Bayesian mechanistic modelling of scRNAseq data using Stan

Requirements:

• Julia (recommended), including JSON.jl, ArgParse.jl
• Python (recommended)
• CmdStan, or any other Stan interface

Description:

The folder stan contains the following models:

• tg_zi_single: Telegraph model, with zero inflation, single allele
• tg_zi_vol_single: Telegraph model, with zero inflation and cell volume, single allele
• tg_zi_vol_pooled: Telegraph model, with zero inflation and cell volume, two alleles (pooled)

These models can be run using CmdStan.

Input

The input to Stan is provided as a JSON file containing the following fields:

• int ncells: Number of cells
• int ngenes: Number of genes
• int N: Number of observations
• int cells[N]: Cell corresponding to each observation
• int genes[N]: Gene corresponding to each observation
• int counts[N]: Measured mRNA numbers for each observation

The script convert_data.jl can be used to convert a count matrix (stored as a CSV file) into this format, skipping missing values.

Output

The sampler computes the posterior over the following quantities:

Gene-specific:

• mu: Mean mRNA numbers
• b: Mean burst size
• dur: Mean burst duration
• p0: Amount of zero inflation
• alpha: Volume dependence (if present)

The following values are computed for convenience:

• rho: Transcription rate
• sigma_on: on switching rate
• sigma_off: on switching rate

Cell-specific:

• vol: Volume scaling factor (if present)

Note: The outputs also contains the latent variables raw_betas, one per observation. This requires a lot of memory or disk space for large batches. These, and other internal variables, can be removed using the script postprocess_csv.py.

Example usage:

1. Convert mRNA count matrix to JSON:

julia scripts/convert_data.jl countmat.csv countmat.json

The input must be a CSV file containing mRNA counts. Each row represents a cell, and each column a gene (except for the first column, which is ignored). Missing values are allowed.

2. Run Stan:

stan/tg_zi_vol_single sample data file=countmat.json output file=raw_samples.csv

See the CmdStan user guide for more information

3. (Optional) Trim output files:

python scripts/postprocess_csv.py <raw_samples_1.csv >samples_1.csv
...