cailab-tamu / sctenifoldnet Goto Github PK

View Code? Open in Web Editor NEW

21.0 21.0 4.0 1.13 GB

A R/MATLAB package to construct and compare gene regulatory networks from single-cell transcriptomic data

R 100.00%

differential-regulation-analysis gene-regulatory-networks manifold-alignment single-cell tensor-decomposition

sctenifoldnet's People

Contributors

Stargazers

Watchers

Forkers

ccshao aceculuses leonrunning

sctenifoldnet's Issues

null hypothesis with significant DE

hello, thank you for developing and maintaining the tool!
I tried to use my dataset for simulating a situation like your outputH0 example (using in X and Y the same input dataset). I am a bit puzzled by the result, as it returns a set of significant DE features (111 genes over a dataset of 6000 genes X 3000 cells).
do you have any idea why is that the case, I would have expected no difference as you showed in the tutorial?

diagonal of adjacency matrix is not zero any more after tensor decomposition

Dear dev,

In the intial outputs of network construction, the diagonal of the square adjacency matrix is 0, whic makes sense.

However, after tensor decomposition, the diagonal of returned consensus matrix is not zero any more. It is difficult
for me to intepret those value on the diagonal, as the value are made up by cp. It seems these value are used in the later steps in manifoldAlignment without setting back to 0 or other preprocessing.

An non zero value on the diagonal would normally suggest self-regulation. How do you think of the resulted regulation network by tensor decomposition?

See an example taken from the package.

nCells = 2000
nGenes = 100
set.seed(1)
X <- rnbinom(n = nGenes * nCells, size = 20, prob = 0.98)
X <- round(X)
X <- matrix(X, ncol = nCells)
rownames(X) <- c(paste0('ng', 1:90), paste0('mt-', 1:10))

# Computing 3 single-cell gene regulatory networks each one from a subsample of 500 cells
mnOutput <- makeNetworks(X = X,
                         nNet = 3, 
                         nCells = 500, 
                         nComp = 3, 
                         scaleScores = TRUE, 
                         symmetric = FALSE, 
                         q = 0.95
                         )

# Computing a K = 3 CANDECOMP/PARAFAC (CP) Tensor Decomposition
tdOutput <- tensorDecomposition(mnOutput, K = 3, maxError = 1e5, maxIter = 1e3)

# All 0 for all matrix in mnOutput
diag(as.matrix(mnOutput[[1]]))
  ng1   ng2   ng3   ng4   ng5   ng6   ng7   ng8   ng9  ng10  ng11  ng12  ng13 
    0     0     0     0     0     0     0     0     0     0     0     0     0 
 ng14  ng15  ng16  ng17  ng18  ng19  ng20  ng21  ng22  ng23  ng24  ng25  ng26 
    0     0     0     0     0     0     0     0     0     0     0     0     0 
 ng27  ng28  ng29  ng30  ng31  ng32  ng33  ng34  ng35  ng36  ng37  ng38  ng39 
    0     0     0     0     0     0     0     0     0     0     0     0     0 
 ng40  ng41  ng42  ng43  ng44  ng45  ng46  ng47  ng48  ng49  ng50  ng51  ng52 
    0     0     0     0     0     0     0     0     0     0     0     0     0 
 ng53  ng54  ng55  ng56  ng57  ng58  ng59  ng60  ng61  ng62  ng63  ng64  ng65 
    0     0     0     0     0     0     0     0     0     0     0     0     0 
 ng66  ng67  ng68  ng69  ng70  ng71  ng72  ng73  ng74  ng75  ng76  ng77  ng78 
    0     0     0     0     0     0     0     0     0     0     0     0     0 
 ng79  ng80  ng81  ng82  ng83  ng84  ng85  ng86  ng87  ng88  ng89  ng90  mt-1 
    0     0     0     0     0     0     0     0     0     0     0     0     0 
 mt-2  mt-3  mt-4  mt-5  mt-6  mt-7  mt-8  mt-9 mt-10 
    0     0     0     0     0     0     0     0     0 

