scAdapt: Virtual adversarial domain adaptation network for single cell RNA-seq data classification across platforms and species

This repository contains the preprocessed data and Python implementation for scAdapt. scAdapt is a virtual adversarial domain adaptation network to transfer single cell labels between datasets with batch effects. scAdapt used both the labeled source and unlabeled target data to train an enhanced classifier, and aligned the labeled source centroid and pseudo-labeled target centroid to generate a joint embedding.

Requirements

• pyTorch>=1.1.0
• umap-learn>=0.3.8
• universal-divergence>=0.2.0

Data preparation

Log-normalized count matrix is recommonded as the input of scAdapt. Raw counts matrix can be normalized by the NormalizeData function in Seurat with default ‘LogNormalize’ normalization method and a scale factor of 10,000. We use top 2000 highly variable genes as input features. We provided an example of processed datasets in the processed_data folder as follows: combine_expression.csv is the expression matrix [gene, cell] combining source and target. combine_labels.csv is the cell type label array combining source and target. domain_labels.csv is the batch/domain label array combining source and target. digit_label_dict.csv is the one-to-one mapping between digital label and cell type label.

Usage

API usage

Here is example API usage of scAdapt in Python:

from scAdapt import scAdapt
from config import * # set hyper-parameters

dataset_path = args.dataset_path 
normcounts = pd.read_csv(dataset_path + 'combine_expression.csv')
labels = pd.read_csv(dataset_path + 'combine_labels.csv')
domain_labels = pd.read_csv(dataset_path + 'domain_labels.csv')
data_set = {'features': normcounts.T.values, 'labels': labels.iloc[:, 0].values,
'accessions': domain_labels.iloc[:, 0].values}

scAdapt(args=args, data_set=data_set)

Python script usage

An example of how to use scAdapt for both classification and batch correction is:

python main.py --dataset_path path/to/input/files
               --result_path path/for/output/files
               --source_name batch name
               --target_name batch name
	       --gpu_id GPU id to run

The dataset_path must contain the four CSV files preprocessed as in processed_data folder. In result_path, there will be three output files: final_model_*.ckpt has the trained model parameters (i.e. weights and biases) and can be loaded for label prediction. pred_labels_*.csv contains the predicted cell label and corresponding confidence score (softmax probability). embeddings_*.csv contains the batch-corrected low-dimensional embeddings (default is 256) for visualization.

Demo

A demo file example.py is provided in the scAdapt folder, and the corresponding datasets are provided in the processed_data folder. In the demo, we use mouse pancreas data (Baron and Tabula Muris) as source and human pancreas data (Segerstolpe) as target. The demo can be run with the following command with default parameters:

python example.py

We can evaluate how well our predicted cell type labels match the true labels with Confusion heatmap. For this demo dataset, we find that there is a perfect agreement (Average acc = 0.98) in cell type prediction.

We can also plot the low-dimensional embeddings of scAdapt model by UMAP. We can find that the cells from mouse and human dataset are well mixed by cell type while the clusters of different cell types are separated clearly.

Reproducibility

All codes and processed datasets used to reproduce the results in the paper will be published after peer-review.

Questions

For questions about the datasets and code, please contact zhoux85@mail2.sysu.edu.cn.