New in version 2.0.1

mrgsolve version 2.0.1 was released to CRAN in May, 2026. This release adds several new features making it easier to convert your model from NONMEM as well as some much-needed refactoring under the hood.

nm-vars
nonmem
pkmodel
new release
Author

Kyle Baron

Published

June 2, 2026

mrgsolve version 2.0.1 was released to CRAN in May, 2026. This release adds several new features making it easier to convert your model from NONMEM as well as some much-needed refactoring under the hood. This blog post will tell you all about it.

Tiptl;dr
  • A new ** operator to calculate pow(a,b)
  • Plugins available to convert Fortran-style IF/ELSE/THEN and terminate lines with semicolons
  • All the conversion stuff can be handled behind the scenes by mrgsolve or use an Rstudio addin to interactively convert block by block
  • 3-compartment PK model now available in closed form
  • Easier to invoke 1-, 2-, and 3-compartment PK models in closed form

Ok … let’s get into it.

1 Use ** rather than pow()

A good portion of the mrgsolve model code is essentially C++. To raise base a to power b in C++, you use a function: pow(a, b). This is what we’ve done since the start of mrgsolve.

Starting with version 2.0.1, mrgsolve recognizes ** to do this calculation. So pow(a, b) becomes a ** b.

Taking a common idiom seen in population PK modeling, this covariate model

CL = THETA(1) * pow(WT/70, 0.75);

can become

CL = THETA(1) * (WT/70) ** 0.75;

Similarly,

pow(IPRED, 2)

can become

IPRED**2

mrgsolve recognizes ** with or without any plugin; if you’re converting from a nonmem model, just leave the ** in your code and mrgsolve will take care of it under the hood. While we recommend letting mrgsolve do this conversion, we also export a function that will do this conversion.

code <- c(
  "CL = THETA(1) * (WT/70) ** 0.75", 
  "V  = THETA(2) * (IBW/60) ** 0.9", 
  "HAZT = BETA * (EDRUGT**BETA) * (T**(BETA - 1)) * EXP(covar)"
)

convert_pow(code) %>% cat(sep = "\n")
CL = THETA(1)*pow(WT/70, 0.75)
V  = THETA(2)*pow(IBW/60, 0.9)
HAZT = BETA*pow(EDRUGT, BETA)*pow(T, BETA-1)*EXP(covar)

You’ll notice the result has some missing whitespace from the original. This is expected given the configuration of the parser and is one reason why we recommend letting mrgsolve handle this conversion on the fly, when the model is loaded.

2 Fortran IF/ELSE/THEN

When invoking the nm-vars plugin in version 2.0.1, mrgsolve will recognize Fortran-style IF/ELSE/THEN constructs. Unlike ** which is always available, you must be using nm-vars to get this syntax recognized.

For example, in your nonmem model, you might write

IF(SEX.EQ.1) THEN
  TVCL = TVCL * THETA(5);
ENDIF

mrgsolve will recognize this syntax and convert to the appropriate C++ construct. This is also done under the hood when the nm-vars plugin is in play, but there is an exported function to let you do it yourself if needed.

nmcode <- '
IF(SEX.EQ.1) THEN
  TVCL = TVCL * THETA(5);
ENDIF
'

nmcode <- strsplit(nmcode, "\n")[[1]]

cpp <- convert_fort_if(nmcode)

cat(cpp, sep = "\n")

if(SEX == 1) {
  TVCL = TVCL * THETA(5);
}

The mrgsolve nm-vars plugin will handle simpler constructs

nmcode <- "IF(FORM.LE.2) FORMF = THETA(9);"

convert_fort_if(nmcode)
[1] "if(FORM <= 2) FORMF = THETA(9);"

or more complex constructs

nmcode <- '
IF(SEX.EQ.1) THEN
  TVCL = TVCL * THETA(5);
ELSEIF (STUDY.NE.5) THEN
  TVCL = TVCL - THETA(12);
ENDIF
'

nmcode <- strsplit(nmcode, "\n")[[1]]

cpp <- convert_fort_if(nmcode)

cat(cpp, sep = "\n")

if(SEX == 1) {
  TVCL = TVCL * THETA(5);
} else if(STUDY != 5) {
  TVCL = TVCL - THETA(12);
}

3 What about the semicolons?

mrgsolve now provides a semicolons plugin that will attempt to place semicolons at the end of each statement in select model blocks. This plugin requires you to also invoke nm-vars. This is required because we only want users to invoke this plugin when coming to mrgsolve from nonmem; the use case is converting from nonmem, not to overcome laziness when writing the mode de novo.

