Simulate individual dosing with covariates #754

tmss1 · 2020-10-19T17:15:32Z

tmss1
Oct 19, 2020

Hi Kyle,

I have been trying to do some simulations in mrgsolve for a population with relatively complex dosing regimens (loading dose followed by multiple doses over weeks with individualised doses) and individual covariates. I've managed to specify the individual regimens in "data_set", and the individual parameters and covariates in "idata_set". However, I noticed that mrgsolve accepts only either the "data_set" or the "idata_set" but not both, and idata allows only one subject per row so I can't specify the dosing regimens in idata. May I ask how can this be done in mrgsolve - specify individual regimens and covariates?

Thank you.

kylebaron · 2020-10-19T17:23:55Z

kylebaron
Oct 19, 2020
Maintainer

HI @tmscs -

You should be able to pass both. But another option is to put the individual parameters / covariates on the data set. See two options below:

library(mrgsolve)
#> 
#> Attaching package: 'mrgsolve'
#> The following object is masked from 'package:stats':
#> 
#>     filter
library(tidyverse)

n <- 30
data <- as_data_set(
  ev(ID = seq(n), amt = 100, ii = 24, addl = 9), 
  ev(ID = seq(n), amt = 300, ii = 24, addl = 9)
) %>% mutate(dose = amt)

idata <- tibble(CL = runif(nrow(data), 0.5, 1.5), ID = seq(length(CL)))

mod <- house()

Option 1:

Pass both data and idata

mod %>% 
  data_set(data) %>% 
  idata_set(idata) %>% 
  carry_out(dose) %>%
  mrgsim() %>% 
  plot(CP~time|factor(dose), scales = "same")

Option 2:

Join idata on to `data to make one data set

new_data <- left_join(data, idata, by = "ID")

Now, individual parameters are combined with dosing data

head(new_data)
#>   ID time cmt evid amt ii addl dose        CL
#> 1  1    0   1    1 100 24    9  100 0.6355913
#> 2  2    0   1    1 100 24    9  100 0.6641948
#> 3  3    0   1    1 100 24    9  100 1.4896229
#> 4  4    0   1    1 100 24    9  100 1.3380746
#> 5  5    0   1    1 100 24    9  100 1.4147902
#> 6  6    0   1    1 100 24    9  100 1.0306401

mod %>% 
  data_set(new_data) %>% 
  carry_out(dose) %>%
  mrgsim() %>% 
  plot(CP~time|factor(dose), scales = "same")

^{Created on 2020-10-19 by the reprex package (v0.3.0)}

4 replies

luyasong Jan 6, 2025

Hi Kyle, I have a situation similar to the case here. Hoping to get your confirmation.

In my mod.cpp, a weight effect is defined as WT_CL = (WT/WTref)^theta, where "WT=70" is specified as a placeholder. My dataset contains a virtual population with various individual weights defined by the WT column. With this script, mod %>% data_set(dataset) %>% mrgsim(), does the WT in the dataset automatically overwrite/update the WT=70 initial in the model such that the simulation does take all individual WTs into calculations?

Thanks,
Yasong Lu

kylebaron Jan 6, 2025
Maintainer

Hi @luyasong -

does the WT in the dataset automatically overwrite/update the WT=70 initial in the model such that the simulation does take all individual WTs into calculations

Yes; if you have WT as a column in the data set, then that ID-specific value will overwrite the WT = 70 placeholder you put in the parameter list.

Can you make sure WT is included in an $INPUT block like here: https://mrgsolve.org/vignette/#coding-model-parameters

Then you can check that WT matches up with a parameter in the model as shown here: https://mrgsolve.org/vignette/#check-if-the-names-match

You don't have to do this but it is a good check to make sure you coded it as WT rather that WTBL or WGT or something like that.

Kyle

luyasong Jan 6, 2025

Thanks for your lightening fast reply, Kyle. It's a comfort to have your confirmation.
I start reading your vignette package just published.
Thanks,
Yasong

kylebaron Jan 6, 2025
Maintainer

Awesome; best wishes for your simulations.

tmss1 · 2020-10-20T08:56:19Z

tmss1
Oct 20, 2020
Author

Thanks Kyle, this is great help.

May I please also ask if there is a way to include uncertainty of the parameters in the simulations, e.g. RSEs from the estimation?

