updated paga

[falexwolf]

Hi Wouter and Robrecht,

this project is awesome! Really cool to have this highly comparable environment... We also cite your review in the revised PAGA preprint...

Here come a few changes to the PAGA container in the context of https://github.com/dynverse/dynmethods/issues/15. Things that might need further discussion:

• you didn't have any preprocessing in the container, I added a basic recipe, which should do a reasonable job for experimental data
• you didn't compute any embedding/dimension reduction, I added the two possibilities that we recommend in the context of PAGA (UMAP and force-directed embeddings); however, an evaluation in a dimension-reduced vector space is a bit against the spirit of PAGA, which makes a case that working with graphs (i.e., dimension-less objects) is the right thing to do; still results should be decent also on UMAP or force-directed embeddings
• you didn't compute any pseudotime... PAGA can be combined with any pseudotime but we strongly recommend using a purely graph-based one... for the purpose of PAGA, I extended DPT to disconnected graphs... I added this in the container. I think that you can then simply discard the praga container, which I didn't fully get anyways (I think things are missing in there)

Thank you for all your hard work and happy to discuss further. Alex

PS: I adapted standard Python coding style for the PAGA run.py... The only thing that is still unorthodox is the 2-space indenting... even though one should use 4-space indenting with Python, I didn't change it as it should probably be consistent across the whole repository

[falexwolf]

Just a couple of more comments 🙂

• I have to admit that I haven't yet run PAGA within https://github.com/dynverse/dyno. I'll do so asap, to make sure that everything behaves correctly.

• On the https://github.com/dynverse/dyno homepage, PAGA seems to have quite high run times. I'm a bit astonished by that as, in our experience, it's way faster than everything else. On 1.3 M cells, it takes 90 s. Maybe preprocessing somehow slow or this is an effect of very small datasets...

• I extensively updated the PAGA homepage this morning: https://github.com/theislab/paga and want to grow this to examples on many different datasets (I've a few others already that are not listed). I'm happy to add a link to your tool once I played around with it myself... ;)

[codecov[bot]]

Codecov Report

Merging #77 into devel will not change coverage. The diff coverage is n/a.

@@          Coverage Diff          @@
##           devel     #77   +/-   ##
=====================================
  Coverage   0.52%   0.52%           
=====================================
  Files         34      34           
  Lines       2076    2076           
=====================================
  Hits          11      11           
  Misses      2065    2065

Impacted Files	Coverage Δ
R/ti_container.R	0% <ø> (ø)	:arrow_up:

---

Continue to review full report at Codecov.

Legend - Click here to learn more Δ = absolute <relative> (impact), ø = not affected, ? = missing data Powered by Codecov. Last update 6e22f7f...606596e. Read the comment docs.

[zouter]

Hi Alex Looks great! Really awesome that you're contributing to the project.

I'm fixing some small things now to the changes (eg. the parameter type of embedding type from string to discrete). Could you perhaps grant us the right to change your pull request (https://help.github.com/articles/allowing-changes-to-a-pull-request-branch-created-from-a-fork/ ) so that we can push those changes?

I think in the current state, even though the pseudotime gets calculated, it will not be used in the output as it still outputs a "cluster_graph" instead of a trajectory. I will also try to fix this. Am I correct that for the calculation of the pseudotime the algorithm needs a root cell? Would it make sense to make this optional prior information and pick a random cell if no root cell is given? Or would prefer it to be a required prior information?

About the run times, you're right that this seems quite high. However, these values are now quite old and were calculated using way earlier versions of both the wrapper and PAGA (heck, we still use the name AGA :stuck_out_tongue: ). @rcannood is actually right now working on a more in depth scalability analyses, rather than a naive "average running time", so stay tuned!

The update homepage: :heart_eyes:

I will keep you updated!

[falexwolf]

Thanks!

• Hope you are now able to do the edits.
• For the pseudotime: right! One needs to choose a root cell! This is prior knowledge. Do you have a way of passing the root cell as a parameter? A random root cell gives you meaningful orderings from that cell, but depending on how you evaluate, this will likely score very badly.