pyphysicell: PhysiCell models in Python, with C++ only as computation kernels

Version: 0.1.2

Release date: 22 October 2026

Based on: PhysiCell 1.14.2 (and the BioFVM vendored in it). This is the version every algorithm, default value and parity test here was written against. PhysiCell changes its defaults and its rules grammar between releases, so a model built for a different version may read differently -- check pyphysicell.PHYSICELL_VERSION before trusting a comparison.

Developed and maintained by: Daniel Mejia.

Supported by: the Network for Computational Nanotechnology and nanoHUB, Purdue University.

Reference:

(a) This repository.

(b) A Ghaffarizadeh, R Heiland, SH Friedman, SM Mumenthaler, and P Macklin.
    "PhysiCell: an Open Source Physics-Based Cell Simulator for Multicellular
    Systems." PLoS Comput. Biol. 14(2): e1005991, 2018.
    DOI: 10.1371/journal.pcbi.1005991

(c) A Ghaffarizadeh, SH Friedman, and P Macklin. "BioFVM: an efficient
    parallelized diffusive transport solver for 3-D biological simulations."
    Bioinformatics 32(8): 1256-8, 2016. DOI: 10.1093/bioinformatics/btv730

pyphysicell is an independent Python reimplementation of PhysiCell, the open-source agent-based simulator for multicellular systems. PhysiCell is a C++ application: a model is an XML file, a run is an executable, and a result is a folder of files. That is a good fit for a cluster and a poor fit for a notebook, an interactive session or an agent. Here the model is a Python object, the simulation stays alive in the process, and C++ is kept for the inner loops only -- diffusion, mechanics, forces -- reached through ctypes over NumPy buffers. No project is compiled, no executable is built, no output folder is parsed.

This is not PhysiCell and not an official PhysiCell product. It tracks one specific release, PhysiCell 1.14.2: its microenvironment matches that version's BioFVM to 4.6e-14 and its model parsing matches that version on 5,826 values across 17 projects, but its agent-side trajectories have never been compared against a compiled PhysiCell run, of any version. Nothing here is guaranteed against 1.13 or earlier, or against a later release. See What is verified below, and cite the PhysiCell papers, not this.

import pyphysicell as pc

model = pc.load("PhysiCell/sample_projects/rules_sample")   # their project, unmodified
print(pc.check(model))                                      # dimensional analysis, milliseconds
sim = pc.Simulation(model, seed=0)
sim.run(until=720, record_every=30)                         # population curve recorded *during* the run
sim.microenvironment.density[..., 0] = 38.0                 # intervene; the state is yours
sim.run(until=900)

No C++ project is compiled, no executable is built, no output folder is parsed.

What the library adds over an XML and an executable

	PhysiCell workflow	here
ask what a model implies before running	read the XML	pc.check(model), cycle.mean_duration(), rule response curves -- milliseconds
change something mid-run	new XML, run again from t=0	edit sim.cells / sim.microenvironment, call sim.run() again
two futures from one state	two runs from the beginning	sim.branch() -- a deepcopy, because the state is Python
save and resume	—	sim.save(path) / pc.Simulation.load(path)
logic outside the rules grammar	custom C++, recompile	@sim.hook("phenotype", cell_type=...), arrays, once per step
stop on a condition	run to max_time, search the output	sim.run(..., stop_when=lambda s: ...)
a parameter sweep	one subprocess and one output folder per point	pc.sweep(model, change, values, replicates=n) in one process
two simulations at once	impossible (process globals)	two objects

Each is covered by a test in tests/test_capabilities.py and demonstrated on every sample project in the notebooks.

Notebooks

Every PhysiCell sample project is a self-contained notebook in notebooks/: the model is Python, generated from the project's XML and rules CSV by pyphysicell.codegen, so a notebook needs no PhysiCell checkout at all. A test asserts that the generated code rebuilds a model with the same content hash as the XML it came from, for all 18 sample projects.

