NMsim

Please make sure to read this at The NMsim website where you can browse several vignettes with examples on specific topics.

NMsim is an R package that can simulate Nonmem models (using the NMsim function) based on just a simulation data set and a path to an estimation control stream. It will retrive and combine output tables with input data once Nonmem has finished and return the results to R.

The interface is “seamless” or fully integrated in R. Run a simulation of the (estimated) model stored in “path/to/file.mod” using the simulation input data set stored in the variable data.sim this way:

simres <- NMsim(file.mod="/path/to/file.mod",
                data=data.sim)

You will quickly learn to do this on your own models, but if you can’t wait to see this working, you can do the following:

data.sim <- read.csv(system.file("examples/derived/dat_sim1.csv",package="NMsim"))
simres <- NMsim(file.mod=system.file("examples/nonmem/xgxr021.mod",package="NMsim"),
                data=data.sim,
                dir.sims=".")

where dir.sims may be needed because the model in this case may be in a read-only location.

Notice, that could be any working Nonmem model as long as the provided simulation data set is sufficient to run it. We are ready to plot:

library(ggplot2)
datl <- as.data.table(simres) |>
    melt(measure.vars=cc(PRED,IPRED,Y))

ggplot(datl,aes(TIME,value,colour=variable))+
    geom_line(data=function(x)x[variable!="Y"])+
    geom_point(data=function(x)x[variable=="Y"])+
    labs(x="Hours since first dose",y="Concentration (ng/mL)")

This example was a simulation of a multiple dose regimen with a loading dose using a model estimated on single dose data. It is from the first vignette NMsim-intro.html. Go there next to get started with NMsim.

Motivation

While NONMEM offers great flexibility for estimation of PK and PK/PD models, many users find the simulation features in NONMEM insufficient and turn to alternative software for simulation. This leads to additional work of model reimplementation, with risk of the simulation model deviating from the estimated model, due to bugs in the reimplementation. For a wide range of model types, the limitation is not in NONMEM’s ability to perform such simulations, but rather in the lack of a simple user-interface to obtain the simulations. NMsim provides such an interface as an R package, allowing the modeler to simulate models as soon as an estimate is available.

The goal for NMsim is to automate the NONMEM simulation workflow and provide a simple, flexible, and powerful R interface. This allows for automation of most simulation-based analyses.

Supported types of simulations

NMsim has a flexible way to define simulation methods. The following methods are currently provided:

• Simulation of new subjects (default or explicitly with method.sim=NMsim_default)
• Simulation of subjects already estimated in Nonmem model (method.sim=NMsim_EBE)
• Simulation with parameter uncertainty based on a Nonmem covariance step (method.sim=NMsim_VarCov)
• Simulation “as is” in case you already prepared a simulation control stream and just want to automate the use of it in combination with simulation data sets (method.sim=NMsim_asis)

In addition, NMsim provides other features to further modify the simulation control stream

• Simulation of typical subjects with all ETAs equal 0 (typical=TRUE)
• Custom modification of control stream sections (modify.sections argument)

To learn how to run these simulations on your Nonmem models, get started with NMsim-intro.html. It is really easy.

In addition, NMsim can simulate multiple models at a time. E.g., if a bootstrap run of a model is available, NMsim can run the simulation with each of the bootstrap models and collect all the results in one dataset. This provides a robust and easy way to simulate a Nonmem model with uncertainty.

You can also write your own methods, if you have some other Nonmem-based simulation (or other job) you want to automate using NMsim.

Many features are available. Prominent ones are:

• Can use submit jobs to clusters. It can wait for the simulations to be done and automatically collect the results like in the example above.
• Simulation replicates using Nonmem SUBPROBLEMS feature avaible through the subproblems argument
• Can modify the simulation control stream on the fly - a powerful feature for studying the effect of varying model parameters
• Simulations of models on transformed observations can be automatically transformed back using the transform argument.

If residual variability is not implemented in the simulated model, NMsim provides a way (addResVar()) to add residual variability in R after the simulation has been run.

