Tricycle Predictions and Integrated scRNA Data

[gomeznick86]

Hello!

I'm trying to apply tricycle to our data and am interested in understanding if I'm choosing the correct workflow.

I have 3 different scRNA-seq data sets that are each normalized by SCT and integrated with Seurat. My question, when and which assay to use in order run tricycle? I can run tricycle on each of the objects individually, prior to integration (top left). I can run it post integration using assay SCT and data slot (top right) or use the integrated RNA assay and data (bottom left) or use the integrated assay data (bottom right). You can see that I get quite different results depending on the assay and whether or not I integrate the data. For example after integration the blue line drops to 0 from 0.5pi to 1.5pi which indicates to me that some transformation is happening that makes it difficult for tricycle to assign the proper score?

I tried looking at the code used in the paper and it wasn't immediately clear to me if you performed tricycle on the integrated object or not. Apologies if I missed it. Any guidance would be appreciated!

Thanks

Nick

[kasperdanielhansen]

Could you send the projection plots (the ones that look like a circle / ellipsoid)?

…

On Mon, Jun 12, 2023 at 7:14 PM Nick Gomez ***@***.***> wrote: Hello! I'm trying to apply tricycle to our data and am interested in understanding if I'm choosing the correct workflow. I have 3 different scRNA-seq data sets that are each normalized by SCT and integrated with Seurat. My question, when and which assay to use in order run tricycle? I can run tricycle on each of the objects individually, prior to integration (top left). I can run it post integration using assay SCT and data slot (top right) or use the integrated RNA assay and data (bottom left) or use the integrated assay data (bottom right). You can see that I get quite different results depending on the assay and whether or not I integrate the data. For example after integration the blue line drops to 0 from 0.5pi to 1.5pi which indicates to me that some transformation is happening that makes it difficult for tricycle to assign the proper score? I tried looking at the code used in the paper and it wasn't immediately clear to me if you performed tricycle on the integrated object or not. Apologies if I missed it. Any guidance would be appreciated! Thanks [image: Screen Shot 2023-06-12 at 4 13 25 PM] <https://user-images.githubusercontent.com/57917990/245300681-d812003e-27f8-466c-b75a-25f2d29bb610.png> Nick — Reply to this email directly, view it on GitHub <#14>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABF2DHYQAXC4VLIH2URTQILXK6PGVANCNFSM6AAAAAAZEBLTXQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

-- Best, Kasper

[gomeznick86]

@kasperdanielhansen Of course. Thanks for the quick reply. I've included the projections below in the same order as above. Top Left is the three data sets independently projected using SCT assay. Top right is integrated dataset but using SCT assay. bottom left is the integrated dataset using the RNA assay, and bottom right is integrated data using the integrated assay. These are cultured hiPSCs which are definitely cycling if thats useful. ( I also appreciate that the top right looks weird. Or maybe they all do?)

[kasperdanielhansen]

My first recommendation would be to run tricycle on the samples prior to your normalization and integration steps (as we generally recommend). I would be quite curious to see what this yields. I don't think there is any issue with estimating cell cycle time per cell prior to integration, but then using the integration to do whatever downstream analysis you need to do.

My comments on your plots

1. The projections - as plotted - are some kind of ellipsoid, but without a hole in the donut. This could either be because of low information content per cell (low sequencing coverage) or because you have lots and lots of cells.
2. Some of your projections look awful. I have never seen something like the "Integrated SCT" plot (the three separate clouds).
3. some of the kernel densities are really messed up as well.

I would assume this is caused by the use of SCT / integration. I don't think this has anything to do with the underlying data.

It also looks like you're interested in comparing the kernel densities across samples (first set of plots). This is something we have been interested in as well - indeed it is a classic approach to estimating cell cycle length - but we have observed some potential issues. More particularly, we still don't really understand the variation between replicates. Sometimes we observe substantial variation between replicates and sometimes we see no variation, depending on the experiment. As of right now, we don't really fully understand this variation. With enough replicates (haha) we can of course overcome such variation, but that is unusual for scRNA. With the very small number of replicates we typically see, we need a better understanding of between-replicate variation before we conclude something about systematic shifts in cell cycle length. We're working on this, but don't hold your breath - I don't feel we are close to understanding it completely.