On another note, I got an error message for non-dosing records without CMT values. Those are records to for observation or to change the covariate values (EIVD=0 or2) so there are no associated CMT numbers. Do I have to put in a CMT number in this case, and does it matter what values I put in? And the same issue with the RATE item, I got an error for missing RATE values when the event is not a dose record. Do I have to put in something for RATE - which I think will get an error in NONMEM when RATE is there without AMT?

0 replies

kylebaron · 2020-10-20T15:25:00Z

kylebaron
Oct 20, 2020
Maintainer

Here's an example of simulating with uncertainty. I have to simulate the parameters here and presumably you'll already have that. But you'll see what are the other mechanics that you need to do:

library(tidyverse)
library(mrgsolve)
#> 
#> Attaching package: 'mrgsolve'
#> The following object is masked from 'package:stats':
#> 
#>     filter

This is a nonmem run inside mrgsolve

mod <- modlib("1005")
#> Building 1005 ... done.

The run stuff is located here

project <- system.file("nonmem", "1005", package = "mrgsolve")

ext_file <- file.path(project, "1005.ext")
est <- read_table(ext_file,skip=1) %>% filter(ITERATION ==-1e9)
#> Parsed with column specification:
#> cols(
#>   ITERATION = col_double(),
#>   THETA1 = col_double(),
#>   THETA2 = col_double(),
#>   THETA3 = col_double(),
#>   THETA4 = col_double(),
#>   THETA5 = col_double(),
#>   THETA6 = col_double(),
#>   THETA7 = col_double(),
#>   `SIGMA(1,1)` = col_double(),
#>   `SIGMA(2,1)` = col_double(),
#>   `SIGMA(2,2)` = col_double(),
#>   `OMEGA(1,1)` = col_double(),
#>   `OMEGA(2,1)` = col_double(),
#>   `OMEGA(2,2)` = col_double(),
#>   `OMEGA(3,1)` = col_double(),
#>   `OMEGA(3,2)` = col_double(),
#>   `OMEGA(3,3)` = col_double(),
#>   OBJ = col_double()
#> )

est <- select(est, matches("THETA|OMEGA|SIGMA"))
mu <- as.numeric(log(est))
#> Warning in FUN(X[[i]], ...): NaNs produced

#> Warning in FUN(X[[i]], ...): NaNs produced

Simulate some parameters

You will already have this from bootstrap, BAYES run etc

pars <- MASS::mvrnorm(1000, mu, dmat(rep(0.01, length(mu))))
pars <- as.data.frame(exp(pars))
names(pars) <- names(est)

head(pars)
#>      THETA1   THETA2     THETA3   THETA4    THETA5    THETA6   THETA7
#> 1  7.424726 22.48143 0.07278706 3.348785 138.38830 0.9693311 1.031972
#> 2  9.957642 19.22409 0.07513595 3.421555  97.47279 0.9794705 1.310917
#> 3  9.779223 21.82776 0.05744524 3.059223 106.25812 0.9090591 1.306647
#> 4  9.900401 20.31143 0.07801912 3.468303 125.49913 0.9647835 1.015118
#> 5 11.001511 25.19532 0.06254278 3.113375 122.54329 1.0089037 1.290575
#> 6 10.888949 20.20446 0.06728664 3.423947 133.32373 0.9415983 1.236596
#>   SIGMA(1,1) SIGMA(2,1) SIGMA(2,2) OMEGA(1,1) OMEGA(2,1) OMEGA(2,2) OMEGA(3,1)
#> 1 0.04679533          0  0.2192475  0.1982942  0.1079991 0.09792380        NaN
#> 2 0.04671150          0  0.2252488  0.2125749  0.1080456 0.09323076        NaN
#> 3 0.04142997          0  0.1769342  0.2383890  0.1271883 0.10542438        NaN
#> 4 0.04085041          0  0.1950293  0.2139424  0.1176727 0.09761303        NaN
#> 5 0.03970289          0  0.2047149  0.1937804  0.1095106 0.10154214        NaN
#> 6 0.07508084          0  0.2404043  0.2206086  0.1262407 0.09860323        NaN
#>   OMEGA(3,2) OMEGA(3,3)
#> 1        NaN 0.05345265
#> 2        NaN 0.05765921
#> 3        NaN 0.03891364
#> 4        NaN 0.04631358
#> 5        NaN 0.04845226
#> 6        NaN 0.04698587

Simplify

pars <- mutate(
  pars, 
  `OMEGA(2,1)` = 0, 
  `OMEGA(3,1)` = 0, 
  `OMEGA(3,2)` = 0
)

