Skip to contents
# Build the model once; the inline citation below reuses this object.
mod <- rxode2::rxode(readModelDb("Augustin_2023_warfarin_qsp"))
#> ℹ parameter labels from comments will be replaced by 'label()'
#> Warning: some etas defaulted to non-mu referenced, possible parsing error: etaiov_vk_input_1, etaiov_vk_input_2, etaiov_vk_input_3, etaiov_vk_input_4, etaiov_vk_input_5, etaiov_vk_input_6, etaiov_vk_input_7, etaiov_vk_input_8, etaiov_vk_input_9, etaiov_vk_input_10, etaiov_vk_input_11, etaiov_vk_input_12, etaiov_vk_input_13, etaiov_vk_input_14, etaiov_vk_input_15, etaiov_vk_input_16, etaiov_vk_input_17, etaiov_vk_input_18, etaiov_vk_input_19, etaiov_vk_input_20, etaiov_vk_input_21, etaiov_vk_input_22, etaiov_vk_input_23, etaiov_vk_input_24, etaiov_vk_input_25, etaiov_vk_input_26, etaiov_vk_input_27, etaiov_vk_input_28, etaiov_vk_input_29, etaiov_vk_input_30, etaiov_vk_input_31, etaiov_vk_input_32, etaiov_vk_input_33, etaiov_vk_input_34, etaiov_vk_input_35, etaiov_vk_input_36, etaiov_vk_input_37, etaiov_vk_input_38, etaiov_vk_input_39, etaiov_vk_input_40, etaiov_vk_input_41, etaiov_vk_input_42, etaiov_vk_input_43, etaiov_vk_input_44, etaiov_vk_input_45, etaiov_vk_input_46, etaiov_vk_input_47, etaiov_vk_input_48, etaiov_vk_input_49, etaiov_vk_input_50, etaiov_vk_input_51, etaiov_vk_input_52, etaiov_vk_input_53, etaiov_vk_input_54, etaiov_vk_input_55, etaiov_vk_input_56
#> as a work-around try putting the mu-referenced expression on a simple line
mod_typ <- rxode2::zeroRe(mod)
#> Warning: some etas defaulted to non-mu referenced, possible parsing error: etaiov_vk_input_1, etaiov_vk_input_2, etaiov_vk_input_3, etaiov_vk_input_4, etaiov_vk_input_5, etaiov_vk_input_6, etaiov_vk_input_7, etaiov_vk_input_8, etaiov_vk_input_9, etaiov_vk_input_10, etaiov_vk_input_11, etaiov_vk_input_12, etaiov_vk_input_13, etaiov_vk_input_14, etaiov_vk_input_15, etaiov_vk_input_16, etaiov_vk_input_17, etaiov_vk_input_18, etaiov_vk_input_19, etaiov_vk_input_20, etaiov_vk_input_21, etaiov_vk_input_22, etaiov_vk_input_23, etaiov_vk_input_24, etaiov_vk_input_25, etaiov_vk_input_26, etaiov_vk_input_27, etaiov_vk_input_28, etaiov_vk_input_29, etaiov_vk_input_30, etaiov_vk_input_31, etaiov_vk_input_32, etaiov_vk_input_33, etaiov_vk_input_34, etaiov_vk_input_35, etaiov_vk_input_36, etaiov_vk_input_37, etaiov_vk_input_38, etaiov_vk_input_39, etaiov_vk_input_40, etaiov_vk_input_41, etaiov_vk_input_42, etaiov_vk_input_43, etaiov_vk_input_44, etaiov_vk_input_45, etaiov_vk_input_46, etaiov_vk_input_47, etaiov_vk_input_48, etaiov_vk_input_49, etaiov_vk_input_50, etaiov_vk_input_51, etaiov_vk_input_52, etaiov_vk_input_53, etaiov_vk_input_54, etaiov_vk_input_55, etaiov_vk_input_56
#> as a work-around try putting the mu-referenced expression on a simple line
net_states <- setdiff(mod$state, c("depot", "central", "fibrin_auc"))

Model and source

  • Citation: Augustin D, Lambert B, Robinson M, Wang K, Gavaghan D. Simulating clinical trials for model-informed precision dosing: using warfarin treatment as a use case. Front Pharmacol. 2023;14:1270443. doi:10.3389/fphar.2023.1270443. Model structure: Section 2.2 Eq 6-9 and Supplementary Appendix S1 (Tables S1.1-S1.3, Eq S1.1-S1.9). Parameter values: the authors’ code and data deposit named in Section 2.2, https://github.com/DavAug/mipd-warfarin (commit 4710c062), models/wajima_coagulation_model.xml (SBML) and models/hartmann_coagulation_model_parameters.csv. Underlying models: Wajima T, Isbister GK, Duffull SB. Clin Pharmacol Ther. 2009;86(3):290-298 (doi:10.1038/clpt.2009.87); Hartmann S et al. CPT Pharmacometrics Syst Pharmacol. 2016;5(10):554-564 (doi:10.1002/psp4.12111) and Front Pharmacol. 2020;11:1041 (doi:10.3389/fphar.2020.01041); Hamberg AK et al. Clin Pharmacol Ther. 2010;87(6):727-734 (doi:10.1038/clpt.2010.37).
  • Description: QSP. Warfarin clinical-trial (CT) model of Augustin 2023: the Wajima 2009 humoral coagulation-network model (51 coagulation / vitamin K states) driven by a one-compartment oral warfarin PK model, with the population model the authors assembled from Hartmann 2016 / 2020 and Hamberg 2010. Warfarin inhibits the two VKORC-catalysed vitamin K conversions (Hill Imax), which lowers vitamin K hydroquinone and with it the synthesis of factors II, VII, IX, X and proteins C and S. Covariates: CYP2C9 2 / 3 allele counts and age on the warfarin elimination rate, VKORC1 -1639A allele count on the warfarin EC50 and on both VKORC-catalysed vitamin K conversion rates. IIV: lognormal on warfarin volume, elimination rate, EC50 and the two conversion rates; normal (20% CV) on ten coagulation- factor production rates. Inter-occasion variability: the vitamin K input rate is redrawn every treatment day (OCC = treatment day 1-56, SD 10%). Simulations start from the Wajima initial concentrations and need a 100-day (2400 h) drug-free run-in before the first dose so each individual reaches its own steady state. The INR is not an algebraic output: it is read from a simulated prothrombin-time test on a blood sample (sample diluted 1:2, 300 nM tissue factor added, PT = time for the fibrin AUC state fibrin_auc to reach 1500 nM*s, INR = PT / 11.8 s), which the validation vignette implements as a second solve of this same model. Solve with method = “lsoda”: under the default liblsoda, round-off near zero can set off a spurious coagulation burst. No parameter was estimated by Augustin 2023; values were adopted from the cited models or hand-calibrated.
  • Article: https://doi.org/10.3389/fphar.2023.1270443
  • Authors’ code and data (SBML model files, parameter tables, simulated trial data): https://github.com/DavAug/mipd-warfarin

Augustin 2023 is a methods paper: it builds a clinical-trial (CT) model that emulates warfarin treatment, uses it to simulate pre-MIPD phase I-III trials, fits three model-informed precision dosing (MIPD) approaches to the simulated data, and then compares those approaches in a simulated MIPD trial. The CT model is the reusable pharmacometric artefact and is what this library entry holds. It has five components (Section 2.1.2, Figure 1):

  1. Mechanistic model – the Wajima 2009 humoral coagulation network (51 coagulation and vitamin K states) driven by one-compartment oral warfarin PK (Eq 6, Supplementary Tables S1.1-S1.3).
  2. Population model – Hartmann 2016 / 2020 normal IIV on ten coagulation-factor production rates, and Hamberg 2010 style lognormal IIV with CYP2C9, VKORC1 and age effects on the warfarin elimination rate, EC50 and volume, extended to the two VKORC-catalysed vitamin K conversion rates (Eq S1.2-S1.9).
  3. Inter-occasion model – the vitamin K input rate is multiplied by a factor drawn from N(1, 0.1^2) each treatment day (Eq 7).
  4. Execution model – dose and sampling times are delayed by an exponentially distributed time with mean 30 min (Eq 8). This is a property of the event table, not of the model, and is implemented below.
  5. Measurement model – lognormal measurement error with scale 0.1 (Eq 9), carried by the model’s expSd residual.

The three MIPD models of the paper (a neural-network regression, a deep reinforcement-learning policy and a Hamberg 2010 PKPD model re-fitted by hierarchical Bayesian inference) are not extracted: all three were trained on data simulated from this CT model rather than on patient data, so they carry no information about warfarin beyond what the CT model already holds.

Population

The CT model is a simulator, not a fit, so its “population” is a virtual one. Supplementary Table S1 lists the simulated cohorts: a phase I trial (N = 60, single 10 mg dose, warfarin concentrations at 10, 35 and 60 h), a phase II trial (N = 100, three weeks of INR-guided daily dosing) and a phase III trial (N = 1000, eight weeks). CYP2C9 and VKORC1 genotype frequencies and ages were modelled after Hamberg 2010 (phase III: CYP2C9 *1*1 66.1%, *1*2 18.4%, *1*3 12.3%, *2*2 1.4%, *2*3 1.2%, *3*3 0.6%; VKORC1 GG 36.5%, GA 48.5%, AA 15%; median age 50 years). The MIPD trial cohort (N = 1000, Figure 6) was built the same way. The values the model uses were adopted from Wajima 2009, Hartmann 2016 / 2020 and Hamberg 2010, and three of them were calibrated by hand: the VKORC1 effect on the vitamin K conversion rates (against a plausible baseline-INR distribution, Figure S1.1) and the VKORC1 GG EC50 and its IIV (against 2543 IWPC maintenance doses, Figure S1.2).

The same information is available programmatically via mod$population.

Source trace

Every ini() value is annotated in inst/modeldb/specificDrugs/Augustin_2023_warfarin_qsp.R with the line of the authors’ parameter file it came from. The paper itself prints the model structure (Tables S1.1-S1.3, Eq S1.1-S1.9) but defers the numeric values to the deposited file models/hartmann_coagulation_model_parameters.csv; the ODE right-hand sides were transcribed reaction by reaction from the deposited SBML file models/wajima_coagulation_model.xml and agree with Table S1.1.

