2
$\begingroup$

My question is similar to this post: What does the long curve-shape t-SNE mean?

but the problem is that my data is not time series, but I also get these long shape line in tSNE plot .

enter image description here

I don't know what that means or how I can get rid of that. I suspect some structure in the data causing that. But I could not find a useful resource explaining what structure in data could cause this.

My data is single cell data, each row is a cell, each column is a biology marker (e.g. CD3, CD4), so the data is the marker expression on each cell.

$\endgroup$
1
  • $\begingroup$ maybe there is some other sequential pattern in your data that is not temporal (maybe spatial.. cells next to each other!?) difficult to say as I'm not familiar with that type of data $\endgroup$
    – oW_
    Dec 4 '19 at 0:19
1
$\begingroup$

Welcome to the community Dalide!

There is a sequential pattern in your data. As a non-Bioinformatics guy, I simply googled and saw that biomarker are also vastly used for learning the Progression of a disease (e.g. Alzheimer). I hope I am right about example and if I am, then it is not a big surprise that you see those curves which might indicate the progress of something. If you need more precise answer you better give some more info about the origin of data (or a link to your data if it is publicly available).

From ML point of view, about which I fortunately know a bit, those curves in tSNE are caused by sequential patterns.

Fun: As an equally surprising example, once I found sequential pattern in characteristics of different Wines!! (each row a Wine and each column a wine property like Alcohole%, density, etc.)

$\endgroup$
4
  • $\begingroup$ Thanks for the answers! The data is basically from many patients, each patients has before and after data, but I don't think that could be a sequential pattern, since there are only two states. I can not think of any sequential pattern that make sense in biological sense. I might need to look into those cells. May I ask, from the wine data, you found sequential pattern, is there any reasonable meaning? $\endgroup$
    – dalide
    Dec 4 '19 at 19:50
  • $\begingroup$ I need to say that if "there exist" a pattern then a statistical significance test should verify it. I didn't try that and saying only from visual inpection. You can see the code here github.com/kasramsh/Feature-Engineering/blob/master/… At the end you see plots in which missing values of a property were replaced by the previous value ("Back Fill") and there seems to be a positive correlation! however the next wine and the previous wine should not logically be in a sequence. But that was a fun example without statistical test $\endgroup$ Dec 5 '19 at 12:55
  • $\begingroup$ can you give more info about the data? so each patients has 2 entries for each marker right? What happens if you plot two separate tSNE for before and after? you still see the curves? $\endgroup$ Dec 5 '19 at 12:56
  • $\begingroup$ the data is like this, each row is a cell, each column is a biomarker, the order of row is cells from patient1 before, cells from patient1 after, cells from patient2 before, cells from patient2 after, ..... the before and after sample for each patient are not from the exact region, before and after could have different cells numbers... I plot tSNE separately, still has the curves. $\endgroup$
    – dalide
    Dec 5 '19 at 21:54

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.