pith. sign in

arxiv: 2606.24427 · v1 · pith:F6K7JP3Lnew · submitted 2026-06-23 · 📊 stat.CO · stat.ML

NoLimits.jl: Flexible and Composable Nonlinear Mixed-Effects Modeling in Julia

Manuel Huth , Jonas Arruda , Nina Schmid , Roy Gusinow , Vincent Wieland , Clemens Peiter , Jan Hasenauer This is my paper

Pith reviewed 2026-06-25 21:52 UTC · model grok-4.3

classification 📊 stat.CO stat.ML
keywords nonlinear mixed-effects modelsJulia packageLaplace approximationstochastic expectation maximizationMarkov chain Monte Carlonormalizing flowsordinary differential equationsneural networks
0
0 comments X

The pith

NoLimits.jl provides a macro-based interface for building and estimating nonlinear mixed-effects models from ODEs, neural networks, and normalizing flows.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces NoLimits.jl to address limits in current open-source tools that restrict the model structures, inference methods, and machine-learning elements usable in nonlinear mixed-effects modeling. It establishes that a macro-based language can compose observation and latent-state models from building blocks such as ordinary differential equations, Markov models, and neural networks while supporting covariate-dependent distributions. A sympathetic reader would care because this setup unifies access to Laplace approximation, stochastic expectation maximization, and Markov chain Monte Carlo estimation in one framework. Demonstrations on case studies show integration of differentiable machine-learning parts and data-driven random-effects estimation via normalizing flows.

Core claim

NoLimits.jl is an open-source Julia package whose macro-based modeling language enables construction of observation and latent-state models from diverse building blocks including ordinary differential equations, Markov models, and neural networks. It supports flexible, covariate-dependent observation and random-effects distributions and provides a unified interface to frequentist inference through Laplace approximation, stochastic expectation maximization, and Bayesian Markov chain Monte Carlo methods. The package is shown on three case studies that include workflows with differentiable machine-learning components and estimation of random-effects distributions using normalizing flows.

What carries the argument

The macro-based modeling language that composes observation and latent-state models from ODEs, Markov models, neural networks, and normalizing flows.

If this is right

  • The same model specification can be estimated with Laplace approximation, stochastic expectation maximization, or MCMC without rewriting the model.
  • Neural networks can be inserted as components for observation or latent states inside a mixed-effects structure.
  • Random-effects distributions can be estimated directly from data using normalizing flows instead of fixed parametric forms.
  • Covariate dependence can be added to both observation and random-effects distributions within the same modeling syntax.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Users may spend less time translating models between separate packages when different inference or component types are needed.
  • The composable structure could make it easier to test and compare model variants that mix mechanistic and data-driven elements.
  • Further building blocks beyond those shown could be added to the macro language to cover additional longitudinal data settings.

Load-bearing premise

The macro-based modeling language can actually construct and estimate the full range of models including ODEs, neural networks, and normalizing flows without hidden restrictions that block the demonstrated workflows.

What would settle it

An attempt to define a model that combines an ODE latent process with a neural-network observation model and a normalizing-flow random effect, then run estimation under all three inference methods, produces errors or fails to complete.

Figures

Figures reproduced from arXiv: 2606.24427 by Clemens Peiter, Jan Hasenauer, Jonas Arruda, Manuel Huth, Nina Schmid, Roy Gusinow, Vincent Wieland.

Figure 1
Figure 1. Figure 1: Overview of the NoLimits.jl workflow. Mechanistic terms, machine-learning com￾ponents, and custom distributions are combined into a single composable model structure comprising fixed effects, random effects, and structural and observation models. Inference yields estimates for fixed and random effects, optionally including uncertainty quantification. Built-in evaluation tools support model validation, comp… view at source ↗
Figure 2
Figure 2. Figure 2: (A) Observed plasma warfarin concentrations versus time for 32 subjects. Each line [PITH_FULL_IMAGE:figures/full_fig_p033_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Model fit diagnostics for the warfarin analysis. (A)–(C) Posterior predictions [PITH_FULL_IMAGE:figures/full_fig_p037_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Marginal posterior distributions of the population log-means from the Bayesian fit [PITH_FULL_IMAGE:figures/full_fig_p038_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Warfarin anticoagulant response and pharmacodynamic model comparison. [PITH_FULL_IMAGE:figures/full_fig_p042_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Random-effect distribution recovery on simulated data. (A) Fitted growth-rate [PITH_FULL_IMAGE:figures/full_fig_p047_6.png] view at source ↗
read the original abstract

Nonlinear mixed-effects models are widely used to analyze longitudinal data, but existing open-source software often supports only a limited subset of the model structures, inference methods, machine-learning components, automatic differentiation techniques, and random-effects distributions required in modern applications. We introduce NoLimits.jl, an open-source Julia package for flexible and composable nonlinear mixed-effects modeling. Its macro-based modeling language enables observation and latent-state models to be constructed from diverse building blocks, including ordinary differential equations, Markov models, and neural networks. NoLimits.jl supports flexible, covariate-dependent observation and random-effects distributions and provides a unified interface to frequentist inference through Laplace approximation, stochastic expectation maximization, and Bayesian Markov chain Monte Carlo methods. We demonstrate the package on three case studies showcasing its workflows, integration of differentiable machine-learning components, and data-driven estimation of random-effects distributions using normalizing flows. Together, these capabilities substantially expand the range of nonlinear mixed-effects models that can be specified, estimated, and compared within a single open-source framework.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper introduces NoLimits.jl, an open-source Julia package for nonlinear mixed-effects modeling. Its macro-based modeling language allows construction of observation and latent-state models from components including ODEs, Markov models, neural networks, and normalizing flows; it supports covariate-dependent distributions for observations and random effects; and it provides a unified interface to Laplace approximation, stochastic expectation maximization, and MCMC inference. Three case studies demonstrate selected workflows, integration of differentiable ML components, and data-driven random-effects estimation.

Significance. If the implementation matches the claims, the package would meaningfully expand the set of NLME models that can be specified and estimated in a single open-source framework, particularly by enabling seamless combination of mechanistic models with neural networks or normalizing flows and by supporting multiple inference paradigms without switching packages. This could reduce fragmentation in pharmacometrics, longitudinal analysis, and related fields.

major comments (2)
  1. [Abstract] Abstract and case-studies section: the central claim that the macro system 'substantially expand[s] the range' of models rests on the assumption that arbitrary combinations of ODEs/Markov/NNs/flows with covariate-dependent observation and random-effects distributions are supported by all three inference back-ends; the three demonstrations cover only selected subsets and do not enumerate or test the Cartesian product, so latent macro-expansion or type constraints could silently invalidate the generality assertion.
  2. [Abstract] Abstract: the statement that the package 'provides a unified interface' to Laplace, SEM, and MCMC is load-bearing for the 'single framework' claim, yet the manuscript supplies no description of the common model representation, automatic-differentiation requirements, or error-handling paths that would allow a reader to verify that all three back-ends can be applied to the same macro-constructed model without hidden restrictions.
minor comments (1)
  1. The manuscript should include a short table or diagram that explicitly lists which combinations of building blocks have been verified to work with each inference method.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which help clarify the scope of our claims about model generality and the unified inference interface in NoLimits.jl. We address each point below and indicate revisions where the manuscript will be updated.

read point-by-point responses

  1. Referee: [Abstract] Abstract and case-studies section: the central claim that the macro system 'substantially expand[s] the range' of models rests on the assumption that arbitrary combinations of ODEs/Markov/NNs/flows with covariate-dependent observation and random-effects distributions are supported by all three inference back-ends; the three demonstrations cover only selected subsets and do not enumerate or test the Cartesian product, so latent macro-expansion or type constraints could silently invalidate the generality assertion.

    Authors: The case studies were selected to highlight distinct capabilities (workflows, ML integration, and data-driven random effects) rather than to exhaustively cover every combination. The macro system generates a unified internal representation that is intended to be compatible with all three back-ends, and compatibility has been verified in the package test suite for the combinations used in the demonstrations. We acknowledge that the manuscript does not enumerate the full Cartesian product or discuss potential type constraints in detail. We will revise the abstract to qualify the generality claim and add a supplementary note or table summarizing tested combinations and their supported inference methods. revision: partial

  2. Referee: [Abstract] Abstract: the statement that the package 'provides a unified interface' to Laplace, SEM, and MCMC is load-bearing for the 'single framework' claim, yet the manuscript supplies no description of the common model representation, automatic-differentiation requirements, or error-handling paths that would allow a reader to verify that all three back-ends can be applied to the same macro-constructed model without hidden restrictions.

