evogytis
commented
10 years ago Yeah, I've thought about this before (hence the option to timeslice the alignment in the LD calculator). Three points on this:
- In the example you give only D' would give strong LD (because aB is not be observed), r^2 (which is roughly what our statistic is) will give LD that depends on the allele frequencies, for example if the transitional form Ab in your example is rare LD will be higher because ab and AB will be more frequent (which is close to what complete LD looks like). I think because r^2 quantifies P(a|b) it makes it resilient to temporal dynamics, as long as alleles arise relatively close in time to each other.
- There's a problem of visualization - I have a script that can produce animated heatmaps and I could animate the LD circles over time, which I bet would look really cool, but would make it difficult to include in a publication in a seamless way. Averaging seems like a good idea,
- By not timeslicing we are saying that the side branches are not that interesting - they're excluded either because they don't persist long enough to accumulate amino acid substitutions or if they do because they are below the 1% minor allele frequency cutoff (in a dataset of 1600 sequences it means that there have to be at least 16 sequences with a particular allele to calculate LD). Partitioning the sequences into time bins will be sensitive to whatever's been sampled within the bin and will also yield high LD between site pairs that belong to divergent lineages, which is not that interesting. Keeping the data as they are gives us an idea of what's persisted for long periods of time and you can think of it as the LD between trunks of each phylogenetic tree.
Having said that point 2 sounds very appealing to me. We could try to address the problems of binning in point 3 by raising the minor allele frequency cutoff and maybe doing sliding window LD rather than time binned LD. I imagine it would look like a plasma globe - a handful of sites on PB1, PB2 and HA will maintain strong links in the data whilst others belonging to transient lineages will appear and fade as you go forwards in time. I'm fine with either binning or keeping the data as it is, if you're more comfortable with binned data I can re-run the LD analyses.
trvrb
Reading your paragraph on D' made me question the LD analysis a bit:
Your analysis of ΔTMRCA nicely controls for differences in sampling date between tips, however, the LD analysis just lumps everything together. Let's say we start with ab and then Ab fixes and then AB fixes, so that most early samples are ab and most late samples are AB. This will appear as strong LD between loci A and B, even though it's just temporal sampling. Would it make sense to bin years before calculating LD and then average across years?