Watch on YouTube Many oral drugs don’t always absorb along one neat first-order path. Some of the dose absorbs quickly, and the rest absorbs more slowly. For example, a modified-release tablet can release part of the dose right away while the rest is released at a constant rate. This is where we can use dual (or parallel) absorption processes. The model for that is easy to write: two depots, each with its own absorption, both feeding the central compartment.
Watch on YouTube nlmixr2scm is on CRAN! Happily, nlmixr2scm, which does stepwise covariate modeling (SCM) for nlmixr2 models, is now on CRAN. It took a while - seems they’re very busy now that everyone is churning out packages with AI - but it’s finally there. install.packages("nlmixr2scm") If you’ve built covariate models in NONMEM, you’ve probably used scm in Perl-speaks-NONMEM. nlmixr2scm follows the same approach (Jonsson & Karlsson, 1998), and it works directly on an nlmixr2 fit.
Watch on YouTube Janice Goh and I wrote a post for the Posit blog: Tiny Doses for Tiny Babies – How much is too much? It is a hands-on introduction to pharmacometrics for R users who have never built a pharmacokinetic model. Posit also published a narrated video walkthrough of the post (above). The question A preterm infant weighing 900 g is having seizures. A doctor has read that one flat dose of phenobarbital might work for every neonate, which would save dose calculations at the bedside.
Watch on YouTube If you are at ACoP 2026 (October 11-14, National Harbor, Maryland), come find the nlmixr2save poster. This post covers the same material, with code you can run. nlmixr2save is on CRAN and version 0.2.0 is out now: install.packages("nlmixr2save") Two problems Problem 1: Fits you can’t open later. Nonlinear mixed-effects models can take minutes to weeks to fit. Once one is done, you want to keep it. The usual way is saveRDS(), but that ties the result to R’s binary format and to how nlmixr2 stored the fit internally.
Watch on YouTube Many pharmacometric reports have the same two items near the top of the methods section: a compartment diagram of the final model and the equations that define it. They are also two of the easiest things to get wrong. The diagram is usually drawn by hand in PowerPoint or Visio early in the analysis. The equations are typed into Word or LaTeX from the control stream. Then the model changes – a transit compartment is added, the effect goes from stimulation to inhibition, a covariate moves – and the code is updated but the figure and the math are not.
Watch on YouTube In the 7.0 post, I talked mostly about parallel solving in focei. This time I want to talk about a feature that has been requested frequently: a full covariance step, and other refinements to the covariance steps. Two things changed there, and they are related: Nearly any covariance method can now be requested from nearly any estimation method, and you can switch a finished fit to a different one without refitting.
Preterm infants are highly susceptible to seizures primarily due to brain immaturity and an increased vulnerability to acute brain injuries and systemic imbalances, so if an infant is delivered pre-term they may be likely to have seizures. So, with a preterm infant with seizures – what can we do in data analytics to support the right drug and dose the first time? As stated by Paracelsus, the dose makes the poison: Too little and the seizures continue, too much and the infant stops breathing.
Watch on YouTube nlmixr2 7.0 This is the release where focei finally solves in parallel on all OpenMP-enabled platforms. Along with that, a few defaults changed, which means that by default focei will not give you the same numbers it gave you in nlmixr2 6. This is a deliberate change and we will talk about the motivation in this post (and how to get the old behavior back if you need it).
Usually, pharmacometric modelling starts with a dataset with individual patient data: one row per observation, in a software-ready format. To model these data, you would first need to get access to the data, which means data-sharing agreements, trial access, or in-house studies. This process is often slow or even not possible due to privacy or competition. However, researchers do publish their model and sometimes summary metrics. The published reports do contain useful information.
