sc3s - efficient scaling of single cell consensus clustering to millions of cells

SC3s, or Single Cell Consensus Clustering with speed, is a package for the unsupervised clustering of single cell datasets. It is an updated version of SC3, now reimplemented in Python, and integrated into the popular Python single cell toolkit Scanpy. The algorithm has also been reengineered to allow it to scale efficiently towards millions of cells.

Installation

Best way to install is to use pip. In a shell environment, run:

pip install sc3s

SC3s is meant to be complemented with Scanpy, making use of the AnnData data structure. You can find its installation instructions here.

Usage

SC3s is designed to be an intermediate step in the single cell workflow. A general workflow would involve importing the data, running some preprocessing and doing dimensionality reduction with PCA. We recommending going through this offical tutorial on Scanpy's documentation, but a minimal barebones workflow could be:

import scanpy as sc
adata = sc.read_h5ad("/path/to/data.h5ad")

# remove lowly expressed genes
sc.pp.filter_genes(adata, min_cells=10)

# log transform the data
sc.pp.log1p(adata)

# dimensionality reduction with PCA
sc.tl.pca(adata, svd_solver='arpack')

Once that is done, you can run SC3s with:

import sc3s
sc3s.tl.consensus(adata, n_clusters=4)

SC3s can quickly evaluate multiple values for n_clusters, useful for exploring structure in data.

sc3s.tl.consensus(adata, n_clusters=[4,8,10])

Parameters

The default settings should suffice for most use cases, but you can get more information about the parameters with:

help(sc3s.tl.consensus)

Output

Running SC3s will modify the obs dataframe of the AnnData object, from which you can assess the assigned cluster labels:

adata.obs             # shows the dataframe
adata.obs['sc3s_4']   # returns the cluster labels

The labels can then be used by subsequent functions in Scanpy, for example, plotted on UMAP and/or PCA plots.

sc.pl.pca(adata, color='sc3s_4')

Citation

Please cite the following publication if you find SC3s useful in your work:

Quah, F.X., Hemberg, M. SC3s: efficient scaling of single cell consensus clustering to millions of cells. BMC Bioinformatics 23, 536 (2022). https://doi.org/10.1186/s12859-022-05085-z