Component Model element Value(s) Source
PK structure depot -> central, first-order ka, kel – Eq 6; Section S1.1
Absorption lka log(2) 1/h csv Pooled dose.absorption_rate; Section S4.2 (Hamberg 2010 fixed ka = 2 1/h)
Volume lvc, etalvc 3.353 (28.6 L), log SD 0.054 csv central_warfarin.size rows; Eq S1.7
Elimination lkel, etalkel -3.716 (0.0243 1/h), log SD 0.071 csv elimination_rate_warfarin rows; Eq S1.5-S1.6
CYP2C9 on kel e_cyp2c9s2_kel, e_cyp2c9s3_kel 0.495, 0.757 (*2*2, *3*3; half per allele) csv Rel. elimination rate shift; Eq S1.5
Age on kel e_age_kel 0.00571 1/year, 1 - tanh(r (AGE - 71)) csv; Eq S1.5
Warfarin EC50 lec50, etalec50 -2.0 (0.135 mg/L), log SD 0.1 csv; Eq S1.3-S1.4; calibrated, Section S1.2
VKORC1 on EC50 e_vkorc1aa_ec50 0.532 (AA vs GG; half per A allele) csv; Eq S1.3
Warfarin effect maximal_inhibitory_effect_warfarin, hill_warfarin 1, 1.15 csv; Table S1.1 vitamin K equations; Section S4.4.2
VK conversions lconversion_rate_vk_vkh2, lconversion_rate_vko_vk -3.781, -1.478; log SD 0.24 each csv; Eq S1.8-S1.9
VKORC1 on VK conversions e_vkorc1aa_vk_vkh2, e_vkorc1aa_vko_vk 0.2 each csv; Section S1.2 (20% total reduction)
Production-rate IIV tvproduction_rate_*, etatvproduction_rate_* ten factors, SD = 20% of mean (V: 19.4%) csv Mean / Std. rows; Eq S1.2
Network rate constants 136 parameters named as in Table S1.3 see file csv Pooled global.* rows
Initial concentrations bl_* csv amounts / 24 L csv Pooled central.*_amount rows
Network ODEs 51 states, 113 reactions – SBML wajima_coagulation_model.xml; Table S1.1
Fibrin integral d/dt(fibrin_auc) <- fibrin – Section S1.1.2
Vitamin K IOV etaiov_vk_input_1 … _56 variance 0.01 Eq 7
Measurement error expSd 0.1 Eq 9; csv Pooled Sigma log
PT test dilution 1:2, 300 nM TF, threshold 1500 nM*s, reference PT 11.8 s – Section S1.1.2; deposited model.py

The prothrombin-time test

The INR is not an algebraic function of the model states. Following Section S1.1.2, a blood sample (the state vector at the sampling time) is diluted 1:2 (all states divided by 3), thrombomodulin is removed, 300 nM tissue factor is added, and the network is integrated in vitro with all external inputs and non-vitamin-K-dependent production switched off. The prothrombin time (PT) is the time at which the fibrin integral reaches 1500 nM*s; INR = PT / 11.8 s.

The authors’ code performs the test with a reduced copy of the network (models/wajima_inr_test_model.xml) that omits the reactions which cannot proceed in the test tube. The helper below reproduces that exactly with a second solve of the same model: the omitted reactions are switched off by setting their rate constants to zero, and the sample is loaded with replacement (evid = 5) events at time 0. The test integrates over 120 s on a 100-point grid and interpolates linearly, as the authors’ code does.

Solver. All solves use method = "lsoda". With rxode2’s default liblsoda, round-off of order 1e-20 nM in the activated factors is amplified by the network’s thrombin feedback loops (several half-maximal constants are 1e-4 nM) into a spurious coagulation burst under warfarin; lsoda does not do this, and it is the configuration under which the gates below reproduce the authors’ own simulations.

solve_method <- "lsoda"

# Rate constants of the reactions absent from the authors' in-vitro test model.
pt_zero_rates <- c(
  "complex_formation_rate_aii_tmod", "complex_formation_rate_apc_ps",
  "degradation_rate_aii_tmod_complex", "degradation_rate_apc_ps_complex",
  "degradation_rate_ca", "degradation_rate_d_dimer", "degradation_rate_fdp",
  "degradation_rate_kallikrein", "degradation_rate_pc", "degradation_rate_pc_activated",
  "degradation_rate_prekallikrein", "degradation_rate_ps", "degradation_rate_tmod",
  "degradation_rate_xf", "degradation_rate_xii", "degradation_rate_xii_activated",
  "input_rate_vk", "maximal_activation_rate_aii_tmod_complex_pc",
  "maximal_activation_rate_axii_xi", "maximal_activation_rate_ca_xii",
  "maximal_activation_rate_kallikrein_xii",
  "maximal_conversion_rate_apc_ps_complex_plasminogen",
  "maximal_conversion_rate_axii_prekallikrein",
  "maximal_degradation_rate_apc_ps_complex_av",
  "maximal_degradation_rate_apc_ps_complex_av_ax_complex",
  "maximal_degradation_rate_apc_ps_complex_aviii",
  "maximal_degradation_rate_apc_ps_complex_xf", "maximal_degradation_rate_plasmin_xf",
  "production_rate_fibrinogen", "tvproduction_rate_pc", "production_rate_plasminogen",
  "production_rate_prekallikrein", "tvproduction_rate_ps", "production_rate_tfpi",
  "production_rate_tmod", "tvproduction_rate_v", "production_rate_viii",
  "tvproduction_rate_xi", "tvproduction_rate_xii", "tvproduction_rate_xiii",
  "transition_rate_vk_central_to_peripheral", "transition_rate_vk_peripheral_to_central"
)

# Reference covariates for solves whose covariate-dependent parameters are
# irrelevant (the in-vitro test) or supplied directly through `params`.
add_reference_covariates <- function(ev) {
  ev$CYP2C9_S2_COUNT <- 0
  ev$CYP2C9_S3_COUNT <- 0
  ev$VKORC1_1639G_COUNT <- 2
  ev$AGE <- 71
  ev$OCC <- 0
  ev
}

# INR of each row of `samples` (one blood sample per row, network states as columns).
inr_test <- function(samples, standard_pt = 11.8, duration_s = 120) {
  n <- nrow(samples)
  grid_h <- seq(0, duration_s, length.out = 100) / 3600
  vals <- as.matrix(samples[, net_states]) / 3
  vals[, "tissue_factor"] <- 300
  vals[, "thrombomodulin"] <- 0
  vals <- cbind(vals, fibrin_auc = 0)
  load <- data.frame(
    id = rep(seq_len(n), each = ncol(vals)), time = 0, evid = 5L,
    cmt = rep(colnames(vals), times = n), amt = as.vector(t(vals))
  )
  obs <- data.frame(
    id = rep(seq_len(n), each = length(grid_h)), time = rep(grid_h, n),
    evid = 0L, cmt = "fibrin_auc", amt = 0
  )
  ev <- rbind(load, obs)
  ev <- add_reference_covariates(ev[order(ev$id, ev$time, -ev$evid), ])
  th <- setNames(rep(0, length(pt_zero_rates)), pt_zero_rates)
  s <- rxode2::rxSolve(mod_typ, ev, params = th, method = solve_method,
                       returnType = "data.frame")
  if (is.null(s$id)) s$id <- 1L
  threshold <- 1500 / 3600 # 1500 nM*s expressed in nM*h
  pt <- vapply(split(s, s$id), pt_from_curve, numeric(1), threshold = threshold)
  unname(pt) / standard_pt
}

# Time (s) at which the fibrin integral crosses the threshold; linear
# extrapolation from the last two grid points when the test is too short.
pt_from_curve <- function(d, threshold) {
  t_s <- d$time * 3600
  a <- d$fibrin_auc
  k <- length(a)
  if (a[k] < threshold) {
    return(t_s[k] + (threshold - a[k]) / ((a[k] - a[k - 1]) / (t_s[k] - t_s[k - 1])))
  }
  stats::approx(a, t_s, xout = threshold, ties = "ordered")$y
}

Drug-free run-in and baseline INR

Simulations start from the Wajima initial concentrations. The authors run each virtual patient drug-free for 100 days before the first dose so that the network settles at the individual’s own steady state (warmup = 100 days in the deposited trial scripts). For the typical patient the initial state is already the steady state, and the typical baseline INR is 1.