Initial conditions are the one thing not inlined: they ship as CSV in notebooks/data/, read with pc.read_cells_csv, so a notebook needs only itself, its data file and the package. (Inlining 1,300 cells cost rules_sample 124 KB of literal tuples; the file is 48 KB now.) A test walks every notebook and fails if it references a data file that is not shipped.

model = pc.load("PhysiCell/sample_projects/rules_sample")   # read a project
print(pc.codegen.to_python(model))                          # write it out as Python

Every project notebook has the same eight sections: the model as an object, answers before simulating, a run, steering it mid-run, two futures from one state, a hook doing something the rules grammar cannot express, an in-process sweep with replicates, and a stopping condition. Sections 3-8 are the ones that are impossible, or mean editing C++, in the file workflow.

Architecture

Python          model (from PhysiCell XML + rules CSV) · state (NumPy, struct-of-arrays)
                time loop · rules engine · cycle/volume/death · analysis · plots
   |  flat arrays, no hidden state
C++ kernels     diffusion (LOD/Thomas) · gradients · secretion-uptake · cell list ·
                mechanics forces · Adams-Bashforth positions
                OpenMP when available, std::thread otherwise
                (python/pyphysicell/kernels.cpp, 330 lines)

Everything PhysiCell keeps in process globals (microenvironment, all_cells, cell_defaults, …) lives in ordinary Python objects here, so several models can exist in one interpreter, state is picklable, and nothing is hidden between steps.

What is verified

Check	Result
Diffusion vs BioFVM itself (2-D, 400 voxels, 2 substrates, Dirichlet, 100 steps)	max relative difference 4.6e-14
Diffusion vs BioFVM itself (3-D, 8000 voxels, 50 steps)	max relative difference 3.5e-14
Pure decay vs analytic exponential	2.5e-4 (the LOD splitting error, same as PhysiCell's)
Mass conservation, no decay, no-flux	4.6e-14
Mechanics force law vs Cell::add_potentials, 20 random separations	exact to 1e-13
Position update vs Cell::update_position (Adams–Bashforth)	exact to 1e-15
Secretion/uptake vs the BioFVM semi-implicit solver	exact to 1e-15
Hill rule response vs multivariate_Hill_response_function	exact

tests/biofvm_reference.cpp builds against the vendored PhysiCell checkout and is the oracle for the first two rows: BioFVM's own solver, same case, binary comparison.

tests/test_regressions.py pins the defects found in review: the cycle ring must close back to phase 0, cells must keep their size across divisions (division halves the volume and the targets, the growth phase doubles the targets again), freed cell slots must be cleared before reuse, internalized substrates are split at division, apply_to_dead decides whether a rule touches a corpse, chained run() calls must not duplicate a recorded time point, and <chemotaxis> must reach the motility bias.

.venv/bin/python -m pytest tests -q        # 16 passed

What it does

• reads a PhysiCell project: PhysiCell_settings.xml, the CBHG rules CSV, cells.csv
• substrates: diffusion, decay, Dirichlet faces, gradients, cell secretion/uptake/export
• cells: volume ODE, cycle (any ring of phases: live, flow-cytometry, Ki67…), apoptosis and necrosis with swelling and lysis, mechanics, motility with chemotaxis, division with PhysiCell's daughter placement
• rules: the CBHG grammar, vectorized over the population (one NumPy pass per cell type per phenotype step, not one Python call per cell)
• check(): diffusion lengths against dx, motility step against the mesh, cycle durations in closed form from the phase graph
• inspect(): what fraction of a given project this can run, naming every rule it skips
• codegen.to_python(): a model written back out as a readable, self-contained script
• initial.auto(): ports a project's C++ setup_tissue placement from its own number_of_* and *_radius parameters, and says so in words
• cell definitions inherit like PhysiCell's: parent_type, else the default definition, else the values from PhysiCell's own constructors

What it does not do (yet)

Measured, not guessed: tools/gaps.py reads every sample project and reports what stands between it and its real model.

Every rule in every sample project now runs -- phagocytosis, attack and damage, transformations, fusion, asymmetric division, contact with <type>, custom:<name>. rules_sample went from 14 of 27 rules to 27 of 27.

