Comparing two sets of CCANs from different cell type populations

[Matt R]

Hi Hannah,

I hope all is well.

I am reaching out to brainstorm some ideas on how to compare two sets of CCANs from different cell type populations. Consider the hypothetical scenario below:

1) Make cell subset 1 and compute coaccessibility between peaks followed by CCAN analysis (Louvain clustering of coaccessible peaks) 

2) Make cell subset 2 and compute coaccessibility between peaks followed by CCAN analysis (Louvain clustering of coaccessible peaks) 

3) Compare the two sets of CCANs identified between subset 1 and subset 2

I am struggling to figure out how to perform #3. Here are some of my ideas:

• Subtract the weighted adjacency matrix of subset 1 from the weighted adjacency matrix of subset 2 to find the differential correlation matrix 
  • Limitations: subset 1 matrix has different number and different combinations of peaks compared to subset 2 matrix
• Differential network graph analysis
  • Limitations: Many different methods to choose from and a lot methods assume equal number of nodes/peaks between the two compared graphs

If you could point me in the direction of some papers or software tools for this task, I would very much appreciate your help!

Best,

Matt

[hpl...@gmail.com]

Hi Matt, 

Sorry for the delay. That's tricky for sure. I was playing around with matching up the CCANs using a maximum bipartite graph algorithm a while back. It's simple but just gives you the best pairs of CCANs between the two datasets rather than doing a quantitative comparison. This paper https://genome.cshlp.org/content/29/5/857.long used an aggregation method to look for CCAN cell-type specificity which may also be relevant.  Sorry I don't have a go-to tool for this task, but hope this helps.

Best,

Hannah

[Daniel Gingerich]

Hi Hannah, 

I also would like to explore this.  I like the bipartite matching concept.  I want to merge the ccan pairs into 'consensus ccans' that have identical peaks and can be quantitatively tested.  However, merging from the maximum bipartite ccans results in consensus ccans that overlap with each other - as in some peaks occur in multiple ccans.  It also results in ccans that have a mean size of about twice that of the original.  

I was thinking of using an alternative approach from Steve Horvath consensus WGCNA networks.  He compares the adjacency matrices of the two datasets and takes the minimum values for each position in the matrix.  I could do this on the conns1 and conns2 dataframes using pmin().  Using the new conns dataframe, feed to generate_ccans() and have consensus_ccans.