ev_runin <- add_reference_covariates(
  data.frame(id = 1L, time = c(0, 600, 1200, 2400), evid = 0L, cmt = "central", amt = 0)
)
runin <- rxode2::rxSolve(mod_typ, ev_runin, method = solve_method, returnType = "data.frame")
#> ℹ omega/sigma items treated as zero: 'etalvc', 'etalkel', 'etalec50', 'etalconversion_rate_vk_vkh2', 'etalconversion_rate_vko_vk', 'etatvproduction_rate_ii', 'etatvproduction_rate_v', 'etatvproduction_rate_vii', 'etatvproduction_rate_ix', 'etatvproduction_rate_x', 'etatvproduction_rate_xi', 'etatvproduction_rate_xii', 'etatvproduction_rate_xiii', 'etatvproduction_rate_pc', 'etatvproduction_rate_ps', 'etaiov_vk_input_1', 'etaiov_vk_input_2', 'etaiov_vk_input_3', 'etaiov_vk_input_4', 'etaiov_vk_input_5', 'etaiov_vk_input_6', 'etaiov_vk_input_7', 'etaiov_vk_input_8', 'etaiov_vk_input_9', 'etaiov_vk_input_10', 'etaiov_vk_input_11', 'etaiov_vk_input_12', 'etaiov_vk_input_13', 'etaiov_vk_input_14', 'etaiov_vk_input_15', 'etaiov_vk_input_16', 'etaiov_vk_input_17', 'etaiov_vk_input_18', 'etaiov_vk_input_19', 'etaiov_vk_input_20', 'etaiov_vk_input_21', 'etaiov_vk_input_22', 'etaiov_vk_input_23', 'etaiov_vk_input_24', 'etaiov_vk_input_25', 'etaiov_vk_input_26', 'etaiov_vk_input_27', 'etaiov_vk_input_28', 'etaiov_vk_input_29', 'etaiov_vk_input_30', 'etaiov_vk_input_31', 'etaiov_vk_input_32', 'etaiov_vk_input_33', 'etaiov_vk_input_34', 'etaiov_vk_input_35', 'etaiov_vk_input_36', 'etaiov_vk_input_37', 'etaiov_vk_input_38', 'etaiov_vk_input_39', 'etaiov_vk_input_40', 'etaiov_vk_input_41', 'etaiov_vk_input_42', 'etaiov_vk_input_43', 'etaiov_vk_input_44', 'etaiov_vk_input_45', 'etaiov_vk_input_46', 'etaiov_vk_input_47', 'etaiov_vk_input_48', 'etaiov_vk_input_49', 'etaiov_vk_input_50', 'etaiov_vk_input_51', 'etaiov_vk_input_52', 'etaiov_vk_input_53', 'etaiov_vk_input_54', 'etaiov_vk_input_55', 'etaiov_vk_input_56'
runin |>
  select(time, coagulation_factor_ii, coagulation_factor_vii, coagulation_factor_x,
         fibrinogen, vitamin_k_hydroquinone) |>
  knitr::kable(digits = 4, caption = "Typical patient, drug-free: network states (nM).")
Typical patient, drug-free: network states (nM).
time coagulation_factor_ii coagulation_factor_vii coagulation_factor_x fibrinogen vitamin_k_hydroquinone
0 1394.400 10.0000 174.3000 8945.5 0.1
600 1394.393 9.9999 174.2991 8945.5 0.1
1200 1394.393 9.9999 174.2991 8945.5 0.1
2400 1394.393 9.9999 174.2991 8945.5 0.1

inr_baseline_typ <- inr_test(runin[runin$time == 2400, ])
#> ℹ omega/sigma items treated as zero: 'etalvc', 'etalkel', 'etalec50', 'etalconversion_rate_vk_vkh2', 'etalconversion_rate_vko_vk', 'etatvproduction_rate_ii', 'etatvproduction_rate_v', 'etatvproduction_rate_vii', 'etatvproduction_rate_ix', 'etatvproduction_rate_x', 'etatvproduction_rate_xi', 'etatvproduction_rate_xii', 'etatvproduction_rate_xiii', 'etatvproduction_rate_pc', 'etatvproduction_rate_ps', 'etaiov_vk_input_1', 'etaiov_vk_input_2', 'etaiov_vk_input_3', 'etaiov_vk_input_4', 'etaiov_vk_input_5', 'etaiov_vk_input_6', 'etaiov_vk_input_7', 'etaiov_vk_input_8', 'etaiov_vk_input_9', 'etaiov_vk_input_10', 'etaiov_vk_input_11', 'etaiov_vk_input_12', 'etaiov_vk_input_13', 'etaiov_vk_input_14', 'etaiov_vk_input_15', 'etaiov_vk_input_16', 'etaiov_vk_input_17', 'etaiov_vk_input_18', 'etaiov_vk_input_19', 'etaiov_vk_input_20', 'etaiov_vk_input_21', 'etaiov_vk_input_22', 'etaiov_vk_input_23', 'etaiov_vk_input_24', 'etaiov_vk_input_25', 'etaiov_vk_input_26', 'etaiov_vk_input_27', 'etaiov_vk_input_28', 'etaiov_vk_input_29', 'etaiov_vk_input_30', 'etaiov_vk_input_31', 'etaiov_vk_input_32', 'etaiov_vk_input_33', 'etaiov_vk_input_34', 'etaiov_vk_input_35', 'etaiov_vk_input_36', 'etaiov_vk_input_37', 'etaiov_vk_input_38', 'etaiov_vk_input_39', 'etaiov_vk_input_40', 'etaiov_vk_input_41', 'etaiov_vk_input_42', 'etaiov_vk_input_43', 'etaiov_vk_input_44', 'etaiov_vk_input_45', 'etaiov_vk_input_46', 'etaiov_vk_input_47', 'etaiov_vk_input_48', 'etaiov_vk_input_49', 'etaiov_vk_input_50', 'etaiov_vk_input_51', 'etaiov_vk_input_52', 'etaiov_vk_input_53', 'etaiov_vk_input_54', 'etaiov_vk_input_55', 'etaiov_vk_input_56'
inr_baseline_typ
#> [1] 0.9984911
stopifnot(
  # Deterministic: same parameters, same solver; measured drift < 1e-4 nM.
  max(abs(runin$coagulation_factor_ii / runin$coagulation_factor_ii[1] - 1)) < 1e-4,
  # The reference PT of 11.8 s is the typical network's own clotting time.
  abs(inr_baseline_typ - 1) < 0.01
)

Gate 1: individual maintenance doses from the authors’ deposit

The deposited file maintenance_dose_distribution.csv lists, for 1000 virtual patients sampled from the population model, every individual parameter value and the daily dose that the authors’ optimiser found to give an INR of exactly 2.5 on day 20 (after the 100-day run-in, no inter-occasion variability, no noise). This is the population that produced Figure S1.2. Twenty of those patients (every 50th row, spanning all three VKORC1 genotypes) are embedded below. Their individual parameters are passed to the model directly, so the check exercises the network, the PK, the warfarin effect and the PT test together: each patient’s day-20 INR must come back as 2.5.

maint <- read.csv(text = "
id,vkorc1,cyp2c9,age,vc,kel,ec50,cvk,cvko,p_ii,p_v,p_vii,p_ix,p_x,p_xi,p_xii,p_xiii,p_pc,p_ps,dose
1,2,0,71.3,27.10146,0.02465334,0.06423569,0.0140048,0.0973433,180.0782,1.249434,11.68906,26.64409,29.33518,2.518083,5.333917,0.1790214,29.6556,29.27996,2.059986
51,2,0,69.6,28.63875,0.02359941,0.06272538,0.01710786,0.077244,108.1862,1.11528,8.488099,22.7082,27.2716,4.204937,3.54223,0.1964962,20.34279,45.74483,1.77651
101,2,1,59.3,27.80313,0.02026781,0.06998511,0.0126112,0.125099,102.248,0.967735,15.27425,25.86967,35.41179,2.365893,4.132094,0.3054835,43.73491,56.60064,1.548424
151,1,1,56.6,28.7435,0.01810668,0.0882218,0.01405224,0.2059379,127.8248,1.168323,12.66359,27.87918,28.89284,2.670567,5.172594,0.2597838,34.86933,58.08755,1.936112
201,1,0,71.5,27.17774,0.02336589,0.1047107,0.01848383,0.1629457,140.7487,1.106938,10.51348,29.12108,24.9715,3.040228,4.190928,0.2236003,21.96359,43.67537,3.646134
251,1,0,76.3,27.56104,0.02694317,0.09815641,0.01528356,0.1736664,141.4789,1.477626,16.25262,40.96252,40.96902,4.050726,4.69445,0.2242317,43.77653,59.18201,4.673672
301,1,0,72.9,26.67942,0.02563038,0.1054005,0.0149227,0.1308483,160.532,0.9420056,13.51426,30.76161,25.24216,2.332201,3.336272,0.2661456,22.92508,47.61534,3.569888
351,1,0,63.2,29.06699,0.0253613,0.1048292,0.01682736,0.122562,129.9306,1.15544,10.27449,18.13984,37.57169,3.32446,5.167022,0.1961203,23.96355,46.85413,4.286764
401,1,0,70.1,30.07927,0.02419014,0.1059045,0.01272577,0.1647716,161.8929,1.145305,13.45872,21.70407,36.03417,3.209166,3.881985,0.2192279,34.19311,42.26476,3.612677
451,1,0,70.3,30.76272,0.02380424,0.09431959,0.01770386,0.1873998,139.2181,1.494419,16.09343,31.0555,27.14099,2.943045,4.737454,0.3276761,25.99615,52.00087,4.397744
501,1,1,63.3,25.16066,0.02033857,0.09775739,0.01665978,0.1340633,129.1696,1.044821,8.005874,29.14384,32.26952,1.632883,5.886331,0.3026575,26.94611,39.69898,2.29052
551,1,2,62.2,27.11011,0.01441869,0.09679855,0.01534478,0.1863798,108.973,1.243703,11.79747,28.12579,32.36393,2.574288,3.344743,0.3027036,29.013,69.62121,1.70041
601,1,0,66.7,25.19835,0.02264737,0.1080836,0.02408123,0.1256708,109.6308,1.693995,15.17402,24.78687,30.23712,2.634519,4.452494,0.1652782,18.60999,49.72054,5.179818
651,0,1,69.7,28.10918,0.01785939,0.1316555,0.02972581,0.2051469,156.0732,1.014278,14.49083,21.41504,30.00312,2.079385,5.050091,0.3197942,36.47128,43.47353,7.649681
701,0,0,55.5,30.5574,0.02989697,0.1418764,0.01613828,0.1736343,103.5597,0.7217821,11.50155,25.39551,32.35633,2.341271,4.20009,0.2502289,25.96333,56.38791,6.156721
751,0,0,60.3,26.84254,0.02955199,0.1502445,0.0285109,0.2028524,110.3251,1.546279,15.47009,25.40005,34.41103,2.62348,4.77843,0.2176517,30.7905,51.06354,12.91268
801,0,2,57.0,28.25769,0.01503652,0.1148707,0.02295149,0.1530914,147.9472,1.678125,11.67181,26.38705,29.89163,2.609481,5.537851,0.1681012,38.05513,38.23518,3.884841
851,0,0,56.9,28.96514,0.03145887,0.1303611,0.01793737,0.2248592,124.0471,1.10964,11.52773,15.94942,22.50598,2.814058,5.198036,0.3094551,29.23469,40.31498,6.006057
901,0,0,70.0,28.29344,0.02276665,0.1279639,0.01513446,0.2149372,157.5709,1.1207,13.90472,27.17955,33.13748,2.551938,4.150063,0.1280127,34.97162,55.15731,4.648408
951,0,0,68.1,31.79349,0.02429616,0.1235917,0.02450908,0.2786605,179.6513,0.9557907,12.84114,35.21889,25.61814,2.651545,5.205054,0.3240496,22.1255,40.14871,8.700386
")
prod_names <- c("ii", "v", "vii", "ix", "x", "xi", "xii", "xiii", "pc", "ps")
pp <- data.frame(
  id = seq_len(nrow(maint)),
  lvc = log(maint$vc), lkel = log(maint$kel), lec50 = log(maint$ec50),
  lconversion_rate_vk_vkh2 = log(maint$cvk), lconversion_rate_vko_vk = log(maint$cvko)
)
for (p in prod_names) pp[[paste0("tvproduction_rate_", p)]] <- maint[[paste0("p_", p)]]

