scDeepCluster_pytorch

The pytorch version of scDeepCluster, a model-based deep embedding clustering for Single Cell RNA-seq data.

Comparing to the original Keras version, I introduced two new features:

1. The Louvain clustering is implemented after pretraining to allow estimating number of clusters.
   
2. A new script "scDeepClusterBatch" uses conditional autoencoder technic to integrate single-cell data from different batches.
   

**updates:

11/27/2023: I updated the model to use float64 precision.

Table of contents

• Network diagram
• Requirements
• Usage
• Parameters
• Outputs
• Reference
• Online app
• Contact

Network diagram

Requirements

Scanpy -- 1.7 (https://scanpy.readthedocs.io/en/stable/)

Pytorch -- 1.8 (https://pytorch.org)

Usage

For single-cell count data:

python run_scDeepCluster.py --data_file data.h5 --n_clusters 0

Set data_file to the destination to the data (stored in h5 format, with two components X and Y, where X is the cell by gene count matrix, and Y is the true labels. Y is optional), n_clusters to the number of clusters (0 for automatically estimating by the Louvain algorithm on the pretrained latent features).

For single-cell count data from multiple batches:

python run_scDeepClusterBatch.py --data_file data.h5 --n_clusters 0

This is the script for clustering analysis of datasets from different batches (stored in h5 format, with three components X, B and Y, where X is the cell by gene count matrix, B is the one-hot encoded batch IDs, and Y is the true labels. Y is optional). Following the idea from scVI paper (https://doi.org/10.1038/s41592-018-0229-2), we use the conditional autoencoder (https://papers.nips.cc/paper_files/paper/2015/hash/8d55a249e6baa5c06772297520da2051-Abstract.html) technic to integrate different batches. n_clusters to the number of clusters (0 for automatically estimating by the Louvain algorithm on the pretrained latent features).

Parameters

--n_clusters: number of clusters, if setting as 0, it will be estimated by the Louvain alogrithm on the latent features after pretraining. If setting as an integer > 0, then the model will use the user defined value as number of clusters.
--knn: number of nearest neighbors, which is used in the Louvain algorithm, default = 20. Not used when setting n_clusters > 0
--resolution: resolution in the Louvain algorith, default = 0.8. Larger value will result to more cluster numbers. Not used when setting n_clusters > 0.
--select_genes: number of selected genes for the analysis, default = 0. It will use the mean-variance relationship to select informative genes. Recommending to select top 2000 genes, but it depends on different datasets.
--batch_size: batch size, default = 256.
--data_file: file name of data.
--maxiter: max number of iterations in the clustering stage, default = 2000.
--pretrain_epochs: pretraining iterations, default = 300.
--gamma: coefficient of the clustering loss, default = 1.
--sigma: coefficient of the random Gaussian noise, default = 2.5.
--update_interval: number of iteration to update clustering targets, default = 1.
--tol: tolerance to terminate the clustering stage, which is the delta of predicted labels between two consecutive iterations, default = 0.001.
--final_latent_file: file name to output final latent representations of the autoencoder, default = final_latent_file.txt.
--predict_label_file: file name to output clustering labels, default = pred_labels.txt.

Outputs

• final_latent: low dimensional representations of scRNA-seq data, default shape (n_cells, 32), which can be visualized by t-SNE or UMAP.
  
• predict_label: predicted clustering labels, shape (n_cells).
  

Reference

Tian, T., Wan, J., Song, Q., & Wei, Z. (2019). Clustering single-cell RNA-seq data with a model-based deep learning approach. Nature Machine Intelligence, 1(4), 191-198. https://www.nature.com/articles/s42256-019-0037-0.

Online app

Online app website: https://app.superbio.ai/apps/107

Contact

Tian Tian tt72@njit.edu