Sunbear

Single-cell multi-modal/multi-condition temporal inference model

Overview

The Sunbear model performs temporal inference alongside cross-condition and cross-modality inference, which can be applied to:

1. infer how a cellular profile changes along a continuous time frame
2. compare condition-differences on time-series data with few matched conditions
3. jointly infer temporal multimodal profile changes for each cell

Installation

Install via docker (recommended):

apptainer pull docker://bearfam/bears
apptainer shell --nv bears_latest.sif

Install via conda:

conda env create -f environment.yml
conda activate sunbear

Example run:

bash ./example.sh

Input data:

The code takes in h5ad format. Here are some example input files.

The h5ad object for scRNA-seq consists of:

• gene expression count matrix (rna_adata.X)
• gene annotation (rna_adata.var)
• cell annotation (rna_adata.obs): rna_adata.obs needs to contain a "time" column. The model also allows for an optional "batch" and/or "condition" column.
  Example input data:

$rna_h5ad=example_single_rna.h5ad

In the multimodal setup, Sunbear takes another scATAC-seq input in h5ad format, which consists of

• a binarized peak accessibility matrix (atac_adata.X)
• peak region annotation (atac_adata.var): must include a column indicating the chromosome so that the model can save memory and include all peak regions
• cell annotation (atac_adata.obs), similar to rna_adata.obs needs to contain a "time" column, an optional batch column, and an optional condition column.
  Example input data:

$rna_h5ad=example_multi_rna.h5ad
$atac_h5ad=example_multi_atac.h5ad

Basic usage:

1. temporal inference of gene expression profiles in a specific $celltype around $timepoint:

rna_h5ad=data/example_single_rna.h5ad
python bin/sunbear.py --domain rna --rna_h5ad $rna_h5ad --batch batch --condition sex --predict temporal --ct_query Muscle_cells --targettime 16 --celltype major_trajectory

2. Cross-condition inference and comparison between biological conditions: For cross-condition comparison, at least two conditions need to be included and both conditions should be seen during training.Here is an example to calculate differential expression between conditon F and M:

python bin/sunbear.py --domain rna --rna_h5ad $rna_h5ad --batch batch --condition sex --targettime 16 --predict diffexp_condition --sourcecondition M --targetcondition F 

3. Cross-modality temporal inference: We first train a multimodal temporal model on existing measurements. Then, for cells in $celltype, we jointly predict their gene expression and chromatin accessibility changes around $timepoint:

$rna_h5ad=example_multi_rna.h5ad
$atac_h5ad=example_multi_atac.h5ad
python bin/sunbear.py --domain multi --rna_h5ad $rna_h5ad --atac_h5ad $atac_h5ad --batch batch --predict temporal --time_range 0.1 --targettime 8.25 --ct_query Hindbrain

other functions:

1. generate simulation and train on simulated data (take linear trend as an example)

Rscript bin/generate_simulation.R
rna_h5ad=data/simulation_linear10.2
python bin/sunbear.py --domain rna --rna_h5ad $rna_h5ad --train_ver simulation --targettime 7.5