DeepSpot v1.0.0

First tagged release of DeepSpot, accompanying the manuscript "DeepSpot infers spot and single-cell virtual spatial transcriptomics from routine histology" (Nonchev et al.; preprint: https://doi.org/10.1101/2025.02.09.25321567).

DeepSpot is a deep-set neural network that combines pathology foundation-model features (UNI, Phikon, H-optimus-0) with multi-level spatial tissue context to predict spatial transcriptomics from routine H&E images, at spot level (10x Visium) and single-cell level (10x Xenium).

What is in this release

• deepspot/ – model, training and inference code (PyTorch / PyTorch Lightning), including the single-cell variant and super-resolution prediction
• example_notebook/, example_data/ – end-to-end tutorial
• scripts/ – helper scripts for feature extraction and prediction
• pretrained_model_weights/ – instructions for the pretrained weights

Related resources

• Pretrained weights (Visium and Xenium models): https://zenodo.org/records/15322099
• TCGA virtual spatial transcriptomics resource (3,780 slides, 56 million spots): https://huggingface.co/datasets/nonchev/TCGA_digital_spatial_transcriptomics
• Reproducibility pipeline (Snakemake): https://github.com/ratschlab/he2st

Installation

pip install git+https://github.com/ratschlab/DeepSpot.git@v1.0.0

Requires Python 3.9 or higher; see requirements.txt / environment.yaml.

Citation

Please cite the manuscript (see CITATION.cff) and this archived release (Zenodo DOI).