camillescott / dissertation Goto Github PK

ref fig-chap1-unitig-frag

This looks pretty interesting! I think some more discussion in the figure caption would be great, though. In particular, is there a takeaway?

Please put axis labels, esp on y axis!

Is there a different way to represent this? Perhaps in an additional figure, since I think this one is also pretty informative? What if you made one where the curves were weighted by unitig size?

Also, I have questions -

• why is the proportion of small nodes increasing for yeast towards the end??

I think this figure would be much easier to parse first time 'round if you could invert the y axis. Alternatively state in the caption that "larger numbers are faster/better".

A few minor questions first -

• so, the sum of the n_ lines should equal one? and kmer_p always ends at x=1, y=1.
• if correct, would suggest adding 'proportion of nodes' as y axis label.
• might also make them a bit bigger so they take up the page width?

I think the discussion of this figure in the Results section is a good start, but I have lots of questions 😆 -

• yeast looks pretty different. this is because of high coverage?
• I still have some trouble understanding the dynamics after looking at this for a while. let's see -
  • n_circular should generally be small, ok
  • n_trivial isn't defined anywhere?
  • the kinks in kmer_p are from non-random sequencing, presumably? maybe verify this with shuffling?
  • wouldn't you expect a fair number of islands in RNAseq?
  • in yeast, it looks like n_islands is converging to a very different point than in the other two data sets; what gives?
  • maybe: if this is due to higher coverage, what happens if you subsample the yeast data set a whole bunch?
  • n_tips behavior is driven by error?

anyway, I think it would be very helpful to develop some intuition, maybe in another (simpler) figure.

also, a different or perhaps complementary tack - what interesting behavior is occurring in this figure that readers should be alerted to, and what behavior is just boring and "trivial"?

First question: is this done on diginormed data, or not?

I still struggle intuitively with the saturation behavior in the leftmost figure. Why would later sketches have high similarity? Can you remind me and then explain it in the caption?

Middle figure, the curve is because you're using flat k-mers w/o abundance, right, so you get a convexity when you have seen a bunch of the k-mers.

Right figure, why on earth is it decreasing a bit towards x=1?

Should you maybe be using containment instead of Jaccard? Seems easier to explain to me.

• The primary and most intensive job of compaction is finding the decision k-mers
  • Extracting sequence is incidental
  • Tracking unitig metrics is potentially useful for downstream
• Streaming compaction builds a foundation for streaming assembly
• Streaming compaction can be used downstream from lighter-weight sub-linear methods
  • Original idea: using compaction metrics for guessing
  • Future idea (actually chapter 2): sketching methods for sub-linear compaction / assembly
• Streaming pipelines enable real-time feedback
  • Stopping early to preserve computational resources
  • Feedback loop w/third generation sequencing instruments

Why switch to Bray-Curtis from Jaccard? Why use log? What does this figure mean? I haz questions!

I think I understand this behavior 🎉

Q -
What does "moderately sized" mean in the caption? Why is it important?

aaaaaand what is the takeaway from this figure?