Add Cook_AIChE2022 PEtab problem, fixes #229

[FFroehlich]

Adds the bone-remodeling model of Cook et al. (2022), "Mathematical Modeling of the Effects of Wnt-10b on Bone Metabolism", AIChE Journal 68(12):e17809 (doi:10.1002/aic.17809), as a new benchmark problem.

Source model and code: https://github.com/ashleefv/Wnt10bBoneCompartment

Model

• Based on the published SimBiology/SBML export (single remodeling cycle).
• SimBiology parameter/compartment IDs renamed to readable names; model id/name set to the problem ID; reference DOI added as an is-described-by annotation.
• The multiple remodeling cycles of the original MATLAB implementation (state reset every 100 days: osteocytes reduced by 20, cell populations below 1 set to 0) are encoded as 11 using v2 condition machinery. Bone volume z is continuous across the resets, so the bone-volume observable is unambiguous at the measurement times.

PEtab problem

• 3 conditions for the Wnt-10b fold changes -1, 5 and 50.
• 4 measurements: the Bennett 2005 / 2007 BV/TV data used for the fit in the paper, taken after 4, 6, 6 and 12 remodeling cycles.
• Observable: relative bone volume change (z - 100), unweighted least squares reproduced with fixed unit noise (normal).
• 4 estimated parameters (beta1adj, alpha3adj, beta2adj, K); nominal values are the published final fit.
• Dose-response visualization (BV/TV change vs Wnt-10b fold change).
• Validated with petablint; imports and simulates with AMICI using default solver settings.

Documentation and figures

• README.md documents the model, data, event-based cycle encoding, nominal-parameter source and differences from the publication.
• make_figures.py reproduces three paper/thesis figures via petab.v1.visualize (dose-response, bone volume over cycles, cell dynamics); the rendered figures are included.

Claude-Session: https://claude.ai/code/session_01RJLtEaShERvSBHRvwCEGSA

Checklist for the submission of new PEtab problems

• The PEtab problem is based on a model that is peer-reviewed and published
• The problem ID is in the format {LAST_NAME_OF_FIRST_AUTHOR}_{ABBREVIATED_JOURNAL_NAME}{YEAR_OF_PUBLICATION}
• The problem ID is in the pull request title
• There is a GitHub issue for this problem
  • The problem ID is in the issue title
  • A brief model description (one or two sentences)
  • A brief data description (one or two sentences)
  • The issue and PR are linked to each other
  • Differences between the implementation and the original publication are described
  • Experience of fitting / uncertainty analysis (e.g. optimizer used, hyperparameters, reproducibility of best fit)
  • Source of nominal parameters (e.g.: taken from the original publication, or from your own fitting)
• The SBML file
  • Annotation with reference to the original publication (example)
  • The model ID and model name attributes in the SBML model file match the problem name (example)
• PEtab files
  • A "simulated data" measurement table is included, using the nominal parameters
  • A visualization table is included, that can be used with the simulated data to reproduce figures from the original publication
  • The PEtab problem is valid (check with e.g. petablint -vy problem.yaml)
  • Labels (e.g. parameterName in the parameter table or xLabel in the visualization table) encode special characters with tex-style strings instead of unicode, e.g. use H$_2$O for water.
• The PEtab problem author(s) are assigned to the GitHub issue
• The README has been updated with bmp-create-overview --update (requires pip install -e src/python from the repository root)
  • The new PEtab problem row in the generated table has the correct reference (and other entries)

[FFroehlich]

Reproduction figures (Cook_AIChE2022)

Generated with make_figures.py, which simulates the PEtab v2 problem with AMICI (reusing src/python/simulate.py's create_v2_simulator). All figures are derived from the four dose experiments.

Figure 1 — dose-response

Simulation (o) vs Bennett 2005/2007 fitting data and Roser-Page 2014 validation data (x).

Figure 2 — relative bone volume over remodeling cycles

Sawtooth trajectory for Wnt-10b fold changes -1, 5 and 50, with literature endpoints.

Figure 3 — cell-population dynamics (fold change 50)

Figure 4 — validation vs Roser-Page 2014

Bone volume at fold change 1.8 with the two held-out Roser-Page endpoints.

Figure 5 — cell dynamics over a single cycle

Figure 6 — cell-population AUC ratios vs Wnt-10b

[Copilot]

Pull request overview

Adds a new PEtab v2 benchmark problem (Cook_AIChE2022) for the Cook et al. (2022) bone-remodeling model and updates the repository’s Python tooling/CI to better support PEtab v2 problems (simulation, metadata/overview utilities).

Changes:

• Add the Cook_AIChE2022 benchmark model + PEtab v2 problem definition (SBML, PEtab tables, README, figure reproduction script, and simulated/measurement data).
• Extend src/python tooling to handle both PEtab v1 and v2 when loading problems, simulating with AMICI, and generating overview/metadata.
• Update CI simulation workflow and Python package extras/dependencies to install AMICI and PEtab v2 requirements.

