search

KrishnaswamyLab / MELD

Quantifying experimental perturbations at single cell resolution
Other
105 stars 9 forks source link

UMAP instead of PHATE embedding #59

Closed michallipinski closed 1 year ago

michallipinski commented 2 years ago

Hello @dburkhardt and the Team! I've just read your beautiful paper and I hope to apply MELD and VFC to my treatment scRNA-seq data. Just to start off, I am not a computational scientist but a humble biologist, so please bear with me.🥴

The one thing I've been a little stuck with is that PHATE is not really improving the visualization of my data (I tried many different knn and density values), as a matter of fact I can no longer see the clustering structure I see with UMAP or even t-SNE. So far it doesn't really seem to be a problem because the MELD and fit_transform don't use PHATE to compute. However if I want to follow your in-depth tutorial (which I do), the benchmarking tool uses an argument "data_phate". Can I use my UMAP embedding instead? Also, is PHATE embedding necessary in the downstream VFC analysis?

Thank you so much for your methods and help. Please correct me if what I misunderstood something.

atong01 commented 2 years ago

Hi Michal,

PHATE embedding is not necessary for either Meld or VFC. You can just substitute data_phate for any embedding and this should work fine.

One caveat, MELD and VFC, like many other methods, are built on top of a graph. By default this is the same graph as is used in the PHATE construction using the graphtools package. If your UMAP graph is substantially different then you may want to build Meld on top of this instead.