#- now the value are non zero.
diag(as.matrix(tdOutput[[1]]))
  ng1   ng2   ng3   ng4   ng5   ng6   ng7   ng8   ng9  ng10  ng11  ng12  ng13 
  0.2   0.0   0.0   0.3   0.0   0.0   0.2   0.3   0.0   0.0   0.0   0.0   0.1 
 ng14  ng15  ng16  ng17  ng18  ng19  ng20  ng21  ng22  ng23  ng24  ng25  ng26 
  0.0   0.0   0.0   0.0   0.1   0.0   0.1   0.0   0.1   0.0   0.0   0.0   0.0 
 ng27  ng28  ng29  ng30  ng31  ng32  ng33  ng34  ng35  ng36  ng37  ng38  ng39 
  0.0   0.0   0.0   0.1   0.0   0.1   0.2   0.0   0.0   0.0   0.0   0.0   0.4 
 ng40  ng41  ng42  ng43  ng44  ng45  ng46  ng47  ng48  ng49  ng50  ng51  ng52 
  0.0   0.0   0.1   0.3   0.0   0.3   0.2   0.1   0.0   0.0   0.0   0.2   0.3 
 ng53  ng54  ng55  ng56  ng57  ng58  ng59  ng60  ng61  ng62  ng63  ng64  ng65 
  0.4   0.0   0.3   0.0   0.0   0.0   0.0   0.0   0.5   0.0   0.0   0.0   0.0 
 ng66  ng67  ng68  ng69  ng70  ng71  ng72  ng73  ng74  ng75  ng76  ng77  ng78 
  0.0   0.0   0.0   0.0   0.0   0.0   0.2   0.0   0.0   0.3   0.0   0.0   0.0 
 ng79  ng80  ng81  ng82  ng83  ng84  ng85  ng86  ng87  ng88  ng89  ng90  mt-1 
  0.0   0.0   0.0   0.0   0.0   0.0   0.4   0.1   0.3   0.2   0.0   0.0   1.0 
 mt-2  mt-3  mt-4  mt-5  mt-6  mt-7  mt-8  mt-9 mt-10 
  0.0   0.0   0.0   0.0   0.0   0.0   0.0   0.4   0.3

Package download size is too large due to `inst/` directory

When using devtools to install directly from GitHub, the package download is quite large due to contents of the inst/ directory. Perhaps that could be refactored somewhere else?

why using a 4d array in cp?

Dear scTenifoldNet dev,

Thanks very much for the interesting method. That is the first time I learn something about tensor low approximation,
ane I would like to try it on my work.

As I am new to the idea of tensor, I was confused on the 4d array used in the rTensor::cp steps. As the N * N * T (N is the number of genes, T is the sampling times) is 3d array, why a 4d array is used in the implementations (example codes taken from CRAN 1.0.0)?

Her are codes used In the script of tensorDecomposition.R,

#- init a 4d array with the 3rd dim is 1
  tensorX <- array(data = 0, dim = c(nGenes,nGenes,1,nNet))
  if(!is.null(yList)){
    tensorY <- array(data = 0, dim = c(nGenes,nGenes,1,nNet))
  }
  

#- Here only the last dim is filled with correlation matrix.
  for(i in seq_len(nNet)){
    tempX <- matrix(0, nGenes, nGenes)
    rownames(tempX) <- colnames(tempX) <- sGenes
    temp <- as.matrix(xList[[i]])
    tGenes <- sGenes[sGenes %in% rownames(temp)]
    tempX[tGenes,tGenes] <- temp[tGenes,tGenes]
    tensorX[,,,i] <- tempX
    
    if(!is.null(yList)){
      tempY <- matrix(0, nGenes, nGenes)
      rownames(tempY) <- colnames(tempY) <- sGenes
      temp <- as.matrix(yList[[i]])
      tGenes <- sGenes[sGenes %in% rownames(temp)]
      tempY[tGenes,tGenes] <- temp[tGenes,tGenes]
      tensorY[,,,i] <- tempY
    }
    
  }

  tensorX <- rTensor::as.tensor(tensorX)
  set.seed(1)
  tensorX <- rTensor::cp(tnsr = tensorX, num_components = K, max_iter = maxIter, tol = maxError)
  tX <- tensorX$est@data[,,,1]
  for(i in seq_len(nNet)[-1]){
    tX <- tX +  tensorX$est@data[,,,i]
  }
  tX <- tX/nNet
  tX <- tX/max(abs(tX))
  tX <- round(tX,1)
  tX <- as(tX, 'dgCMatrix')
  rownames(tX) <- colnames(tX) <- sGenes

Actually I have tried tensorDecomposition with 3d array, i.e., simly remove the 3rd dim; and replace the "1" in the 3rd to other values. In the former senario I got different results compared to your codes, and the latter case there was an error:

Error in as(tX, "dgCMatrix") :
no method or default for coercing “array” to “dgCMatrix”

Thanks in advance.

Pseudotime based subsampling

Hi,

I tried scTenifoldNet and it gives interesting results. Is there already some R implementation of a pseudotime-guided subsampling, as you have described in your preprint?

All the best,
Kevin

cailab-tamu / sctenifoldnet Goto Github PK

sctenifoldnet's People

Contributors

Stargazers

Watchers

Forkers

sctenifoldnet's Issues

null hypothesis with significant DE

diagonal of adjacency matrix is not zero any more after tensor decomposition

Package download size is too large due to `inst/` directory

why using a 4d array in cp?

Pseudotime based subsampling

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent