deepcell-types documentation#

DeepCell Types is a novel approach to cell phenotyping for spatial proteomics that addresses the challenge of generalization across diverse datasets with varying marker panels.

For details on running the cell-type inference workflow, see the Tutorial.

The source code can be found at the github repo.

For access to pre-trained models and datasets, see Model and Datasets.

Installation#

The development version of deepcell-types can be installed with pip:

pip install git+https://github.com/vanvalenlab/deepcell-types@master

Pre-trained models#

A pre-trained model is required to run the cell-type inference pipeline. Pre-trained models are available via the users.deepcell.org portal according to the license terms. See Model and Datasets for details.

The latest version can be downloaded like so:

from deepcell_types.utils import download_model

download_model()  # No argument == latest released version

Running#

The Tutorial demonstrates how to set up, run, and visualize the outputs of the cell-type inference pipeline.

This tutorial can also be run locally with jupyter.

Start by cloning the sorce repository:

git clone https://github.com/vanvalenlab/deepcell-types.git && cd deepcell-types

Ensure that all necessary extra dependencies are installed in your virtual environment (along with deepcell-types itself):

pip install -r docs/requirements.txt .

Launch a jupyter lab instance:

jupyter lab

Then open the tutorial at docs/site/tutorial.md.

Note

Depending on your jupyterlab version, you may need to right-click the tutorial.md and select Open with -> Jupytext notebook.

Limitations#

  1. Maximum channel limit

    The model currently supports a maximum of 75 channels per dataset. If you have a dataset with more than 75 marker channels, it will be necessary to remove channels to get below this limit. Note that (by default) nuclear and non-recognized channel names are automatically dropped.

  2. Recognized channels

    The model maintains a registry of natively-supported channels at deepcell_types/dct_kit/config/master_channels.yaml. If your data contains markers not found in this listing, they will be ignored at inference time. There are two ways to add support for additional channels:

    • To add a new alias for a marker name that is currently supported, add the alias to deepcell_types/dct_kit/config/channel_mapping.yaml. For example, if your data contains a channel named FP3 representing the FoxP3 marker, add the following line to channel_mapping.yaml:

      FP3: FoxP3

      Note that the target name (FoxP3 in this example) must be one of the names already found in master_channels.yaml

    • Adding new markers to the model can be achieved by manually acquiring embeddings for additional channels via DeepSeek (model and prompt details can be found in the paper).

  3. Image preprocessing

    The model requires users to preprocess input images to align with the distribution of our training data for optimal performance and generalization. See the paper for details.