deepspot-m

Load DeepSpot-M with DeepSpotM.from_pretrained("ratschlab/DeepSpotM", source="scgpt"), process a 224x224 PIL tile with the returned image_processor, and call model.predict_genes with a list of HGNC gene symbols. Values come back in log1p-CPM aligned with your gene list.

Tile the slide into 224x224 patches at the level nearest 0.5 microns per pixel using histolab's GridTiler, then run batched predict_genes calls and assemble the tiles-by-genes matrix into an AnnData object with spatial coordinates. Batch size 32 on GPU is a reasonable starting point.

The model covers a panel of roughly 19,000 protein-coding genes queried by HGNC symbol, exposed as model.gene_names. Symbols outside the panel raise a KeyError, so check membership before querying and map aliases like CD45 to current symbols such as PTPRC.

The weights at ratschlab/DeepSpotM are gated, so you must request access on the model page and run huggingface-cli login before downloading. A CUDA GPU speeds up batched inference, but CPU inference works with smaller batches of around 8 tiles.

The source argument selects which frozen gene embedding the router uses: evo2 (genomic sequence), orthrus (RNA), prott5 (protein sequence), scgpt (single-cell expression), or apertus (language model). Pick one per run and keep it fixed across a slide or cohort so values stay comparable.

Tiles must be exactly 224x224 RGB pixels at roughly 20x magnification, about 0.5 microns per pixel. Crops of other sizes or resolutions change the texture the backbone reads, so re-tile at the correct slide level rather than resampling from a coarser level.