Installation

Requirements: Python 3.10, CUDA 12.1+, GPU with 40 GB+ VRAM (A100/H100 recommended).

Conda Environment

The simplest way is to create the environment from the provided environment.yml:

conda env create -f environment.yml
conda activate spatia

Or build manually:

conda create -n spatia python=3.10
conda activate spatia

pip install torch==2.4.0 torchvision==0.19.0 torchaudio==2.4.0 \
    --index-url https://download.pytorch.org/whl/cu121

Package Installation

Install in this order — data_processing must come before SPATIA-scprint.

# Flash Attention (required for the scPRINT transformer)
pip install flash-attn --no-build-isolation

# Data processing package
pip install -e data_processing

# SPATIA-scprint representation model
pip install -e gene_encoders/SPATIA-scprint

Generation Dependencies

For Stage 3 (flow-matching image generation), install extra packages:

pip install torchdiffeq flow-matching torch-fidelity

Notes

numba cache — On shared filesystems with read-only site-packages, export a writable cache before running anything that imports scanpy:

export NUMBA_CACHE_DIR=/tmp/numba_cache_$USER && mkdir -p $NUMBA_CACHE_DIR

Flash Attention — Requires CUDA 11.6+ and GLIBC 2.17+. If the build fails, the SPATIA-scgpt path has a built-in fallback to standard nn.MultiheadAttention.

lamindb (pretraining only) — Required for building the MIST pretraining dataset. Not needed for downstream evaluation.

pip install "lamindb[bionty]==0.76.12"
lamin init --storage ./data --name scprint_db
python -c "import bionty as bt; bt.base.reset_sources()"