Cell-type Annotation for Single-cell Transcriptomics using Deep Learning with a Weighted Graph Neural Network
Document detailed in https://scdeepsort.readthedocs.io/en/master/index.html
Recent advance in single-cell RNA sequencing (scRNA-seq) has enabled large-scale transcriptional characterization of thousands of cells in multiple complex tissues, in which accurate cell type identification becomes the prerequisite and vital step for scRNA-seq studies.
To addresses this challenge, we developed a pre-trained cell-type annotation method, namely scDeepSort, using a state-of-the-art deep learning algorithm, i.e. a modified graph neural network (GNN) model. In brief, scDeepSort was constructed based on our weighted GNN framework and was then learned in two embedded high-quality scRNA-seq atlases containing 764,741 cells across 88 tissues of human and mouse, which are the most comprehensive multiple-organs scRNA-seq data resources to date. For more information, please refer to https://doi.org/10.1093/nar/gkab775
We provide CPU and CUDA builds, If you want to install scDeepSort with a CPU build, please download scDeepSort-v1.0-cpu.tar.gz
from the release page and execute the following command:
pip install scDeepSort-v1.0-cpu.tar.gz
For more details, see installation guide in the document.
The test single-cell transcriptomics csv data file should be pre-processed by first revising gene symbols according to NCBI Gene database updated on Jan. 10, 2020, wherein unmatched genes and duplicated genes will be removed. Then the data should be normalized with the defalut LogNormalize
method in Seurat
(R package), detailed in pre-process.R
.
Predict using pre-trained models DeepSortPredictor
Train your own model and predict DeepSortClassifier
Please refer to the document of scDeepSort for detailed guidence using scDeepSort as a python package.
Human tissues and cell types
and Mouse tissues and cell types
~anaconda3/deepsort-pretrained/celltype2subtype.xlsx
with the cell types of the scRNA-seq reference when using DeepSortClassifier
to avoid errordev branch
All pre-processed data are available in the form of readily-for-analysis for researchers to develop new methods. Please refer to the release page called Pre-processed data
Shao et al., scDeepSort: a pre-trained cell-type annotation method for single-cell transcriptomics using deep learning with a weighted graph neural network. Nucleic Acids Research, 2021. PMID:34500471
Should you have any questions, please contact Xin Shao at [email protected], Haihong Yang at [email protected], or Xiang Zhuang at [email protected]