environment.yml

# CellSim Campaign-1 Python environment (non-AI)
#
# Usage:
#   mamba env create -f environment.yml    # or conda
#   conda activate cellsim
#
# Pinned to known-good versions as of the restart. Bump via PR only.
# Every tool below is BUY (integrate) rather than BUILD per the plan.
#
# Per the 2026-04-20 non-AI amendment (see MISSION.md), ML potentials
# (mace-torch, openmm-ml, NequIP, OrbNet, Allegro) and GNINA CNN-
# scored docking as the primary path are explicitly excluded. UQ
# comes from parameter sweeps + Sobol (SALib) + Monte-Carlo and
# MAPIE conformal as a post-hoc non-parametric wrapper.

name: cellsim
channels:
  - conda-forge
  - bioconda
  # `omnia` removed 2026-10: the legacy OpenMM channel supplies none of
  # these packages (everything resolves from conda-forge / bioconda), and
  # a timeout fetching its index failed CI before any test ran.
dependencies:
  - python=3.11
  - pip

  # Cheminformatics
  - rdkit>=2024.03.1
  - openff-toolkit>=0.16
  - openff-forcefields>=2024
  - openff-interchange>=0.3
  - ambertools>=23            # for AM1-BCC, antechamber

  # MD engines (no ML potentials)
  - openmm>=8.1
  - openmmtools>=0.23
  - openmmforcefields>=0.15  # SMIRNOFFTemplateGenerator for the
                             # hybrid AMBER14-protein + SMIRNOFF-
                             # ligand path in Milestone B Phase-2
                             # (pure-SMIRNOFF via Interchange hit
                             # O(N²) memory on mid-size proteins;
                             # see BENCHMARKS.md Known Phase-2
                             # issue)
  - pdbfixer>=1.9
  - mdanalysis>=2.6

  # Quantum
  - xtb-python>=22.1
  - pyscf>=2.5

  # Coarse-grained
  - vermouth                  # martinize2

  # Docking + FEP (non-AI path)
  - vina>=1.2.5               # AutoDock Vina — primary docking engine
  - meeko>=0.7                # PDBQT prep for Vina (ligand + receptor)
  - fpocket>=4.2              # geometric binding-pocket detection
                              # (classical Voronoi/alpha-sphere; non-AI)

  # FEP — choderalab/perses, installed from conda-forge (NOT PyPI;
  # `pip install perses` resolves to a different ML-tracking library
  # by Vincent Primault that has nothing to do with alchemical FEP).
  # On Apple-silicon the conda package may need the omnia channel or
  # a GitHub source install (see docs/perses_install.md when the
  # Layer 1.3 FEP integration lands).
  # - perses                  # choderalab/perses — commented out
                              # pending a cross-platform-verified
                              # install path. openmmtools (above)
                              # provides enough alchemical primitives
                              # for a manual FEP driver in the
                              # meantime.

  # Metabolism (Campaign 2 — NOT pinned here)
  # cobra + python-libsbml were previously listed in anticipation of
  # Campaign 2 pathway modelling, but (a) they have zero callers in
  # Campaign 1 (grep src/ tests/ scripts/ → none) and (b) the pinned
  # python-libsbml>=5.20 had no satisfier on bioconda (tops out at
  # 5.19.x), which blocked fresh env creation on the M5 Max during
  # the Milestone A gate. Dropped from this env; re-introduce in
  # Campaign 2 with version pins verified at that time.

  # UQ + stats (non-AI)
  - salib>=1.4                # Sobol, Morris, FAST sensitivity
                              # (conda-forge pkg is "salib"; Python
                              # import name is "SALib")
  - mapie>=0.8                # post-hoc conformal prediction wrapper
  - numpy>=1.26
  - scipy>=1.11
  - pandas>=2.1
  - scikit-learn>=1.3         # calibration helpers (no neural use)

  # Storage
  - h5py>=3.10
  - pyarrow>=14
  - sqlite>=3.44

  # Pipeline
  - snakemake>=8
  - pyyaml
  - jinja2

  # Dev
  - pytest>=7
  - ruff
  - mypy
  - pre-commit

  - pip:
    # NOTE: "perses" was listed here previously; PyPI's `perses`
    # package is Vincent Primault's ML experiment tracker and is
    # NOT the choderalab/perses FEP engine the code expects. See
    # the Docking + FEP section above for the correct install path.
    - posebusters             # pose validity