pyproject.toml

[build-system]
requires = ["setuptools>=77", "wheel"]
build-backend = "setuptools.build_meta"

[project]
name = "pybnf"
dynamic = ["version"]
description = "An application for parallel fitting of BioNetGen and SBML models using metaheuristics"
readme = { file = "README.md", content-type = "text/markdown" }
requires-python = ">=3.11"
license = "BSD-3-Clause"
license-files = ["LICENSE"]
authors = [
    { name = "Eshan Mitra" },
    { name = "Ryan Suderman" },
    { name = "Alex Ionkov" },
    { name = "Jacob Neumann" },
    { name = "Bill Hlavacek" },
]
keywords = ["BioNetGen", "BNGL", "SBML", "parameter-estimation", "systems-biology"]
classifiers = [
    "Development Status :: 4 - Beta",
    "Intended Audience :: Science/Research",
    "Operating System :: MacOS",
    "Operating System :: POSIX :: Linux",
    "Programming Language :: Python :: 3",
    "Programming Language :: Python :: 3 :: Only",
    "Programming Language :: Python :: 3.11",
    "Programming Language :: Python :: 3.12",
    "Programming Language :: Python :: 3.13",
    "Programming Language :: Python :: 3.14",
    "Topic :: Scientific/Engineering :: Bio-Informatics",
]
dependencies = [
    "bngsim>=0.12.2,<1",
    "dask>=2024.1.0",
    "distributed>=2024.1.0",
    "libroadrunner>=1.6.0",
    "msgpack>=0.6.2",
    "numpy>=1.24",
    "paramiko>=2.7",
    "pydantic>=2,<3",
    "pyparsing>=2.4,<4",
    "scipy>=1.10",
    "tornado>=6.1",
]