ev_maint <- bind_rows(
  data.frame(id = rep(seq_len(nrow(maint)), each = 21),
             time = 2400 + 24 * rep(0:20, nrow(maint)), evid = 1L, cmt = "depot",
             amt = rep(maint$dose, each = 21)),
  data.frame(id = seq_len(nrow(maint)), time = 2400 + 20 * 24, evid = 0L,
             cmt = "central", amt = 0)
) |>
  arrange(id, time, desc(evid)) |>
  add_reference_covariates()

sim_maint <- rxode2::rxSolve(mod_typ, ev_maint, params = pp, method = solve_method,
                             returnType = "data.frame")
#> ℹ omega/sigma items treated as zero: 'etalvc', 'etalkel', 'etalec50', 'etalconversion_rate_vk_vkh2', 'etalconversion_rate_vko_vk', 'etatvproduction_rate_ii', 'etatvproduction_rate_v', 'etatvproduction_rate_vii', 'etatvproduction_rate_ix', 'etatvproduction_rate_x', 'etatvproduction_rate_xi', 'etatvproduction_rate_xii', 'etatvproduction_rate_xiii', 'etatvproduction_rate_pc', 'etatvproduction_rate_ps', 'etaiov_vk_input_1', 'etaiov_vk_input_2', 'etaiov_vk_input_3', 'etaiov_vk_input_4', 'etaiov_vk_input_5', 'etaiov_vk_input_6', 'etaiov_vk_input_7', 'etaiov_vk_input_8', 'etaiov_vk_input_9', 'etaiov_vk_input_10', 'etaiov_vk_input_11', 'etaiov_vk_input_12', 'etaiov_vk_input_13', 'etaiov_vk_input_14', 'etaiov_vk_input_15', 'etaiov_vk_input_16', 'etaiov_vk_input_17', 'etaiov_vk_input_18', 'etaiov_vk_input_19', 'etaiov_vk_input_20', 'etaiov_vk_input_21', 'etaiov_vk_input_22', 'etaiov_vk_input_23', 'etaiov_vk_input_24', 'etaiov_vk_input_25', 'etaiov_vk_input_26', 'etaiov_vk_input_27', 'etaiov_vk_input_28', 'etaiov_vk_input_29', 'etaiov_vk_input_30', 'etaiov_vk_input_31', 'etaiov_vk_input_32', 'etaiov_vk_input_33', 'etaiov_vk_input_34', 'etaiov_vk_input_35', 'etaiov_vk_input_36', 'etaiov_vk_input_37', 'etaiov_vk_input_38', 'etaiov_vk_input_39', 'etaiov_vk_input_40', 'etaiov_vk_input_41', 'etaiov_vk_input_42', 'etaiov_vk_input_43', 'etaiov_vk_input_44', 'etaiov_vk_input_45', 'etaiov_vk_input_46', 'etaiov_vk_input_47', 'etaiov_vk_input_48', 'etaiov_vk_input_49', 'etaiov_vk_input_50', 'etaiov_vk_input_51', 'etaiov_vk_input_52', 'etaiov_vk_input_53', 'etaiov_vk_input_54', 'etaiov_vk_input_55', 'etaiov_vk_input_56'
#> Warning: multi-subject simulation without without 'omega'
day20 <- sim_maint[sim_maint$time == 2880, ]
maint_check <- maint |>
  transmute(
    id_deposit = id,
    VKORC1 = c("GG", "GA", "AA")[vkorc1 + 1],
    CYP2C9 = c("*1/*1", "*1/*2", "*1/*3", "*2/*2", "*2/*3", "*3/*3")[cyp2c9 + 1],
    dose_mg = dose,
    inr_day20 = inr_test(day20)
  )
#> ℹ omega/sigma items treated as zero: 'etalvc', 'etalkel', 'etalec50', 'etalconversion_rate_vk_vkh2', 'etalconversion_rate_vko_vk', 'etatvproduction_rate_ii', 'etatvproduction_rate_v', 'etatvproduction_rate_vii', 'etatvproduction_rate_ix', 'etatvproduction_rate_x', 'etatvproduction_rate_xi', 'etatvproduction_rate_xii', 'etatvproduction_rate_xiii', 'etatvproduction_rate_pc', 'etatvproduction_rate_ps', 'etaiov_vk_input_1', 'etaiov_vk_input_2', 'etaiov_vk_input_3', 'etaiov_vk_input_4', 'etaiov_vk_input_5', 'etaiov_vk_input_6', 'etaiov_vk_input_7', 'etaiov_vk_input_8', 'etaiov_vk_input_9', 'etaiov_vk_input_10', 'etaiov_vk_input_11', 'etaiov_vk_input_12', 'etaiov_vk_input_13', 'etaiov_vk_input_14', 'etaiov_vk_input_15', 'etaiov_vk_input_16', 'etaiov_vk_input_17', 'etaiov_vk_input_18', 'etaiov_vk_input_19', 'etaiov_vk_input_20', 'etaiov_vk_input_21', 'etaiov_vk_input_22', 'etaiov_vk_input_23', 'etaiov_vk_input_24', 'etaiov_vk_input_25', 'etaiov_vk_input_26', 'etaiov_vk_input_27', 'etaiov_vk_input_28', 'etaiov_vk_input_29', 'etaiov_vk_input_30', 'etaiov_vk_input_31', 'etaiov_vk_input_32', 'etaiov_vk_input_33', 'etaiov_vk_input_34', 'etaiov_vk_input_35', 'etaiov_vk_input_36', 'etaiov_vk_input_37', 'etaiov_vk_input_38', 'etaiov_vk_input_39', 'etaiov_vk_input_40', 'etaiov_vk_input_41', 'etaiov_vk_input_42', 'etaiov_vk_input_43', 'etaiov_vk_input_44', 'etaiov_vk_input_45', 'etaiov_vk_input_46', 'etaiov_vk_input_47', 'etaiov_vk_input_48', 'etaiov_vk_input_49', 'etaiov_vk_input_50', 'etaiov_vk_input_51', 'etaiov_vk_input_52', 'etaiov_vk_input_53', 'etaiov_vk_input_54', 'etaiov_vk_input_55', 'etaiov_vk_input_56'
#> Warning: There was 1 warning in `transmute()`.
#> ℹ In argument: `inr_day20 = inr_test(day20)`.
#> Caused by warning:
#> ! multi-subject simulation without without 'omega'
knitr::kable(maint_check, digits = 4,
             caption = "Day-20 INR at the deposited maintenance dose (target 2.5).")
Day-20 INR at the deposited maintenance dose (target 2.5).
id_deposit VKORC1 CYP2C9 dose_mg inr_day20
1 AA 1/1 2.0600 2.5000
51 AA 1/1 1.7765 2.4999
101 AA 1/2 1.5484 2.5000
151 GA 1/2 1.9361 2.5000
201 GA 1/1 3.6461 2.5000
251 GA 1/1 4.6737 2.5000
301 GA 1/1 3.5699 2.5000
351 GA 1/1 4.2868 2.5000
401 GA 1/1 3.6127 2.5000
451 GA 1/1 4.3977 2.5000
501 GA 1/2 2.2905 2.4999
551 GA 1/3 1.7004 2.5000
601 GA 1/1 5.1798 2.5000
651 GG 1/2 7.6497 2.5000
701 GG 1/1 6.1567 2.4999
751 GG 1/1 12.9127 2.5000
801 GG 1/3 3.8848 2.5000
851 GG 1/1 6.0061 2.5000
901 GG 1/1 4.6484 2.5000
951 GG 1/1 8.7004 2.5000
stopifnot(
  nrow(day20) == nrow(maint),
  # Same individual parameters on both sides, so only solver and optimiser
  # tolerance separate them: measured maximum deviation 8e-5.
  max(abs(maint_check$inr_day20 - 2.5)) < 0.002
)

Gate 2: phase I trial (warfarin PK)

The phase I trial (Algorithm S1) gives 60 patients a single 10 mg dose and measures warfarin concentrations at 10, 35 and 60 h, with exponentially delayed dosing and sampling and lognormal measurement error. The patients’ covariates and the 5th-95th percentiles of the deposited simulated concentrations (trial_phase_I.csv) are embedded below; the packaged model is simulated with fresh random draws for the same patients.