Introduction Survival or time-to-event (TTE) analysis models when something happens, not just whether it happens. The endpoint might be time to tumour progression, time to an adverse event, or time to treatment discontinuation, for instance. In this post we’ll illustrate how to fit parametric TTE models in nlmixr2 using its custom log-likelihood interface. For the purposes of this post, we’re going to simulate a two-arm randomised trial from a known Gompertz model, then recover the parameters with nlmixr2.
The nlmixr2 development team, represented by Bill Denney, Lisa Hendricks, and I (Matt Fidler), recently gave a invited seminar to the FDA quantitative clinical pharmacology reviewers to discuss nlmixr2 and how open, community-built modeling tools can be helpful in regulatory settings. The conversation was constructive and forward-looking, and a clear theme emerged: anything that lets regulatory reviewers trust, reproduce, and exchange models across tools would be very useful. As we go through these reflections, I must add that these are my takeaways and are not necessarily the views of the FDA, nor are they necessarily the views of the companies or individuals represented by the nlmixr2 working group, or the R Consortium.
VPCs – a new method for VPCs in nlmixr2 I am excited to announce a new VPC method available in nlmixr2 – tidyvpc. This method can be built on-top of previous code and provides slightly different visualizations than the vpc package. Below is an example that shows the two different methods: library(nlmixr2) ## ── Attaching packages ───────────────────────────────────────────────────────── nlmixr2 5.0.0 ── ## ✔ lotri 1.0.2 ✔ monolix2rx 0.0.6 ## ✔ nlmixr2data 2.
Inter-occasion variability Happy new year! With the holidays over we had the opportunity to celebrate many special occasions. I invite you to celebrate a new occasion for nlmixr2 – inter-occasion variability. We now support inter-occasion variability estimation in nlmixr2! Implementing inter-occasion variability in models To implement inter-occasion variability in nlmixr2 models you simply have to specify that the variable is an occasion variable, say occ: iov.cl ~ 0.1 | occ The ~ syntax is similar to how between subject variability is specified, the addition of | tells that this a between occ variability.
nlmixr2 5.0 As time progresses, dependencies update, dependencies come and go, and sometimes these dependencies change what happens to rxode2/nlmixr2 models. One dependency change (qs being dropped) means that models from before nlmixr2/rxode2 5.0 have the UI/fit models saved to disk will may not be able to load completely in nlmixr2 5.0. This is because internally nlmixr2/rxode2 encoded its models in binary qs format. Why is qs going away? While the maintainer wanted to keep qs and its file format around, R is changing and what is acceptable to access within C/C++ from R has changed.
nlmixr2 log-likelihood In 2022 we announced the focei log-likelihood. However, in our last advisory committee meeting Mats Karlsson pointed out that focei log-likelihood may not be the best approach. He believed that Stuart L. Beal and Lewis B. Sheiner did not include this method since they may have been trying to protect users from methods that may not make sense (for example, maybe focei log-likelihood is not an accurately enough approximation of the likelihood).
nlmixr2 4.0 verse As promised, I am discussing some other change that are breaking and makes nlmixr2 a major release. First is a very visible change as you load nlmixr2: library(nlmixr2) ## ── Attaching packages ───────────────────────────────────────────────────────── nlmixr2 4.0.1 ── ## ✔ lotri 1.0.2 ✔ monolix2rx 0.0.6 ## ✔ nlmixr2data 2.0.9 ✔ nlmixr2lib 0.3.0 ## ✔ nlmixr2est 4.1.0 ✔ nlmixr2rpt 0.2.1 ## ✔ nlmixr2extra 3.0.3 ✔ nonmem2rx 0.1.8 ## ✔ nlmixr2plot 3.
nlmixr2 4.0 We at the nlmixr2 team are committed to quality results. This is seen in line with our vision: The vision of nlmixr2 is to develop a R-based open-source nonlinear mixed-effects modeling software package that can compete with commercial pharmacometric tools and is suitable for regulatory submissions. This release furthers the quality by changing linear compartment models the ADVAN solutions to the WNL solutions(both claim to be something that NONMEM is using for the linear solved systems).
nlmixr2 ecosystem I have seen a few new pharmacometrics tools integrated in the nlmixr2 ecosystem recently. I thought I would point out the tools I know that integrate in the nlmixr2 ecosystem. Some are maintained by our nlmixr2 team, and many are not. For each category, these are ordered alphabetically. Tools that Enhance nlmixr2’s language nlmixr2lib In addition to a model library, it has tools to change model components (like add Weibull absorption, add transit compartments, change standard elimination to Michaleis-Menton absorption; maintained by Bill Denney and developed by the nlmixr2 team).
