geosketch is a Python package that implements the geometric sketching algorithm described by Brian Hie, Hyunghoon Cho, Benjamin DeMeo, Bryan Bryson, and Bonnie Berger in "Geometric sketching compactly summarizes the single-cell transcriptomic landscape", Cell Systems (2019). This repository contains an example implementation of the algorithm as well as scripts necessary for reproducing the experiments in the paper.

You should be able to install from PyPI:

pip install geosketch

Parameter documentation for the geometric sketching gs() function is in the source code at the top of geosketch/sketch.py.

For an example of usage of geosketch in R using the reticulate library, see example.R. WARNING: The indices returned by geosketch are 0-indexed, but R uses 1-indexing, so the one_indexed parameter should be set to TRUE when called from R.

Here is example usage of geosketch in Python. First, put your data set into a matrix:

X = [ sparse or dense matrix, samples in rows, features in columns ]

Then, compute the top PCs:

# Compute PCs.
from fbpca import pca
U, s, Vt = pca(X, k=100) # E.g., 100 PCs.
X_dimred = U[:, :100] * s[:100]

Now, you are ready to sketch!

# Sketch.
from geosketch import gs
N = 20000 # Number of samples to obtain from the data set.
sketch_index = gs(X_dimred, N, replace=False)

X_sketch = X_dimred[sketch_index]

All of the data used in our study can be downloaded from http:https://cb.csail.mit.edu/cb/geosketch/data.tar.gz. Download and unpack this data with the command:

wget http:https://cb.csail.mit.edu/cb/geosketch/data.tar.gz
tar xvf data.tar.gz

We can visualize a large data set of cells from different regions of the mouse brain collected by Saunders et al. (2018).

To visualize the sketches obtained by geometric sketching and other baseline algorithms, download the data using the commands above and then run:

python bin/mouse_brain_visualize.py

This will output PNG files to the top level directory visualizing different sketches produced by different algorithms, including geometric sketching.

For those interested, the algorithm implementation is available in the file geosketch/sketch.py.

For questions, please use the GitHub Discussions forum. For bugs or other problems, please file an issue.