As an example, this model compiles as-is and we can run it.

code <- '
$PLUGIN nm-vars autodec semicolons

$PKMODEL advan = 1

$PARAM TVCL = 1, V = 20, FLAG = 1, THETA2 = 1.2

$PK
CL = TVCL

IF(FLAG.EQ.2) THEN 
  CL = TVCL * THETA2 
ENDIF
'

mod <- mcode("semicolons-example", code)
Building semicolons-example ... done.
mrgsim(mod, ev(amt = 100), end = 2)
Model:  semicolons-example 
Dim:    4 x 3 
Time:   0 to 2 
ID:     1 
    ID time     A1
1:   1    0   0.00
2:   1    0 100.00
3:   1    1  95.12
4:   1    2  90.48

The semicolons are added via convert_semicolons().

sp <- modelparse(code, split = TRUE)

convert_semicolons(sp$PK) %>% cat(sep = "\n")
CL = TVCL;
IF(FLAG.EQ.2) THEN 
  CL = TVCL * THETA2;
ENDIF

mrgsolve version 2.0.1 also provides a plugin called nm-like that combines the following.

  1. nm-vars
  2. autodec
  3. semicolons

This will give you the most nonmem-like coding experience. See the following modlib() model for an example.

mod <- modlib("nm-like")
Building nm-like ... done.
$PROB Model written with some nonmem-like syntax features

$PLUGIN nm-like

$PARAM
THETA1 = 1, THETA2 = 21, THETA3 = 1.3, WT = 70, F1I = 0.5, D2I = 2
KIN = 100, KOUT = 0.1, IC50 = 10, IMAX = 0.9

$CMT @number 3

$PK
CL = THETA(1) * (WT/70) ** 0.75
V  = THETA(2)
KA = THETA(3)

F1 = F1I
D2 = D2I
A_0(3) = KIN / KOUT

$DES 
CP = A(2)/V
INH = IMAX*CP/(IC50 + CP)
  
DADT(1) = -KA*A(1)
DADT(2) =  KA*A(1) - (CL/V)*A(2)
DADT(3) =  KIN * (1-INH) - KOUT * A(3)

$SIGMA 0.0025

$ERROR
CP = A(2)/V
DV = CP*EXP(ERR(1))

$CAPTURE CP

We recommend you only use the semicolons plugin when porting model code from and existing nonmem run. The reason is that it can be difficult to discern which lines need semicolons and which lines don’t. We’ve set up the rules with the expectation that the code we’re working on came from nonmem; working with this guidance is the only way we can for sure know where to put the semicolons. So if you are coding a model in C++ on your own, take the time to add the semicolons yourself.

4 NONMEM to mrgsolve Rstudio addin

If you’re using Rstudio for coding your model, you will have access to a addin that will let you select a chunk of code (or the whole model) and mrgsolve will do the conversion, including addition of semicolons, into your model file. This lets you review what happened and save the semicolons into the model source file rather than relying on mrgsolve to always get the conversion right.

Check out this video to see the addin in action. Watch on YouTube

5 Closed-form 3-compartment model

mrgsolve has provided 1- and 2-compartment pharmacokinetic models for a while now. Starting with mrgsolve 2.0.1, a 3-compartment model is also provided.

mod <- modlib("pk3", compile = FALSE)

param(mod)

 Model parameters (N=7):
 name value . name value
 CL   1     | V2   20   
 KA   1     | V3   10   
 Q3   2     | V4   50   
 Q4   0.5   | .    .

Using the analytic solutions can provide a large speed advantage over using the equivalent model written in differential equations.

6 PKMODEL gains advan argument

PKMODEL is the code block that you use to ask for 1, 2, or 3-compartment PK models in your model file. Starting with mrgsolve 2.0.1, you can use the advan argument to select the compartmental structure as well as automatically specifying the compartments.

For example, I can get a 1-compartment model with first-order absorption with

code <- '
$PKMODEL advan = 2

$PARAM CL = 1, V = 20, KA = 1.2
'

mod <- mcode("advan-example", code)
Building advan-example ... done.
param(mod)

 Model parameters (N=3):
 name value . name value
 CL   1     | V    20   
 KA   1.2   | .    .    
init(mod)

 Model initial conditions (N=2):
 name     value . name     value
 A1 (1)   0     | A2 (2)   0

Because I left the compartments unspecified, $PKMODEL will automatically give you A1 and A2. You can still ask for compartments named the way you want, but you don’t have to.

  • advan = 1 - one-compartment, bolus
  • advan = 2 - one-compartment, first-order input
  • advan = 3 - two-compartment, bolus
  • advan = 4 - two-compartment, first-order input
  • advan = 11 - three-compartment, bolus
  • advan = 12 - three-compartment, first-order input

7 ERR(n)

When using the nm-vars plugin, ERR(n) is now available as a synonym for EPS(n).

$PLUGIN nm-vars autodec

$ERROR 

IPRED = A(2) / V;

Y = IPRED * EXP(ERR(1));

8 Work with objects in the model environment

Did you know that the model object contains an R environment where you can assign and get R objects before, during or after a simulation? New in mrgsolve 2.0.1, you can assign() an object to that environment from inside your model. We’ll show you that here; but first, let lay some groundwork.

8.1 Initialize the model environment

You can use the $ENV block to initialize the model with an object in the model environment.

code <- '
$ENV vec <- c(1,2,3)
'
mod <- mcode("env", code)
Building env ... done.

