Cells with no neighbors

[shahrozeabbas]

Hoping to find someone else having a similar problem....has anyone encountered scATAC datasets where after calculating the neighborhood graph, you find that some cells have no neighbors? If so, are there reasons or solutions to this? I suspect that it may have to do with the features/number of dimensions used for neighborhood graph construction.

[timoast]

We use a k-nearest neighbor algorithm, so each cell will have k neighbors by definition

[shahrozeabbas]

Yeah that's what I expected, that's why I've been confused recently.

I had to switch over to using scanpy because the size of the dataset is over a million cells. The neighbors method relies on UMAP to calculate neighbors, but the idea should still be the same. It just doesn't make sense.

[shahrozeabbas]

I am performing the TFIDF normalization and opted to subset features per dataset by using the FindVariableFeatures function to reduce the size of the dataset after merging. I then take the union across all datasets. You think that may have something to do with it?

[timoast]

Typically we do not run FindVariableFeatures on scATAC-seq data as the mean and variance are very low across all peaks. You can run FindTopFeatures to select features based on total number of counts instead

[shahrozeabbas]

It turns out this was the reason for the problem I was seeing! Good to know for anyone else who might experience it as well.

[shahrozeabbas]

Hi Tim,

Any thoughts on poisson gene selection from scvi for selecting top features in ATAC?

Thanks!