Cell-BLAST 0.3.7

Single-cell transcriptome querying tool

Cell BLAST

Cell BLAST is a cell querying tool for single-cell transcriptomics data.

For each query cell, it searches for most similar cells in the reference database. Annotations in reference cells, e.g. cell type, can then be transfered to query cells based on cell-to-cell similarities. See our preprint for details.

Installing the python package

We only support installation via pip right now.

Installation within virtual environments are recommended, see virtualenv or conda.

For conda, here's a one-liner to set up an empty environment for installing Cell BLAST:

conda create -n cb python=3.6 && source activate cb

Now follow the instructions below to install Cell BLAST:

1. Make sure you have a working version of tensorflow or tensorflow-gpu (version >= 1.5). You can follow the official instructions about how to install tensorflow (and dependencies like CUDA and CuDNN for the GPU version), or just install via anaconda, which handles dependencies automatically:

   For installing the GPU supported version: conda install tensorflow-gpu=1.8

   For installing the CPU only version: conda install tensorflow=1.8

2. Install Cell BLAST by running: pip install Cell-BLAST

3. Check if the package can be imported in python interpreter: import Cell_BLAST as cb

Online documentation can be found here.

Web-based interface

We also provide a Web-based service for off-the-shelf querying of our ACA reference panels.

Repository structure

• The Cell_BLAST directory contains the Cell BLAST python package.
• The Datasets directory contains scripts used for building the ACA database.
• The Evaluation directory contains scripts used for benchmarking and producing some figures of the manuscript.
• The docs directory contains files used to generate the online documentation.
• The test directory contains unit tests for the python package.

Reproduce results

Obtain required data files

Environment setup

Python

First create a conda environment and install most packages via:

conda env create -n cb-gpu -f env.yml && conda activate cb-gpu

Optionally, if GPU does not work properly (possibly due to inconsistent conda channels used), reinstalling tensorflow should solve the problem:

conda install tensorflow=1.8.0 tensorflow-base=1.8.0 tensorflow-gpu=1.8.0 --force-reinstall

Finally, install customized packages or packages unavailable in conda. All dependencies have already been installed via env.yml, so --no-deps is added to prevent overwriting conda installed packages:

pip install Cell-BLAST==0.3.7 --no-deps
pip install local/scScope-0.1.5.tar.gz --no-deps  # Add random seed setting
pip install local/DCA-0.2.2.tar.gz --no-deps  # Allow GPU memory growth, suppress integer warning
pip install local/DCA_modpp-0.2.2.tar.gz --no-deps  # Modify preprocessing
pip install local/ZIFA-0.1.tar.gz --no-deps  # Remove fixed random seeds
pip install local/Dhaka-0.1.tar.gz --no-deps
pip install local/scvi-0.2.3.tar.gz --no-deps  # Fix torch bugs
tar xf local/SAUCIE.tar.gz -C ${CONDA_PREFIX}/lib/python3.6/site-packages/  # Add random seed setting
pip install fcswrite  # Dependency of SAUCIE not available in conda

For scPhere, we use a separate environment because of conflicting dependencies (the environment should be named "scphere" for it to be found in the benchmarking pipeline):

conda create -n scphere 'python>=3.6' 'numpy>=1.16.4' 'scipy>=1.3.0' \
   'pandas>=0.21.0' 'matplotlib>=3.1.0' 'tensorflow=1.14.0' \
   'tensorflow-probability=0.7.0' 'ipykernel' && conda activate scphere
pip install local/scPhere-0.1.0.tar.gz --no-deps
pip install Cell-BLAST==0.3.7  # Still need cb.data to read data (tf dependent functions may not work properly)

R

Start R (tested on version 3.6.0) at project root and run:

packrat::restore()

Then install the customized version of Seurat by:

install.packages("local/seurat-2.3.3.tar.gz", repos=NULL, type="source")  # Remove fixed random seeds

For CCA anchor (Seurat v3) and Harmony, we used a separate packrat environment.

To build this dedicated environment, start R at directory "packrat/envs/seurat_v3" and run:

packrat::restore()

Then install the customized version of Seurat v3 by:

install.packages("../../../local/seurat-3.0.2.tar.gz", repos=NULL, type="source")  # Remove fixed random seeds

Run all benchmarks and reproduce figures

Make sure the conda environment create above is activated. Go to directory "Evaluation" and run the following command:

snakemake -prk

Some jobs will likely fail, e.g. due to timeout or memory issues, and cause downstream steps, including result plotting, to fail as well.

These failing jobs will be blacklisted in future runs, so just run the above command for a second time, and results for the successful jobs should be summarized and plotted without error.

Contact

Feel free to submit an issue or contact us at [e-mail] for problems about the python package, website or database.