Restarting a previous discussion on here:
How convinced are #SingleCell people by cell-cell communication methods? I'm still at "not very" 😅
I see them used everywhere, but I don't know of any really good way to validate them...
More importantly, I don't know of any *non-simulated* training data.
How convinced are #SingleCell people by cell-cell communication methods? I'm still at "not very" 😅
I see them used everywhere, but I don't know of any really good way to validate them...
More importantly, I don't know of any *non-simulated* training data.
Comments
In astrocytes? or elsewhere as well?
As I said elsewhere, I don't think I would have a problem with these methods if they were called "cell-cell association" or "cell-cell correlation" methods 😅 Reduces the risk of over-interpretation...
I also think single cell stuff is finally getting to the point where it can be "comprehensively descriptive", and that's much more useful.
Recently Julio Aguirre-Ghiso's lab validated predictions from our NATMI tool in vivo.
https://www.cell.com/cell/abstract/S0092-8674(24)01034-1
What kind of validation do you do? How do you prove that there is communication, and not just correlation?
But the abstract has statements like “learns and informs about the signaling events” and “represent diverse mechanisms of cell-cell communication”. You just can’t properly show that, no?
But I think claims like this are overselling, which follows a long and established tradition for the single cell field… 😅
The evaluations that we (mainly D. Dimitrov) have proposed within LIANA (https://www.nature.com/articles/s41556-024-01469-w) are mainly testing ...
It is why we can do doublet ID really well, and why there isn't any real justification for widely used QC thresholds beyond "they make the QC plots look more pretty".
(It's why we need more gold-standard datasets...)
You need to have good truth for which pairs of genes are interacting between all pairs of celltypes, and maybe harder, which genes *are not* interacting. How do you prove this?
I know nothing about the topic, and I can think of: ligand-receptor (cytokine release); gap junctions; extracellular vesicles; nanotubules 😎. I'm sure there are more.
This also seems hard, particularly showing an *absence* of interaction. Also v hard to do at scale.
I guess spatial data can really help. This is likely a blind spot for me - I don't follow spatial work very closely, so I've likely missed useful things.
It would be great to hear thoughts of others 😊
I think I wouldn't have a problem if they were described as "cell-cell association" methods. It's the claim of communication, often without *any* validation, that annoys me 😅
Bottom line: mRNA L-R pairs give a good experimental shortlist but aren’t always biologically functional. You still need to do the experiment!
Last time I looked, they differed quite a lot between CellChat and NicheNet so there’s definitely a need for more manual curation.