Add text

[msubirana]

added sections 2 and 3 and figures 2 and 3

[miltondp]

Looks good. See my suggestion to fix a DOI (remember to click on "Commit suggestion" and verify in the "Files changed" tab of the PR that the change is included).

[miltondp]

• In the inferred patterns/scatter plots, you have points/dots that represent samples and red lines that highlight the pattern you want people to see. The linear pattern in the real biological networks does not have points/samples.
• There is a large white gap on the right of the real network. I think we could fill it with some additional examples of how gene-gene relationships might look like in the transcriptome data. I suggest adding the quadratic example you have below (usually, though, it has a U shape, not the other way around). Another example is noncoexistence patterns (see Figure 1 in the CCC paper).
• For the current nonlinear pattern, which we call "two lines" in the CCC paper, with the horizontal line of points at the bottom and the linear pattern embedded, I would highlight that the horizontal line has female samples while the embedded linear pattern has male samples, and the two genes could be from the X and Y chromosomes (gene in X chromosome in x-axis, and gene in the Y chromosome in y-axis). An example is KDM6A (X chr) and UTY (Y). This is one of the examples from the CCC paper.
• Maybe it's not necessary to say "Gene module 3"; we can have "Gene module" only.
• The "Lost" and "Captured" should match the linear and nonlinear patterns from the real network, right? Then, PLIER should capture everything except the nonlinear patterns (quadratic and two lines). Vanilla VAE should capture everything (but it's not interpretable). And the Interpretable VAE should have all the benefits (captures all patterns and it is interpretable).
  • The inferred patterns column should include all the examples that the method captures.
• Instead of crossing out "Biological network" for the Vanilla VAE, maybe we can leave the text in red font and add a question mark at the end to indicate uncertainty. Since the module might represent a real network, but we cannot interpret it.

[msubirana]

about this: "Vanilla VAE should capture everything", sutanu told me that VAE only captures non-linear

[miltondp]

A VAE also captures linear patterns since it's a more general approach for dimensionality reduction. A VAE is also more flexible: you can design your VAE (the architecture, hidden layers, etc) to mimic PCA and find equivalent, linear-only spaces.

[msubirana]

perhaps we can only add some of the linear patterns to show that decomposition methods are superior in inferring linear patterns (if it is true)

[miltondp]

It's a good point. Not sure how to convey that idea.

[msubirana]

to commit the new fig3.svg do I have to git pull the branch and modify in my repo or change it through the website?

[miltondp]

Before doing anything, just in case, make sure to back up the figure file before doing anything.

If you have a new fig3.svg in your computer and you want to push it to the PR, then checkout the branch add_text, pull (just in case someone else pushed to the same branch), modify the figure in your computer (or override with the new one), commit, and push (this will update branch add_text). The PR will be automatically updated here since you pushed to the linked branch.

[miltondp]

@msubirana I skimmed through the changes here. I plan to read all entirely after all the content is pushed.

[msubirana]

ok