,
12 tweets,
4 min read
Read on Twitter
The community structure of functional brain networks exhibits scale-specific patterns of inter- and intra-subject variability | with @max_bertolero @DaniSBassett @gordonneuro @GrattonCaterina @ndosenbach | Now in press @NeuroImage_EiC | bit.ly/30oWFxV | 1/
We use multi-layer modularity maximization to estimate communities for multiple subjects simultaneously. This avoids the sometimes challenging issue of mapping communities between subjects, gives us consensus communities for free, and makes it possible to assess variability 
We can recover known community structure, at multiple scales, without imposing any thresholds on network data.
Most importantly, we saw that the variability in community structure across subjects was scale dependent. The way subjects differed from one another depended on whether the user looks for big communities or smaller communities.
This means that to some extent the user (who chooses the parameters for the community detection algorithm) can control inter-subject variability, which has implications for studies that try to relate community variability to behavior/cognition/disease.
We saw this when we looked at HCP data and related it to behavior -- depending on whether we looked for small/large communities, the brain-behavior correlation patterns looked different.
We replicated the patterns of inter-subject variability using @club_scan data, but also found that patterns of INTRA-subject variability looked very different.
So what does all of this mean? 1) community detection is hard, but we have made some incremental steps towards making it easier. 2) the number/size of communities (which are partially user-determined) shapes brain/community-behavior correlations.
3) the implication is that we (as a field) have probably missed out on multi-scale brain-behavior associations in past studies, and this should motivate investigation community structure at different scales in future studies.
And finally, a massive 'thank you' to the three reviewers whose comments and suggestions dramatically improved this manuscript.
It's worth noting that our approach is hardly the last word on this topic. There are many other very good methods (many that are less computationally expensive) for mapping communities across individuals -- some statistical/inferential, some heuristic, etc.
Ours has an explicit "network flavor," is data-driven, and is multi-scale. All of which appeal to me, but might make this approach less palatable for others or wholly unsuitable for certain research questions.
