Incoherent cell positions between different visualization layers ?

[philmar1]

Hello,

I am working on scATACseq and I tried MS Phate. It seems that some of my clusters switch positions for different visualization layers. I know the ground truth labels of my cells. 4 of them are:

• intermediate mono
• CD4+ memory T
• CD8+ naive T
• CD4+ naive T

For the upper plots, I use the visualization level 0, i.e each dot represents a cell For the lower plots, I use a coarser visualization level, i.e each dot represents cells that were merged during the condensation phase. For each pair of plots within the same column, I try to confirm that cells belonging to the same orange cluster are found in the same region of the associated condensed 2D plot.

As it can be observed, it is not the case. It is very obvious for CD8+ naive and Intermediate mono cells also seem to have moved their position in the 2D plot after condensation

Is this behavior expected?

Thanks a lot for your help

[mkuchroo]

Hi Phillipe,

Thank you for reaching out and using our tool. I apologize, I don't completely understand your question.

The position of points in multiscale PHATE embeddings change from layer to layer and this is expected due to the MDS optimization process for embedding condensed diffusion potential. The cluster/groupings should not change in this way - the clusters form a condensation tree which should have a relatively logical agglomeration structure.

Not sure if this helped. Please feel free to clarify your question if this was not helpful! Manik

On Mon, Dec 4, 2023 at 11:08 AM Philippe Martin @.***> wrote:

Hello,

I am working on scATACseq and I tried MS Phate. It seems that some of my clusters switch positions for different visualization layers. I know the ground truth labels of my cells. 4 of them are:

• intermediate mono
• CD4+ memory T
• CD8+ naive T
• CD4+ naive T

For the upper plots, I use the visualization level 0, i.e each dot represents a cell For the lower plots, I use a coarser visualization level, i.e each dot represents cells that were merged during the condensation phase. For each pair of plots within the same column, I try to confirm that cells belonging to the same orange cluster are found in the same region of the associated condensed 2D plot.

As it can be observed, it is not the case. It is very obvious for CD8+ naive and Intermediate mono cells also seem to have moved their position in the 2D plot after condensation

image.png (view on web) https://github.com/KrishnaswamyLab/Multiscale_PHATE/assets/43274675/cc5daea4-47cb-4d03-8bb9-e8f8934fb68e image.png (view on web) https://github.com/KrishnaswamyLab/Multiscale_PHATE/assets/43274675/56b2b45e-8a19-4f5e-95b1-823ddc458fe3

Is this behavior expected?

Thanks a lot for your help

— Reply to this email directly, view it on GitHub https://github.com/KrishnaswamyLab/Multiscale_PHATE/issues/16, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAWDLTBIAZQB47R452CJAODYHXYOFAVCNFSM6AAAAABAGHISOCVHI2DSMVQWIX3LMV43ASLTON2WKOZSGAZDIMRRGE4DAMA . You are receiving this because you are subscribed to this thread.Message ID: @.***>

[philmar1]

Hello Manik,

Thanks for the quick feedback. Let me try to clarify my question. I expect the clusters to condense around a central point, like a gravitational point. For instance, I expect all points in the upper right corner to condense and stay in the upper right zone of the image. It would be surprising that they condense and end up in the lower left corner zone, am I right?

This is proper behavior is observed in most of the data points, as below:

However, for some points, the unexpected behavior appears, such as here:

Is the problem more clear now? Therefore, my question is: is my intuition wrong on how should the points condense? Can it theoretically happen that for instance, points from an upper right zone will condense and end in lower left zone of a 2D plot ?

Thanks a lot for your help, I hope it's easier to understand now

Philippe

[mkuchroo]

Hi Philippe,

Of course, glad you are using our tool!

Yes as I suspected this is an expected phenomena due to our use of MDS as a means to embed diffusion potential into 2 coordinate axes!

Nothing to be concerned about. PHATE and Multiscale PHATE unlike UMAP and tSNE produce largely reproducible embeddings but they often are not exactly similar and may occasionally flip structures that are disconnected from the remainder of the graph.

I hope this was helpful! Manik

On Wed, Dec 6, 2023 at 5:38 AM Philippe Martin @.***> wrote:

Hello Manik,

Thanks for the quick feedback. Let me try to clarify my question. I expect the clusters to condense around a central point, like a gravitational point. For instance, I expect all points in the upper right corner to condense and stay in the upper right zone of the image. It would be surprising that they condense and end up in the lower left corner zone, am I right?

This is proper behavior is observed in most of the data points, as below: Capture.d.ecran.2023-12-06.a.11.23.52.png (view on web) https://github.com/KrishnaswamyLab/Multiscale_PHATE/assets/43274675/2c161534-5ae7-4db2-b024-ca6e69f700cd

However, for some points, the unexpected behavior appears, such as here: Capture.d.ecran.2023-12-06.a.11.26.41.png (view on web) https://github.com/KrishnaswamyLab/Multiscale_PHATE/assets/43274675/96046ed6-65c2-42c4-a558-60a95dd7735e

Is the problem more clear now? Therefore, my question is: is my intuition wrong on how should the points condense? Can it theoretically happen that for instance, points from an upper right zone will condense and end in lower left zone of a 2D plot ?

Thanks a lot for your help, I hope it's easier to understand now

Philippe

— Reply to this email directly, view it on GitHub https://github.com/KrishnaswamyLab/Multiscale_PHATE/issues/16#issuecomment-1842613965, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAWDLTCIGQOW2HHGJ6ST5ILYIBDI5AVCNFSM6AAAAABAGHISOCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQNBSGYYTGOJWGU . You are receiving this because you commented.Message ID: @.***>

[philmar1]

Hello Manik,

Thanks for you answer ! I believe this behavior may induce to errors and misunderstandings. Indeed, I didn't know that and I first thought that some cells start to emit a signal only when they are aggregated (i.e after condensation). I believe it is important that cells in a specific layer condense in a new layer close to their original location. Therefore, we won't have this phenomenon of flipping or moving positions.

Do you plan to correct for that behavior ?

Thanks again a lot for your time

Philippe