Cyclum

Introduction

Cyclum is a package to tackle cell cycle. It provides methods to recover cell cycle information and remove cell cycle factor from the scRNA-seq data. The methodology is to rely on the circular manifold, instead of the marker genes. We provide an Auto-Encoder based realization at this time, and we are adding Gaussian Process Latent Variable Model soon.

Tutorials

We provide a one-click self-contained demo ships with its dataset, which shows how to start with an expression matrix, then decide the optimal dimensionality, and finally calculate the circular pseudotime.

More examples are available in test/notebooks, where there is a detailed table of contents.

Our paper was published: Liang S, Wang F, Han J, Chen K. Latent periodic process inference from single-cell RNA-seq data. Nat Commun 11(1):1441, 3/2020. e-Pub 3/2020. PMID: 32188848. old preprint was able @ BioRxiv. Documentation for submodules and classes in the Cyclum Python module is available as a website. Explanations of other files in test are available as README.md in the corresponding folders.

Installation

You can install cyclum by running the following commands, in a directory you desire.

conda create -n cyclum python=3.7 pip
conda activate cyclum
git clone https://github.com/KChen-lab/Cyclum.git
cd Cyclum
pip install .

⚠️ This will install cyclum and its dependencies to your system, which affects more than just this directory. Thus, we highly recommend virtual environment such to avoid messing up your system / breaking dependencies of other software. Miniconda may be a good choice if you have not already installed one.

You can then run Jupyter notebook simple using the following command.

jupyter notebook

A browser window will open showing the directory, where you can go to tests/notebokks to view/run the exmaples.

⚠️ Please be advised that pip will attemp to install the latest version of the dependencies, which is different from where this program is tested (shown in the table below). TensorFlow < 1.15.2 has a known security issue and are not recommended. However, you can force pip to install specific versions by change install_requires=[...] in setup.py to install_requires=['keras==2.2.4', 'numpy==1.16.5', 'pandas==0.25.2', 'scikit-learn==0.21.3', 'h5py==2.9.0', 'jupyter==1.0.0', 'matplotlib==3.1.1', 'tensorflow==1.14.0'].

Use as portable software

You can also use cyclum as a portable software, without installing. All the notebooks contains code that add cyclum to the path, so that you can run them directly. However, please make sure the dependencies are fulfilled.

Software dependencies:

Cyclum was tested on these package versions. Please make sure that you have TensorFlow 1.x. Cyclum is compatible with newer versions shown in the "Latest tested" column.

Software	Version	Latest tested
python	3.7.4	3.7.6
keras	2.2.4	2.3.1
tensorflow	1.14.0	1.15.2
numpy	1.16.5	1.18.1
pandas	0.25.2	1.0.1
scikit-learn	0.21.3	0.22.1
h5py	2.9.0	2.10.0
jupyter	1.0.0	1.0.0
matplotlib	3.1.1	3.1.3

We recommend Miniconda to manage the packages. The code should work on packages of newer versions, but in case it fails, you can return to the specific version by, for example, conda install python=3.7.4.

System requriment:

The code is on Debian GNU/Linux 10 (buster) with both CPU and GPU. The code should run on most mainstream systems (Linux, Mac, Windows) supporting Tensorflow.

Collaborate with other tools

Refine labels from other sources:

If you have labels from other sources such as scanpy.tl.score_genes_cell_cycle or Seurat::CellCycleScoring, you can use Cyclum to refine the result along the inferred pseudotime.

from cyclum.postproc import refine_labels
refine_labels(pseudotime=[4, 5, 1, 2, 3, 6, 7, 8], original_labels=['G1', 'G1', 'G1', 'G1', 'S', 'S', 'G2M', 'G2M'])

It will output ['G1', 'G1', 'G1', 'G1', 'G1', 'S', 'G2M', 'G2M'] as the refined result. Note that an S is replaced by G1 as it is surrounded by G1s.

Transferring data between python and R

Although Python is a good data analysis tool in addition to a general programing language, researchers may want to use R, which is more focused on statistics. Cyclum is implemented in python, but in order to help use both languages, we implemented mat2hdf and hdf2mat in both Python and R, to help transferring data back and forth rapidly. In general, the correspondence of data structures in R and Python are: unnamed matrices -- 2D numpy.array, named matrices -- pandas.DataFrame, data.frame -- pandas.DataFrame. (Prerequisites: hdf5r in R, h5py in python.)

Transferring data to GSEA

GSEA is a powerful tool to perform downstream gene enrichment analysis. We implemented in R...

• mat2txt, which writes a expression matrix to a GSEA compatible .txt file (Prerequisite: data.table, for much faster writing than write.table),
• vec2cls, which writes phenotypes (either discrete, e.g., cell type, or continuous, e.g., pseudotime) to a GSEA compatible .cls file,
• mat2cls, which writes multiple sets of phenotypes (continuous only, e.g., multiple PCs) to a GSEA compatible .cls file.

The old version

We revised almost everything, except for the concept of using sinusoidal function in an autoencoder to find circular biological processes ab initio. The autoencoder is now rewritten using keras, in a more readable way. We hope this will help researchers who want to experiment similar network structures. We also implemented class cyclum.tuning.CyclumAutoTune, which automatically select the proper number of linear components to help find the "most circular" manifold. The old version is kept in old-version.