deepspot-m – skill Tài liệu & Office cho agent AI

DeepSpot-M

Overview

DeepSpot-M is a multimodal foundation model that maps a 224x224 H&E histology tile to

spatial gene expression in log1p-CPM. The output is virtual spatial transcriptomics: one

value per queried gene per tile, laid out on the grid the tiles came from.

A LoRA-adapted pathology foundation backbone (Midnight) tokenises the tile. A

cross-attention gene decoder lets each gene query attend to the patch tokens, and a gene

router hypernetwork builds gene-specific projections from frozen biological embeddings

(Evo 2, Orthrus, ProtT5, scGPT, Apertus). Genes enter the model as queryable embeddings

rather than fixed output slots, so the released model covers a ~19k protein-coding gene

panel including genes unseen in training. The panel ships with the weights as

tokens.csv and is exposed as model.gene_names; genes outside it cannot be queried in

this release.

Applied to TCGA, the model produced a virtual spatial transcriptomics atlas of 28,664

slides across 32 cancer types.

Licensing

The code is PolyForm Noncommercial 1.0.0 and the weights are CC-BY-NC-SA-4.0.

The current model access form additionally restricts eligibility to academic/public

nonprofit researchers without concurrent commercial affiliations and excludes

commercially funded/collaborative use. Read the live declarations before requesting

access; a noncommercial intention alone does not establish eligibility. Research use

only, not clinical or diagnostic use. Preserve the applicable attribution and licenses.

Installation

uv pip install deepspotm==1.0.0

PyPI 1.0.0 declares Python >=3.10 without an upper bound; that is not proof of

compatibility with every future dependency release. Use an isolated environment.

Install the PyTorch build that matches your CUDA runtime if using a GPU.

The optional histolab 0.7.0 tiler requires Python <3.12 and an older scientific stack:

run tiling separately and exchange PNG tiles plus a coordinate manifest.

Model access

The weights are gated:

1. Open <https://huggingface.co/ratschlab/DeepSpotM> and request access.

2. Once access is granted, authenticate the machine that will download them:

hf auth login

The Hugging Face downloader uses cached credentials or HF_TOKEN; approval and

authentication are separate. Set HF_HOME before importing Hub libraries if choosing

a cache location. Do not put tokens in code. An approved local model directory can

be loaded offline.

Quick start

Illustrative gated inference; the loading/query API and local preprocessing were

checked without downloading the weights:

import torch
from PIL import Image
from deepspotm import DeepSpotM

device = "cuda" if torch.cuda.is_available() else "cpu"
revision = "48be27af436a50e5c74175680ac2b7b2596a506b"
model, image_processor = DeepSpotM.from_pretrained(
    "ratschlab/DeepSpotM", source="scgpt", device=device, revision=revision
)

with Image.open("tile.png") as image:
    if image.size != (224, 224):
        raise ValueError("Expected a pre-tiled 224x224 image at verified ~0.5 um/px")
    batch = image_processor(image.convert("RGB")).unsqueeze(0).to(device)
genes = ["EPCAM", "CD3D"]
vals = model.predict_genes(batch, genes).cpu()  # shape (1, 2)

image_processor turns a PIL image into a tensor; unsqueeze(0) adds the batch

dimension, and predict_genes takes the batch plus a

list of HGNC gene symbols. Values come back in log1p-CPM, aligned with the gene list you

passed, so keep that list beside the output to keep the columns labelled. Symbols must be

in the released ~19k-gene panel (model.gene_names); an unknown symbol raises KeyError

naming the offending genes.

Tile requirements

Tiles must be 224x224 RGB at roughly 20x magnification (about 0.5 microns per pixel). Check

the size at the boundary of your pipeline rather than passing an unchecked crop through:

TILE_PX = 224

def require_tile(tile):
    """Return an RGB 224x224 tile, or raise if the crop is the wrong size."""
    if tile.size != (TILE_PX, TILE_PX):
        raise ValueError(
            f"DeepSpot-M expects a {TILE_PX}x{TILE_PX} tile at about 20x "
            f"(~0.5 microns per pixel); got {tile.size[0]}x{tile.size[1]}. "
            "Re-tile at the matching level or resample the crop."
        )
    return tile.convert("RGB")

Pixel size alone does not verify magnification. Read both MPP axes and the actual

level downsample factors; use a native level only if its physical resolution is

close enough to the chosen target. Otherwise explicitly resample from sufficiently

fine source pixels and record the transform. Never infer 20x from a level number.