Neural Network ODEs and nlmixr2 I have had some requests to talk about nlmixr2 using neural network ODEs, since neural networks are something that more people are exploring with the explosion of artificial intelligence LLMs. There is a package, pmxNODE, by Dominic Bräm that adds neural network ODEs to pharmacometric modeling tools like NONMEM, Monolix and nlmixr2. In addition to the code that Dominic has added, I extended this package to allow Neural Networks directly in a rxode2 or nlmixr2 model.
This month I will on a single nlmixr2’s plot function that is shared with nlme, augPred(). I think this is useful but also harder to find like the rxode2 plots discussed last month. The example of this feature is the phenobarbitol data: library(nlmixr2) pheno <- function() { ini({ tcl <- log(0.008) # typical value of clearance tv <- log(0.6) # typical value of volume ## var(eta.cl) eta.cl + eta.v ~ c(1, 0.
This month, I will talk about about a new iteration of mu-referencing in nlmixr2, which I call mu3 and mu4. What is mu referencing in nlmixr2 – Review from another post From the last blog post about mu-referencing, I will give a brief overview of mu-referencing and what mu-2 referencing is and how it is expanded a bit more. mu-referencing is combining a fixed effect, random effect and possibly a covariate in the form:
nlmixr2 3.0 is here nlmixr2 3.0 has been released and most of the related packages in the nlmixr2 ecosystem have been updated as well. Since there were a few non-backward compatible changes and breaking changes, the version has been bumped from 2 to 3. Most code will run the same, but because of the breaking change, we changed the major version. The big changes are: Non abi binary linkages between every package.
One of the things that can be useful from time to time is to calculate different PK parameters based on whatever parameters you have estimated. There is a great function, calc_derived() in pmxTools that allows this calculation of the derived parameters (by my collaborators Justin Wilkins and Bill Denney). I think this is an underrated function that can help many people with typical calculations. rxode2 has the same type of function, which can be helpful to test the linCmt() functions, rxDerived().
This month, I will talk about about a new iteration of mu-referencing in nlmixr2, which I call mu2. What is mu referencing in nlmixr2 mu-referencing is combining a fixed effect, random effect and possibly a covariate in the form: [ \theta_\mathsf{pop}+\eta_\mathsf{individual}+\theta_\mathsf{covariate}\times \mathsf{DataCovariate} ] Often they are placed in exponentials for these to be log-normally distributed like: [ \exp\left(\theta_\mathsf{pop}+\eta_\mathsf{individual}+\theta_\mathsf{covariate}\times \mathsf{DataCovariate}\right) ] In optimization routines like saem, these are switched out with a single parameter during optimization classically called (\phi) in both NONMEM and Monolix.
Both nlmixr2 and rxode2 have been updated, the below describes all the nlmixr2 related packages (maintained by the nlmixr2 team). Most of items in this release are bug fixes. One of the changes will make random number generation platform independent. Unfortunately, this means simulations from within rxode2/nlmixr2 will have different numbers drawn from random distributions but I think platform independence is important enough to push this change through. Versions of new packages nlmixr2 2.
One of the exciting new features of the recent rxode2 is user functions. This allows you to define your own R functions for use in nlmixr2 or rxode2. This new feature can really help make your code more concise by reusing custom transformations or apply more complex routines. This can call R functions directly, but at a cost – single threaded and slower execution. However, you can reduce the cost by converting the R functions to C automatically with rxFun().