p1_cov <- read.table(text = gsub(" ", "\n", "
0:0:2:72.3 1:2:2:61.4 2:0:2:79.5 3:2:2:66.9 4:1:2:64.3 5:0:2:75.3 6:0:2:75.1 7:0:2:64.8 8:0:2:69.1 9:0:1:55.3
10:0:1:77.3 11:0:1:76.1 12:0:1:75.8 13:0:1:56.6 14:0:1:66.8 15:2:1:69.0 16:0:1:71.8 17:0:1:71.0 18:1:1:78.6 19:0:1:59.3
20:2:1:67.6 21:0:1:64.7 22:0:1:70.3 23:0:1:72.9 24:0:1:78.6 25:0:1:65.9 26:0:1:73.8 27:1:1:57.7 28:0:1:67.8 29:1:1:76.6
30:1:1:66.3 31:3:1:71.3 32:0:1:76.5 33:1:1:76.5 34:0:1:75.1 35:0:1:68.2 36:2:1:69.9 37:1:1:66.8 38:1:1:75.0 39:0:0:86.4
40:0:0:68.5 41:2:0:64.3 42:0:0:68.2 43:0:0:55.3 44:1:0:69.7 45:0:0:62.2 46:2:0:67.1 47:1:0:68.1 48:0:0:73.3 49:2:0:69.5
50:1:0:74.0 51:4:0:58.8 52:0:0:59.1 53:0:0:67.3 54:0:0:64.4 55:0:0:67.0 56:1:0:70.5 57:5:0:67.8 58:0:0:72.0 59:0:0:79.4
"), sep = ":", col.names = c("id", "cyp2c9", "vkorc1", "age"))
p1_ref <- data.frame(
  time = rep(c(10, 35, 60), each = 5),
  q = rep(c(0.05, 0.25, 0.5, 0.75, 0.95), 3),
  conc = c(0.2420, 0.2580, 0.2782, 0.3039, 0.3328,
           0.1275, 0.1498, 0.1668, 0.1850, 0.2393,
           0.0667, 0.0758, 0.0939, 0.1138, 0.1633)
)
cyp_s2 <- c(0, 1, 0, 2, 1, 0)
cyp_s3 <- c(0, 0, 1, 0, 1, 2)

set.seed(2023)
rxode2::rxSetSeed(2023)
n1 <- nrow(p1_cov)
delay_dose <- rexp(n1, rate = 1 / 0.5)
ev_p1 <- bind_rows(
  data.frame(id = seq_len(n1), time = delay_dose, evid = 1L, cmt = "depot", amt = 10),
  data.frame(id = rep(seq_len(n1), each = 3),
             time = rep(c(10, 35, 60), n1) + rexp(3 * n1, rate = 1 / 0.5),
             evid = 0L, cmt = "central", amt = 0)
) |>
  arrange(id, time, desc(evid)) |>
  mutate(
    CYP2C9_S2_COUNT = cyp_s2[p1_cov$cyp2c9[id] + 1],
    CYP2C9_S3_COUNT = cyp_s3[p1_cov$cyp2c9[id] + 1],
    VKORC1_1639G_COUNT = 2 - p1_cov$vkorc1[id],
    AGE = p1_cov$age[id], OCC = 0
  )
sim_p1 <- rxode2::rxSolve(mod, ev_p1, method = solve_method, returnType = "data.frame") |>
  filter(!is.na(Cc)) |>
  mutate(nominal = c(10, 35, 60)[findInterval(time, c(0, 30, 55))],
         conc = Cc * exp(rnorm(n(), 0, 0.1)))
p1_sim_q <- sim_p1 |>
  group_by(nominal) |>
  reframe(q = c(0.05, 0.25, 0.5, 0.75, 0.95), conc_sim = quantile(conc, q))
p1_cmp <- p1_ref |>
  left_join(p1_sim_q, by = c("time" = "nominal", "q")) |>
  mutate(ratio = conc_sim / conc)
p1_cmp |>
  pivot_wider(id_cols = time, names_from = q, values_from = c(conc, conc_sim)) |>
  knitr::kable(digits = 4, caption = "Warfarin concentration (mg/L) percentiles: deposited phase I data (conc_) vs this model (conc_sim_).")
Warfarin concentration (mg/L) percentiles: deposited phase I data (conc_) vs this model (conc_sim_).
time conc_0.05 conc_0.25 conc_0.5 conc_0.75 conc_0.95 conc_sim_0.05 conc_sim_0.25 conc_sim_0.5 conc_sim_0.75 conc_sim_0.95
10 0.2420 0.2580 0.2782 0.3039 0.3328 0.2414 0.2589 0.2784 0.2964 0.3283
35 0.1275 0.1498 0.1668 0.1850 0.2393 0.1238 0.1407 0.1585 0.1822 0.2171
60 0.0667 0.0758 0.0939 0.1138 0.1633 0.0691 0.0794 0.0966 0.1122 0.1538

ggplot(sim_p1, aes(nominal, conc)) +
  geom_point(alpha = 0.3, position = position_jitter(width = 1)) +
  geom_line(data = filter(p1_ref, q == 0.5), aes(time, conc), colour = "red") +
  geom_line(data = filter(p1_ref, q %in% c(0.05, 0.95)), aes(time, conc, group = q),
            colour = "red", linetype = 2) +
  labs(x = "Time after dose (h)", y = "Warfarin concentration (mg/L)",
       caption = "Points: this model. Red: median and 5th/95th percentiles of the authors' phase I data (Figure 5, left panel).")


stopifnot(
  # Centre: a wrong volume, clearance or covariate coding moves the medians by
  # tens of percent; with n = 60 the sampling error of a median is ~3-5%.
  all(abs(log(p1_cmp$ratio[p1_cmp$q == 0.5])) < 0.10),
  # Spread: interquartile percentiles within 15%.
  all(abs(log(p1_cmp$ratio[p1_cmp$q %in% c(0.25, 0.75)])) < 0.15)
)

Gate 3: phase II trial (INR under INR-guided dosing)

The phase II trial (Algorithm S2, Figure 5 middle panel) treats 100 patients for three weeks: 10, 7.5 and 5 mg on days 0-2, then the dose is adjusted on days 3, 5, 7 and 13 whenever the INR measured that morning is outside 2-3, using dose * 2.5 / INR rounded to what tablets allow (0, 1, 2 mg or a multiple of 0.5 mg, at most 30 mg; adjust_dosing_regimen() in the authors’ 3_perform_trial_phase_2.py). INR is measured before dosing on days 0, 1, 2, 3, 5, 7, 13 and 20. The trial is re-run here for the same 100 patients (covariates from the deposited trial_phase_II.csv, embedded below) with fresh draws of the IIV, the daily vitamin K factor, the execution delays and the measurement error, and with the dose adjustments made by the same rule on this model’s own INR measurements. The INR quartiles on each measurement day and the quartiles of the dose chosen on day 13 are compared with the deposited trial.

p2 <- read.csv(text = "
id,cyp2c9,vkorc1,age,d3,d5,d7,d13
0,0,2,60.8,7.5,7.5,7.5,6
1,2,2,61.8,5,3,2,1
2,0,2,69.1,5,3.5,2,1
3,0,2,68.8,5,5,5,4
4,0,2,74.4,6.5,6.5,5,5
5,0,2,76.5,5,3,2.5,2
6,0,2,67.0,5,3.5,3,2.5
7,4,2,68.0,5,5,3,1
8,0,2,62.8,7,7,7,5.5
9,1,2,63.0,5,3.5,2.5,2.5
10,0,2,53.1,5,5,5,3
11,1,2,58.1,5,3.5,2.5,2.5
12,1,2,58.2,5,6.5,5,3.5
13,0,2,66.6,5,5,5,3
14,2,2,62.9,5,2.5,2,1
15,0,1,79.0,7.5,10.5,10.5,10.5
16,1,1,78.7,8.5,11.5,11.5,11.5
17,0,1,71.3,5,5,5,3.5
18,3,1,71.4,7,7,5.5,3.5
19,0,1,73.0,8,11.5,11.5,8
20,0,1,64.3,7,10,10,10
21,0,1,64.4,8.5,8.5,8.5,8.5
22,0,1,59.0,6.5,6.5,6.5,4.5
23,1,1,62.6,8,10,10,6.5
24,0,1,72.8,5,5,5,4
25,0,1,65.9,5,5,5,5
26,0,1,62.7,8.5,8.5,8.5,6
27,2,1,63.1,7,7,7,4.5
28,2,1,65.7,8.5,11.5,11.5,7.5
29,0,1,72.9,7,7,7,7
30,0,1,57.0,5,5,5,5
31,1,1,63.2,5,5,5,5
32,1,1,70.9,6.5,6.5,6.5,6.5
33,1,1,56.3,7,10.5,10.5,7.5
34,1,1,74.2,7,7,7,4.5
35,0,1,59.8,7.5,7.5,7.5,7.5
36,2,1,69.5,5,4,2.5,2
37,0,1,65.6,7,9,9,5.5
38,0,1,67.3,8.5,12,15.5,15.5
39,0,1,57.6,5,6.5,6.5,6.5
40,1,1,65.8,10,18,18,12
41,0,1,62.7,5,5,5,5
42,0,1,73.3,5,5,5,4
43,1,1,61.6,5,5,5,5
44,0,1,64.9,8,11,14,10.5
45,1,1,63.1,5,5,4,4
46,1,1,67.6,5,5,4,4
47,0,1,73.3,8,12.5,12.5,9.5
48,0,1,65.1,5,7,7,6
49,0,1,68.0,7.5,10,10,10
50,0,1,58.6,8,8,8,6.5
51,1,1,65.4,6.5,5,3.5,2.5
52,0,1,57.6,5,5,5,5
53,0,1,59.4,7.5,7.5,7.5,6
54,0,1,73.6,7.5,7.5,7.5,5.5
55,2,1,64.5,6.5,6.5,9.5,6.5
56,0,1,61.0,9,13.5,13.5,13.5
57,4,1,65.1,6.5,6.5,5,2.5
58,0,1,56.5,5,5,5,3.5
59,0,1,77.2,8.5,11.5,11.5,9
60,2,1,61.9,6.5,8,6.5,4
61,0,1,66.7,7,7,7,7
62,1,1,63.3,6.5,6.5,6.5,5
63,0,1,64.6,9.5,9.5,9.5,7
64,0,0,58.7,9,14.5,19.5,19.5
65,0,0,62.9,8,8,8,8
66,0,0,69.5,8,13,13,10.5
67,2,0,61.0,8,11.5,11.5,8
68,3,0,64.1,7,7,6,4
69,0,0,77.9,7,7,7,7
70,0,0,65.5,7,7,7,5.5
71,1,0,59.5,8,11,11,8.5
72,0,0,72.9,7.5,7.5,9.5,9.5
73,2,0,70.8,8.5,8.5,8.5,8.5
74,0,0,71.0,5,5,5,4
75,2,0,66.9,11,19.5,26,16.5
76,1,0,66.6,5,5,5,4
77,0,0,74.7,9,15,19,14
78,5,0,63.4,8,8,6.5,3.5
79,2,0,71.9,5,5,5,4
80,0,0,67.0,5,5,5,5
81,1,0,70.5,7.5,7.5,7.5,7.5
82,1,0,65.5,7.5,7.5,7.5,7.5
83,0,0,63.4,7.5,9.5,9.5,9.5
84,0,0,60.2,6.5,6.5,6.5,6.5
85,0,0,59.6,5,7.5,7.5,7.5
86,0,0,72.8,9.5,9.5,9.5,9.5
87,0,0,60.1,5,5,5,5
88,0,0,68.7,7.5,10.5,10.5,10.5
89,0,0,63.2,9,9,9,9
90,0,0,60.2,7,9.5,9.5,9.5
91,2,0,72.5,6.5,8,8,8
92,0,0,76.3,7,7,7,7
93,0,0,53.7,5,5,5,5
94,0,0,65.8,8.5,11,11,8
95,2,0,69.2,7,7,7,5.5
96,0,0,65.5,6.5,8.5,11,8
97,0,0,82.0,6.5,11,11,7
98,0,0,81.7,9,12,12,12
99,0,0,67.6,7,9.5,9.5,7.5
")
p2_ref <- data.frame(
  day = rep(c(0, 1, 2, 3, 5, 7, 13, 20), each = 3),
  q = rep(c(0.25, 0.5, 0.75), 8),
  inr = c(0.954, 1.083, 1.187, 1.241, 1.396, 1.629, 1.491, 1.683, 1.970,
          1.603, 1.815, 2.101, 1.893, 2.170, 2.486, 2.203, 2.510, 2.833,
          2.751, 3.122, 3.525, 2.698, 2.935, 3.180)
)
# Dose update of the authors' phase II script: only outside the therapeutic
# range, rounded to tablet strengths, capped at 30 mg.
adjust_dose <- function(dose, inr, target = 2.5) {
  out <- dose
  move <- inr < 2 | inr > 3
  d <- dose[move] * target / inr[move]
  d <- ifelse(d < 0.5, 0, ifelse(d < 1.5, 1, ifelse(d < 2.25, 2, round(2 * d) / 2)))
  out[move] <- pmin(d, 30)
  out
}

