FAQ  •  Register  •  Login

concatenatings samples from different runs-viSNE

Forum rules
Please be as geeky as possible. Reference, reference, reference.
Also, please note that this is a mixed bag of math-gurus and mathematically challenged, so choose your words wisely :-)
<<

narulam

Participant

Posts: 9

Joined: Sun Mar 29, 2020 10:28 am

Post Mon Mar 30, 2020 1:43 am

concatenatings samples from different runs-viSNE

Hi,

Pretty new with Cytof analysis...

I'm trying to concatenate samples from 2 different cytof experiments (same staining panel and protocol, experiments done on different days) for viSNE analysis. Unfortunately, I don't have the same technical control ran in both experiments but what I notice is that the visne plots of all samples ran in experiment 1 have the same basic structure, i.e., same populations on the same islands but all samples ran in experiment 2 show these populations on a different island. I wonder why do I get different visne structure for 2 different experiments even when they are analyzed together in the same visne analysis run. Does anyone else have this issue too?

Thanks!
M
<<

dtelad11

Master

Posts: 129

Joined: Mon Oct 31, 2016 6:26 pm

Post Tue Mar 31, 2020 12:52 am

Re: concatenatings samples from different runs-viSNE

viSNE and other dimensionality reduction techniques are highly sensitive to batch effects. This is by design: the purpose of these algorithms is to maintain distances from the high-dimensional space in the low-dimensional space. In this case, distances are strongly influenced by the batch. In a strict mathematical sense, the method is doing its job.

There are several ways in which you could proceed.

One, you could look into batch effect correction. Personally, I am not a fan of these methods, even when a technical control is available. Since you don't have one here, you will not have a dependable benchmark to validate that correction's accuracy.

Two, you could visualize the batches separately. Generate the viSNE map for each batch, go over the maps to make your observations, and keep those that appear in both maps (or figure out why a relevant observation appears in only one map).

Three, it's possible that the divergence in the viSNE map is caused by one marker (or a small set of markers). You could try to generate the map using less markers from your experiment. Start by using the minimal set (for example, CD3/19/14/56, if you're looking at the immune system). Hopefully, there will be no batch-dependent separation, and you can try adding more markers.

As an aside, I would refrain from referring to the method as "viSNE analysis". Dimensionality reduction algorithms are visualization tools, not analysis tools. Anything you do with the resulting map is the analysis part. I think that "viSNE map" or "viSNE visualization" is more suitable.

EDIT: Four, you could try a meta-clustering approach. Cluster each sample or each batch separately. Then use meta-clustering to align the clusters and visualize them using viSNE or MDS.

You might want to check out my webinar at https://www.youtube.com/watch?v=njKvRUwaYxA for more ideas on how to deal with the data. Full disclosure: I run a company that provides a commercial solution for analysis of CyTOF data and the webinar includes a sales pitch to our service.
<<

narulam

Participant

Posts: 9

Joined: Sun Mar 29, 2020 10:28 am

Post Tue Mar 31, 2020 4:24 am

Re: concatenatings samples from different runs-viSNE

Thank you very much, this is very helpful. I don't quite understand the 4th option right now but will go through the papers mentioned in your webinar.

I'm also curious if batch effect also affects the FlowSOM analysis downstream?

Return to CyTOF data analysis

Who is online

Users browsing this forum: No registered users and 7 guests