One of the things that I changed in the last release was steady state. Once I created the nonmem2rx package, I searched for NONMEM control streams that we could test the import from, especially those with attached data. I ran across nmtests that uses NONMEM to test against mrgsolve and how it behaves. During that test I noticed that rxode2 did not follow the convention of NONMEM in lagged steady state doses.
Running large population PK/PD analyses on laptops and desktops often requires long computational times. This is quite tedious. In addition, when using parallel computing on your machine, it can slow it down for a while, creating further nuisances. Outsourcing computation to the cloud is a solution to this problem. Among the various cloud providers, Amazon Web Service (AWS) is one of the most famous and used by industries in various fields.
nlmixr2 2.1.0 was released and I promised to talk about the new features. One of the things that can impact many peoples work-flow is new estimation methods for population-only data. Many people use population-only estimation methods before changing the model to a mixed effect model, so I believe these can be useful for many people trying to find the best model to the data at hand. I will talk about the new ones (and why you may want to use them).
nlmixr2 2.1.0 was released 9 Jan 2024 which requires the newest (and recently released) engine packages nlmixr2est 2.2.0 (released 12 Dec 2023) and rxode2 2.0.10 (released 13 Dec 2023). I will be blogging about a few new features but want to mention a few now: More flexible mu referencing User defined functions can now be used with nlmixr2 and rxode2 Event handling changes Many new estimation methods for population only fitting The truncated changelog for these packages are below:
nonmem2rx/babelmixr2 I am really excited to announce that the nlmixr2 team has released a new version of babelmixr2 and a new package nonmem2rx that allows you to convert NONMEM to rxode2 or even a nlmixr2 object. To install, simply upgrade babelmixr2 with: install.packages(c("nonmem2rx", "babelmixr2")) What you can do with nonmem2rx/babelmixr2 You can do many useful tasks directly converting between nlmixr2 and NONMEM models; you can: Convert a NONMEM model to a rxode2 model
This is another release of a group of nlmixr2-related packages. Feature Highlights There are a few things I would like to highlight in this release: Highly requested feature(s) A much requested feature has been added for rxode2; Diagonal zeros in the omega and sigma matrices are treated as zeros in the model. The corresponding omega and sigma matrices drop columns/rows where the diagonals are zero to create a new omega and sigma matrix for simulation.
This is more of a methodology post, pointing out how things are done in nlmixr2 and how it likely doesn’t match what is done in NONMEM (and at least one reason why a drop-in replacement of rxode2 by another tool like PKPDsim, mrgsolve, or deSolve is not an easy project). For the impatient, adding focei lag time (and other dose-based events) have improved in stability for this release of nlmixr2.
Last time I blogged promised to talk about a few other things, including: Likelihood based on each observation (and how to get it) Standard Errors / Hessians, etc for between subject variabilities or etas (and how to get them) Hessians for the individual between subject variability is also used for the focei calculation. So, if you are impatient, I will give you brief instructions on where to get each component of the likelihood:
I am pretty excited abut the new nlmixr2 release (2.0.8). When I joined the the nlmixr2 team, I wanted to do a fancy heavy tailed, skewed model in an open source tool so I could figure out how to do even more with it. With this release, it is possible to do a heavy tailed (t-distribution dt()) skewed (coxBox(lambda)) distribution: my old wish is now possible with focei! A few other things that people may be interested in are:
Over the past half year, a lot of changes have been happening behind the scenes, and the time has finally come to reveal them! nlmixr2 nlmixr2 will be the version in active development going forward, taking over from nlmixr, starting with the current CRAN version, 2.0.6. Our new home on GitHub is here, and on CRAN, we’re here. The reasons for the name and format change are many, but most importantly, we’ve taken this step to improve overall user experience and to help us maintain the project more effectively.