Jvis

Emerging single-cell technologies profile multiple types of molecules within individual cells. A fundamental step in the analysis of the produced high-dimensional data is their visualization using dimensionality reduction techniques such as t-SNE and UMAP. We introduce j-SNE and j-UMAP as their natural generalizations to the joint visualization of multimodal omics data. Our approach automatically learns the relative contribution of each modality to a concise representation of cellular identity that promotes discriminative features but suppresses noise. On eight datasets, j-SNE and j-UMAP produce unified embeddings that better agree with known cell types and that harmonize RNA and protein velocity landscapes. j-SNE and j-UMAP are available in the JVis Python package.

The details for the underlying mathematics can be found in https://genomebiology.biomedcentral.com/articles/10.1186/s13059-021-02356-5.

Van Hoan Do and Stefan Canzar. A generalization of t-SNE and UMAP to single-cell multimodal omics. Genome Biology. 2021;22(1):130. doi:10.1186/s13059-021-02356-5

Installing

Requirements:

• Python 3.6 or greater
• numpy
• scipy
• scikit-learn >= 0.23.0
• numba

Install Options

PyPI install, presuming you have numba and sklearn and all its requirements (numpy and scipy) installed:

pip install Jvis-learn

If you have a problem with pip installation then we'd suggest installing the dependencies manually using anaconda followed by pulling umap from pip:

conda install numpy 
conda install scipy==1.5.3
conda install scikit-learn==0.24.1
conda install numba==1.19.2
pip install Jvis-learn

How to use Jvis

The Jvis package inherits from sklearn TSNE, and UMAP. Therefore, all parameters of tSNE and UMAP are naturally extended for Jvis.

An example of making use of these options:

from Jvis import JUMAP, JTSNE
import numpy as np

# Create a toy example from a random distribution (n_cells = 500)
rna_rand = np.random.rand(500, 100)
adt_rand = np.random.rand(500, 15)
data = {'rna': rna_rand, 'adt': adt_rand} # create a dictionary of modalities.

# Run joint TSNE of the two "random" modalities.
embedding_jtsne = JTSNE(n_components=2).fit_transform(data)

# Run joint UMAP of the two "random" modalities.
embedding_jumap = JUMAP(n_neighbors=20,
                        min_dist=0.3,
                        metric='correlation').fit_transform(data)

For more realistic examples and Python scripts to reproduce the results in our paper are available at GitHub: https://github.com/canzarlab/JVis_paper

Tunning parameters of t-SNE and UMAP can be found here: https://scikit-learn.org/stable/modules/generated/sklearn.manifold.TSNE.html

https://umap-learn.readthedocs.io/en/latest/api.html

FAQ

1/ How to print the weight of modalities? #3

License

The JVis package is 3-clause BSD licensed.

Jvis package is inherited from scikit-learn and UMAP package under 3-clause BSD license.

This code was tested on Python 3.6, 3.7; scikit-learn version 0.24.1; numpy version 1.19.2; scipy version 1.5.3; numba version 0.52.0