scAEGAN- Unification of Single-Cell Genomics Data by Adversarial Learning of Latent Space Correspondences

This repository contains the scAEGAN code and online data for the single-omics and multi-omics integration. It also contains the code to evaluate and visualize the integration results. Metrics are available for quantifying outputs quality.

• Summary
• Installation Requisites
• Datasets
• Usage
• Running Example

Summary

scAEGAN is a python based deep learning model that is designed for single-cell-omics and multi-omics integration. scAEGAN performs this by using an Autoencoder which learns a low-dimensional embedding of each experiment independently, respecting each sample's uniqueness, protocol. Next, cycleGAN learns a non-linear mapping between these two Autoencoder representations.

scAEGAN Workflow

Installation Requisites

The required libraries are included in environment file. In order to install these libraries, follow the following steps:

• Creating the conda environment with the following command. This will create and install the libraries included in the environment.yml file for training the scAEGAN.

conda env create --prefix ./env --file environment.yml --force

• The second step is to activate the conda envirnoment.

conda activate ./env      

• Optionally, run R_libraries.R, in tested R version 4.1.1, to install automatically the libraries and dependencies required for the scripts in the Evaluation folder.

• scAEGAN is simply installed by cloning the repository.

git clone https://github.com/sumeer1/scAEGAN.git

cd scAEGAN/

Datasets

Simulated data : Two datasets containing 600 cells from 5 populations and with 3000 genes each were simulated using SymSim (Zhang, X et al.).

Real data: The pre-processed mouse hematopoietic stem cell dataset of young and old individuals downloaded from https://github.com/quon-titative-biology/scalign.

Usage

There are two steps for the basic usage after activating the conda environment.

• Training the autoencoder with the given parameters to get the latent representation by running.

python AE.py --input_file1 <Specifies the domainA input file (cell by gene matrix in csv format)> \
             --input_file2 <Specifies the domainB input file (cell by gene matrix in csv format)>  \
             --output_file1 <Specifies the low dimensional representation of the input1 from the autoencoder> \
             --output_file2 <Specifies the low dimensional representation of the input2 from the autoencoder> \
             --batch_size <Specifies the batch size to train the autoencoder, default=16>  \
             --epochs <Specifies  the number of epochs for which autoencoder is trained, default=200> \
             --dropout <Specifies the dropout rate used to train the autoencoder, default=0.2> \
             --learning_rate <Specifies the learning rate, default=0.0001>

• Training the cyclegan with the given parameters on latent representations obtained from the Autoencoder by running.

python cGANtrain.py --data_path <Specifies the folder path to the training and testing data> \
                    --train_file <Specifies the training files for training the cGAN for both domains (A and B) that are to be integrated. 
                    For instance --train_file domain_A.csv domain_B.csv \
                    --test_file <Specifies the testing files. For instance --test_file domain_A.csv domain_B.csv> \
                    --save_path <Specifies the folder path where the output from the cGAN in the csv format will be saved> \
                    --input_shape <Specifies the shape of the input, default=50> \
                    --batch_size <Specifies the batch size, default=4> \
                    --epochs <Specifies the number of epochs for training cGAN, default=400>

Running Example

• In this tutorial we show how to run scAEGAN on the Simulated Data. We have prepared the required input dataset which you can find in the Simulated_Data folder.
• We created a command-line interface for scAEGAN that allows it to be run in a high-performance computing environment. Because scAEGAN is built with tensorflow/keras, we recommend running it on GPUs to significantly reduce run-time. It has been tested on Linux and OS X platforms.
• The experiments were performed on a Linux server using an Intel Xeon CPU E5-2680 v4 @ 2.40GHz processor with 128 GB RAM and an NVIDIA Tesla V100 GPU.
• For model training and evaluation, a vignette presents an example how to run the scAEGAN and carry out the benchmarking using the Evaluation folder scripts.