terra-st 0.1.13

TERRA is a self-supervised foundation model for spatial transcriptomics. It serializes each cell together with its spatial neighbors into a sequence of gene tokens, then trains with a Joint-Embedding Predictive Architecture (JEPA): some tokens are masked and the model predicts their representations in latent space — rather than reconstructing raw expression — to infer the molecular and spatial context of the neighboring cells. This yields hierarchical embeddings at the gene, cell, and neighborhood scales, capturing both a cell's own expression and its tissue microenvironment.

Pretrained on HST-Corpus-112M (>100M cells at single-cell resolution spanning human spatial-transcriptomics datasets), TERRA produces cell- and neighborhood-level embeddings that transfer to downstream tasks such as niche and cell-type identification, batch-integrated atlasing, spatial gene-pair scoring, and in-silico perturbation — without task-specific retraining.

• 📖 Documentation: https://terra-st.readthedocs.io
• 🤗 Pretrained models: https://huggingface.co/Lotfollahi-lab (TERRA-96M, TERRA-112M)
• 📓 Tutorial: end-to-end walkthrough

Key features

• Spatially-aware embeddings — cell and neighborhood representations learned in latent space via JEPA.
• Pretrained and ready to use — download a model from the Hugging Face Hub and embed your own AnnData in a few lines.
• Self-contained model bundles — each release ships the checkpoint, tokenizer, and gene-reference files needed to reproduce its training-time harmonization.
• Downstream analyses — niche/cell-type clustering, gene-pair spatial scoring, EMD-based spatial structure, and perturbation.

Installation

TERRA is published on PyPI as terra-st (the import name is terra) and requires an NVIDIA GPU. Install PyTorch first (so it matches your GPU), then TERRA — we recommend uv.

1. Install PyTorch for your hardware. Run nvidia-smi, read the "CUDA Version" in the top-right, and install the matching CUDA build (see the PyTorch install guide), e.g.:

uv pip install torch --index-url https://download.pytorch.org/whl/cu124

2. Install TERRA.

uv pip install terra-st

Plain pip install terra-st works too. For a development install from a clone of this repository (after step 1): uv pip install -e ".[dev,test,doc]".

Verify the install (PyTorch sees your GPU, and TERRA imports):

python -c "import torch; print(torch.__version__, torch.version.cuda, torch.cuda.is_available())"
python -c "import terra; print(terra.__version__)"

The last value from the first command should be True — TERRA requires a GPU. If it prints False, PyTorch can't see your GPU (usually a CUDA build that doesn't match your driver — revisit step 1).

Quickstart

Download a pretrained model and embed your own spatial AnnData with the end-to-end pipeline. Each downloaded bundle contains the gene-reference files needed for harmonization, so no external paths are required:

import anndata as ad
from terra import download_pretrained, harmonize_tokenize_embed_pipeline

adata = ad.read_h5ad("my_spatial_data.h5ad")   # raw counts in adata.X

model_dir = download_pretrained("Lotfollahi-lab/TERRA-96M")

adata = harmonize_tokenize_embed_pipeline(
    adata=adata,
    sample_key="sample",            # column in adata.obs identifying samples
    batch_key="batch",              # column to store the batch identifier
    model_folder_path=model_dir,
    cache_directory_path="./terra_cache",
)

# Cell- and neighborhood-level embeddings are now in adata.obsm.

See the documentation for the step-by-step pipeline, downstream analyses (niche identification, gene-pair scoring, perturbation), and the full tutorial.

Citation

If you use TERRA in your research, please cite the manuscript (in preparation). A BibTeX entry and DOI will be added here on publication.

License

The TERRA code is released under the BSD 3-Clause License. Pretrained model weights distributed on the Hugging Face Hub are released under CC-BY-NC-4.0 (non-commercial use).