GRN-Transformer

About

This directory contains the code and resources of the following paper: Boosting single-cell gene regulatory network reconstruction via bulk-cell transcriptomic data which is published in Briefings in Bioinformatics (https://doi.org/10.1093/bib/bbac389).

Overview of the Model

Dependencies

• python==3.7
• einops==0.3.0
• pytorch==1.7.0+cu101
• pandas==1.1.4
• numpy==1.19.4
• scikit-learn==0.23.2
• argparse==1.4.0

All dependencies can be installed within a few minutes.

Description of directories and files

• [src] contains the detailed implementation for the self-supervised step of the GRN-Transformer

• [demo_data] contains the example of the used dataset.

  • BEELINE
    • 500_ChIP-seq_mDC
      • inputs:
        • Including the gene expression data 'data.csv' (cells in columns and genes in rows),
        • ground truth cell-type-specific GRN 'label.csv' (two columns, TF names and target names),
        • GRN inference result by using PIDC 'PIDC_output_file.txt'
        • train-test-split index files 'split_non_*.pkl'(two list, first list is the training index and second list is the test index)
        • TF-target names 'train_z.npy' (two columns, TF names and target names)
        • ground truth cell-type-specific GRN labels 'train_y.npy' (1 is for edge in cell-type-specific GRN while 0 is for edge not in cell-type-specific GRN, generated from label.csv)
        • cell-type-non-specific ChIP-seq training labels 'train_y_non.npy'(1 is for edge in non-specific GRN while 0 is for edge not in non-specific GRN, generated from other_data/human_Non-specific-ChIP-seq-network.csv (for hESC and hHep) and other_data/mouse_Non-Specific-ChIP-seq-network.csv (for mHSC, mESC, mDC))
      • outputs:
        • Including example pretrain result 'pretrain_output.pkl' by running code self_supervised_main.py
        • example performance output file 'performance.pkl' by running code supervised_main.py
    • other_data Including TFs and cell-type-non-specific GRN collected by BEELINE benchmark.
  • Simulation
    • ER/SF
      • inputs: Similar with inputs directory of 500_ChIP-seq_mDC except train-test-split index files. 'split_*_0.5.pkl' denotes to the split file for train-test-split (without label flipped).
      • outputs: Similar with outputs directory of 500_ChIP-seq_mDC

• [self_superivsed_main.py] The main function for the self-supervised learning step.

• [supervised_main.py] The main function for the supervised learning step.

Usage

We take 500_ChIP-seq_mDC data as an example.

• self-supervised step

  python self_superivsed_main.py --data_file ./demo_data/BEELINE/500_ChIP-seq_mDC/inputs/data.csv --PIDC_file  
  ./demo_data/BEELINE/500_ChIP-seq_mDC/inputs/PIDC_output_file.txt --save_name pretrain_output --attention conventional_attention

  where '--data_file' denotes the input expression data (cells in rows and genes in columns),
  '--PIDC_file' denotes the input pre-calculated PIDC file (three columns (TF name, Target gene name, and importance values),
  '--attention' denotes the attention mechanism (selection from {conventional_attention, efficient_attention, performer, reformer}),
  '--save_name' denotes the output filename used in the following supervised step.

  We recommend using conventional_attention when the number of genes is not too large (e.g., less than 2000 for GTX1080 Ti). Please run self_superivsed_main.py with GPU. It should be finished within 6 hours for the given demo dataset.

• supervised step

  python supervised_main.py --split_file split_non_1.pkl  --pre_GRN_file  
  ./demo_data/BEELINE/500_ChIP-seq_mDC/outputs/pretrain_output.pkl  --data_dir 
  ./demo_data/BEELINE/500_ChIP-seq_mDC/inputs/ --train_y_file  train_y_non.npy --output_file 
  ./demo_data/BEELINE/500_ChIP-seq_mDC/outputs/performance.pkl

  where '--pre_GRN_file' denotes the output file of the previous self-supervised step,
  '--data_dir' denotes the directory which includes all training data,
  '--train_y_file' denotes the name of the training label file,
  '--output_file' denotes the output file name.
  It should be finished in several minutes for the given demo dataset.
  
  If you have any questions, please feel free to contact me.
  Email: shuht96@gmail.com

Other baseline single-cell RNA-seq packages

• BEELINE https://github.com/Murali-group/Beeline
• DeepSEM https://github.com/HantaoShu/DeepSEM
• CNNC https://github.com/xiaoyeye/CNNC
• DeepDRIM https://github.com/jiaxchen2-c/DeepDRIM