[NCP Pilot 3A] Ideal seeding density for neurons is 2500 or 5000; not clear if GFP interferes

[gwaybio]

@mtegtmey outlined this proposed experiment during our last checkin.

Goals

• To determine the ideal seeding density for imaging neurons
• To determine if GFP^ interferes in imaging neurons

^ The cell lines are transfected to express GFP, and it would take a lot of time/effort/money to reprocess without GFP

Experimental Design

Expected date for imaging: Done
Dyes: Cell Painting dyes
Cell Type: Day 28 Neurons
Plates: 384 well (Glia on all wells) plates in #2 without Glia were poor

Strategy

1. Plating neurons at a lower density (to form monolayer) will address multi-plane issues
2. Inhibit endogenous GFP and test if it helps with segmentation and neuron/glia classification
   • Previously GFP overlapping Alexa 488 made segmentation difficult
3. @mtegtmey will consider more transparent/direct communication with CDOT or perform the antibody staining himself to address low concentration issue

Conclusion

• 2500 and 5000 seeding densities look promising. See [NCP Pilot 3A] Ideal seeding density for neurons is 2500 or 5000; not clear if GFP interferes #6 (comment)
• Our conclusion about GFP from this experiment: @mtegtmey said: I think there wasn't much interference. The endogenous GFP wasn't very bright at D28. The GFP and phalloidin emit to the same channel, so we were
  lucky there.

[mtegtmey]

Update:
@shntnu @gwaygenomics @raldanehme

The next rounds of pilots are currently on d14, and will be fixed/stained in two weeks!

Should we try catching up briefly this week or next to discuss compound treatments?

[shntnu]

The plate has been imaged and is now being transferred to S3
s3://imaging-platform/projects/2019_05_28_Neuronal_Cell_Painting/NCP_PILOT_3/

[shntnu]

The images are now available at s3://imaging-platform/projects/2019_05_28_Neuronal_Cell_Painting/NCP_PILOT_3/images/BR_NCP_PILOT_3. Over to @bethac07 's team to process these. Please tag @mtegtmey for any clarifications about the data.

Meanwhile @mtegtmey can you share the plate metadata?

[mtegtmey]

Attached is the metadata.

NCPpilot3_metadata.xlsx

[shntnu]

• @shntnu delete /imaging/analysis/2019_05_28_Neuronal_Cell_Painting/NCP_PILOT_3/images/BR_NCP_PILOT_3 once Beth et al. process the data

[shntnu]

@bethac07 said that she ended up doing MAJOR revisions on the NCP pipeline improving it considerably, which works well because this is effectively the first dataset that we are going to analyze. She said that the images were extremely dim, but she knows a few tricks that should help. More when we do the analysis.

[mtegtmey]

Great! I did notice they where a little dim when I peeked through the images. Hopefully her tricks will work! :)

…

On Sep 18, 2020, at 4:40 PM, Shantanu Singh ***@***.***> wrote: @bethac07 <https://github.com/bethac07> said that she ended up doing MAJOR revisions on the NCP pipeline improving it considerably, which works well because this is effectively the first dataset that we are going to analyze. She said that the images were extremely dim, but she knows a few tricks that should help. More when we do the analysis. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#6 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AMSE5ERCCL24K3LARH2O6WTSGPAVTANCNFSM4MWD6FVQ>.

[bethac07]

@shntnu Apparently I can't check a check box that's assigned to you (that's weird, GH), but I deleted that folder on i_a while i was on there deleting other stuff for other projects.

[bethac07]

So processing finished over the weekend; as I stated up thread, it was reasonably dim, but not outside the scale of some other projects I've worked on. I've attached here a representative output image - Nuclei are in blue (and circled in teal), the actin/wga stain is in red, and the ER channel is in green. Cell bodies and neurites are outlined in white.

One thing I did notice- we seem to have a mixture of morphologies present, in the actin/wga staining we seem to have a fibroblast-y morphology, with neurites showing up mostly/entirely in the other channels (ER, RNA, and Mito). I'm not sure if this is expected, from either partial differentiation or glia or something else, but for this segmentation I ignored the areas that seemed to be actin/wga only and traced out neurites based on the other channels. If we think that's correct, then these are set to go (Shantanu, I'll let you make the backends, since we have extra compartments than usual); if not, let me know and I can revisit the segmentation.

[raldanehme]

Thanks for the update, Beth! The neurons in this experiment are co-cultured with mouse astrocyte, so it would be make sense that these would have a more flat, "fibroblast-y" morphology....I agree that it would make sense to ignore them here!

[mtegtmey]

The wga stain was one of the ones Francesca left out in her approach, as well as SYTO14. @bethac07 , in what you see from her data, as well as our current dataset, do you think there's a strong reason for us to try analyzing our dataset without these stains as well?

[shntnu]

Checking platemaps

BR_NCP_PILOT_3.txt

@mtegtmey does this look right?

[mtegtmey]

@shntnu density map is perfect. There's something slightly off on the treatment map. I'll give an update ASAP - almost to Broad and will check the map on my computer. Stay tuned...

[shntnu]

@mtegtmey great, thanks for checking. If you can also confirm the condition map, that would be great. Perfectly ok to do this on Monday!

[mtegtmey]

I think there's only one error above, a red square to the left. I'm not sure the best way to fix this i.e. i resend a meta data sheet, or just give you a snapshot of the positions of compounds. Below is a screenshot (i tried to make the colors match the ones you've used.) If you want a new meta data sheet just let me know! Thanks!

[shntnu]

@mtegtmey Sounds good. Will fix it at my end. Can you clarify whether the NA's are DMSO-treated or untreated?

[mtegtmey]

NA’s are all untreated!

…

On Nov 16, 2020, at 1:09 PM, Shantanu Singh ***@***.***> wrote: @mtegtmey <https://github.com/mtegtmey> Sounds good. Will fix it at my end. Can you clarify whether the NA's are DMSO-treated or untreated? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#6 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AMSE5ESPZK3RC73C5WTGKCLSQFTD3ANCNFSM4MWD6FVQ>.

[shntnu]

A high-level summary of the dataset is here
https://github.com/broadinstitute/neuronal-cell-painting/blob/master/1.main-run-workflows/1.inspect-profiles-pilot3.md

I haven't dug into it yet but overall data quality looks good!

[mtegtmey]

@shntnu any updates on the neuronal cell type here? If we have an idea about the best density, I could get started on the putting the entire cohort through the upstream cell biology/culture workflow and have them ready for profiling late Feb.

[shntnu]

@mtegtmey Sorry this got pushed out again!

We're pretty close – inspecting the data again, I think we can make a decision based on Grit.

@gwaygenomics In this notebook, I'm trying to compute grit, grouped by Metadata_line_condition and Metadata_plating_density.

There's a Metadata_group column that combines Metadata_line_condition and Metadata_compound_ID.
The Metadata_replicate column is simply the Metadata_compound_ID.

grit doesn't seem to be cooperating but I bet I'm doing something wrong. Can you have a look at that notebook?

[shntnu]

For @mtegtmey, here's a one-pager explaining Grit

Source: https://docs.google.com/drawings/d/1qwUysTDkKZfFYf9o24nNd8GG2kTlWOy1BV3gFDQI16o/edit

[shntnu]

@mtegtmey once we have this figured out, we will have a grit value for each compound, for each condition-density pair.

We can make a call on which seeding density with that info.

Although we have something similar from past analysis, where we compare replicate correlation values against a null, I'm increasingly confident that Grit is a better measure to report, especially when we have such few perturbations to compute a null distribution.

For our reference, here is the metric I'd have previously used

"Pick the density that has the most number of points to the right of the null (vertical bar), averaged across the two control and deletion condition."

(from https://github.com/broadinstitute/neuronal-cell-painting/blob/ncp-pilot-3/1.main-run-workflows/1.inspect-profiles-pilot3.md#audits)

But that's a bad idea because the null is not going to be very effective here (the conditions are no independent, too few conditions).

