Deep count autoencoder for denoising scRNA-seq data

A deep count autoencoder network to denoise scRNA-seq data and remove the dropout effect by taking the count structure, overdispersed nature and sparsity of the data into account using a deep autoencoder with zero-inflated negative binomial (ZINB) loss function.

See our manuscript and tutorial for more details.

Installation

pip

For a traditional Python installation of the count autoencoder and the required packages, use

$ pip install dca

conda

Another approach for installing count autoencoder and the required packages is to use Conda (most easily obtained via the Miniconda Python distribution). Afterwards run the following commands.

$ conda install -c bioconda dca

Usage

You can run the autoencoder from the command line:

dca matrix.csv results

where matrix.csv is a CSV/TSV-formatted raw count matrix with genes in rows and cells in columns. Cell and gene labels are mandatory.

Results

Output folder contains the main output file (representing the mean parameter of ZINB distribution) as well as some additional matrices in TSV format:

• mean.tsv is the main output of the method which represents the mean parameter of the ZINB distribution. This file has the same dimensions as the input file (except that the zero-expression genes or cells are excluded). It is formatted as a gene x cell matrix. Additionally, mean_norm.tsv file contains the library size-normalized expressions of each cell and gene. See normalize_total function from Scanpy for the details about the default library size normalization method used in DCA.

• pi.tsv and dispersion.tsv files represent dropout probabilities and dispersion for each cell and gene. Matrix dimensions are same as mean.tsv and the input file.

• reduced.tsv file contains the hidden representation of each cell (in a 32-dimensional space by default), which denotes the activations of bottleneck neurons.

Use -h option to see all available parameters and defaults.

Hyperparameter optimization

You can run the autoencoder with --hyper option to perform hyperparameter search.

---

Fork notes (jkobject/dca)

This is a fork of theislab/dca maintained for use as a submodule in scPRINT-2.

What was changed and why

DCA was written for Keras 1.x / early Keras 2.x and uses several APIs that have been removed or relocated in modern versions of Keras (2.12+):

File	Old import	New import	Reason
dca/network.py	from keras.objectives import mean_squared_error	from keras.losses import mean_squared_error	keras.objectives module was removed; losses moved to keras.losses
dca/layers.py	from keras.engine.topology import Layer	from keras.layers import Layer	keras.engine.topology was deprecated and removed; Layer is now directly in keras.layers
dca/layers.py	from keras.engine.base_layer import InputSpec	from keras.layers import InputSpec	Same — keras.engine.base_layer no longer exists

These are minimal, non-breaking patches — no logic was changed, only import paths updated to match Keras 2.12 conventions.

Usage in scPRINT-2

import sys
sys.path.insert(0, "tools/dca")
from dca.api import dca
result = dca(adata, copy=True)

Requires a separate Python environment with tensorflow==2.12.0 and keras==2.12.0 (incompatible with the main scPRINT-2 environment due to TF version conflicts).