Single-cell sequencing technologies provide a revolutionary understanding of cellular heterogeneity at the single-cell resolution. However, the high-dimensional and high-noise nature of single-cell data presents challenges for downstream analyses, and feature selection has thus become a critical step in processing and analyzing single-cell data. Feature selection approaches are mature for single-cell RNA sequencing (scRNA-seq) data, but there is a significant gap in the availability of effective methods for feature selection in single-cell chromatin accessibility (scCAS) data. Here, we present Cofea, a correlation-based framework, for selecting informative features by exploring the interaction between features derived from scCAS data.

1. We recommend you to build a python virtual environment with Anaconda. If Anaconda (or miniconda) is already installed with Python3, skip to 2.

2. Create and activate a new virtual environment:

$ conda create -n cofea python=3.8
$ conda activate cofea

Python packages required by Cofea are listed below:

1. Python 3.8.13
2. Packages for Cofea and tutorial
  anndata==0.8.0
  episcanpy==0.3.2
  h5py==3.7.0
  hdf5storage==0.1.18
  jupyter-contrib-core==0.4.0
  loess==2.1.2
  louvain==0.7.1
  matplotlib==3.5.2
  matplotlib-inline==0.1.3
  memory-profiler==0.61.0
  numba==0.55.2
  numpy==1.22.4
  pandas==1.4.3
  patsy==0.5.2
  progress==1.6
  scanpy==1.9.1
  scikit-learn==1.1.1
  scipy==1.8.1
  seaborn==0.11.2
  statsmodels==0.13.2
  umap-learn==0.5.3

Install the package and other requirements:

Package installation:
$ git clone https://github.com/likeyi19/Cofea   
$ cd Cofea   
$ pip install -r requirements.txt

We provide a quick-start notebook notebook which describes the fundamentals in detail and reproduces the results of Cofea.

Six parameters are required, including the path of dataset, TFIDF implementation method, number of PCS, correlation coefficient calculation method, number of selected features, and random seed. TFIDF is 'tfidf2' by default, the number of PCS is 100 by default, and the correlation coefficient is 'PCC' by default.

For exsample:

$ cd code/
$ python cofea.py -l ../data/scanpy.h5ad  -n 20000 -s 2
$ cd ..

Or you can get help in this way:

$ python code/cofean.py -h
usage: cofea.py [-h] 
                [-l LOAD_PATH] 
                [-t TFIDF] 
                [-p PC] 
                [-c CORR] 
                [-n SELECT_NUMBER] 
                [-s SEED_BASE]

optional arguments:
  -h, --help           show this help message and exit
  -l, --load_path      str, default=None
                       storage path of the h5ad file, which contains the peak-by-cell data to be processed.
  -t, --TFIDF          str, default='tfidf2', options=['tfidf1', 'tfidf2', 'tfidf3']
                       TF-IDF implementation, 'tfidf1' represents the original version of TF-IDF transformation, 'tfidf2' indicates the TF-IDF transformation used by Signac, and 'tfidf3' represents the TF-IDF transformation used by scOpen.
  -p, --PC             int, default=100
                       Dimension of cell-wise PCA.
  -c, --corr           str, default='PCC', options=['PCC', 'CSC', 'SPCC']
                       Correlation coefficient calculation method. PCC represents Pearson correlation coefficient, CSC indicates Cosine correlation coefficient and SPCC represents Spearman correlation coefficient.
  -n, --select_number  int, default=20000
                       Number of selected features.
  -s, --seed_base      int, default=2
                       Random seed.

The overlapped proportion between the identified features and the ground-truth informative features offers a intuitive and accurate metric than evaluating the performance based on downstream analysis tasks. To this end, we generated five high-fidelity scCAS datasets to enable a quantitative evaluation of feature selection methods.

Taking S4 and S5 datasets as an example, we draw the visualization and UMAP of the dataset, and compare the overlapped proportion between the features selected by each method and the cell type-specific features.

We provide a notebook to reproduce the generation and testing process for simulation datasets.

From the perspective of revealing biological insights, we evaluated Cofea via several numeric experiments including cell type-specific peaks annotation and candidate enhancers identification.

We provide a notebook to analyze the sample dataset.

If you have any questions, you can contact me from the email: [email protected]