Pull out OMEGA and SIGMA

We make a list of matrices for each

omegas <- as_bmat(pars, "OMEGA")

For example

omegas[[1]]
#>           [,1]      [,2]       [,3]
#> [1,] 0.1982942 0.0000000 0.00000000
#> [2,] 0.0000000 0.0979238 0.00000000
#> [3,] 0.0000000 0.0000000 0.05345265

sigmas <- as_bmat(pars, "SIGMA")

Dummy data

data <- expand.ev(ID = 1:100, amt = c(100, 300, 1000))

Simulate

The first argument will be i, the replicate number
We update parameters, omegas, and sigmas
Then return a data frame, tagged with the replicate number

do_sim <-  function(i) {
  
  mod <- update(
    mod, 
    param = slice(pars,i), 
    omega = omegas[[i]], 
    sigma = sigmas[[i]]
  )

  mrgsim_df(mod, data = data) %>% mutate(rep = i)
}

out <- map_dfr(seq(100), do_sim)

head(out)
#>   ID time       GUT      CENT    PERIPH      CL        Q       V2       V3
#> 1  1    0   0.00000  0.000000 0.0000000 9.00391 3.348785 29.85156 138.3883
#> 2  1    0 100.00000  0.000000 0.0000000 9.00391 3.348785 29.85156 138.3883
#> 3  1    1  93.51326  5.302249 0.3191565 9.00391 3.348785 29.85156 138.3883
#> 4  1    2  87.44731  8.475473 1.0914230 9.00391 3.348785 29.85156 138.3883
#> 5  1    3  81.77483 10.270128 2.1146738 9.00391 3.348785 29.85156 138.3883
#> 6  1    4  76.47031 11.177728 3.2593804 9.00391 3.348785 29.85156 138.3883
#>          KA     ETA_1     ETA_2       ETA_3     IPRED rep
#> 1 0.0670669 0.1928431 0.2835473 -0.08184758 0.0000000   1
#> 2 0.0670669 0.1928431 0.2835473 -0.08184758 0.0000000   1
#> 3 0.0670669 0.1928431 0.2835473 -0.08184758 0.1776205   1
#> 4 0.0670669 0.1928431 0.2835473 -0.08184758 0.2839206   1
#> 5 0.0670669 0.1928431 0.2835473 -0.08184758 0.3440399   1
#> 6 0.0670669 0.1928431 0.2835473 -0.08184758 0.3744437   1

^{Created on 2020-10-20 by the reprex package (v0.3.0)}

0 replies

kylebaron · 2020-10-20T15:28:25Z

kylebaron
Oct 20, 2020
Maintainer

On another note, I got an error message for non-dosing records without CMT values. Those are records to for observation or to change the covariate values (EIVD=0 or2) so there are no associated CMT numbers. Do I have to put in a CMT number in this case, and does it matter what values I put in? And the same issue with the RATE item, I got an error for missing RATE values when the event is not a dose record. Do I have to put in something for RATE - which I think will get an error in NONMEM when RATE is there without AMT?

Right; you'll have to make RATE non-missing on every record. This was a feature request from another user.

When EVID==0 you can have CMT be non-negative. When it is 2, you'll have to reference an existing compartment. This might be different than what nonmem does. No claim that it behaves exactly the same.

Thanks,
Kyle

0 replies

tmss1 · 2020-10-21T09:20:14Z

tmss1
Oct 21, 2020
Author

Thanks Kyle. I do most simulations in mrgsolve, but can't completely avoid using nonmem for benchmarking or file sharing etc. so try to have the same dataset in both if possible or with known discrepancies for reproducibility and reduce the risk of error. I suppose these non-dosing CMT and RATE values are needed to be there but do not impact on the simulations?

Thanks you so much for the example. If I understand it correctly, I'd have to first generate the parameters for each replicate using, e.g. the mvrnorm function or bootsrap etc. before inputting them to mrgsolve for the simulations, and I'd run through the replicates by updating the model parameters in each iteration for each replicate?

If I may clarify a few things, is it possible to use the uncertainty values from the nonmem output directly when using mvrnorm to generate the parameters (instead of 0.01)? I noticed that you log-transformed the parameters for mu and then back transformed them during the process. Is it necessary/beneficial for a reason?

