/review-scRNA-seq-DENOISING

Primary LanguageJupyter Notebook

review-scRNA-seq-DENOISING

This repository contains the source code that was used for the paper A review of computational strategies for denoising and imputation of single-cell transcriptomic data.

Folder organization

In order to download the datasets used in the analyses, after cloning this repository in your local machine, download the zip files located here and extract them in the local folder that contains this repository. After downloading all files, you will have the following folders:

  • ground_truth: this folder contains the rds files with the datasets used as ground truth.

  • final_GT: this folder contains the rds files with the datasets used as ground truth, respectively for UMI and nonUMI simulations. It contains the down-sampled datasets from folder ground_truth.

  • noise_mixed: this folder contains the rds files with the full datasets with noise.

After simulating the counts, for each dataset we applied a basic preprocessing, removing genes expressed in less than 5% of cells and cells with more than 90% of zero counts. Then we also selected for each dataset the top 500, 1000, 2000 and 5000 most variable genes.

  • final_data/csv/noise_filt contains the csv of the pre-processed noise files.

  • final_data/csv_mostvar/noise_filt contains the csv files of the pre-processed noise files with the selection of the most variable genes.

In Both csv_mostvar and csv there is a folder named noise_filt_normalized: this folder contains the noise datsets that were also library size normalized (with a scaling factor of 1e4) and then log-transformed (log(X+1)).

Both in csv_mostvar and csv there is a folder named true_filt. This folder contains the ground truth for each dataset with noise, without those genes and cells that were removed during the pre-processing of the noise files.

IMPORTANT: in each csv, the information about each cell's subpopulation is stored in the cell name. In particular, each cell has a name cellxxxxpx, the number after p indicates the population.

Reproducibility

To reproduce the generation of synthetic datasets, the scripts need to be run in the following order:

  • GT_simulation.R, which produces the files located in the folder final_GT.

  • noise_simulation.R, which produces the files located in the folder final_noise.

  • mix_noise.R, which produces the files located in folder noise_mixed.

  • preprocess_dataset.R, which produces the files located in folder final_data/csv.

  • preprocess_dataset_highlyvar.R, which produces the files located in folder final_data/csv_mostvar.

To reproduce the results of the different denoising methods, follow these steps.

  • For all denoisers written in R, execute the corresponding script located in folder denoising_script. Note that, before running the script, you need to change to working directory to the folder that contains the folder noise_filt (e.g. review-scRNA-seq-DENOISING/final_data/csv to denoise synthetic datasets).

  • For all denoisers wirtten in Python, you need to execute the script all_python.py, executing it inside the directory denoising_script.