STELLAR - Spatially Resolved Single-Cell Data Analysis

STELLAR (SpaTial cELl LeARning) is a geometric deep learning tool for cell-type discovery and identification in spatially resolved single-cell datasets. Developed at Stanford University, it uses graph convolutional neural networks to analyze spatial relationships between cells in tissue samples.

Overview

STELLAR addresses the challenge of annotating cells in spatial single-cell datasets by combining molecular features with spatial organization. The method takes two inputs:

1. A reference dataset of annotated spatially resolved single-cell data
2. An unannotated dataset with unknown cell types

Using these inputs, STELLAR:

• Learns low-dimensional cell embeddings using graph convolutional networks
• Assigns cells to known cell types from the reference dataset
• Identifies novel cell types not present in the reference dataset
• Preserves spatial information about cell neighborhoods

Applications

STELLAR has been successfully applied to:

• CODEX multiplexed imaging data: Used for analysis of Barrett’s esophagus and tonsil tissue
• HuBMAP datasets: Applied to healthy intestine tissues across 8 donors, 64 tissues, and 2.6 million cells
• Higher-order tissue structure analysis: Capturing multicellular structural features within tissues

Implementation

STELLAR is implemented in PyTorch and leverages PyTorch Geometric for graph neural network operations. The codebase includes:

• Core STELLAR algorithm for cell annotation
• Demo notebooks for example usage
• Utilities for dataset processing
• Pre-trained models

Citation

When using STELLAR, please cite:

@article{stellar2022,
  title={Annotation of spatially resolved single-cell data with STELLAR},
  author={Brbić, Maria and Cao, Kaidi and Hickey, John W and Tan, Yuqi and Snyder, Michael P and Nolan, Garry P and Leskovec, Jure},
  journal={Nature Methods},
  volume={19},
  number={11},
  pages={1411--1418},
  year={2022},
  publisher={Nature Publishing Group}
}