What is left is C++ that projects write themselves. Five of the eight Cell_Functions slots have hook equivalents, covering 54 of the 59 real (non-stub) assignments in the sample projects -- the logic still has to be written in Python, but it no longer needs a compiler:

slot	assignments across projects	here
update_phenotype	23	@sim.hook("phenotype")
custom_cell_rule	11	@sim.hook("custom_rule")
update_migration_bias	11	@sim.hook("migration_bias")
contact_function	8	@sim.hook("contact") over touching pairs
cell_division_function	1	@sim.hook("division") over parent/daughter pairs
instantiate_cell	3	no -- needs a Cell subclass (PhysiMeSS fibres)
update_velocity	1	no -- custom mechanics (PhysiMeSS)
pre/post_update_intracellular	2	no -- PhysiBoSS, libRoadrunner, dFBA

Also still missing, and known:

• attachments and springs, so contact hooks run over neighbours rather than attached pairs, and cell attachment rate / cell detachment rate do nothing
• basement-membrane functions, is movable as a behavior
• intracellular models of any kind (the cell-level model runs; its inner model does not)
• MultiCellDS output and SVG export (results stay as arrays and dataframes)

Layout

python/pyphysicell/
  kernels.cpp                   the C++ computation kernels (C ABI, OpenMP/std::thread)
  _kernels.py                   builds and binds them on first import (ctypes, no build step)
  model.py  io_xml.py           the model object and the PhysiCell file readers
  state.py                      Microenvironment and Cells (struct-of-arrays)
  phenotype.py  rules.py        volume/cycle/death, and the CBHG engine
  signals.py                    the signal/behavior dictionary
  simulation.py                 the time loop and the tier-1 scalars
  check.py  viz.py              model checks and plots
tests/                          unit tests + the BioFVM golden test
notebooks/01_rules_sample.ipynb the demo

Requires Python ≥ 3.10, NumPy, a C++17 compiler; matplotlib and pandas for the notebook.

Credit

This library is an independent reimplementation. It contains no PhysiCell source: the algorithms were reimplemented in Python and C++ from the published papers and the PhysiCell 1.14.2 source, and several sample projects' custom C++ modules were ported to Python (python/pyphysicell/ports/). Everything it can do, it can do because the people below did the work first. If you use it, cite their papers — not this package.

PhysiCell and BioFVM — the simulator this reimplements

• PhysiCell — Ahmadreza Ghaffarizadeh, Randy Heiland, Samuel H. Friedman, Shannon M. Mumenthaler, Paul Macklin. PhysiCell: an open source physics-based cell simulator for 3-D multicellular systems. PLoS Comput. Biol. 14(2): e1005991 (2018). doi:10.1371/journal.pcbi.1005991
• BioFVM — Ahmadreza Ghaffarizadeh, Samuel H. Friedman, Paul Macklin. BioFVM: an efficient, parallelized diffusive transport solver for 3-D biological simulations. Bioinformatics 32(8): 1256–8 (2016). doi:10.1093/bioinformatics/btv730
• Copyright (c) 2015–2024, Paul Macklin and the PhysiCell Project, BSD-3-Clause. PhysiCell is developed by Paul Macklin's lab and a wider community of contributors: github.com/MathCancer/PhysiCell

PhysiCell Studio — the source of this library's default values

• R. Heiland, D. Bergman, B. Lyons, G. Waldow, J. Cass, H. B. Mohammad Nia, M. Ruscone, V. Noël, A. Shutt, P. Macklin. PhysiCell Studio: a graphical tool to make agent-based modeling more accessible. Gigabyte (2024).