The returned processor resizes/crops images but cannot establish their physical MPP.

Keep the dependency optional

deepspotm and its weights are a heavy, gated dependency. Import it inside the function

that needs it so the surrounding project installs, imports and tests without it, and turn

an ImportError into a message that names every step:

DEEPSPOTM_HELP = (
    "DeepSpot-M is unavailable. Install it with `uv pip install deepspotm==1.0.0`, request "
    "access to the gated weights at https://huggingface.co/ratschlab/DeepSpotM, then "
    "authenticate with `hf auth login`."
)

def load_deepspotm(source="scgpt"):
    try:
        from deepspotm import DeepSpotM
    except ImportError as exc:
        raise RuntimeError(DEEPSPOTM_HELP) from exc
    return DeepSpotM.from_pretrained("ratschlab/DeepSpotM", source=source)

Embedding sources

source selects which frozen gene embedding the router builds projections from. It is one

of five values:

| source | Gene embedding |

| --------- | --------------------------------- |

| evo2 | genomic sequence |

| orthrus | RNA |

| prott5 | protein sequence |

| scgpt | single-cell expression |

| apertus | language model |

Each gives a different view of gene identity. Pick one per run, and run the same tiles

through more than one source when the choice matters to your analysis. The released

multi-source model can switch using model.set_source(name); no reload is required.

See

references/api.md for the inference call surface, batching and device placement, gene symbol

handling and output units.

Whole slide workflow

Prediction is per tile, so a slide-scale run is a tiling step followed by batched

inference:

1. Verify slide MPP, extract 224x224 tiles on a tissue grid with the histolab skill,

and keep each tile's level-0 bounding box and measured output MPP.

2. Process and stack tiles into batches with torch.stack.

3. Call predict_genes once per batch with the same gene list.

4. Concatenate batches in a recorded tile-ID order; join coordinates by those IDs,

checking uniqueness and missing tiles rather than assuming file/report row order.

Keep outputs labeled as model predictions, not measured transcript counts. Validate

on held-out slides/patients with paired assays for the intended tissue and processing

conditions; tiles from one slide are not independent biological replicates. The

released base model also differs from cancer-specific fine-tuned TCGA atlas models.

That matrix is the virtual spatial transcriptomics map for the slide, and it drops

into AnnData for downstream analysis. [Whole-slide workflow](references/whole_slide.md)

has a worked loop, batch sizing and an AnnData assembly step.

Verification scope

Reviewed the current PyPI 1.0.0 wheel and matching upstream model source on

2026-09-30. CPU checks used Python 3.12.10, Torch 2.7.1, Torchvision 0.22.1,

Transformers 5.18.0, PEFT 0.21.1, Lightning 2.6.6 and Hugging Face Hub 1.33.0.

Synthetic checks cover loading/query contracts, normalization, resolution rejection,

coordinate joins and H5AD round-trip. Gated weights, CUDA, real-slide tiling and

biological prediction accuracy were not tested; full inference examples are illustrative.

Common use cases

• Spatial expression maps for marker genes across a tumour section.
• Transcriptome-wide prediction over a slide cohort with no matching assay run.
• Querying any of the ~19k panel genes by symbol, including genes unseen in training —

far beyond the few hundred genes of a typical spatial assay panel.

• Adding an expression channel to a morphology-only histology pipeline.
• Building a slide-level cohort atlas, as done for TCGA.

Detailed references

• references/api.md: from_pretrained and predict_genes in full, the five embedding

sources and how to choose, batching, device placement, gene symbol handling, and

converting log1p-CPM output.

• references/whole_slide.md: tiling with histolab, a slide-scale prediction loop,

assembling and storing a tiles-by-genes matrix, and cohort-scale runs.

Primary sources

• Paper: <https://doi.org/10.64898/2026.06.19.26356060> (medRxiv, 2026)
• Code: <https://github.com

… (nội dung đầy đủ có trong gói tải về)