Read on Twitter
Actually, not transforming the data outperforms log(y/s+1). 1/
https://twitter.com/const_ae/status/1645455772468273153
The "performance" in this analysis boils down to checking consistency of the kNN graph after transformation. That's certainly a property one can optimize for, but it's by no means the only one. In fact, if it was the only property of interest, one could just not transform. 2/
Of course that is trivial and uninteresting. The purpose of normalization is to remove technical noise and stabilize variance. But then one should check how well that is done. And as it turns out, log(y/s+1) actually removes too much "noise". 3/
What does it mean to remove too much "noise"? It means that log(y/s+1) removes biological signal along with cleaning out technical noise. A 🧵on this here with @GorinGennady:
https://twitter.com/lpachter/status/16532600457345187844/
BTW more on the consistency argument in the @const_ae and @wolfgangkhuber paper here:
https://twitter.com/JanLause/status/1648265065420800000?s=205/6
A final point: what does the kNN graph even mean, and why has it become an obsession in #scRNAseq? Perhaps, as a colleague recently noted to me (in another context), the kNN graph is the last resort when all else fails... 6/6
• • •
Missing some Tweet in this thread? You can try to
force a refresh
