Golas:

• identify optimal settings for each instrument
• quantify the magnitude of impact of these variations (are they relatively small, or substantial?)
• assess the ability to match signatures across samples imaged by different instruments

Conclusions:

Scope Vendors:

• Molecular Devices
• Nikon
• Perkin Elmer
• Yokogawa US
• Yokogawa Japan

Additional information on scopes and imaging details can be found here.

Experimental details:

Plate map and compounds: JUMP-MOA plate map and compounds were used (90 compounds from 47 distinct MOA classes, with 4 replicates per compound)

We have a qualitative difference between conditions that have higher FOV counts (20X, typically) those with lower FOV counts (10X, typically, though not exclusively). This matches previous work

A couple of complicating factors here though -

1. 4 fields of 10X vs 9 fields of 20X may have equal numbers of cells, even with lower numbers of fields. We've always speculated that more fields is better data because more cells, but maybe it isn't - maybe it's more chances to get a really good segmentation, or some other thing we haven't thought of. Our previous experiments were all tested at the same magnification, with cell count and FOV count essentially perfectly correlated, so we have no data addressing this directly.
2. It may truly be that other magnifications really are worse than our standard 20X, but right now we can't tell because we know that FOV count MAY be a contributing factor.

To address these, I think we need to do 2 things -

• Subsample all the existing batches to all the applicable FOV counts we can - we have batches with 1, 2, 3, 4, and 9 sites, so we should subsample 9 site batches to 4, 3, 2, and 1 sites, subsample for site batches to 3, 2, 1 sites, etc. Because we are currently pretty sure that we can use the same recipe run for all the batches being subsampled to 4 sites, another one for all the batches subsampled to 3, etc, this is actually only 4 config files to generate and run.
• Once we've done the above, for each vendor look at a 2x2 matrix - percent matching / percent replicating vs FOV count/ cell count. I think this will help us be more definitive about both questions above.

The question was raised as to whether the feature composition of each batch is the same; I think it's going to be hard to do this on a per feature level because of the random dropout of highly correlated features, but I do think there are a few metrics we can quite easily generate for each batch based just on the columns present in each CSV:

• How many total features did this batch use?
• What percent of features are Cells vs Nuclei vs Cytoplasm? (These should add to 100)
• What percent of features are Texture vs Neighbors vs AreaShape etc? (These should add to 100)
• What percent of features are RNA vs DNA vs ER vs Mito vs AGP vs BF? (These should not typically add to 100 but may coincidentally, since AreaShape features have no channels and Colocalization have 2)