MHAST: Morphology-guided Hierarchical cell type reAssignment of Spatial Transcriptomics

This repository contains the method and demos for the paper Learned morphological features guide cell type assignment of deconvolved spatial transcriptomics.

Installation

We recommend creating a conda environment for running and testing the method pipeline:

conda env create -n celltyping_env -f environment.yml

To activate the environment:

conda activate celltyping_env

Method

The function containing the hierarchical permutation method is found in celltype_permutation.py. To run it, simply pass:

A: one-hot encoded matrix (N cells x M spots) indicating the belonging of each cell to a spot
B: matrix (N cells x K features) indicating morphological features per cell
X_perm: one-hot encoded matrix (N cells x L types) indicating the initial assigned cell type per cell

to X_global = hierarchical_permutations(A, X_perm, B), where X_global will be the rearranged cell types.

Demos

• simulated_data.ipynb shows how the simulated Visium data was generated
• synthetic_data.ipynb shows how Visium data was synthesized from Xenium data
• real_data.ipynb shows a real use case using the Tangram cell type deconvolution method run_tangram.py

Reference

Chelebian, E., Avenel, C., Leon, J., Hon, C. C., & Wahlby, C. Learned morphological features guide cell type assignment of deconvolved spatial transcriptomics. In Medical Imaging with Deep Learning. https://openreview.net/forum?id=QfYXJUmIit

@inproceedings{chelebian2024learned,
  title={Learned morphological features guide cell type assignment of deconvolved spatial transcriptomics},
  author={Chelebian, Eduard and Avenel, Christophe and Leon, Julio and Hon, Chung-Chau and Wahlby, Carolina},
  booktitle={Medical Imaging with Deep Learning, 2024, Paris, France},
  year={2024},
}