Which method do you use to build adjacency matrix of a genes co-expression network?

Question

[arronar]

Hello.

I'm interested in creating adjacency matrix for genes-coexpression network from a microarray data and I would like to know which methods you guys are using for such a procedure.

Till now, I figured that there is the WGCNA but I don't know if it can calculate the adjacency matrix and also there is a package called coexnet that support that can calculate the adjacency matrix but nobody is referring to it and I cannot understand if it's robust or not.

Answer 1

[Kevin Blighe]

Hi arronar,

There are various methods and they are all fundamentally based on the construction of correlation matrices.

iGraph

For one, you can follow my tutorial here, which utilises the igraph package in R:

Network plot from expression data in R using igraph (start from Step 2).

-----------

WGCNA

WGCNA has been and still is quite popular, but there are very robust tutorials that you can follow:

WGCNA: an R package for weighted correlation network analysis

-----------

STRINGdb

STRINGdb is increasingly popular and it can build your network using known protein-to-protein interactions, I believe:

STRING: functional protein association networks

-----------

Finally, if you want to follow an entire 'pipeline', then you could follow recent work that I completed with a colleague, which has just been placed on bioRxiv: New insights in Tibial muscular dystrophy revealed by protein-protein interaction networks - it's a small study but the network methods employed, utilising STRINGdb and Cytoscape + Cytoscape plugins mainly, are standardised network parameters that the biological community are only now beginning to use more and more, such as:

• Hub score
• Betweenness centrality
• Closeness centrality
• Vertex degree

*these can also be calculated via igraph in R

Kevin

[arronar]

Thank you very much for that extended list. After reading about them, I decided to use the WGCNA and more accurately its function called bicor(). Now let me ask you something about the procedure. So far on my microarray experiment, I ran RMA algorithm, completed the annotation step and divide the matrix into 3 separate matrices, one for each sample. The initial one was [3xControl, 3xTreatment1, 3xTreatment2] while now I have one matrix with 3xControl, one with 3xTreatment1 and one with 3xTreatmen2.

What I want to do as I said in my I.P is to calculate the adjacency matrix (with beta = 1) for each one of them. My question now is if there is a need for summarizing somehow the 3 repetition samples in each condition. Should I calculate the mean or something else - like MAD - for each gene (row) for the same conditions?

[Kevin Blighe]

I have not used bicor() - sorry. I also believe that there are more flexible network methods than WGCNA, despite the enormous popularity of this program. This said, I do not believe that you have to summarise your replicate (repetition) samples. If you want to summarse them, however, then summarise them by the mean value per gene.

[1769mkc]

@Kevin this is the kind of analysis which i was looking for as i kind of not sure how to go with differential expression between various stages ,the go for network analysis and show the differences or similarity , "selected from proteins with degree > 10, BC > 0.05 and CC > 0.3" so for this did you just discard those nodes which were below your prescribed threshold ? and then form a network

Second one "titin network with first neighbourhood was constructed using titin as seed in the two " so you constructed this one from Biogrid ?as a reference network ,"titin network was then merged with the TMD network using " ? this is something im looking into what kind of similarities or differences that comes with each stage of differentiation ,but im not sure how to proceed ,so far im using WGCNA but your approach is also another of doing...

[Leite]

Dear krushnach,

With regard to the parameters to filter the key nodes of the network, they come according to the number of nodes and vertices of the network. If you have a large network, you put a high degree, BC and the CC, and the same if it is a small network, you can put a smaller cut-off.

For the construction of the titin network, We used STRING v 10.5 database (Szklarczyk et al., 2015), Genemania (Montojo et al., 2014) and BioGRID v. 3.4 (Chatr-aryamontri et al., 2017), as it was just one seed, it was easier to use several databases.

And then all the files generated in those databases were exported to Cytoscape and this allows the merger of two or more networks.

Tools → Merge → Networks....

As you can see here.

Best regards,

Leite

[Kevin Blighe]

Hello krushnach. Regarding the degree, BC, and CC, these were just used to identify key nodes in the pre-existing network. There was no new network formed after this. By identifying the key nodes, one can then build a 'discussion' and 'hypothesis' around these key nodes.

For the other part on the titin network, let me see if I can get help from someone (I will ask them to reply here).

[1769mkc]

Hello @kevin yes for titin " titin network" in the paper you mentioned "titin" as seed sequence "aforementioned interaction databases as well as the BioGRID v. 3.4 " so im not clear about the seed sequence ? so as an input you gave just that single gene to BioGrid and found out the interactions ?

[Leite]

Dear krushnach80,

In our study, we did not use the "TTN" sequence as seed but rather the gene itself. And yes, just put the name of the gene of interest in BIOGRID as well as in the other databases and find the interactions.

[1769mkc]

" "TTN" sequence a" sorry for the mistake since i didn't have the concept of seed in networks so that error happened ..