GreS: Resources for Semantic-Guided Spatial Domain Identification

This repository hosts the resources needed to run GreS, a graph-based framework that incorporates gene-level semantic priors into spatial representation learning for spatial domain identification.

It contains two parts:

1. embeddings/ — pretrained gene embeddings and vocabulary used to build per-spot semantic descriptors.
2. DLPFC/ — example 10x Visium spatial transcriptomics data (human dorsolateral prefrontal cortex).

Contents

ylu99/Gres/
├── embeddings/
│   ├── pretrained_gene_embeddings.pt   # pretrained gene embedding matrix
│   └── vocab.json                      # gene -> index vocabulary
└── DLPFC/
    ├── 151507/
    │   ├── metadata.tsv                # per-spot annotations (incl. layer labels)
    │   ├── 151507_truth.txt            # ground-truth layer labels
    │   └── spatial/                    # tissue positions + scale factors
    ├── 151508/
    └── ...                             # 12 samples in total

embeddings/

• pretrained_gene_embeddings.pt: pretrained semantic embeddings for genes, aggregated to the spot level (weighted by expression) to form each spot's semantic descriptor.
• vocab.json: maps gene symbols to embedding indices so genes can be aligned to the embedding matrix.

DLPFC/

The DLPFC dataset comprises 12 tissue sections (151507–151510, 151669–151676) from the human dorsolateral prefrontal cortex, a widely used benchmark for spatial domain identification with manually annotated cortical layers (layers 1–6 and white matter).

Each sample folder contains:

• metadata.tsv: per-spot metadata, including the ground-truth layer annotation (layer_guess_reordered).
• <id>_truth.txt: ground-truth labels.
• spatial/: tissue_positions_list.csv and scalefactors_json.json for spatial coordinates.

Note: the raw gene-expression matrices (filtered_feature_bc_matrix.h5) and full-resolution tissue images are not included here due to size. Obtain them from the original spatialLIBD / 10x Genomics release and place them next to the provided files.

Usage with GreS

Download the resources and place them under the GreS project tree:

# Gene embeddings  ->  GreS/embedding/text_embedd_large/
huggingface-cli download ylu99/Gres --repo-type dataset \
  --include "embeddings/*" --local-dir ./gres_resources

# Example DLPFC data  ->  GreS/data/raw_h5ad/<dataset_id>/
huggingface-cli download ylu99/Gres --repo-type dataset \
  --include "DLPFC/*" --local-dir ./gres_resources

Then follow the three-step pipeline in the GreS repository:

1. preprocess/generate_data.py — build the graph-augmented data.h5ad.
2. preprocess/generate_raw_gene_concat_spot_embedding.py — build per-spot semantic embeddings.
3. tools/train.py — train and cluster.

See the project tutorial.ipynb for an end-to-end walkthrough.

License

Released under the MIT License. The DLPFC data originates from the spatialLIBD project; please also cite the original data source when using it.