Graph Convolutional Networks for Inferring Cell-Cell Communication from Spatial Transcriptomics Data

SPatial Inference of Communication Effects (SPICE)

Installation

SPICE uses poetry as its package manager.

pip install poetry
poetry install
poetry self add poetry-plugin-shell
poetry shell
pre-commit install

Please note that at present, SPICE builds pytorch-geometric (PyG) libraries from whl files which are version dependent. By default we use Python3.10. To make the poetry config compatible for other versions, please replace the whls for PyG and the python version in the pyproject.toml file and then run poetry lock in the commannd line.

You can also force poetry to use your Python3.10 via the command poetry env use [python3.10 PATH HERE].

Tutorial

Here, we briefly demonstrate how to use SPICE on a random subset of 20000 cells from the MERFISH Hypothalamus Dataset.

Configuration

SPICE uses hydra for configuring datasets, optimizers, training, testing, and models. Hydra builds upon OmegaConf to offer CLI overrides, multiruns, syncing multiple configs together, managing outputs, etc.

Each configuration file starts with a defaults list that determines which sub-configs are loaded for key components such as the model, dataset, and optimizer. The global structure follows this layout:

defaults:
  - _self_
  - optimizer:
  - datasets:
  - model:
  - training:
  - predict:

mode:

n_neighbors:

radius:

gpus:

paths:
  root:
  data: ${paths.root}/data
  output: ${paths.root}/output

hydra:
  job:
    env_set:
      CUDA_VISIBLE_DEVICES:
  output_subdir:
  run:
    dir: .

For the tutorial we use the configDemo.yaml. Creating new config files is recommended for new settings with substantial changes.

Non-Response Indexes

At present, SPICE datasets need access to txt files that contain the column indexes of ligand and receptor features. We provide non_response_blank_removed.txt for the MERFISH dataset and non_response_blank_removed_xenium.txt for the Fresh Frozen Mouse Brain Dataset. For this demo, we use the former.

In the event you want to use your own dataset, you would need to create your own class in merfish_dataset.py and populate your ligand and receptor index in a txt to be saved in the spatial/ directory.

Training

spatial mode=train

Training with CLI Hyperparameter Overrides

We can also edit hyperparameters directly from the command line!

spatial model.kwargs.kernel_size=15 optimizer.params.lr=0.0001 radius=10

Multiple Runs

We can also create a multirun (-m) that will sequentially train multiple models with one command call.

spatial -m radius=0,5,10,15,20,25,30

Evaluation

spatial mode=predict

Contributing

Finally, if you are planning to contribute code to this repository, consider installing our pre-commit hooks so that your code commits will be checked locally for compliance with our coding conventions:

pre-commit install