8.2 Get objects from the environment

Once you have the model object, you can extract the environment.

env_get_env(mod)
<environment: 0xbb3b2b888>
env_get_env(mod) %>% as.list()
$vec
[1] 1 2 3

Or you could get() the object with env_get().

env_get(mod, "vec")
[1] 1 2 3

If I wanted to change the value of vec in the environment, run this.

mod <- env_update(mod, vec = c(3,4,5))

env_get(mod, "vec")
[1] 3 4 5

Or just use assign.

assign("vec", c(7,8,9), env_get_env(mod))

env_get(mod, "vec")
[1] 7 8 9

The simulation output object also has a copy of the model object; so you can get() an object from the simulation output as well.

out <- mrgsim(mod)

env_get(out, "vec")
[1] 7 8 9

But you might coerce the simulated output to data frame; then you would have to look at the model object to find the objects in that environment.

8.3 Create and assign objects en model

New in mrgsolve 2.0.1, the mrgx plugin gives you access to an assign() function that will let you put R objects into this environment from inside your model.

The call is mrgx::assign("name", value, self)

  • "name" - a string literal with the name for the object
  • value - the object you want to assign
  • self - the model self object; this isn’t the environment, but it knows where to find the environment

Consider this model where we run dynamic dosing and collect pieces of output data in C++ deques. Once we hit the final row of the simulation, we will take that data, create a data frame, and assign() it into the model environment.

model/assign-2-0-1.solv
[ plugin] mrgx evtools

[ global ]
evt::regimen reg; 
std::deque<double> Time;
std::deque<double> Dv;

[ param ] CL = 1, V = 20

[ pkmodel ] advan = 1

[ error ] 
if(NEWIND <=1) {
  reg.init(self);
  reg.ii(24); 
  reg.amt(100); 
  reg.until(168); 
  evt::ev obs = evt::tgrid(0, 240, 1);
  self.push(obs);
}

reg.execute();

Time.push_back(TIME);
Dv.push_back(A1/V);

if(FINAL_ROW) {
  Rcpp::DataFrame data = Rcpp::DataFrame::create(
    Rcpp::Named("TIME") = Time, 
    Rcpp::Named("DV") = Dv
  );
  mrgx::assign("data", data, self);
}

Let’s compile and run that model.

mod <- mread("model/assign-2-0-1.solv")
Building assign-2-0-1_solv ... done.
out <- mrgsim(mod, end = -1, add = c(0,336))

I specifically set up the simulation to output only the first and last time of the simulation, ignoring all the doses and observations in between. We’ll get the rest of the data from the assigned object.

out
Model:  assign-2-0-1_solv 
Dim:    2 x 3 
Time:   0 to 336 
ID:     1 
    ID time       A1
1:   1    0 1.00e+02
2:   1  336 9.69e-03

Remember, we assigned the output data frame to data in the model; so we look there once the simulation is complete.

data <- env_get(mod, "data")

head(data)
  TIME       DV
1    0 0.000000
2    0 5.000000
3    1 4.756147
4    2 4.524187
5    3 4.303540
6    4 4.093654
library(ggplot2)

ggplot(data, aes(TIME, DV)) + geom_line() + theme_bw()

9 Last record for an individual or dataset

Two macros were added:

  • LAST_ROW is true when processing the last row in the entire simulation
  • LAST_IROW is true when processing the last row for an individual

10 Internal improvements

  • Prevent an unneeded copy of the data when returning simulated data.

  • During population simulation, and individual’s data records are cleared prior to moving on to the next individual; this provides a theoretical improvement in memory utilization for big simulations.

11 Breaking changes

  • mrgsolve now depends on R >= 4.1

  • data_set() and idata_set() no longer accept .subset, .select, object, or need arguments. These arguments allowed filtering and column selection inside the call; that processing should now be done on the data frame before passing it to data_set() or idata_set(). (#1374)

  • The n argument to simeta() and simeps() has been discouraged for a while and is now removed. Calling simeta(n) or simeps(n) to resimulate a single ETA or EPS value is no longer supported. Use simeta() or simeps() (no argument) to resimulate all values. (#1373)

  • knobs(), wf_sweep(), and render() have been removed. These functions were previously deprecated. (#1369)

  • The default value of root in $NMXML and $NMEXT has changed from "working" to "cppfile". The root directory for locating NONMEM output files now defaults to the directory containing the model .cpp file rather than the R working directory. Pass root = "working" to restore the previous behavior; but users are highly encouraged to use the new default. (#1368)

  • summarise.each() is removed; it has not been practically reachable by the user for several years (#1352).

  • Several modlib model library models have had parameter and compartment names standardized (#1361):

    • The first extravascular compartment is now named EV (was EV1) in pk1cmt, pk2cmt, irm1-irm4, and emax.
    • The first absorption rate constant is now named KA (was KA1) in models that previously used numbered names.
    • Code that references these compartments or parameters by name (e.g., init(mod, EV1 = 0) or param(mod, KA1 = 1)) will need to be updated.
    • We continue to discourage users from using these models in production code.