How NMsim works

NMsim does not simulate, translate or otherwise interpret a NONMEM model. Instead, it automates the NONMEM simulation workflow (including execution of NONMEM) and wraps it all into one R function. Provided with a path to a NONMEM control stream and a data.frame to simulate, NMsim will do the following:

• Save the simulation input data in a csv file for NONMEM
• Create a simulation input control stream based on file.mod ($INPUT and $DATA matching the saved simulation data set; $SIMULATE instead of $ESTIMATION and $COVARIANCE)
• Update and fix initial values based on estimate (from file.ext)
• Run NONMEM on the generated simulation control stream
• Collect output data tables, combine them, and merge with the simulation input data
• Return the collected data in R

NMsim can call NONMEM directly or via PSN. If NMsim is run on a system where NONMEM cannot be executed, NMsim can still prepare the simulation control stream and datafile.

NMsim is in itself a relatively small R package. It makes extensive use of functionality to handle NONMEM data and control streams provided by the R package NMdata.

Supported model types

The methods currently provided by NMsim will work with (many or most) Pop PK models and most continuous-scale PD models. Methods are currently not provided for for time-to-event models. Also, depending on the coding of the models, other censored data models may not work out of the box, because the model may not have a single variable (in Nonmem) that simulates the wanted information for all data rows, as their interpretation may depend on other values.

The input data set must contain whatever variables are needed by the Nonmem model. A common issue is if the Nonmem model uses a covariate that is not in the simulation input data set. NMdata’s NMcheckData is a good help identifying input data issues before running Nonmem - and when Nonmem acts unexpectedly.

NMsim and speed

Nonmem may not be the fastest simulator out there. But actually most often, the reason Nonmem is slow at providing a simulation result is that it takes a long time writing the $TABLE files (yes, that can account for 90% or more of the time Nonmem spends). NMsim provides a simple way to get around this. The argument text.table can be used to define only the columns needed in the simulation output (which may be as little as PRED, IPRED, and a couple more - remember the input data is merged back automatically). As a result, NMsim may still be slower than a re-implementation in a different framework. But it’s extremely easy to do.

Requirements

NMsim is dependent on running Nonmem. Often, that will mean Nonmem must be available on the same system as the one running R. However, if Nonmem is run on a separate system through qsub or in another way initiates Nonmem on another system, that will work too. Then however, only if R can read the file system where Nonmem writes the results, it can retrieve the results.

NMsim does not need PSN but can use it. However, not all features are available with PSN, so for some features you will have to specify the path to the Nonmem executable (say path.nonmem=/path/to/nmfe75 or any Nonmem executable you want to use). Specifically of the simulation types currently available, simulation of known subjects is not possible using PSN (but works if a Nonmem executable is provided).

If PSN is used, NMsim uses PSN’s execute to run models. In addition, NMsim by default uses PSN’s update_inits to update initial values in control streams, if PSN is available. NMsim does also include its own simple function to do this if PSN is not available.

Is NMsim reliable?

Importantly, NMsim does not (at least not by default) modify, translate or simulate the model itself. It does modify control stream sections $INPUT, $DATA, $ESTIMATION, $SIMULATION, $THETA, $OMEGA, $SIGMA, $TABLE as needed. The fact that NMsim allows for skipping the re-implementation but just uses Nonmem to simulate the Nonmem model as is, eliminates the risk of discrepancies between the estimated model and the simulated model.

The produced control stream is saved together with simulation data set open for manual inspection and can obviously be run with Nonmem independently of NMsim.

Easily create simulation datasets

NMsim includes functions (NMcreateDoses and addEVID2) to very easily create simulation data sets. While one certainly does not need to use these functions to use NMsim, they do add to the package providing a framework that enables a complete simulation workflow in only 5-15 simple lines of R code.

Run Nonmem from R

There are several other packages out there that can do this, and NMsim may not be your best choice if this feature is all you are looking for. However, running Nonmem using the NMexec() function provided by NMsim has one important advantage in that it saves the input data together with the Nonmem control streams. This ensures that output data can be merged with input data as it went into the model, even if the input data file should be modified or lost.

• Saves input data with Nonmem model
• Provides a simple R command for submission of Nonmem jobs
• Optionally handles cluster configuration
• Saves the xml file by default

NMexec will submit model runs to a cluster by default. This can be switched off for running Nonmem locally. Please notice the jobs are submitted to a cluster in a very specific way using PSN. If your setup is different, this is for now not supported. Please use NMexec(sge=FALSE) in that case (which may not be desirable). Notice that simulations are not done on a cluster by default so you may still be able to use NMsim.

Install

NMsim is on CRAN, MPN and github:

## From CRAN/MPN repositories
install.packages("NMsim")
## From github
library(remotes)
install_github("NMautoverse/NMsim")