CeSpGRN: Inferring cell specific GRN using single-cell gene expression data

Zhang's Lab, Georgia Institute of Technology

Developed by Ziqi Zhang, Jongseok Han

Preprint: available soon

Description

CeSpGRN is a package that is able to infer cell specific GRN using single-cell gene expression data

• src stores the inference algorithms.
• test stores the scripts (testing scripts) and results (testing results are directly generated by the scripts, too large to be pushed onto github).
  • scripts_GGM: testing script for the GGM data
  • scripts_softODE: testing script for the softODE data
  • scripts_THP-1: testing script for the THP-1 data
• simulator stores the simulation code:
  • GGM: simulator for GGM data
  • soft_boolODE: simulator for the softODE data
• data stores the real and simulated data (available upon requests)

Dependency

(required)
pytorch >= 1.15.0 
numpy >= 1.19.5
scipy >= 1.7.1
networkx >= 2.5
sklearn >= 0.24.2

(optional)
matplotlib >= 3.4.3
statsmodels >= 0.12.2

Usage

• Load in the count matrix as a numpy ndarray, the matrix should be of the shape (ncells, ngenes). e.g.

  import sys, os
  sys.path.append('./src/')
  import numpy as np 
  
  # load CeSpGRN
  import g_admm as CeSpGRN
  import kernel
  
  # read in count matrix
  counts = np.load("counts.npy")

• Set the hyper-parameter including: bandwidth, neighborhoodsize, and lambda. e.g.

  # smaller bandwidth means that GRN of cells are more heterogeneous.
  bandwidth = 1
  # number of neighbor being considered when calculating the covariance matrix.
  n_neigh = 30
  # sparsity regulatorization, larger lamb means sparser result.
  lamb = 0.1

• Calculate the kernel function, and covariance matrix for each cell, e.g.

  # calculate PCA of count matrix
  from sklearn.decompose import PCA
  pca_op = PCA(n_components = 10)
  X_pca = pca_op.fit_transform(counts)
  
  # using X_pca to calculate the kernel function
  K, K_trun = kernel.calc_kernel_neigh(X_pca, k = 5, bandwidth = bandwidth, truncate = True, truncate_param = n_neigh)
  
  # estimate covariance matrix, output is empir_cov of the shape (ncells, ngenes, ngenes)
  empir_cov = CeSpGRN.est_cov(X = counts, K_trun = K_trun, weighted_kt = True)

• Estimating cell-specific GRN, e.g.

  # estimate cell-specific GRNs, thetas of the shape (ncells, ngenes, ngenes)
  cespgrn = CeSpGRN.G_admm_minibatch(X=counts[:, None, :], K=K, pre_cov=empir_cov, batchsize = 120)
  thetas = cespgrn.train(max_iters=max_iters, n_intervals=100, lamb=lamb)

An example run is shown in demo.py.