[project.optional-dependencies]
antimony = [
    "bngsim[antimony]>=0.12.2,<1",
]
# The PEtab math translator (pybnf.petab.formula, ADR-0035): an expression
# observableFormula is parsed/serialized through petab's sympy-backed math grammar
# (petab.v2.math). OPTIONAL runtime extra -- only the expression import/export path
# imports petab; the bare-name observableFormula path and core stay dependency-free
# (ADR-0019). Stock PyPI petab carries the math grammar, so this is a normal extra a
# user installs with `pip install pybnf[petab]` (the test extra below overrides the
# pin with the dogfooded fork for the BnglModel validation oracle).
petab = [
    "petab>=0.8,<1",
]
# The gradient-based reference sampler (job_type = hmc, ADR-0059): blackjax NUTS on
# the analytical model's JAX log-density, for evaluating PyBNF's gradient-free
# samplers (am/dream/p_dream) against reference-grade NUTS. OPTIONAL runtime extra --
# only the HMC path (pybnf.algorithms.samplers.hmc, plus the priors' logpdf_jax and
# the analytical model's nll_jax) imports jax/blackjax, lazily, with a clear install
# hint; core stays numpy/scipy (ADR-0019). A missing extra raises a pointed
# PybnfError naming `pip install pybnf[jax]`, never a bare ImportError (the house
# pattern, mirroring pybnf[petab]). HMC also needs the analytical-expression backend,
# but the canned gaussian/rotated_gaussian targets used as the closed-form oracle are
# hand-written in jax.numpy and need only this extra.
jax = [
    "jax>=0.4",
    "jaxlib>=0.4",
    "blackjax>=1.2",
]
# The ArviZ InferenceData bridge (pybnf.inference_data, ADR-0055): from_pybnf()
# maps a finished MCMC run's Results/samples.txt onto an arviz.InferenceData so
# PyBNF's posterior output is first-class in the ArviZ/bayesplot/loo ecosystem.
# OPTIONAL runtime extra -- only the bridge imports arviz, lazily, with a clear
# install hint; core stays dependency-free (ADR-0019). h5netcdf gives the netCDF
# engine the config-gated auto-emit of Results/inference_data.nc writes through.
# Uncapped across the arviz major boundary: arviz 1.x is a 2026 rewrite onto
# xarray's DataTree (InferenceData -> DataTree, a changed from_dict signature). The
# bridge supports BOTH lines -- the only API difference is the one from_dict
# construction call (pybnf.inference_data._build_idata branches on its signature);
# every downstream access (.posterior/.sample_stats/.attrs/to_netcdf) is identical.
# So we neither pin to the superseded 0.x nor force 1.x, and pip install
# pybnf[arviz] never downgrades or conflicts with a user's existing arviz.
arviz = [
    "arviz>=0.17",
    "h5netcdf>=1",
    # h5netcdf declares h5py only as an optional extra, but its NETCDF4 write path
    # (to_netcdf -> Results/inference_data.nc) needs it, so require it explicitly.
    "h5py",
]
# Profile-likelihood plots (job_type = profile_likelihood, #467): the driver renders
# Results/profile_likelihood.png (one Delta chi2 panel per parameter) at the end of a run.
# OPTIONAL runtime extra -- matplotlib is imported lazily only in the plotting path, which
# fails soft when it is absent (the profile_*.txt curves + the summary table are written
# regardless); core stays plot-free and dependency-light (ADR-0019).
plot = [
    "matplotlib>=3.5",
]
tests = [
    "pytest>=7,<10",
    "hypothesis>=6,<7",
    # Parallel test execution. CI runs `pytest -m "not slow" -n auto`; the local
    # pre-push hook mirrors CI, so xdist is a declared test dependency (the CI
    # composite action installs it directly too).
    "pytest-xdist",
    # PEtab v2 export oracle (ADR-0025/0026) + the math translator (ADR-0035): the
    # exporter tests validate emitted problems against petab's own checks, and the
    # formula tests use petab.v2.math. Test-only -- core stays dependency-free.
    # Stock PyPI petab, so this extra (and therefore the published wheel's
    # Requires-Dist) carries NO direct URL reference -- PyPI rejects any distribution
    # whose metadata names a `git+` ref. The #420 Step B fork that teaches
    # libpetab-python to load `language: bngl` problems natively (validating the
    # exporter oracle at MODEL level rather than via the register_bngl() monkeypatch)
    # is NOT pinned here: it is injected only on the dedicated CI leg through the
    # setup-pybnf action's `petab-spec` input (see .github/workflows/tests.yml). Once
    # upstream ships native BNGL (PEtab-dev/PEtab#436) even that override goes away.
    # register_bngl() stays a working fallback for stock petab, and the oracle/formula
    # tests `pytest.importorskip('petab')`, so this is not a hard test requirement.
    "petab>=0.8,<1",
    # ArviZ InferenceData bridge oracle (ADR-0055): tests/test_inference_data.py
    # builds real containers and round-trips them through netCDF, so the bridge is
    # exercised in CI rather than skipped. importorskip('arviz') keeps it non-fatal
    # for a bare checkout. Uncapped like the runtime extra (the bridge supports both
    # the 0.x InferenceData and 1.x DataTree lines); CI exercises whichever resolves.
    "arviz>=0.17",
    "h5netcdf>=1",
    # h5netcdf's NETCDF4 write path needs h5py (declared only as an optional extra
    # of h5netcdf); without it the netCDF round-trip tests fail rather than skip.
    "h5py",
    # HMC reference sampler oracle (ADR-0059): tests/test_hmc.py runs blackjax NUTS on
    # the analytical gaussian/rotated_gaussian targets and checks closed-form posterior
    # moment recovery, so the gradient path is exercised in CI rather than skipped. The
    # module guards on jax/blackjax importability (skipif), so a bare checkout still
    # passes. Uncapped like the runtime extra.
    "jax>=0.4",
    "jaxlib>=0.4",
    "blackjax>=1.2",
    # Profile-likelihood plot oracle (#467): tests/test_profile_likelihood.py renders the
    # profile panels and asserts the PNG lands in Results/, so the plotting path is exercised
    # in CI rather than skipped. The plot tests importorskip('matplotlib'), so a bare checkout
    # still passes; this mirrors the optional `plot` runtime extra above.
    "matplotlib>=3.5",
]

[project.scripts]
pybnf = "pybnf.pybnf:main"

[project.urls]
Homepage = "https://github.com/lanl/PyBNF"
Documentation = "https://lanl.github.io/PyBNF/"
Source = "https://github.com/lanl/PyBNF"
Issues = "https://github.com/lanl/PyBNF/issues"