n2 <- nrow(p2)
meas_days <- c(0, 1, 2, 3, 5, 7, 13, 20)
adjust_days <- c(3, 5, 7, 13)

# One fixed random draw per patient, re-used by every sequential solve below so
# the patient is the same person each time the trial is re-simulated.
set.seed(4)
omega_sd <- sqrt(diag(mod$omega))
etas <- as.data.frame(matrix(rnorm(n2 * length(omega_sd)), n2) %*% diag(omega_sd))
names(etas) <- names(omega_sd)
delay <- matrix(rexp(n2 * 21, rate = 1 / 0.5), n2, 21) # days 0-20, one per day
eps <- matrix(rnorm(n2 * length(meas_days), 0, mod$theta[["expSd"]]), n2)
pars_p2 <- cbind(id = seq_len(n2), etas)

p2_events <- function(doses, last_day) {
  dose_rows <- data.frame(id = rep(seq_len(n2), each = 20), day = rep(0:19, n2),
                          evid = 1L, cmt = "depot", amt = as.vector(t(doses)))
  obs_rows <- data.frame(id = rep(seq_len(n2), each = 8), day = rep(meas_days, n2),
                         evid = 0L, cmt = "central", amt = 0)
  bind_rows(dose_rows, obs_rows) |>
    filter(day <= last_day) |>
    # INR is sampled just before that day's (equally delayed) dose
    mutate(time = 2400 + 24 * day + delay[cbind(id, day + 1)] - ifelse(evid == 0L, 1e-6, 0),
           OCC = day + 1) |>
    bind_rows(data.frame(id = seq_len(n2), day = -1, evid = 0L, cmt = "central",
                         amt = 0, time = 0, OCC = 0)) |>
    mutate(CYP2C9_S2_COUNT = cyp_s2[p2$cyp2c9[id] + 1],
           CYP2C9_S3_COUNT = cyp_s3[p2$cyp2c9[id] + 1],
           VKORC1_1639G_COUNT = 2 - p2$vkorc1[id],
           AGE = p2$age[id]) |>
    arrange(id, time, desc(evid))
}

p2_inr <- function(doses, last_day) {
  s <- rxode2::rxSolve(mod_typ, p2_events(doses, last_day), params = pars_p2,
                       method = solve_method, returnType = "data.frame", keep = "day")
  s <- s[s$day >= 0, ]
  stopifnot(!anyNA(s$coagulation_factor_ii))
  s$inr <- inr_test(s) * exp(eps[cbind(s$id, match(s$day, meas_days))])
  s
}