0 replies

kylebaron · 2020-10-22T14:03:00Z

kylebaron
Oct 22, 2020
Maintainer

Hi @tmscs -

Right ... don't use the code I had there for simulating from the posterior; that was just to get something for me to work with to show you the mechanics of simulation assuming you already have the posterior. If you don't have that, you can bootstrap the NONMEM run or I'd highly recommend running BAYES if you really want an easy way to generate that uncertainty distribution. I know you can't just switch to BAYES any time, but if it's a possibility, I'd consider it.

If you want to simulate from the covariance matrix of the estimate, you can check this vignette out:
https://github.com/mrgsolve/examples/blob/master/PrTS/pts.pdf

See Section 9 ... this uses the function simpar from the metrumrg package. Unfortunately, that's not on CRAN at this point, but you can get to it here:
https://mpn.metworx.com/docs/packages/metrumrg

install.packages("metrumrg", repos = "https://mpn.metworx.com/snapshots/stable/2020-09-20")

There is also some code related to that vignette here:
https://github.com/mrgsolve/examples/tree/master/PrTS

On the question of log-transformation: if you do simulate from the covariance matrix, I think you do need to estimate the parameters on log scale; you will simulate negative values for THETAs that are supposed to be positive etc and just in general the distribution of THETAs that you simulate won't be correct.

As far as the data set goes, I know it's hard to balance all of the competing interests. CMT and RATE do not factor in to the simulation for non-dosing records; we just put a check in for missing values at someone's request. I put zeros in there in the data set so it works both ways. For CMT, I don't know how it works in NONMEM, but mrgsolve needs a valid compartment number in there. We might be able to refactor that, but for now it needs a valid compartment number. It doesn't matter which compartment.

I can help you with code to make the transformation if you can't do it in the source data.

Kyle

0 replies

tmss1 · 2020-10-26T21:49:17Z

tmss1
Oct 26, 2020
Author

Thanks for the clarification Kyle. May I ask when you said running BAYES, did you mean running an additional step with $EST BAYES using the final model to get a distribution of parameters or else?

0 replies

kylebaron · 2020-10-26T21:54:02Z

kylebaron
Oct 26, 2020
Maintainer

Hi @tmscs -

I was suggesting to estimate the model using $EST METHOD = BAYES from the start; just estimate the model that way.
That may or may not be acceptable for your situation, but I always at least consider it especially when I know having that posterior distribution is important for subsequent simulation steps.

If the model has to be estimates with FOCEI, I'd bootstrap or simulate from the covariance matrix (examples in the previous comment).

Kyle

0 replies

tmss1 · 2020-10-28T23:19:41Z

tmss1
Oct 28, 2020
Author

Thanks very much for your clarification. It's really great help.

0 replies

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Simulate individual dosing with covariates #754

{{title}}

Replies: 9 comments 4 replies

{{title}}

{{title}}

{{title}}

{{title}}

{{title}}

{{title}}

{{editor}}'s edit

{{editor}}'s edit

{{title}}

{{title}}

{{title}}

{{title}}

{{title}}

{{title}}

{{title}}

{{editor}}'s edit

{{editor}}'s edit

Select a reply

Simulate individual dosing with covariates #754

tmss1 Oct 19, 2020

Replies: 9 comments · 4 replies

kylebaron Oct 19, 2020 Maintainer

Option 1:

Option 2:

luyasong Jan 6, 2025

kylebaron Jan 6, 2025 Maintainer

luyasong Jan 6, 2025

kylebaron Jan 6, 2025 Maintainer

tmss1 Oct 20, 2020 Author

kylebaron Oct 20, 2020 Maintainer

This is a nonmem run inside mrgsolve

Simulate some parameters

Pull out OMEGA and SIGMA

Dummy data

Simulate

kylebaron Oct 20, 2020 Maintainer

tmss1 Oct 21, 2020 Author

kylebaron Oct 22, 2020 Maintainer

tmss1 Oct 26, 2020 Author

kylebaron Oct 26, 2020 Maintainer

tmss1 Oct 28, 2020 Author

tmss1
Oct 19, 2020

Replies: 9 comments 4 replies

kylebaron
Oct 19, 2020
Maintainer

kylebaron Jan 6, 2025
Maintainer

kylebaron Jan 6, 2025
Maintainer

tmss1
Oct 20, 2020
Author

kylebaron
Oct 20, 2020
Maintainer

kylebaron
Oct 20, 2020
Maintainer

tmss1
Oct 21, 2020
Author

kylebaron
Oct 22, 2020
Maintainer

tmss1
Oct 26, 2020
Author

kylebaron
Oct 26, 2020
Maintainer

tmss1
Oct 28, 2020
Author