scAce: an adaptive embedding and clustering method for scRNA-seq data

Overview

scAce is consisted of three major steps, a pre-training step based on a variational autoencoder, a cluster initialization step to obtain initial cluster labels, and an adaptive cluster merging step to iteratively update cluster labels and cell embeddings. In the pre-training step, scAce takes the single-cell gene expression matrix as its input to train a VAE network. For each gene, the VAE learns and outputs three parameters of a ZINB distribution (mean, dispersion, and proportion of zero). In the cluster initialization step, scAce offeres two manners. With de novo initialization, Leiden is used to obtain initial cluster labels; with clustering enhancement, initial cluster labels are obtained by applying a cluster splitting approach to a set of existing clustering results. In the adaptive cluster merging step, given the pre-trained VAE network and the initial cluster labels, the network parameters, cell embeddings, cluster labels and centroids are iteratively updated by alternately performing network update and cluster merging steps. The final results of cell embeddings and cluster labels are output by scAce after the iteration process stops.

Installation

Please install scAce from pypi with:

pip install scace

Or clone this repository and use

pip install -e .

in the root of this repository.

Quick start

Load the data to be analyzed:

import scanpy as sc

adata = sc.AnnData(data)

Perform data pre-processing:

# Basic filtering
sc.pp.filter_genes(adata, min_cells=3)
sc.pp.filter_cells(adata, min_genes=200)

adata.raw = adata.copy()

# Total-count normlize, logarithmize and scale the data  
sc.pp.normalize_per_cell(adata)
adata.obs['scale_factor'] = adata.obs.n_counts / adata.obs.n_counts.median()

sc.pp.log1p(adata)
sc.pp.scale(adata)

Run the scAce method:

from scace import run_scace
adata = run_scace(adata)

The output adata contains cluster labels in adata.obs['scace_cluster'] and the cell embeddings in adata.obsm['scace_emb']. The embeddings can be used as input of other downstream analyses.

Please refer to tutorial.ipynb for a detailed description of scAce's usage.