Overview

This project allows the generation and calibration of cellml (and soon to be more) circulatory system models from an array of module/vessel names and connections.

The repository is circulatory_autogen; the package it installs is libcuflynx. They are the same project under two names: papers, issues and this repository say circulatory_autogen, while PyPI, pip install and every import say libcuflynx. Searching for either name should find you this page.

Note: Test results and pass percentage are displayed in the GitHub Actions workflow summary. The badge above shows the overall test status (passing/failing) for master of this repository, which is where pull requests are merged.

Installing, and what each extra costs

The package is libcuflynx. A plain install gives you generation, simulation, calibration, sensitivity analysis and the built-in (emcee) MCMC:

pip install libcuflynx

Everything beyond that is an extra, because the optional parts are not small. These are installed sizes on disk, not wheel sizes — measured by installing into a fresh venv on Linux and taking du of site-packages, so they include the shared libraries that show up as companion directories (scipy.libs, numpy.libs) and every transitive dependency:

[emulation] is the one to think about before typing: autoemulate pulls in torch, which is 734 MB on its own — more than the whole default install — and requires Python >=3.10,<3.13. On Linux, pip's default torch wheel also drags in the bundled NVIDIA CUDA libraries, which can take it past 2 GB; install a CPU-only torch first if you do not want them. Because [all] includes [emulation], [all] inherits that Python range too.

The per-package numbers below do not add up to the total: the long tail of small transitive dependencies (packaging, dateutil, six, typing-extensions, cycler, joblib, threadpoolctl, ...) accounts for the rest.

What makes up the default install: scipy 109 (+27 in scipy.libs), pandas 65, statsmodels 49, scikit-learn 46, numpy 41 (+27 in numpy.libs), matplotlib 28 with fontTools 25 / pillow 21, myokit 13, lxml 12, libcellml 8, kiwisolver 7, nevergrad 5, SALib 5, rdflib 5, libcuflynx itself 4, pint 3, seaborn 3, and emcee / corner / tqdm / numdifftools at about 1 MB each.

The two that are larger than they look are statsmodels (49 MB) and scikit-learn (46 MB). Neither is optional today: statsmodels supplies the Geweke diagnostic run after every emcee MCMC chain, and scikit-learn fits the quadratic used by the Laplace identifiability analysis. Both are on paths a plain install is expected to reach.

[mpi] is an extra, and that is the deliberate part. mpi4py itself is only 5 MB, but it compiles against a system MPI toolchain (libopenmpi-dev, mpich) at install time, and that is the commonest pip install failure on macOS and Windows. A serial calibration needs none of it: libcuflynx.utilities.mpi_utils answers rank/size and supplies the one-rank collectives without ever importing mpi4py. Install the toolchain first, then the extra:

sudo apt install libopenmpi-dev          # or: brew install open-mpi
pip install "libcuflynx[mpi]"

Ask for a model type, solver or analysis whose extra is missing and the error names the extra to install rather than reporting a bare missing module.

Developing on a checkout instead: pip install -e ".[dev]", which adds the test and lint tooling plus mpi4py (tests/conftest.py imports it at module scope, so the suite will not even collect without it) and casadi (whose tests otherwise importorskip themselves into silence).

Quickstart

pip install libcuflynx

No import path setup and no checkout are required — the package is importable, and its commands runnable, from any directory:

from libcuflynx.utilities.utility_funcs import get_default_inp_data_dict
from libcuflynx.scripts.script_generate_with_new_architecture import generate_with_new_architecture
from libcuflynx.solver_wrappers import get_simulation_helper_from_inp_data_dict
from libcuflynx.param_id.paramID import CVS0DParamID

inp = get_default_inp_data_dict(file_prefix, input_param_file, resources_dir)
generate_with_new_architecture(inp_data_dict=inp)     # CSV arrays -> CellML
sim = get_simulation_helper_from_inp_data_dict(inp)   # simulate it
sim.run()
pid = CVS0DParamID.init_from_dict(inp)                # calibrate it

Each pipeline stage also has a console command, all configured from user_run_files/user_inputs.yaml and all taking --help:

CUFLYNX_USER_DIR — where an installed libcuflynx reads and writes

Every stage needs a directory holding user_run_files/user_inputs.yaml, and (unless the config overrides them) resources/, module_config_user/, funcs_user/, generated_models/ and param_id_output/. It is $CUFLYNX_USER_DIR if set; otherwise the circulatory_autogen checkout being run from, if this is one — so a clone or a pip install -e . needs no configuration; otherwise the current working directory.

After a plain pip install libcuflynx there is no checkout, so either run from your working directory or point the variable at it:

export CUFLYNX_USER_DIR=/path/to/my_study
cuflynx-param-id

Nothing is ever written inside the installed package.

Imports without the libcuflynx. prefix are deprecated. from param_id.paramID import CVS0DParamID still works in 0.4.0 and emits a DeprecationWarning; the shims are removed in 0.5.0. See CHANGELOG.md for the migration, including the one for anyone who edited funcs_user/*_funcs_user.py in place.

Tutorial

Follow the instructions in the tutorial to run the project: https://physiomelinks.github.io/circulatory_autogen/

AI-generated interactive tutorial

BETA MODE: This AI-generated tutorial can be used to further understand the code base: https://deepwiki.com/FinbarArgus/circulatory_autogen/1-overview

The CVODE_opencor solver requires OpenCOR

Every solver works from a plain install except CVODE_opencor. That backend needs the opencor Python module, which is supplied by an OpenCOR installation and is not published on PyPI — so it cannot be shipped in a wheel and no pip install can provide it. Asking for it without OpenCOR raises an error naming the alternative rather than a bare ModuleNotFoundError.

• Use solver: CVODE_myokit instead. It is a drop-in replacement: the same CellML model, integrated by CVODE, with no OpenCOR involved. It is what user_inputs.yaml ships with, and nothing else in the project needs OpenCOR.
• Or run inside OpenCOR, whose bundled interpreter provides opencor. That route is deprecated — see Deprecated: OpenCOR-based setup.

This is expected to be replaced by a plain pip install libopencor once libOpenCOR reaches PyPI; the bundled-interpreter route and the scripts that support it will be removed then.

The handful of tests that exercise this backend are marked need_opencor. They have no auto-skip, so without OpenCOR they fail rather than skip — deselect them with -m "not need_opencor", which is what CI does.

Releasing

Releases are cut from a v* tag and published to PyPI by .github/workflows/release.yml using trusted publishing (OIDC) — no API token is stored anywhere. The procedure, the release-notes checklist, and the rules that cannot be undone (the tag must match version in pyproject.toml; a PyPI version can never be re-uploaded) are in CONTRIBUTING.md.

Citing this work

Cite the project as circulatory_autogen — that is the name used by the publications, the repository and any archived DOI. The PyPI distribution libcuflynx is the same software under its packaging name, so a paper citing circulatory_autogen and an environment listing libcuflynx==0.4.0 refer to one artifact, not two. When it helps reproducibility, record both: the citation for the project and the exact package version installed, e.g. "circulatory_autogen (PyPI package libcuflynx, version 0.4.0)".

License

circulatory_autogen is fully open source and distributed under the very permissive Apache License 2.0. See LICENSE for more information.

Optional third-party backends (not part of circulatory_autogen)

circulatory_autogen is complete and fully open source on its own. Every feature works without installing any proprietary software.

Separately, the project ships optional adapters that let users who already hold a licence for certain third-party products plug them in. Those products are not part of circulatory_autogen, are not bundled or installed with it, are not covered by the Apache-2.0 licence above, and are not required by any feature.

• See Optional third-party backends in the tutorial for more info.