hESC screening


I am new to this forum, yet I have been using CP for over 2 years, in very basic ways, but recently I had to boost up my skills a bit…
I am optimising a “directed differentiation assay” on human embryonic stem cells (hESC), looking at the capacity of small molecules (of marine origin) to induce cell differentiation. Our marker is a “dim” GFP captured along with an Hoechst channel for the nuclei. Yes, cells are alive.
Because GFP expression occurring with cell differentiation is a “rare” event, I have to acquire 5X5 montages of my wells in the two channels aforementioned, to pick up statistically relevant data.
I use a BD Pathway 855 fitted with 20X NA0.75 obj (Epifluorescence). Each image is ~65Mb! There could also be some problems of alignment, but I haven’t tried addressing them for the moment…
Light sources are of quite poor quality on the BD, and even with flat field reference correction, in some areas of the montage, when I use “CorrectIllumination”, I have background almost as strong as my dimmer cells (see attached images - these have been resized/re-scaled to post them - If you want the biggies on which the pipeline is run, can you explain me how to send them to you).
My first immediate problems at the moment is i/can I and ii/how should I: use CorrectIllumination on this sort of dim montage? or is there any other way I can reliably pick up the GFP cells? I have attached some output of correct illumination and you can see it’s quite heterogeneous; good in most part and terrible in handling background in some other parts. To compensate for the artefacts I pick up, I have put some size filters but they can only do so much
Basically, the output I am after is simply a ratio of GFP cells/total cells (from Hoechst).
As you can see, hESC are not the most sexy looking cells!!!
At the moment, I have used two different strategies: one is to simply use the Hoechst mask over the GFP image: Even if segmentation is not very faithful, quantitatively, it’s fairly OK. The second is to pick up Nuclei, shrink them to a pixel and look around them for GFP object using propagation/intensity/intensity on my secondaries…I then run two filters, intensity and size based…
I have been running screenings for 4 years now using state of the art assays for various indications (carbohydrates, lipids metabolism imbalances, cancer, antiviral, anti-infective, Central Nervous System, Inflammation, Organogenesis). This is my second screening on stem cells. The first one directed us towards very new very exciting chemistry. I say that because, if you guys would be interested in some collaboration and a number of screenings, I’d be interested to pitch you up in more details and see where we could go from there.
mmmh, that was a long post - sorry :smile:
Sorry, I can’t upload the images or pipelines at the moment… I’ll talk to the IT guys and fix that ASAP…
My specs are: CellProfiler 1.0.6025 Compiled on iMac 10.5.6 with 2Gb of RAM & 2 GHz Intel Core Duo…


I’ll try to address some of your questions:

[quote=“ffontaine”]… I have to acquire 5X5 montages of my wells in the two channels aforementioned, to pick up statistically relevant data.
I use a BD Pathway 855 fitted with 20X NA0.75 obj (Epifluorescence). Each image is ~65Mb! [/quote]

Just a bit of clarification: are the images 65MB because you are stitching together a 5 x 5 array of images (so each image is actually ~2MB), or is the image that big because of the high resolution you are using? What are the pixel dimensions of your images?

In any case, CP isn’t well built to handle images that size. Are you certain the images need to be this big?

We recommend using YouSendIt for large files like this.

i think both of these are good strategies, at least without seeing any images first. My first instinct would be something along the lines of the latter.

Also, in the same vein, have you tried IDSecondary from the nuclei using Distance-N? If you are more interested in a simple count, then as long as the cells of interest are brighter in the vicinity of the nucleus (and not necessarily the entire cytoplasm), then you don’t have to worry about identifying the whole cell; just enough of the cell to identify it correctly.

I will private message you about this in a little bit.


Hello Mark,

Thanks for your reply. Bear with me re: sending some images, I should be able very soon, once I am getting more proficient with YouSendIt.

To answer your questions:

Indeed, the image output is a massive array of 25 images stitched together. The camera moves across the well in a 5X5 array and all images are stitched together by the BD acquisition software (Attovision).
The pixel dimension of the original motage is 6720 pixels wide by 5120 pixels tall… We could make the image smaller, possibly a 3X3, now that we manage to get homogeneous cell seeding, would that help?

I will try your suggestion regard Distance-N on Secondary…

However, here I believe the big problem is running a Correct illumination on images stitched together. I guess ideally I should run Correct illumination on each image separately. It would work better would it not?

Thanks for your help!


That’s right, it probably would. You might save on processing time and memory overhead, if each image is handled separately.

Smaller images would help too, provided that the objects you are looking for don’t get lost in the downsampling.

It is typical to ensemble average all images in all the wells to generate a per-plate illumination correction, so that would be a good first approach. However, if the illumination varies in a consistent way from site to site (i.e, all images at site 2 have the same illumination function), it may be possible to compute and apply an illumination correction on a per-site basis rather than per-well.