Grit should be much more informative 🤞

[gwaybio]

Here are the grit metrics

perturbations are not very gritty, but 7,500 control appear to the best

Updated notebook in #11 - @shntnu this exercise raised a potential enhancement for grit (broadinstitute/grit-benchmark#12)

[shntnu]

Thanks @gwaygenomics !

I think I may also have misunderstood the meaning of replicate_id and group_id; it makes sense now (the potential enhancement makes sense to me).

What I was actually looking to compute was a grit score for each replicate, which I'd then aggregate as I like. I've updated the notebook in #11 to compute this.

This is what you computed

This is what I computed

And then further aggregated per compound (using median)

or using mean

I'd love to chat about what drives the difference between first (compound grit) and last plot (replicate-level grit aggregated using mean) are different

[gwaybio]

this is so cool!

[shntnu]

Sidebar about Grit:

This is how my brain views it

And the first thing that popped is that we should give some thought to the z-scoring because the similarities to null are likely going to have skewed distributions. This is what the distributions look like for this dataset.

[shntnu]

@gwaygenomics in the version of grits computed here, grit is configured like this:

        control_perts = ["Untreated"]

        replicate_groups = {
            "replicate_id": "Metadata_compound_ID",
            "group_id": "Metadata_group",
        }

Is that a valid way to configure it, given that there are multiple rows with the same Metadata_compound_ID within each Metadata_group? Just trying to understand how this works.

My current understanding of the functionality and the method is that replicate_ids should be unique within each group.

[shntnu]

@mtegtmey and I chatted about the plot below and concluded that 2500 and 5000 look promising

We decided to do this additional analysis – comparing control and deletion lines, at baseline, for all densities.

@mtegtmey This ^^ is consistent with our decision to go with either 2500 or 5000. In the above, 2500 looks a bit better. One could do a KS-test to quantify but that's not necessary; eyeballing is probably sufficient here.

Here is the histogram version (not included in the notebook)

[mtegtmey]

@shntnu @gwaygenomics This is great! Exciting to see all of this data coming through. So interesting!

Would it be possible to access some of these images? I'm thinking in terms of displaying a representative image of each cell type with the aggregated stains.

[shntnu]

See #14

[shntnu]

This is what the distributions look like for this dataset.

@gwaygenomics bump

[gwaybio]

Thanks @shntnu - i added to the discussion in cytomining/cytominer-eval#38 (comment)

[mtegtmey]

A high-level summary of the dataset is here
https://github.com/broadinstitute/neuronal-cell-painting/blob/master/1.main-run-workflows/1.inspect-profiles-pilot3.md

I haven't dug into it yet but overall data quality looks good!

@shntnu I get error when following this link, has this summary disappeared?

[gwaybio]

It looks like the path was renamed. A good reminder to use versioned links! https://github.com/broadinstitute/neuronal-cell-painting/blob/9798025892e07327ee24c0b606c0f6a8504db966/1.run-workflows/1.inspect-profiles-pilot3.md

[shntnu]

• Not clear if we concluded anything about GFP from this experiment

@mtegtmey just wrapping up our notes here. Do you happen to recollect if we made any conclusions about GFP interference when analyzing Cell Painting profiles?

[mtegtmey]

I think there wasn't much interference. The endogenous GFP wasn't very bright at D28. The GFP and phalloidin emit to the same channel, so we were lucky there.

…

On Thu, Jun 24, 2021 at 11:52 AM Shantanu Singh ***@***.***> wrote: - Not clear if we concluded anything about GFP from this experiment @mtegtmey <https://github.com/mtegtmey> just wrapping up our notes here. Do you happen to recollect if we made any conclusions about GFP interference when analyzing Cell Painting profiles? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#6 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AMSE5EUAMSCT4HVGOOXT743TUNIDNANCNFSM4MWD6FVQ> .

[shntnu]

Great, I've updated #6 (comment) and all set here.