K1 - disconnected nodes in a cluster

[ewysocka]

Hi, While analysing the output of K1 method, I've noticed that cluster nodes are disconnected when I take their induced subgraph on the same network used to produce clusterings. Does the method assume this? Thanks!

[mattiat]

Hi, thank you for your question. I find this surprising: I will contact the algorithm developer. In order for us to reproduce, could you share the input network, the parameters and the output that was produced? Thank you, Mattia

[ewysocka]

Hi, I just forked your repo and created branch with the data (input and output). I added some code (R) to show the problem on a specific cluster that is disconnected. As for the parameters, I run K1 with the default ones. thanks Emilia

[jjc2718]

Hi @ewysocka - thanks for the question and for the code to reproduce. Our K1 algorithm (and in particular the DSD distance metric underlying it) is intended to capture similarities in global network structure rather than local interactions, so there's nothing about it that requires clusters to be connected in the original network.

I'd be a bit surprised if what you're showing happens often, although we haven't looked at this in detail, but occasionally seeing disconnected clusters doesn't necessarily concern me. I'd say this is one of the advantages of our method, as opposed to relying mostly on first- and second-order network interactions as some other module detection methods do.

Let me know if that doesn't fully answer your question. Happy to chat about potential ways to fix this in your analysis if necessary, or to look further into things on our end if you'd like.

[ewysocka]

Hi @jjc2718, Thanks for very much for your explanations and reassurance that this is not something you would see as forbidden. Indeed, after reading about DSD, I thought it is the explanation but I wasn't sure if you reassemble missing links between cluster members afterwards because, surely, this doesn't happen very often with other clusters. This reassemblage would probably result with much larger clusters.

I found K1 results very interesting from the perspective of enrichment analysis. In some cases, where protein interactions do not connect nodes, they become connected when biological terms come into play. Poorly connected proteins are suddenly found in medium sized clusters.

Thanks again, Emilia

[jjc2718]

Great! Glad to see you're getting use out of our method and finding it interesting.

I'm going to close this issue for now, since it sounds like there isn't any action we need to take at the moment, but if you have additional questions feel free to reopen it and/or ping me.