Unimodal Embeddings for SpatialFusion

These WDL workflows generate the unimodal embedding inputs used by SpatialFusion:

• GenerateUnimodalEmbeddingsForSpatialFusion generates both scGPT.parquet and UNI.parquet from one Terra submission.
• GenerateScgptEmbeddingsForSpatialFusion generates scGPT.parquet from spatial transcriptomics data.
• GenerateUniEmbeddingsForSpatialFusion generates UNI.parquet from H&E / whole-slide imaging data.

Use the combined workflow if you want both outputs from one Terra submission. Use the modality-specific workflows for scGPT-only or UNI-only runs.

Combined Inputs

• adata: AnnData (.h5ad) file used for scGPT embeddings and for the spatial coordinates consumed by UNI. Spatial coordinates are expected in adata.obsm["spatial"].
• input_is_log_normalized: whether the AnnData expression values in the selected layer are already log-normalized.
• wsi: H&E / whole-slide image in TIFF / OME-TIFF format.
• uni_weights: UNI model weights file (pytorch_model.bin) from Mahmood Lab.

scGPT Inputs

• adata: AnnData (.h5ad) file used for scGPT embeddings.
• input_is_log_normalized: whether the AnnData expression values in the selected layer are already log-normalized.

scGPT weights are bundled in the workflow Docker image at /app/scgpt_weights, so users do not need to provide a scGPT weights input.

UNI Inputs

• adata: AnnData (.h5ad) file with spatial coordinates stored in adata.obsm["spatial"].
• wsi: H&E / whole-slide image in TIFF / OME-TIFF format.
• uni_weights: UNI model weights file (pytorch_model.bin) from Mahmood Lab.

UNI weights are not bundled with the workflow. Users must request access to the UNI2-h weights from Mahmood Lab at https://huggingface.co/MahmoodLab/UNI2-h, upload the weights file to an accessible gs:// bucket, and provide that path as the uni_weights input.