Reviewed changes

Copilot reviewed 17 out of 23 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
src/python/simulate.py	Adds PEtab v2 simulation path via AMICI PetabImporter, and a v1/v2 dispatch simulate() helper.
src/python/pyproject.toml	Adds scipy dependency (for PEtab v2 import chain) and a simulate extra for AMICI.
src/python/benchmark_models_petab/overview.py	Makes overview stats and SBML4Humans links work for PEtab v1 and v2 layouts.
src/python/benchmark_models_petab/check_sbml_metadata.py	Uses a v1/v2 SBML accessor to validate model metadata across formats.
src/python/benchmark_models_petab/base.py	Loads PEtab problems as v1 or v2 based on format_version; keeps simulation TSV loading via v1 helper.
README.md	Adds the Cook_AIChE2022 row to the repository-wide benchmark overview table.
Benchmark-Models/Cook_AIChE2022/README.md	Documents the new benchmark problem, data sources, v2 structure, cycle-reset encoding, and figure reproduction.
Benchmark-Models/Cook_AIChE2022/model_Cook_AIChE2022.xml	Adds the SBML model (no SBML events; intended to be used with PEtab v2 experiment-period resets).
Benchmark-Models/Cook_AIChE2022/Cook_AIChE2022.yaml	Declares the PEtab v2 problem and associated tables/files.
Benchmark-Models/Cook_AIChE2022/conditions_Cook_AIChE2022.tsv	Defines dose conditions and the cycle_reset state-change condition.
Benchmark-Models/Cook_AIChE2022/experiments_Cook_AIChE2022.tsv	Encodes repeated cycles via experiment periods applying cycle_reset at 100-day boundaries.
Benchmark-Models/Cook_AIChE2022/measurementData_Cook_AIChE2022.tsv	Adds fitting + validation measurements.
Benchmark-Models/Cook_AIChE2022/simulatedData_Cook_AIChE2022.tsv	Adds nominal-parameter simulated outputs for the measurements.
Benchmark-Models/Cook_AIChE2022/parameters_Cook_AIChE2022.tsv	Defines estimated parameters, bounds, and nominal values.
Benchmark-Models/Cook_AIChE2022/observables_Cook_AIChE2022.tsv	Defines the bone-volume observable and its noise placeholder wiring.
Benchmark-Models/Cook_AIChE2022/make_figures.py	Reproduces paper/thesis figures by simulating the PEtab v2 problem with AMICI.
.github/workflows/simulate.yml	Installs AMICI via the project’s simulate extra and adds required system deps for building AMICI.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[dilpath]

Are you able to get a similarly-good or better fit from fitting this PEtab problem, as the original paper?

Since this is a bit different to the normal submission, I'll wait for feedback on some of my comments before approving.

[dilpath]

Regarding a few of these package files, it's a bit messy with the combined v1/v2 functions but not sure what else to do. Does it make sense to instead maintain one overview table for v1 problems, and a separate overview table for v2 problems?

@m-philipps @dweindl

[dweindl]

I think we first should decide on how we want to maintain this collection. Have a distinct set of v1 and v2 problems, or have them implemented in both versions? Due to the lack of resources I would just upconvert the old problems, but not backport anything. In that case, having different tables and modularizing the code makes more sense.

[dilpath]

@dweindl do you think these parameters should rather be exp(parameter) in the model, or the log handled elsewhere, in some tool-specific manner? In general, for parameterScale functionality.

[dweindl]

In the original study, no log-scaling was used as far as I understand. It's the question how much close we want to be to the original. Not sure.

[dilpath]

Until now, this kind of information belongs in the GitHub repo issue. Fine to keep here but then add a link to this from the github issue.

[dweindl]

I think it's good to have it in the repository, this way, it will also be included in the zenodo archives.

[dweindl]

I think also the scripts are helpful for better understand what has been done initially. For other problems, this turned out to be very difficult... However, I would clearly separate the PEtab problem from the rest. The Python package should ideally only include the PEtab files, not all the other import/conversion/validation scripts.

[dilpath]

If you move the measurement data out, I would suggest moving the 1 noiseParameters to the training data observable table to simplify the specification of the training problem overall.

I'm also confused why standard deviation is considered for the validation data, but ignored for the training data.

[dilpath]

Should these be parameters?

[dilpath]

This seems like problem-specific information. Does it substantially change any of the computed LLHs?

[dweindl]

I don't think we need those

[dweindl]

I think it's good to have it in the repository, this way, it will also be included in the zenodo archives.

[dweindl]

Isn't that supposed to be [1, 100]? https://github.com/ashleefv/Wnt10bBoneCompartment/blob/5a8a7da42968e713516eda82394ab1512a722bba/BoneCompartmentUp4Para.m#L50

[dweindl]

Ah, there are different configurations below. But the paper says

We assume that the saturation parameter KM could take on a value between 1 and 100, and we bound the parameter fitting accordingly.

[dweindl]

In the original study, no log-scaling was used as far as I understand. It's the question how much close we want to be to the original. Not sure.

[dweindl]

I think we first should decide on how we want to maintain this collection. Have a distinct set of v1 and v2 problems, or have them implemented in both versions? Due to the lack of resources I would just upconvert the old problems, but not backport anything. In that case, having different tables and modularizing the code makes more sense.

[dweindl]

I think also the scripts are helpful for better understand what has been done initially. For other problems, this turned out to be very difficult... However, I would clearly separate the PEtab problem from the rest. The Python package should ideally only include the PEtab files, not all the other import/conversion/validation scripts.