Domain generalization enables general cancer cell annotation in single-cell and spatial transcriptomics

Abstract

Single-cell and spatial transcriptome sequencing, two recently optimized transcriptome sequencing methods, are increasingly used to study cancer and related diseases. Cell annotation, particularly for malignant cell annotation, is essential and crucial for in-depth analyses in these studies. However, current algorithms lack accuracy and generalization, making it difficult to consistently and rapidly infer malignant cells from pan-cancer data. To address this issue, we present Cancer-Finder, a domain generalization-based deep learning algorithm that can rapidly identify malignant cells/spots in single-cell and spatial transcriptomics data. Additionally, Cancer-Finder integrated an interpretability module, which can utilizes a saliency map to identified important genes related to prognosis and tumor microenvironment.

Set Up Environment

Prerequisites:

System: Ubuntu 18.04
python: 3.9.16
CUDA: 11.6
torch: 1.13.1

You can build the environment with Anaconda:

conda create -n scf python==3.9.16
conda activate scf
pip install -r requirements.txt

Running the Code

Usage and Options - Inference

A input count matrix should be :

tsv, csv and h5ad format are supported.

It can be used for new inference by executing the following command:

python -u infer.py --ckp=checkpoints/sc_pretrain_article.pkl --matrix=sample_data/sample_data_matrix.tsv --out=out.csv

The purpose of the above command is to infer the malignancy status of cells in the expression matrix sample_data/sample_data_matrix.tsv.

The parameter ckp denotes the checkpoints of pre-trained model used for inference. The checkpoints used in article for scRNA-seq data and spatial transcriptomics (ST) data are available for download.

The out.csv file contains examples of the expected output.

This is a sample dataset consisting of 10 cancer cell lines and 10 healthy human peripheral blood cells, and its output should typically appear as follows:

If you wish to perform inference on your own dataset, simply replace sample_data/sample_data_matrix.txt with your own expression matrix.
tsv, csv and h5ad format are supported.

More usage for inference:

python -u infer.py \     
    --ckp=<ckp_file> \   # path for pre-trained model
    --matrix=<data_file> \ # path for data, "tsv", "csv" and  "h5ad" format are supported. 
    --out=<output_file> \ # out path
    --threshold=<threshold> # threshold of inference, default=0.5. Recommended to use 0.5 for scRNA-seq, 10x Visium, legacy ST and slide-seq data. Recommended to use 0.9766 for MERFISH data.

Additionally, a pre-trained model trained with 476,562 cells (148,332 more cells than the article.) from 13 tissues can also be downloaded. Its internal performance is as follows.

And more pre-trained models will be updated in the future.

Usage and Options - Train

If you want train a gene set, please run the command:

python -u train_gene_set.py

If you want to train without a gene set, just to train a pre-trained model for inference. Please run the command:

python -u train.py

By default, the preceding command will use the data within data/train/* as from the training domain and data/val/* as from the validation domain for training.

More usage for training:

python -u train.py  \
    --train_dir=<train_dir> \     # Directory of training data
    --val_dir=<val_dir> \         # Directory of val data
    --batch_size=<batch_size> \   # batch size
    --lr=<learning_rate> \        # learning rate
    --max_epoch=<max_epoch> \     # max epoch
    --output=<output_dir> \       # The output directory of training, including checkpoint and the gene list, and logs files
    --gpu_id=<id>                 # Not necessary, Specify the No. of the gpu if it is available

License

This project is licensed under the MIT License - see the LICENSE file for details.