CZA EC50 there is little CZA kill to enhance. For NARA864 colistin lowers the CZA Emax, so the combination kills less (the paper finds antagonism at CZA 0.25-0.5 mg/L with CST 0.25-0.5 mg/L, Fig. 5).

dir_case <- function(i, caz, col) {
  p <- data.frame(caz = caz, col = col)
  obs <- solve_typical(mods[[i]], p, times = c(0, 8, 24, 30))
  add <- solve_typical(rxode2::ini(mods[[i]], int_col_caz = 0), p, times = c(0, 8, 24, 30))
  data.frame(isolate = i, caz = caz, col = col, time = obs$time,
             observed = obs$log_cfu, no_interaction = add$log_cfu)
}
dir_tab <- bind_rows(
  dir_case("NARA1295", 8, 0.125),
  dir_case("NARA864", 0.5, 0.5)
) |>
  mutate(difference = observed - no_interaction)
#> ℹ change initial estimate of `int_col_caz` to `0`
#> ℹ change initial estimate of `int_col_caz` to `0`
knitr::kable(dir_tab, digits = 2)
isolate caz col time observed no_interaction difference
NARA1295 8.0 0.12 0 5.63 5.63 0.00
NARA1295 8.0 0.12 8 2.88 6.10 -3.22
NARA1295 8.0 0.12 24 5.36 8.60 -3.24
NARA1295 8.0 0.12 30 6.47 8.60 -2.13
NARA864 0.5 0.50 0 5.68 5.68 0.00
NARA864 0.5 0.50 8 2.22 0.78 1.44
NARA864 0.5 0.50 24 2.90 -0.44 3.35
NARA864 0.5 0.50 30 3.17 -0.90 4.06
d30 <- dir_tab |> filter(time == 30)
stopifnot(
  d30$difference[d30$isolate == "NARA1295"] < -1,
  d30$difference[d30$isolate == "NARA864"] > 1
)

Clinical-profile simulations (Figure 6)

Figure 6 and Figs. S6-S7 drive the NARA1295 model with typical patient concentration-time profiles from two external popPK models (ceftazidime from Sy et al. 2019, colistin from Kristoffersson et al. 2020) that are not in this library, so they are not reproduced here. The model accepts time-varying CONC_CAZ_MGL / CONC_CST_MGL columns for that purpose: rxode2 interpolates covariates between records, and the adaptation switch follows the instantaneous concentration exactly as IF(A(1).GT.0) does in Code S2. Note that avibactam is assumed throughout to be present at a fully inhibiting, non-bactericidal concentration, which the paper flags as an assumption for these simulations (Discussion).

Assumptions and deviations

  • Residual error is a variance. Table 2 RES_ADD is the NONMEM $SIGMA variance (see the residual section); each file carries addSd = sqrt(RES_ADD). The paper fitted below-quantification counts with Beal’s M3 method (LLOQ 200 CFU/mL); the model files carry the plain additive error, and a user re-estimating with censored data should set CENS/LIMIT accordingly.
  • Linear CFU state. Eq 5 calls B a log10 count, but the control streams integrate DADT(3) = (KG*PLATEAU - E_COMB)*A(3) with A_0(3) = 10**INOC and PLATEAU = 1 - A(3)/10**BMAX, i.e. logistic growth of the linear CFU/mL count. The files follow the control streams.
  • Template control stream. Only the NARA1295 control streams are published. The paper states that the same structure was used for all four isolates with isolate-specific estimates, so the other three files reuse the Code S2 structure with their Table 2 column.
  • NARA864 Emax interaction. No control stream is printed for an interaction on Emax. The file multiplies the CZA Emax by the GPDI factor (1 + INT * C_CST / (0.39 + C_CST)) and, as in Code S2, keeps the Bliss normaliser at the larger of the two mono-drug Emax values (9.17 1/h). The alternative of normalising by the interaction-reduced Emax was also checked against the digitised Fig. S3 medians and fits marginally worse (90th percentile absolute difference 0.30 versus 0.29 log10).
  • Avibactam has no covariate. Every experiment used 4 mg/L, which the authors assume fully inhibits the beta-lactamases with no bactericidal effect of its own; CONC_CAZ_MGL is the ceftazidime concentration in the presence of that avibactam.
  • Interaction gating. Code S2 sets the interaction to zero unless both drugs are present. The files compute it whenever colistin is present; without ceftazidime the ceftazidime kill term is zero, so the two are identical.
  • Observation floor. log_cfu uses the 1e-6 CFU/mL floor of the control-stream error block; deep-kill predictions below about -6 log10 are reported as -6.
  • Adaptive-resistance state names. aron_col / aroff_col use the registered colistin suffix; aron_caz / aroff_caz are declared as paper-specific compartments because caz is not a registered suffix token.