mims-harvard/SPATIA_MIST

MIST Dataset Construction For SPATIA (ICML 2026)

MIST (Multimodal Imaging and Spatial Transcriptomics) is the pretraining dataset. It combines cell-level gene expression with morphology image crops from Xenium spatial transcriptomics data. Construction has three stages:

Stage A: Crop cell images into LMDB

Crop cell-centered patches from Xenium morphology TIFF images and store them in an LMDB database.

Input layout (per Xenium dataset):

dataset_dir/
├── morphology_mip.ome.tif    # or morphology.ome.tif / DAPI.tif
├── cells.parquet             # cell centroids + boundaries
└── cell_feature_matrix.h5    # (optional, for building h5ad)

Crop pipeline:

cd gene_encoders/SPATIA-scprint

# Standard Xenium format (cells.parquet + morphology.ome.tif)
python scripts/0510_crop_images_cell_refactored.py \
    --output-lmdb /path/to/output/dataset_name.lmdb \
    --output-size 256 \
    --cache /path/to/cache

# SPATCH format (adata.h5ad + DAPI.tif, for COAD/HCC/OV datasets)
python scripts/0510_crop_images_cell_spatch.py \
    --input-dir /path/to/SPATCH/Xenium-5K \
    --output-lmdb /path/to/output/lmdb \
    --dataset-name HCC

Image processing details:

• TIFF max intensity projection across channels
• Normalize to uint8 (0-255)
• Crop around cell centroid (adaptive size from cell boundaries, or default 32px radius)
• Resize to 256x256
• Store as raw bytes in LMDB with key format: {dataset_name}/{cell_id}
• Coordinate mapping: pixel_x = spatial_y / 0.2125, pixel_y = spatial_x / 0.2125 (Xenium coordinate swap)

Stage B: Build annotated h5ad and register in lamindb

Annotate each dataset with ontology metadata and add to a lamindb Collection:

cd gene_encoders/SPATIA-scprint

python scripts/0512_add_single_dataset.py \
    /path/to/adata.h5ad \
    --tissue lung --disease normal \
    --dataset_name xenium_lung \
    --collection_name xenium_all_0212

Required metadata columns (added automatically by the script):

• organism_ontology_term_id (e.g., NCBITaxon:9606)
• cell_type_ontology_term_id, tissue_ontology_term_id, disease_ontology_term_id
• assay_ontology_term_id, sex_ontology_term_id, development_stage_ontology_term_id
• donor_id, dataset_name, index (cell ID matching LMDB keys)

Stage C: Merge per-dataset LMDBs (optional)

Consolidate multiple per-dataset LMDBs into a single file:

python scripts/0514_merge_lmdb.py \
    --input-dir /path/to/per_dataset_lmdbs/ \
    --output /path/to/merged/all.lmdb

Data loading at training time

The training data loader (scdataloader.data_spatial.Dataset) reads:

1. Gene expression from a lamindb Collection (multiple h5ad files)
2. Cell images from LMDB files (supports multiple scales)

LMDB environments are mapped to image keys in the batch:

• 1st LMDB path -> image (cell-level crop)
• 2nd LMDB path -> region_image (niche-level, optional)
• 3rd LMDB path -> tissue_image (tissue-level, optional)

Images are preprocessed at load time: 256x256 grayscale -> stack to RGB -> AutoImageProcessor (ViT-MAE) -> (3, 224, 224) float tensor.