SeNNet deepcell types integration Senoquant

[anthomontan]

Hi,

I am part of the SenNet consortium, I was talking with Katy at the last SenNet meeting in Washington and since I talk with Yash.

As part of SenNet we design a plugin in Napari that does segmentation ,spot detection and quantification using mutliplex proteomics, I sent the whole code to Hubmap and they are looking to implement it in HubMap / SenNet.

The plugin gives an excel table with spatial coordinates , expression of all markers per cells (intensities various) but it doesn't include the classification since Senescence is not defined by a single marker but many (and the definition varies between labs).

Hence, we discussed to check if we could connect deepcell types with Senoquant, to get a full end to end pipeline by which :

• cell types can be defined using your model
• adding Senescent markers expression on top each cell (p21, p16, laminb1, hmgb1 .... ) since None of them alone can define senescence since it is not a cell type.

That way we can have cell types + Sen marker expression at a single cell level , then users can export data and define senescent per cell type per markers and visualize the expression all at once.

Senoquant outputs raw intensities and later we define the thresholds manually for each marker to define positives and negatives then getting a binary for each marker to perform downstream analysis.

Happy to discuss more and exchange ideas on how to make this the best way possible,

Thanks
antho

[rossbar]

Hi @anthomontan - thanks for the ping, the work with SeNNet sounds very cool!

Re: how deepcell-types can be used with the senescense analysis pipelines:

cell types can be defined using your model

Yes - in principle you can try this out already! Although I will caveat that I would not expect great results out-of-the-box with the pre-release model. This model incorporates as much training data as we've been able to find re: publicly labeled data, but is still quite limited given the variability in marker panels and tissues. We have been dutifully annotating every bit of data we can ingest and these updates will be incorporated in the near future, which will hopefully greatly improve prediction results across a wider range of tissues. Nevertheless, going through the process of getting deepcell-types working on your data is still a worthwhile effort from the perspective setup and testing. The deepcell-types inference pipeline is designed to be quite flexible¹: I'd recommend starting with the tutorial to see if you can't map those patterns onto your own images.

That way we can have cell types + Sen marker expression at a single cell level , then users can export data and define senescent per cell type per markers and visualize the expression all at once.

Re: incorporating senescent data - I would characterize this description as the "classic" approach - i.e. extract some "features" (e.g. average expression levels) on a per-cell basis then run some downstream analysis (often dimensionality reduction + clustering) to try to identify patterns in this hand-selected, pre-determined feature space. The approach taken by deepcell-types differs in that there are no pre-determined features; rather the features themselves are learned from the data. In other words, rather than trying to quantify senescent marker expression explicitly, you could in principle use the multi-channel attention architecture pioneered by deepcell-types to perform senescense prediction directly on the spatial data. To do so, you'd need to have labeled images where the labels would denote "senescense states" (apologies for the imprecise terminology) rather than cell types. Anyways - this is a more research-y direction but if you were interested in pursuing it I'm sure we'd be happy to help!

Footnotes

1. Though here too there's still plenty of work to do! Especially when it comes to handling variability in marker aliases. Improvements incoming... ↩

[anthomontan]

Hi Ross,

Thanks a lot,
I think combining the cell types with the Senescent markers will be great, since any cell type could be senescent or expressing some Senescent markers (one or many). Labelling could be the culprit and the most time consuming factor but something that can still be done with time and efforts.

I will dig into the tutorials to get familiarize and start from there , hopefully cell annotation could be done downstream of senoquant and be highly valuable and integrated in the CODCC SenNet :D .

Thanks a lot,
antho