SGAE: Deciphering spatial domains from spatially resolved transcriptomics with Siamese Graph Autoencoder

Overview

Spatial transcriptomics (ST) is a newly emerging field that facilitates a comprehensive characterization of tissue organization and architecture. By profiling the spatially-resolved gene expression patterns, ST technologies allow scientists to gain an in-depth understanding of the complex cellular dynamics and within tissue. Graph neural network (GNN) based methods usually suffer from representations collapse, which tends to map spatial spots into same representation. To address this issue, we proposed a Siamese Graph Autoencoder (SGAE) framework to learn discriminative spot representation and decipher accurate spatial domains. SGAE outperformed those spatial clustering methods across multiple platforms derived datasets based on the evaluation of ARI, FMI, NMI. Moreover, the clustering results derived from the SGAE model can be further utilized in 3D Drosophila Embryo reconstruction.

Schematic Diagram

Dependences

Quick Start

Install Dependencies via bash

### Python enviroment constructed by Conda
conda create -n SGAE python=3.8
conda activate SGAE
pip install -r requirements.txt

Run on Code Ocean

We also upload our code to Code Ocean. Please check it for easier compilation.

Running SGAE Script from the Command-line

cd SGAE
python3 run_sgae.py --n_epochs 1000 --name xxx  --data_file xxx.h5ad

Please specify your own data name and data_file via the arguments showed above. You can also check the tutorial below to get a quick start.

Data

We used data from various of platform and samples to benchmark our method. Here is a table for the data mentioned in article.

Dataset	Platform	Samples	Species	Tissue	Source
DLPFC	12	10X Visium	Human	Dorsolateral prefrontal cortex	http://research.libd.org/spatialLIBD
MG	1	seqFISH	Mouse	Gastrulation	https://crukci.shinyapps.io/SpatialMouseAtlas/
MC	1	MERFISH	Mouse	Cortex data	https://doi.brainimagelibrary.org/
MOB	1	SLIDE-seq v2	Mouse	Olfactory bulb	https://singlecell.broadinstitute.org/single_cell/study/SCP815/highly-sensitive-spatial-transcriptomics-at-near-cellular-resolution-with-slide-seqv2#study-summary
DE	16	Stereo-seq	Drosophila	Embryo	https://db.cngb.org/stomics/flysta3d/spatial/
MB	1	Stereo-seq	Mouse	Brain	https://zenodo.org/record/7340795

Codes

The foundation functions of SGAE is deposited at models directory.

Parameter settings

General setting

Parameter	Type	Defination	Default
name	str	name to save result to indicate the data or sample	dblp
modelname	str	name to save result to indicate the model	SGAE
project_dir	str	directory to save result	./
cuda	bool	whether to use GPU	True
gpu_id	str	choose a specific GPU	0
seed	int	determine random seed	1
n_clusters	int	number of clustering	20

Graph setting

Parameter Type Defination Default

k_nn int number of neighbors to construct graph 3

alpha_value float alpha value for graph diffusion 0.2

Parameter	Type	Defination	Default
k_nn	int	number of neighbors to construct graph	3
alpha_value	float	alpha value for graph diffusion	0.2

Training setting

Parameter	Type	Defination	Default
n_epochs	int	total epoch	1000
patience	float	denote the early stopping point	0.2
batch_size	int	the size of a single batch	256
lr	float	learning rate	1e-4
lambda_value	float	weight for clustering guidance loss	10

Tutorial

Reproduce the result of article via run_case.py.

DLPFC:

python3 run_case.py --n_epochs 1000 --dataset dlpfc

seqFISH Mouse Gastrulation:

python3 run_case.py --n_epochs 1000 --dataset seqfish

MERFISH Mouse Cortex data:

python3 run_case.py --n_epochs 1000 --dataset merfish

SLIDE-seq v2 Mouse Olfactory bulb:

python3 run_case.py --n_epochs 1000 --dataset slideseq

Stereo-seq Drosophila Embryo:

python3 run_case.py --n_epochs 1000 --dataset drosophila_14_16
python3 run_case.py --n_epochs 1000 --dataset drosophila_16_18
python3 run_case.py --n_epochs 1000 --dataset drosophila_l1

Stereo-seq Mouse Brain:

python3 run_case.py --n_epochs 1000 --dataset mousebrain

Contact

Any questions or suggestions on EAGS are welcomed! Please report it on issues, or contact Lei Cao (caolei2@genomics.cn) or Shuangsang Fang (fangshuangsang@genomics.cn). We recommend using STOmics Cloud Platform to access and use it.

zhongmicai/STOmics_SGAE