doses <- matrix(rep(c(10, 7.5, 5, rep(5, 17)), each = n2), n2)
for (d in adjust_days) {
  s <- p2_inr(doses, d)
  inr_d <- s$inr[s$day == d][order(s$id[s$day == d])]
  doses[, (d + 1):20] <- adjust_dose(doses[, d], inr_d)
}
#> Warning: multi-subject simulation without without 'omega'
#> ℹ omega/sigma items treated as zero: 'etalvc', 'etalkel', 'etalec50', 'etalconversion_rate_vk_vkh2', 'etalconversion_rate_vko_vk', 'etatvproduction_rate_ii', 'etatvproduction_rate_v', 'etatvproduction_rate_vii', 'etatvproduction_rate_ix', 'etatvproduction_rate_x', 'etatvproduction_rate_xi', 'etatvproduction_rate_xii', 'etatvproduction_rate_xiii', 'etatvproduction_rate_pc', 'etatvproduction_rate_ps', 'etaiov_vk_input_1', 'etaiov_vk_input_2', 'etaiov_vk_input_3', 'etaiov_vk_input_4', 'etaiov_vk_input_5', 'etaiov_vk_input_6', 'etaiov_vk_input_7', 'etaiov_vk_input_8', 'etaiov_vk_input_9', 'etaiov_vk_input_10', 'etaiov_vk_input_11', 'etaiov_vk_input_12', 'etaiov_vk_input_13', 'etaiov_vk_input_14', 'etaiov_vk_input_15', 'etaiov_vk_input_16', 'etaiov_vk_input_17', 'etaiov_vk_input_18', 'etaiov_vk_input_19', 'etaiov_vk_input_20', 'etaiov_vk_input_21', 'etaiov_vk_input_22', 'etaiov_vk_input_23', 'etaiov_vk_input_24', 'etaiov_vk_input_25', 'etaiov_vk_input_26', 'etaiov_vk_input_27', 'etaiov_vk_input_28', 'etaiov_vk_input_29', 'etaiov_vk_input_30', 'etaiov_vk_input_31', 'etaiov_vk_input_32', 'etaiov_vk_input_33', 'etaiov_vk_input_34', 'etaiov_vk_input_35', 'etaiov_vk_input_36', 'etaiov_vk_input_37', 'etaiov_vk_input_38', 'etaiov_vk_input_39', 'etaiov_vk_input_40', 'etaiov_vk_input_41', 'etaiov_vk_input_42', 'etaiov_vk_input_43', 'etaiov_vk_input_44', 'etaiov_vk_input_45', 'etaiov_vk_input_46', 'etaiov_vk_input_47', 'etaiov_vk_input_48', 'etaiov_vk_input_49', 'etaiov_vk_input_50', 'etaiov_vk_input_51', 'etaiov_vk_input_52', 'etaiov_vk_input_53', 'etaiov_vk_input_54', 'etaiov_vk_input_55', 'etaiov_vk_input_56'
#> Warning: multi-subject simulation without without 'omega'
#> Warning: multi-subject simulation without without 'omega'
#> ℹ omega/sigma items treated as zero: 'etalvc', 'etalkel', 'etalec50', 'etalconversion_rate_vk_vkh2', 'etalconversion_rate_vko_vk', 'etatvproduction_rate_ii', 'etatvproduction_rate_v', 'etatvproduction_rate_vii', 'etatvproduction_rate_ix', 'etatvproduction_rate_x', 'etatvproduction_rate_xi', 'etatvproduction_rate_xii', 'etatvproduction_rate_xiii', 'etatvproduction_rate_pc', 'etatvproduction_rate_ps', 'etaiov_vk_input_1', 'etaiov_vk_input_2', 'etaiov_vk_input_3', 'etaiov_vk_input_4', 'etaiov_vk_input_5', 'etaiov_vk_input_6', 'etaiov_vk_input_7', 'etaiov_vk_input_8', 'etaiov_vk_input_9', 'etaiov_vk_input_10', 'etaiov_vk_input_11', 'etaiov_vk_input_12', 'etaiov_vk_input_13', 'etaiov_vk_input_14', 'etaiov_vk_input_15', 'etaiov_vk_input_16', 'etaiov_vk_input_17', 'etaiov_vk_input_18', 'etaiov_vk_input_19', 'etaiov_vk_input_20', 'etaiov_vk_input_21', 'etaiov_vk_input_22', 'etaiov_vk_input_23', 'etaiov_vk_input_24', 'etaiov_vk_input_25', 'etaiov_vk_input_26', 'etaiov_vk_input_27', 'etaiov_vk_input_28', 'etaiov_vk_input_29', 'etaiov_vk_input_30', 'etaiov_vk_input_31', 'etaiov_vk_input_32', 'etaiov_vk_input_33', 'etaiov_vk_input_34', 'etaiov_vk_input_35', 'etaiov_vk_input_36', 'etaiov_vk_input_37', 'etaiov_vk_input_38', 'etaiov_vk_input_39', 'etaiov_vk_input_40', 'etaiov_vk_input_41', 'etaiov_vk_input_42', 'etaiov_vk_input_43', 'etaiov_vk_input_44', 'etaiov_vk_input_45', 'etaiov_vk_input_46', 'etaiov_vk_input_47', 'etaiov_vk_input_48', 'etaiov_vk_input_49', 'etaiov_vk_input_50', 'etaiov_vk_input_51', 'etaiov_vk_input_52', 'etaiov_vk_input_53', 'etaiov_vk_input_54', 'etaiov_vk_input_55', 'etaiov_vk_input_56'
#> Warning: multi-subject simulation without without 'omega'
#> Warning: multi-subject simulation without without 'omega'
#> ℹ omega/sigma items treated as zero: 'etalvc', 'etalkel', 'etalec50', 'etalconversion_rate_vk_vkh2', 'etalconversion_rate_vko_vk', 'etatvproduction_rate_ii', 'etatvproduction_rate_v', 'etatvproduction_rate_vii', 'etatvproduction_rate_ix', 'etatvproduction_rate_x', 'etatvproduction_rate_xi', 'etatvproduction_rate_xii', 'etatvproduction_rate_xiii', 'etatvproduction_rate_pc', 'etatvproduction_rate_ps', 'etaiov_vk_input_1', 'etaiov_vk_input_2', 'etaiov_vk_input_3', 'etaiov_vk_input_4', 'etaiov_vk_input_5', 'etaiov_vk_input_6', 'etaiov_vk_input_7', 'etaiov_vk_input_8', 'etaiov_vk_input_9', 'etaiov_vk_input_10', 'etaiov_vk_input_11', 'etaiov_vk_input_12', 'etaiov_vk_input_13', 'etaiov_vk_input_14', 'etaiov_vk_input_15', 'etaiov_vk_input_16', 'etaiov_vk_input_17', 'etaiov_vk_input_18', 'etaiov_vk_input_19', 'etaiov_vk_input_20', 'etaiov_vk_input_21', 'etaiov_vk_input_22', 'etaiov_vk_input_23', 'etaiov_vk_input_24', 'etaiov_vk_input_25', 'etaiov_vk_input_26', 'etaiov_vk_input_27', 'etaiov_vk_input_28', 'etaiov_vk_input_29', 'etaiov_vk_input_30', 'etaiov_vk_input_31', 'etaiov_vk_input_32', 'etaiov_vk_input_33', 'etaiov_vk_input_34', 'etaiov_vk_input_35', 'etaiov_vk_input_36', 'etaiov_vk_input_37', 'etaiov_vk_input_38', 'etaiov_vk_input_39', 'etaiov_vk_input_40', 'etaiov_vk_input_41', 'etaiov_vk_input_42', 'etaiov_vk_input_43', 'etaiov_vk_input_44', 'etaiov_vk_input_45', 'etaiov_vk_input_46', 'etaiov_vk_input_47', 'etaiov_vk_input_48', 'etaiov_vk_input_49', 'etaiov_vk_input_50', 'etaiov_vk_input_51', 'etaiov_vk_input_52', 'etaiov_vk_input_53', 'etaiov_vk_input_54', 'etaiov_vk_input_55', 'etaiov_vk_input_56'
#> Warning: multi-subject simulation without without 'omega'
#> Warning: multi-subject simulation without without 'omega'
#> ℹ omega/sigma items treated as zero: 'etalvc', 'etalkel', 'etalec50', 'etalconversion_rate_vk_vkh2', 'etalconversion_rate_vko_vk', 'etatvproduction_rate_ii', 'etatvproduction_rate_v', 'etatvproduction_rate_vii', 'etatvproduction_rate_ix', 'etatvproduction_rate_x', 'etatvproduction_rate_xi', 'etatvproduction_rate_xii', 'etatvproduction_rate_xiii', 'etatvproduction_rate_pc', 'etatvproduction_rate_ps', 'etaiov_vk_input_1', 'etaiov_vk_input_2', 'etaiov_vk_input_3', 'etaiov_vk_input_4', 'etaiov_vk_input_5', 'etaiov_vk_input_6', 'etaiov_vk_input_7', 'etaiov_vk_input_8', 'etaiov_vk_input_9', 'etaiov_vk_input_10', 'etaiov_vk_input_11', 'etaiov_vk_input_12', 'etaiov_vk_input_13', 'etaiov_vk_input_14', 'etaiov_vk_input_15', 'etaiov_vk_input_16', 'etaiov_vk_input_17', 'etaiov_vk_input_18', 'etaiov_vk_input_19', 'etaiov_vk_input_20', 'etaiov_vk_input_21', 'etaiov_vk_input_22', 'etaiov_vk_input_23', 'etaiov_vk_input_24', 'etaiov_vk_input_25', 'etaiov_vk_input_26', 'etaiov_vk_input_27', 'etaiov_vk_input_28', 'etaiov_vk_input_29', 'etaiov_vk_input_30', 'etaiov_vk_input_31', 'etaiov_vk_input_32', 'etaiov_vk_input_33', 'etaiov_vk_input_34', 'etaiov_vk_input_35', 'etaiov_vk_input_36', 'etaiov_vk_input_37', 'etaiov_vk_input_38', 'etaiov_vk_input_39', 'etaiov_vk_input_40', 'etaiov_vk_input_41', 'etaiov_vk_input_42', 'etaiov_vk_input_43', 'etaiov_vk_input_44', 'etaiov_vk_input_45', 'etaiov_vk_input_46', 'etaiov_vk_input_47', 'etaiov_vk_input_48', 'etaiov_vk_input_49', 'etaiov_vk_input_50', 'etaiov_vk_input_51', 'etaiov_vk_input_52', 'etaiov_vk_input_53', 'etaiov_vk_input_54', 'etaiov_vk_input_55', 'etaiov_vk_input_56'
#> Warning: multi-subject simulation without without 'omega'
samp <- p2_inr(doses, 20)
#> Warning: multi-subject simulation without without 'omega'
#> ℹ omega/sigma items treated as zero: 'etalvc', 'etalkel', 'etalec50', 'etalconversion_rate_vk_vkh2', 'etalconversion_rate_vko_vk', 'etatvproduction_rate_ii', 'etatvproduction_rate_v', 'etatvproduction_rate_vii', 'etatvproduction_rate_ix', 'etatvproduction_rate_x', 'etatvproduction_rate_xi', 'etatvproduction_rate_xii', 'etatvproduction_rate_xiii', 'etatvproduction_rate_pc', 'etatvproduction_rate_ps', 'etaiov_vk_input_1', 'etaiov_vk_input_2', 'etaiov_vk_input_3', 'etaiov_vk_input_4', 'etaiov_vk_input_5', 'etaiov_vk_input_6', 'etaiov_vk_input_7', 'etaiov_vk_input_8', 'etaiov_vk_input_9', 'etaiov_vk_input_10', 'etaiov_vk_input_11', 'etaiov_vk_input_12', 'etaiov_vk_input_13', 'etaiov_vk_input_14', 'etaiov_vk_input_15', 'etaiov_vk_input_16', 'etaiov_vk_input_17', 'etaiov_vk_input_18', 'etaiov_vk_input_19', 'etaiov_vk_input_20', 'etaiov_vk_input_21', 'etaiov_vk_input_22', 'etaiov_vk_input_23', 'etaiov_vk_input_24', 'etaiov_vk_input_25', 'etaiov_vk_input_26', 'etaiov_vk_input_27', 'etaiov_vk_input_28', 'etaiov_vk_input_29', 'etaiov_vk_input_30', 'etaiov_vk_input_31', 'etaiov_vk_input_32', 'etaiov_vk_input_33', 'etaiov_vk_input_34', 'etaiov_vk_input_35', 'etaiov_vk_input_36', 'etaiov_vk_input_37', 'etaiov_vk_input_38', 'etaiov_vk_input_39', 'etaiov_vk_input_40', 'etaiov_vk_input_41', 'etaiov_vk_input_42', 'etaiov_vk_input_43', 'etaiov_vk_input_44', 'etaiov_vk_input_45', 'etaiov_vk_input_46', 'etaiov_vk_input_47', 'etaiov_vk_input_48', 'etaiov_vk_input_49', 'etaiov_vk_input_50', 'etaiov_vk_input_51', 'etaiov_vk_input_52', 'etaiov_vk_input_53', 'etaiov_vk_input_54', 'etaiov_vk_input_55', 'etaiov_vk_input_56'
#> Warning: multi-subject simulation without without 'omega'

p2_cmp <- samp |>
  group_by(day) |>
  reframe(q = c(0.25, 0.5, 0.75), inr_sim = quantile(inr, q)) |>
  left_join(p2_ref, by = c("day", "q")) |>
  mutate(ratio = inr_sim / inr)
p2_cmp |>
  pivot_wider(id_cols = day, names_from = q, values_from = c(inr, inr_sim)) |>
  knitr::kable(digits = 3, caption = "INR quartiles by trial day: deposited phase II data (inr_) vs this model (inr_sim_).")
INR quartiles by trial day: deposited phase II data (inr_) vs this model (inr_sim_).
day inr_0.25 inr_0.5 inr_0.75 inr_sim_0.25 inr_sim_0.5 inr_sim_0.75
0 0.954 1.083 1.187 0.944 1.047 1.205
1 1.241 1.396 1.629 1.258 1.408 1.592
2 1.491 1.683 1.970 1.465 1.653 1.907
3 1.603 1.815 2.101 1.604 1.908 2.158
5 1.893 2.170 2.486 1.887 2.152 2.457
7 2.203 2.510 2.833 2.312 2.542 2.887
13 2.751 3.122 3.525 2.851 3.104 3.452
20 2.698 2.935 3.180 2.760 2.976 3.370

dose13 <- data.frame(q = c(0.25, 0.5, 0.75),
                     deposited = quantile(p2$d13, c(0.25, 0.5, 0.75)),
                     simulated = quantile(doses[, 14], c(0.25, 0.5, 0.75)))
knitr::kable(dose13, digits = 2, row.names = FALSE,
             caption = "Quartiles of the dose (mg) given from day 13 onwards.")
Quartiles of the dose (mg) given from day 13 onwards.
q deposited simulated
0.25 4 4.00
0.50 6 6.25
0.75 8 8.00

ggplot(samp, aes(day, inr, group = id)) +
  geom_line(alpha = 0.15) +
  geom_line(data = filter(p2_ref, q == 0.5), aes(day, inr, group = NULL), colour = "red") +
  geom_line(data = filter(p2_ref, q != 0.5), aes(day, inr, group = q), colour = "red",
            linetype = 2) +
  geom_hline(yintercept = c(2, 3), colour = "blue", linetype = 3) +
  labs(x = "Day of treatment", y = "INR",
       caption = "Grey: this model, one line per patient. Red: median and quartiles of the authors' phase II data (Figure 5, middle panel). Blue: therapeutic range.")


