MARVEL: Microenvironment Annotation by supeRVised Graph ContrastivE Learning

Description

Recent advancements in in situ molecular profiling technologies, including spatial proteomics and transcriptomics, have enabled detailed characterization of the microenvironment at cellular and subcellular levels. While these techniques provide rich information about individual cells' spatial coordinates and expression profiles, how to effectively identify biologically meaningful spatial structures remains a significant challenge. Current methodologies rely on aggregating features from neighboring cells for spatial environment annotation, which is labor-intensive and demands extensive domain knowledge.

To address these challenges, we propose a supervised graph contrastive learning framework: Microenvironment Annotation by supeRVised Graph ContrastivE Learning (MARVEL). This framework combines supervised contrastive learning with graph contrastive learning methods to effectively map local microenvironments, represented by cell neighbor graphs, into a continuous representation space, facilitating various downstream microenvironment annotation scenarios. By leveraging partially annotated samples as strong positives, our approach mitigates the common issues of false positives encountered in conventional graph contrastive learning.

Using real-world annotated data, we demonstrate that MARVEL outperforms existing methods in three key microenvironment-related tasks:

• Transductive microenvironment annotation
• Inductive microenvironment querying
• Identification of novel microenvironments across different tissue slices.

Environment Setup

To set up the environment for MARVEL, please follow the steps below:

1. Clone the repository:

   git clone https://github.com/C0nc/marvel.git
   cd marvel

2. Create a new virtual environment (optional but recommended):

   conda create -n MARVEL
   conda activate MARVEL

3. Install the required dependencies:

   pip install pygcl (required install torch_geometric and dgl (https://github.com/PyGCL/PyGCL/tree/main)
   pip install scanpy squidpy 

Running MARVEL

To run the MARVEL framework, follow the steps below and the CRC_CODEX dataset can be downloaded from "https://drive.google.com/file/d/1w5oB4r7EWSiRQFdeTtPZeHtceOCEtSRy/view?usp=sharing":

1. Run transductive label transfer and meanwhile learn embedding

   python run.py
   python run.py --baseline

2. Run inductive label query:

   python run_inductive.py

3. Run inductive novel ME identification:

   python run_news.py

Citation