Will it be beneficial to use all genomics bins, instead of most variable feature selection?

[wangmeijiao]

Hi Kai and all,
It seems that it is quite important to select features for downstream dimension reduction and other. If I understood correct the hints you raised in the discussion thread "#116", both top ranked variable features (needs tunning) and iteratively selection method can do good to the next dimension reduction step. Here I ask that why not just select all genomic bins (with some filtering of black list regions) as SnapATAC version 1 does? if CPU and memory are not the problems.

Best,
Meijiao

[kaizhang]

SnapATAC version 1 also filters bins. Some bins, especially those with low counts, may be very noisy. If you don't filter them, you will see a lot of noisy structures in your clustering result.

[wangmeijiao]

Thanks Kai. Yes snapatac v1 does a lot of pre-filtering works with its 5000bp genomic bins. I am only feel confused with my own data: a very homogeneous cell population with similar cell types (unlike tremendous diversity among cell types in your paper Kai et al 2021 ): ). So could you please help me with feature selection suggestions to improve the cell inner struction? We do observe some clusters distributing far away from others clusters in snapatac v1 results, however, it is disappeared to mix with others in snapatac v2.2.0 results. We do snapatac2.2.0 with default parameters and about 1M 500bp genomic variable bins produced). I decide to reanalysis with snapatac v1 before your suggest and will report the difference for you then. Hope this will help you to improve this method and help others too.

Best
Meijiao

[wangmeijiao]

And will it be possible to rewrite the original dimension reduction algorithm (jaccard similarity + diffusion map ) to snapatac2 for backward compatibility, in case of someone's specific demands (for homogeneous data)?

[kaizhang]

SnapATAC2's dimension reduction algorithm should produce better results than SnapATAC version 1 in any situation. But the feature selection algorithm in v2.2 has some issues, which have been improved in v2.3. I will release v2.3 by the end of this month. Or you can try the latest developmental version to see the difference. If you feel SnapATAC version 1 produces better results, please feel free to upload the codes and results here. I will investigate it. Thank you!

[wangmeijiao]

Thanks Kai, I will report the result (and code) after my trials. SnapATAC2 is definitely fast, simple to use and creative. I will try the v2.2.3 dev version then.

[kaizhang]

v2.3 has been released. You can now install it using pip.