stopifnot(
  # Centre: an EC50, VKORC1 or network transcription error shifts the median
  # INR trajectory by far more than the ~5% sampling error of n = 100.
  all(abs(log(p2_cmp$ratio[p2_cmp$q == 0.5])) < 0.10),
  # Quartiles within 15% on every measurement day.
  all(abs(log(p2_cmp$ratio)) < 0.15),
  # The titrated dose is the trial's integrated readout of sensitivity.
  abs(log(dose13$simulated[2] / dose13$deposited[2])) < 0.2
)

PKNCA: single-dose warfarin PK by CYP2C9 genotype

The paper reports no NCA. As a check of the PK layer, a single 10 mg dose is simulated without IIV for a 71-year-old in each CYP2C9 genotype group, and the NCA half-life is compared with the model’s own log(2) / kel.

geno <- data.frame(
  treatment = c("*1/*1", "*1/*2", "*1/*3", "*2/*2", "*2/*3", "*3/*3"),
  CYP2C9_S2_COUNT = cyp_s2, CYP2C9_S3_COUNT = cyp_s3
)
tgrid <- c(0, 0.25, 0.5, 1, 2, 3, 4, 6, 8, 12, 24, 36, 48, 72, 96, 120, 168, 240, 336)
ev_nca <- bind_rows(
  data.frame(id = seq_len(6), time = 0, evid = 1L, cmt = "depot", amt = 10),
  data.frame(id = rep(seq_len(6), each = length(tgrid)), time = rep(tgrid, 6), evid = 0L,
             cmt = "central", amt = 0)
) |>
  arrange(id, time, desc(evid)) |>
  mutate(CYP2C9_S2_COUNT = geno$CYP2C9_S2_COUNT[id], CYP2C9_S3_COUNT = geno$CYP2C9_S3_COUNT[id],
         VKORC1_1639G_COUNT = 2, AGE = 71, OCC = 0)
sim_nca <- rxode2::rxSolve(mod_typ, ev_nca, method = solve_method, returnType = "data.frame") |>
  mutate(treatment = geno$treatment[id])
#> ℹ omega/sigma items treated as zero: 'etalvc', 'etalkel', 'etalec50', 'etalconversion_rate_vk_vkh2', 'etalconversion_rate_vko_vk', 'etatvproduction_rate_ii', 'etatvproduction_rate_v', 'etatvproduction_rate_vii', 'etatvproduction_rate_ix', 'etatvproduction_rate_x', 'etatvproduction_rate_xi', 'etatvproduction_rate_xii', 'etatvproduction_rate_xiii', 'etatvproduction_rate_pc', 'etatvproduction_rate_ps', 'etaiov_vk_input_1', 'etaiov_vk_input_2', 'etaiov_vk_input_3', 'etaiov_vk_input_4', 'etaiov_vk_input_5', 'etaiov_vk_input_6', 'etaiov_vk_input_7', 'etaiov_vk_input_8', 'etaiov_vk_input_9', 'etaiov_vk_input_10', 'etaiov_vk_input_11', 'etaiov_vk_input_12', 'etaiov_vk_input_13', 'etaiov_vk_input_14', 'etaiov_vk_input_15', 'etaiov_vk_input_16', 'etaiov_vk_input_17', 'etaiov_vk_input_18', 'etaiov_vk_input_19', 'etaiov_vk_input_20', 'etaiov_vk_input_21', 'etaiov_vk_input_22', 'etaiov_vk_input_23', 'etaiov_vk_input_24', 'etaiov_vk_input_25', 'etaiov_vk_input_26', 'etaiov_vk_input_27', 'etaiov_vk_input_28', 'etaiov_vk_input_29', 'etaiov_vk_input_30', 'etaiov_vk_input_31', 'etaiov_vk_input_32', 'etaiov_vk_input_33', 'etaiov_vk_input_34', 'etaiov_vk_input_35', 'etaiov_vk_input_36', 'etaiov_vk_input_37', 'etaiov_vk_input_38', 'etaiov_vk_input_39', 'etaiov_vk_input_40', 'etaiov_vk_input_41', 'etaiov_vk_input_42', 'etaiov_vk_input_43', 'etaiov_vk_input_44', 'etaiov_vk_input_45', 'etaiov_vk_input_46', 'etaiov_vk_input_47', 'etaiov_vk_input_48', 'etaiov_vk_input_49', 'etaiov_vk_input_50', 'etaiov_vk_input_51', 'etaiov_vk_input_52', 'etaiov_vk_input_53', 'etaiov_vk_input_54', 'etaiov_vk_input_55', 'etaiov_vk_input_56'
#> Warning: multi-subject simulation without without 'omega'

conc_obj <- PKNCAconc(filter(sim_nca, !is.na(Cc)), Cc ~ time | treatment + id)
dose_obj <- PKNCAdose(
  data.frame(id = seq_len(6), time = 0, amt = 10, treatment = geno$treatment),
  amt ~ time | treatment + id
)
intervals <- data.frame(start = 0, end = Inf, cmax = TRUE, tmax = TRUE,
                        aucinf.obs = TRUE, half.life = TRUE)
nca_res <- pk.nca(PKNCAdata(conc_obj, dose_obj, intervals = intervals))
nca_tab <- as.data.frame(nca_res) |>
  filter(PPTESTCD %in% c("cmax", "tmax", "aucinf.obs", "half.life")) |>
  select(treatment, PPTESTCD, PPORRES) |>
  pivot_wider(names_from = PPTESTCD, values_from = PPORRES)
kel_typ <- exp(mod$theta[["lkel"]]) *
  (1 - (mod$theta[["e_cyp2c9s2_kel"]] * geno$CYP2C9_S2_COUNT +
          mod$theta[["e_cyp2c9s3_kel"]] * geno$CYP2C9_S3_COUNT) / 2)
nca_tab <- nca_tab |>
  left_join(data.frame(treatment = geno$treatment, thalf_model = log(2) / kel_typ),
            by = "treatment")
nca_tab |>
  rename("CYP2C9" = treatment, "Cmax (mg/L)" = cmax, "Tmax (h)" = tmax,
         "AUC0-inf (h*mg/L)" = aucinf.obs, "t1/2 NCA (h)" = half.life,
         "t1/2 = log(2)/kel (h)" = thalf_model) |>
  knitr::kable(digits = 3, caption = "Single 10 mg dose, typical 71-year-old, VKORC1 GG.")
Single 10 mg dose, typical 71-year-old, VKORC1 GG.
CYP2C9 Cmax (mg/L) Tmax (h) t1/2 NCA (h) AUC0-inf (h*mg/L) t1/2 = log(2)/kel (h)
1/1 0.331 2 28.490 14.363 28.488
1/2 0.334 2 37.862 19.092 37.858
1/3 0.336 3 45.839 23.119 45.838
2/2 0.338 3 56.413 28.455 56.412
2/3 0.341 3 76.173 38.426 76.171
3/3 0.344 3 117.240 59.149 117.235
stopifnot(
  # Terminal phase is mono-exponential in kel once absorption (ka = 2 1/h) is over.
  all(abs(nca_tab$half.life / nca_tab$thalf_model - 1) < 0.02),
  # AUC = dose / (kel * V) for a one-compartment model.
  all(abs(nca_tab$aucinf.obs / (10 / (kel_typ * exp(mod$theta[["lvc"]]))) - 1) < 0.02)
)

Assumptions and deviations

  • Parameter source. The paper prints the model structure but not the parameter values; it points to the authors’ GitHub repository for them. All values were taken from models/hartmann_coagulation_model_parameters.csv (repository commit 4710c062), using the population-model rows (“Pooled”, “Log mean”, “Log std.”, “Mean”, “Std.”, “Rel. … shift”) rather than the deterministic rows at the top of the file. The two differ only in the absorption rate (pooled 2 1/h, the Hamberg 2010 value cited in Section S4.2, vs 1 1/h).
  • Fixed parameters. Every value is wrapped in fixed() because the CT model was assembled and hand-calibrated, never estimated.
  • Production-rate IIV. Eq S1.2 gives normal IIV with SD = 20% of the mean. It is encoded as rate = mean * (1 + eta) with variance (SD / mean)^2, which is the same distribution. The deposited SD for factor V is 0.223 for a mean of 1.15 (19.4%), and that value is used.
  • Inter-occasion variability. The daily vitamin K factor is encoded as one eta per treatment day for days 1-56 (OCC = treatment day), the longest simulated trial. Records with OCC = 0, or beyond day 56, carry none.
  • Execution model. The exponential dose and sampling delays (Eq 8) belong to the event table and are generated in the gates above, not in the model.
  • Dose input. The authors approximate an oral dose by a 0.01 h infusion into the dose compartment (Section S1.1.1); the gates use a bolus into depot, which differs by less than the solver tolerance after a few minutes.
  • INR. The PT test needs a second solve, so the INR cannot be an output of the model itself; the helper inr_test() above implements Section S1.1.2 as the authors’ code does, including their use of a reduced reaction set in the test tube, rather than the prose description that sets all production rates to zero (the vitamin K-dependent production retained by the code changes the 120 s test negligibly).
  • Volume of distribution. The typical volume, 28.6 L, is twice the Hamberg 2010 S-warfarin volume of 14.3 L. The paper does not say whether the dose should be read as racemic warfarin or as S-warfarin; the gates use the doses exactly as the authors’ trial scripts did.
  • VKORC1 coding. The paper writes the genotype as GG / GA / AA; this is mapped to VKORC1_1639G_COUNT (2 / 1 / 0).