Dear Kurban, 

I have never personally worked with NCA, but as far as I know, it requires (besides gene expression data) a prior knowledge about the network structure to work properly (such as an expression-independent blueprint gene regulatory network based on ChIP, DapSeq or promoter binding predictions using transcription factor PWMs). These blueprint networks are well established for model species, such as E. coli or S. cerevisiae, but to a lesser degree for other species. Do you have such data with reasonable coverage for your species of interest?

You might want to have a look at gene regulatory network inference approaches solely based on gene expression as well as multiple dataset integration. If you are interested, we have recently published a review on such methods:

Computational inference of gene regulatory networks: Approaches, limitations and opportunities 

http://www.sciencedirect.com/science/article/pii/S1874939916301882

Methods that might be of particular interest for your problem could be random forest regression approaches (such as the GENIE3 approach), as it evaluates not only individual TF-TG relationships, but also compares the combined effect of multiple TFs on the TG's expression level (therefore, gives an estimate of possible cooperative target gene regulation). It uses an exhaustive bootstrap strategy to perform inference on subsamples of the entire gene expression datasets with make it ideal to identify specific regulation. If you use R, I have implemented a faster random forest regression GRN inference approach similar to GENIE3 (https://github.com/mbanf/GRACE/). I use it to construct an initial GRN that is later post-processed to increase accuracy using additional datasets. There are also Matlab and Python versions available by the author of GENIE3 (http://www.montefiore.ulg.ac.be/~huynh-thu/software.html)

Other relatively fast approaches would be CLR (context likelihood of relatedness). You could apply the approach to both, coexpression networks as well as relevance networks based on mutual information. CLR implements a genome wide background correction per TF-TG pair which can have a huge effect. 

Best, 

Michael

If your number of samples are more than 17 then you can perform WGCNA, that is more powerful than the statistics you are mentioning. Its extremely easy to handle in R.

It would also give you nodes and edges.

you can then use cytoscape or 'igraph' package from R to construct the network with directions (It has many features, just go through it)

If  you already have TF and TG, then its even easier, just make two coloumns as nodes and edges, make third coloumn as power of driver and draw it through cytoscape or igraph

hth

The WGCNA suite could indeed be nice, but it will not provide any directional information. It is basically a coexpression network approach, using soft thresholding, which is to exponentiate correlations afterwards and then apply a cutoff. Directionality from transcriptomics data alone, unless time series data, is more or less infeasible (although regression approaches tend to provide at least weak directionality indication)