    Authors: The manuscript focuses on the user-facing macro language and case studies. The common model representation is an expression graph that supports automatic differentiation through Julia's AD ecosystem and is shared across back-ends, with error handling managed via standard Julia mechanisms; these internals are documented in the package source and README. We agree that a brief description in the paper would strengthen the unified-interface claim. We will add a short paragraph in the methods section outlining the shared representation, AD requirements, and compatibility approach. revision: yes

Circularity Check

0 steps flagged

No circularity: software package description with no derivations

full rationale

This is a software introduction paper presenting the NoLimits.jl package. Its claims concern the existence and demonstrated functionality of modeling macros, inference back-ends, and case-study workflows. No mathematical derivation chain, predictions, or first-principles results exist that could reduce to inputs by construction. No equations, fitted parameters presented as predictions, uniqueness theorems, or ansatzes are invoked. Self-citations, if any, are not load-bearing for any derivation. The paper is self-contained against external benchmarks (the package itself) and receives the default non-finding.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

This is a software tool paper; the central claim rests on the existence and correct implementation of the described package rather than on mathematical axioms or fitted parameters.

pith-pipeline@v0.9.1-grok · 5722 in / 1065 out tokens · 21981 ms · 2026-06-25T21:52:02.140414+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

146 extracted references · 5 canonical work pages

  1. [1]

    1974 , journal =

    A New Look at the Statistical Model Identification , author =. 1974 , journal =

    1974

  2. [2]

    2015 , journal =

    A nonlinear mixed effects approach for modeling the cell-to-cell variability of Mig1 dynamics in yeast , author =. 2015 , journal =

    2015

  3. [3]

    2015 , journal =

    Using sensitivity equations for computing gradients of the FOCE and FOCEI approximations to the population likelihood , author =. 2015 , journal =

    2015

  4. [4]

    Proceedings of the 41st International Conference on Machine Learning , publisher =

    An amortized approach to non-linear mixed-effects modeling based on neural posterior estimation , author =. Proceedings of the 41st International Conference on Machine Learning , publisher =. 2024 , month =

    2024

  5. [5]

    2024 , booktitle =

    An amortized approach to non-linear mixed-effects modeling based on neural posterior estimation , author =. 2024 , booktitle =

    2024

  6. [6]

    2018 , journal =

    Mechanistic models versus machine learning, a fight worth fighting for the biological community? , author =. 2018 , journal =

    2018

  7. [7]

    Fitting Linear Mixed-Effects Models Using

    Bates, Douglas and M. Fitting Linear Mixed-Effects Models Using. 2015 , journal =

    2015

  8. [8]

    2019 , journal =

    NONMEM tutorial part I: description of commands and options, with simple examples of population analysis , author =. 2019 , journal =

    2019

  9. [9]

    2011 , journal =

    Prediction-Corrected Visual Predictive Checks for Diagnosing Nonlinear Mixed-Effects Models , author =. 2011 , journal =

    2011

  10. [10]

    2021 , journal =

    Besan. 2021 , journal =

    2021

  11. [11]

    2017 , journal =

    Julia: A fresh approach to numerical computing , author =. 2017 , journal =

    2017

  12. [12]

    2024 , journal =

    DataInterpolations.jl: Fast Interpolations of 1D Data , author =. 2024 , journal =

    2024

  13. [13]

    2026 , url =

    BlackBoxOptim.jl: Black-Box Optimization in Julia , author =. 2026 , url =

    2026

  14. [14]

    2017 , journal =

    Variational inference: A review for statisticians , author =. 2017 , journal =

    2017

  15. [15]

    2023 , journal =

    Universal differential equations for glacier ice flow modelling , author =. 2023 , journal =

    2023

  16. [16]

    Journal of Open Source Software , volume =

    Borisov, Ivan and Demin, Aleksander and Metelkin, Evgeny , year =. Journal of Open Source Software , volume =

  17. [17]

    2023 , journal =

    A multidimensional ODE-based model of Alzheimer’s disease progression , author =. 2023 , journal =

    2023

  18. [18]

    2023 , journal =

    DataFrames.jl: Flexible and Fast Tabular Data in Julia , author =. 2023 , journal =

    2023

  19. [19]

    2022 , journal =

    Using scientific machine learning to develop universal differential equation for multicomponent adsorption separation systems , author =. 2022 , journal =

    2022

  20. [20]

    2024 , journal =

    Low-dimensional neural ODEs and their application in pharmacokinetics , author =. 2024 , journal =

    2024

  21. [21]

    Low-Dimensional Neural Ordinary Differential Equations Accounting for Inter-Individual Variability Implemented in

    Br. Low-Dimensional Neural Ordinary Differential Equations Accounting for Inter-Individual Variability Implemented in. 2025 , journal =

    2025

  22. [22]

    2001 , publisher =

    Foundations of mathematical economics , author =. 2001 , publisher =

    2001

  23. [23]

    2025 , journal =

    A nonparametric approach to practical identifiability of nonlinear mixed effects models , author =. 2025 , journal =

    2025

  24. [24]

    2018 , journal =

    Neural ordinary differential equations , author =. 2018 , journal =

    2018

  25. [25]

    1995 , journal =

    Understanding the Metropolis-Hastings Algorithm , author =. 1995 , journal =

    1995

  26. [26]

    2017 , journal =

    Parameter estimation in nonlinear mixed effect models using saemix, an R implementation of the SAEM algorithm , author =. 2017 , journal =

    2017

  27. [27]

    Extending the

    Comets, Emmanuelle and Delattre, Maud and Karimi, Belhal , year =. Extending the

  28. [28]

    2024 , note =

    Improving Simulations by Learning the True Random Effects' Distribution from a Population Using Generative Machine Learning , author =. 2024 , note =

    2024

  29. [29]

    jl , author =

    Interpretable machine learning for science with PySR and SymbolicRegression. jl , author =. 2023 , journal =

    2023

  30. [30]

    2003 , journal =

    Nonlinear models for repeated measurement data: an overview and update , author =. 2003 , journal =

    2003

  31. [31]

    1997 , publisher =

    Bootstrap methods and their application , author =. 1997 , publisher =

    1997

  32. [32]

    1993 , journal =

    Comparison of four basic models of indirect pharmacodynamic responses , author =. 1993 , journal =

    1993

  33. [33]

    2025 , journal =

    Conditional universal differential equations capture population dynamics and interindividual variation in c-peptide production , author =. 2025 , journal =

    2025

  34. [34]

    2025 , journal =

    Physiology-informed regularisation enables training of universal differential equation systems for biological applications , author =. 2025 , journal =

    2025

  35. [35]

    1999 , journal =

    Convergence of a stochastic approximation version of the EM algorithm , author =. 1999 , journal =

    1999

  36. [36]

    1977 , journal =

    Maximum likelihood from incomplete data via the EM algorithm , author =. 1977 , journal =

    1977

  37. [37]

    doi:10.5281/zenodo.7738525 , url =

    Vaibhav Kumar Dixit and Christopher Rackauckas , year =. doi:10.5281/zenodo.7738525 , url =

    work page doi:10.5281/zenodo.7738525

  38. [38]

    2021 , journal =

    Practical identifiability in the frame of nonlinear mixed effects models: the example of the in vitro erythropoiesis , author =. 2021 , journal =

    2021

  39. [39]

    2025 , doi =

    nlmixr2: Nonlinear Mixed Effects Models in Population PK/PD , author =. 2025 , doi =

    2025

  40. [40]

    1925 , journal =

    Theory of Statistical Estimation , author =. 1925 , journal =

    1925

  41. [41]

    Huber sandwich estimator

    On the so-called “Huber sandwich estimator” and “robust standard errors” , author =. 2006 , journal =

    2006

  42. [42]

    1990 , journal =

    Sampling-based approaches to calculating marginal densities , author =. 1990 , journal =

    1990

  43. [43]

    1995 , publisher =

    Bayesian data analysis , author =. 1995 , publisher =

    1995

  44. [44]

    Gelman, J.B

    Gelman, Andrew and Carlin, John B and Stern, Hal S and Dunson, David B and Vehtari, Aki and Rubin, Donald B , year =. doi:10.1201/b16018 , url =

    work page doi:10.1201/b16018

  45. [45]

    2015 , journal =

    Partial Migration: Growth Varies between Resident and Migratory Fish , author =. 2015 , journal =

    2015

  46. [46]

    2016 , publisher =

    Deep Learning , author =. 2016 , publisher =

    2016

  47. [47]

    1998 , booktitle =

    The general approximation theorem , author =. 1998 , booktitle =

    1998

  48. [48]

    2013 , publisher =

    Foundations of Applied Mathematics , author =. 2013 , publisher =

    2013

  49. [49]

    2002 , publisher =

    Ordinary Differential Equations , author =. 2002 , publisher =

    2002

  50. [50]

    2019 , journal =

    A Physiological Perspective on the Ecology and Evolution of Partial Migration , author =. 2019 , journal =

    2019

  51. [51]

    , author =

    The No-U-Turn sampler: adaptively setting path lengths in Hamiltonian Monte Carlo. , author =. 2014 , journal =

    2014

  52. [52]

    and Staatz, Christine E

    Hooker, Andrew C. and Staatz, Christine E. and Karlsson, Mats O. , year =. Conditional Weighted Residuals (. Pharmaceutical Research , publisher =

  53. [53]

    1989 , journal =

    Multilayer Feedforward Networks Are Universal Approximators , author =. 1989 , journal =

    1989

  54. [54]

    2018 , journal =

    Flux: Elegant Machine Learning with Julia , author =. 2018 , journal =

    2018

  55. [55]

    2012 , booktitle =

    Soft Decision Trees , author =. 2012 , booktitle =

    2012

  56. [56]

    1985 , journal =

    Numerical parameter identifiability and estimability: Integrating identifiability, estimability, and optimal sampling design , author =. 1985 , journal =

    1985

  57. [57]

    2022 , journal =

    Deep compartment models: a deep learning approach for the reliable prediction of time-series data in pharmacokinetic modeling , author =. 2022 , journal =

    2022

  58. [58]

    2016 , journal =

    Parameter identifiability of fundamental pharmacodynamic models , author =. 2016 , journal =

    2016

  59. [59]

    1995 , journal =

    Three new residual error models for population PK/PD analyses , author =. 1995 , journal =

    1995

  60. [60]

    2015 , journal =

    Nonlinear mixed-effects modelling for single cell estimation: when, why, and how to use it , author =. 2015 , journal =

    2015

  61. [61]

    2024 , journal =

    Longitudinal single-cell data informs deterministic modelling of inflammatory bowel disease , author =. 2024 , journal =

    2024

  62. [62]

    2016 , journal =

    Improved variational inference with inverse autoregressive flow , author =. 2016 , journal =

    2016

  63. [63]

    2013 , journal =

    Profile Likelihood in Systems Biology , author =. 2013 , journal =

    2013

  64. [64]

    2004 , journal =

    Coupling a stochastic approximation version of EMwith an MCMC procedure , author =. 2004 , journal =

    2004

  65. [65]

    2005 , journal =

    Maximum Likelihood Estimation in Nonlinear Mixed Effects Models , author =. 2005 , journal =

    2005

  66. [66]

    and Deloukas, Panos and Daly, Ann K

    Lane, Steven and Al-Zubiedi, Sameh and Hatch, Ellen and Matthews, Ivan and Jorgensen, Andrea L. and Deloukas, Panos and Daly, Ann K. and Park, B. Kevin and Aarons, Leon and Ogungbenro, Kayode and others , year =. The Population Pharmacokinetics of. British Journal of Clinical Pharmacology , publisher =

  67. [67]

    2014 , publisher =

    Mixed effects models for the population approach: models, tasks, methods and tools , author =. 2014 , publisher =

    2014

  68. [68]

    2016 , journal =

    What do we mean by identifiability in mixed effects models? , author =. 2016 , journal =

    2016

  69. [69]

    2023 , journal =

    Improved modeling of crystallization processes by universal differential equations , author =. 2023 , journal =

    2023

  70. [70]

    1990 , journal =

    Nonlinear mixed effects models for repeated measures data , author =. 1990 , journal =

    1990

  71. [71]

    and Nocedal, Jorge , year =

    Liu, Dong C. and Nocedal, Jorge , year =. On the Limited Memory. Mathematical Programming , publisher =

  72. [72]

    2018 , journal =

    A hierarchical, data-driven approach to modeling single-cell populations predicts latent causes of cell-to-cell variability , author =. 2018 , journal =

    2018

  73. [73]

    1982 , journal =

    Finding the observed information matrix when using the EM algorithm , author =. 1982 , journal =

    1982

  74. [74]

    2021 , journal =

    Neural-ODE for pharmacokinetics modeling and its advantage to alternative machine learning models in predicting new dosing regimens , author =. 2021 , journal =

    2021

  75. [75]

    2002 , journal =

    Bayesian analysis of population PK/PD models: general concepts and software , author =. 2002 , journal =

    2002

  76. [76]

    2012 , journal =

    A longitudinal model for magnetic resonance imaging lesion count data in multiple sclerosis patients , author =. 2012 , journal =

    2012

  77. [77]

    1993 , journal =

    Bootstrap and wild bootstrap for high dimensional linear models , author =. 1993 , journal =

    1993

  78. [78]

    2024 , journal =

    Data-Driven Discovery of Feedback Mechanisms in Acute Myeloid Leukaemia: Alternatives to classical models using Deep Nonlinear Mixed Effect modeling and Symbolic Regression , author =. 2024 , journal =

    2024

  79. [79]

    2011 , journal =

    Mixed hidden Markov models for longitudinal data: An overview , author =. 2011 , journal =

    2011

  80. [80]

    1979 , journal =

    A Comparison of Three Methods for Selecting Values of Input Variables in the Analysis of Output from a Computer Code , author =. 1979 , journal =

    1979

Showing first 80 references.