Spectral embedding with mixed high/low depth nuclei samples

[wangmeijiao]

Hi, Kai and all,
I understand that SnapATACv2 uses spectral embedding for snATAC-seq datasets and performs greatly even for low depth as few as 1000 fragments per nuclei. Here is my question: in my case we have both high and low sequence depth nuclei (tsse > 5, min_counts = 3000) in twelve dataset. We removed batch effect with harmony (harmonypy.run_harmony(data.obsm['X_spectral'][:,range(18)], data.obs, 'sample', max_iter_harmony = 20)) . The resulting umap seems sequence-biased (but not sample biased!): the adjacent clusters show a contiuous depth destribution and one extremlly low-depth cluster separated at the farest point. This may cause problem when one try to identify differentially accessible regions (DARs). Please help with some comments (is it a common case? ) and suggestions (how to deal with high/low depth mixed dataset) and thank you!!

Meijiao

[kaizhang]

This is uncommon. You can try to use distance_metric='jaccard' in tl.spectral to see if it gives better result.

[wangmeijiao]

Hi Kai,
It seems that tl.spectral takes distance_metric = 'jaccard' as default (SnapATAC v2.2.0). I did all analysis steps as the manual you provide (https://kzhang.org/SnapATAC2/tutorials/integration.html). It is indeed quite strange to give this pattern, but, the snRNA-seq part of this same dataset shows the same pattern (one low UMI cluster far from the center), what is more, a published dataset show the same one (see below). Besides, this low UMI/fragment cluster show higer Mt gene percentage. Taking together, do you think this may be the truth but a apoptosis cell/nuclei population?

Thanks, Kai!!

Meijiao

[kaizhang]

It is certainly possible that they represent different cell states. One way to test this may be aggregating these two populations into pseudo-bulk and see if there are significantly differential genes.

[wangmeijiao]

Thanks, Kai, I will have a try.