environment.yml

# Conda/mamba environment for embpy with CUDA 12.4 PyTorch.
# -----------------------------------------------------------------------------
# Usage:
#   mamba env create -f environment.yml
#   mamba activate embpy
#
# For a fully reproducible (lockfile-backed) install, prefer pixi.toml instead:
#   pixi install -e gpu && pixi shell -e gpu
# -----------------------------------------------------------------------------
name: embpy
channels:
  - conda-forge
  - bioconda
  - pytorch
  - nvidia
dependencies:
  - python>=3.11,<3.14
  - pip>=24.0

  # Scientific stack -- aligned with pyproject.toml (numpy 2.x).
  - numpy>=1.26,<3
  - pandas>=2.3
  - scipy>=1.11
  - scikit-learn>=1.3
  - matplotlib-base
  - seaborn
  - pyarrow>=15

  # AnnData / scverse
  - anndata
  - scanpy>=1.10

  # PyTorch (GPU -- CUDA 12.4). Comes pre-built from the pytorch channel,
  # so we do NOT install torch via pip afterwards.
  - pytorch::pytorch>=2.5.1
  - pytorch::pytorch-cuda=12.4

  # Binary bioinformatics deps (pip versions are fragile on HPC).
  - conda-forge::biopython
  - conda-forge::pyensembl>=2.3.13
  - conda-forge::rdkit
  - bioconda::pysam>=0.22

  # Pure-Python extras added via pip. We deliberately use a minimal "-e ."
  # so heavy compile-from-source extras (evo, caduceus, boltz, evo2) are
  # opt-in -- add them later with:
  #   pip install ".[helical]"           # single-cell foundation models
  #   pip install ".[pertpy]"            # pertpy metadata
  #   pip install ".[boltz]"             # Boltz-2 (needs CUDA)
  #   pip install ".[caduceus]"          # mamba-ssm (needs CUDA nvcc)
  - pip:
      - broad-babel>=0.1
      - -e .