# Auto-discover every subpackage under pybnf/ rather than hand-maintaining the
# list. A hand-maintained list silently dropped pybnf.priors and pybnf.noise
# from the built wheel when those subpackages were added (the source tree still
# imported them, so it only broke an installed wheel, never a source checkout).
# tests/test_packaging_metadata.py guards that every on-disk subpackage is
# matched by this include pattern.
[tool.setuptools.packages.find]
include = ["pybnf*"]

[tool.setuptools.dynamic]
version = { attr = "pybnf.__version__" }

[tool.ruff]
target-version = "py311"
# Pins ruff's default rule set (pycodestyle E4/E7/E9 + pyflakes F) so local runs
# and CI agree. Line length (E501) and the broader style/lint families are
# intentionally left off for v1; revisit once the codebase is reformatted.

[tool.ruff.lint]
# UP032 (use an f-string over str.format) is enforced as a cheap regression
# guard: the package source is f-string-clean after the UP031/UP032 sweep and
# there are zero UP032 violations repo-wide. UP031 (printf % -> f-string) is
# deliberately NOT gated -- ~290 %-format sites remain on purpose (e.g. '%d'
# over divmod floats, which truncate where an f-string would not, plus the test
# suite's established %-style), so enforcing it would force incorrect rewrites.
extend-select = ["UP032"]

# Pre-existing legacy style that PyBNF does not enforce. Each is either purely
# cosmetic or a behavioral change to "fix", so it is out of scope for the lint
# gate (which must stay mechanical). Items tied to dev/PUNCHLIST.md are tracked
# and fixed there, not here.
ignore = [
    "E402",  # module-import-not-at-top: intentional (lazy backend guards, test bootstrap)
    "E712",  # ==True/False: legacy; rewriting can change truthiness semantics (e.g. numpy bools)
    "E721",  # type(x) == T comparisons: legacy, semantically fine in context
    "E722",  # bare except: tracked as ROB-4 in dev/PUNCHLIST.md (behavioral fix)
    "E741",  # ambiguous variable names (l/I/O): legacy
    "F841",  # unused locals: several are CQ-5 dead code (tracked); common bind-then-assert in tests
]

[tool.ruff.lint.per-file-ignores]
"tests/**" = [
    "E702",  # multiple-statements-on-one-line (semicolons): established test-file style
    "F402",  # import shadowed by loop variable
]
# context.py is a re-export shim: test modules do `from .context import data`,
# so its imports are "unused" only within the file itself.
"tests/context.py" = ["F401"]
# Teaching notebooks: compact grouped imports (`import os, tempfile`) and one-line
# plot/setup statements (`plt.tight_layout(); plt.show()`) are idiomatic in a cell
# and read better there than split across lines -- the same E702 stance as tests/**.
# Substantive checks (unused imports F401, empty f-strings F541, undefined names)
# stay enforced on the cells.
"examples/notebooks/*.ipynb" = ["E401", "E702"]

[tool.pytest.ini_options]
minversion = "7.0"
testpaths = ["tests"]
addopts = "-m 'not slow and not recovery'"
markers = [
    "slow: long-running statistical recovery tests; deselected by default, run with `-m slow`",
    "recovery: synthetic-data parameter-recovery fits through the real bngsim backend; opt-in, run with `-m recovery` (needs bngsim + BNG2.pl)",
    "newera: fast new-era (edition-2) validation surface (#436) -- backend-free import/export round-trips (run in default CI) plus a tiny new-era recovery sub-tier (runs by default where bngsim is present, skips otherwise); select with `-m newera`",
    "integration: end-to-end algorithm runs against analytical targets (no simulation backend)",
    "bngsim: requires bngsim installed and PYBNF_NO_BNGSIM unset",
    "bngsim_nfsim: requires bngsim built with NFsim",
    "bngsim_rulemonkey: requires bngsim built with RuleMonkey",
    "bngsim_sbml: requires bngsim built with libsbml",
    "bngsim_antimony: requires bngsim + antimony Python package",
    "roadrunner: requires libroadrunner",
    "bionetgen: requires a working BNG2.pl on PATH",
    "arviz: requires the optional arviz extra (the InferenceData bridge); skips otherwise",
]