Addons

• PhysiMeSS (ECM fibres — reimplemented here as kernels) — V. Noël, M. Ruscone, R. Shuttleworth, C. K. Macnamara. PhysiMeSS — a new PhysiCell addon for extracellular matrix modelling. bioRxiv 2023.10.27.564365 (2023). doi:10.1101/2023.10.27.564365
• PhysiBoSS (not implemented here; the cell-level model of PhysiBoSS projects runs, its Boolean layer does not) — Gaelle Letort, Arnau Montagud, Gautier Stoll, Randy Heiland, Emmanuel Barillot, Paul Macklin, Andrei Zinovyev, Laurence Calzone. Bioinformatics 35(7): 1188–96 (2019). doi:10.1093/bioinformatics/bty766 — and PhysiBoSS 2.0: Miguel Ponce-de-Leon, Arnau Montagud, Vincent Noël, Annika Meert, Gerard Pradas, Emmanuel Barillot, Laurence Calzone, Alfonso Valencia. npj Syst. Biol. Appl. 9(1): 54 (2023). doi:10.1038/s41540-023-00314-4
• libRoadRunner (not implemented here) — Endre T. Somogyi, Jean-Marie Bouteiller, James A. Glazier, Matthias König, J. Kyle Medley, Maciej H. Swat, Herbert M. Sauro. Bioinformatics 31(20): 3315–21 (2015). doi:10.1093/bioinformatics/btv363
• MaBoSS — the Boolean engine behind PhysiBoSS, by the Institut Curie Computational Systems Biology of Cancer group. Not vendored or used here.

The cell behavior hypothesis grammar, and the published models replicated here

The rules engine implements the grammar from, and the published/ corpus replicates the models of:

• Johnson JAI, Bergman DR, Rocha HL, Zhou DL, Cramer E, Mclean IC, Dance YW, Booth M, Nicholas Z, Lopez-Vidal T, Deshpande A, Heiland R, Bucher E, Shojaeian F, Dunworth M, Forjaz A, Getz M, Godet I, Kurtoglu F, Lyman M, Metzcar J, Mitchell JT, Raddatz A, Solorzano J, Sundus A, Wang Y, DeNardo DG, Ewald AJ, Gilkes DM, Kagohara LT, Kiemen AL, Thompson ED, Wirtz D, Wood LD, Wu PH, Zaidi N, Zheng L, Zimmerman JW, Phillip JM, Jaffee EM, Gray JW, Coussens LM, Chang YH, Heiser LM, Stein-O'Brien GL, Fertig EJ, Macklin P. Human interpretable grammar encodes multicellular systems biology models to democratize virtual cell laboratories. Cell (2025). doi:10.1016/j.cell.2025.06.048 — preprint: Digitize your Biology! Modeling multicellular systems through interpretable cell behavior, doi:10.1101/2023.09.17.557982
• collective-invasion / cell-ECM framework — Metzcar J, Duggan BS, Fischer B, Murphy M, Heiland R, Macklin P. A simple framework for agent-based modeling with extracellular matrix. Bull. Math. Biol. 87(3): 43 (2025). doi:10.1007/s11538-024-01408-8
• breastcancer-hypoxia — Heber Lima da Rocha, with the Macklin and Gilkes labs; supported by the JKTG Foundation, BCRF, the IU Engineered nanoBIO Node and NCI U01CA232137.
• pc3types, three_types — Paul Macklin.
• The sample projects ported in python/pyphysicell/ports/ are the work of the PhysiCell project and its contributors.

Related Python work in this ecosystem, which this project learned from

• pcdl / physicelldataloader — Elmar Bucher et al.
• UQ-PhysiCell — Heber L. Rocha et al.
• PhysiCOOL — I. G. Gonçalves et al.
• PhysiCell-X (MPI) — Gaurav Saxena, Miguel Ponce-de-Leon, Arnau Montagud, David Vicente Dorca, Alfonso Valencia.

License

BSD 3-Clause, the same license as PhysiCell — see LICENSE. The PhysiCell copyright notice is retained there, as condition 1 of that license requires for the ported modules and the reimplemented algorithms.

Per condition 3 of that license, neither the PhysiCell name nor the names of its contributors are used to endorse or promote this package. This is not an official PhysiCell product and the PhysiCell authors are not responsible for it; please do not send them bug reports about it.

Acknowledgements

pyphysicell is developed and maintained by Daniel Mejia, with support from the Network for Computational Nanotechnology and nanoHUB, Purdue University.

The PhysiCell project itself has been supported by the National Science Foundation, the National Cancer Institute, the Breast Cancer Research Foundation and the Jayne Koskinas Ted Giovanis Foundation, among others; see the upstream repositories for their own acknowledgements.