scBiMapping

Fast and Accurate Non-linear Dimensionality Reduction and Cell Annotation for Large and High-dimensional Single-Cell Datasets:

1. In terms of dimension reduction, the proposed method demonstrates high effectiveness and efficiency on very large datasets, such as Tahoe-100M (the largest publicly available single-cell dataset to date, comprising over 100 million single-cell transcriptomes collected from 50 cancer cell lines);

2. For automatic reference-based cell type annotation, the proposed method exhibits strong performance on extremely large spatial omics datasets (ranging from millions to tens of millions of cells), including MerFish and Stereo-seq data.

$\text{\color{red}Note}$:

• For both dimension reduction and cell type annotation, neither GPU nor subsampling of cells is needed.

• The initial version of this project was released on https://github.com/scBGI/scBiMapping, and subsequent maintenance of the project will be conducted here, due to the login issue.

Install

pip install scBiMapping

$\text{\color{red}Note}$:

• the current version (v0.0.8) on PyPI has been compiled in python 3.11, and python 3.11 is necessary for using this package.*

• if you are a BGIer, you can directly use the public image (named scBiMapping) on the cloud platform.*

How to use

There are two major functions in scBiMapping, scBiMapping_DR and scBiMapping_annotation, corresponding to the following two tasks.

Task 1: Dimension reduction

scBiMapping_DR(adata,n_embedding):

• Input:

  • adata: anndata format (cell-feature sparse matrix is stored in adata.X);

  • n_embedding: an integer, denoting embedding dimensions (default to 30; slight adjustment may lead to better performance in practice);

• Output:

  • Embedded matrix is stored in adata.obsm['U'], where each row denotes the embedded vector of one cell;

Task 2: reference-based cell type annotation

scBiMapping_annotation(adata_ref,adata_query,n_embedding, K, K_majority, CellType_Key_for_ref)

• Key Inputs:

  • adata_ref: referenc dataset (anndata format);

  • adata_query: query dataset (anndata format);

  • n_embedding: an integer, denoting the number of embeddings (default to 30; slight adjustment may lead to better performance in practice);

  • K: an integer, denoting how many features are used as the new vector representation of each cell in the embedding (default to 30; adjustment may be needed in practice);

  • K_majority: an integer, denoting how many reference cells are used for majority voting (default to 10; adjustment may be needed in practice);

  • CellType_Key_for_ref: key in adata_ref.obs that stores the cell type labels of the reference cells (IMPORTANT!!!);

• Output:

  • the predicted cell types for all query cells are stored in adata_query.obs['cell_type_predicted']

Tutorials for tasks 1 and 2

We provide several demos to further demonstrate how to conduct dimension reduction and reference-based cell type annotation using scBiMapping; see Details in the Turtorials file (the test datasets are available at the following "Pre-procssed datasets" section).

Scripts to reproduce primary experimental results

See the reproducing_results file. See also reproducible program in codeOcean: https://codeocean.com/capsule/3904732/tree.

Pre-processed datasets

Processed small and large public datasets in this study can be downloaded at (all in .h5ad format):

• https://doi.org/10.5281/zenodo.15871971 (Pre-processed Tahoe-100M scRNA-seq cancer cell line datasets for scBiMapping (Part I))

• https://doi.org/10.5281/zenodo.15871981 (Pre-processed Tahoe-100M scRNA-seq cancer cell line datasets for scBiMapping (Part II))

• https://doi.org/10.5281/zenodo.15826658 (Pre-processed cell type annotation datasets for scBiMapping (Part I))

• https://doi.org/10.5281/zenodo.16396750 (Pre-processed cell type annotation datasets for scBiMapping (Part II))

• https://doi.org/10.5281/zenodo.15380061 (Pre-processed datasets for scBiMappinig evaluation)