Improving Optical Pooled Screening for Functional Genomics

We are big believers in pooled screening approaches, i.e., approaches in which many different gene knockouts grow in one Petri dish side by side. Yet, none of the approaches we currently have at hand are compatible with imaging: capturing the morphology of a cell or assessing how certain biomarkers change their subcellular localization allows us to access novel areas of biology.

Together with Element Bio, we were able to pull off a pooled CRISPR screen in which we looked at signal transduction of a key transcription factor, NF-kB, which regulates inflammation and immunity. 

Conventional pooled screens collapse each cell down to a single number: did it survive or did it cross a fluorescence threshold? You learn that a gene matters, but not what it actually does.

Optical pooled screening adds imaging on top of that, which helps, but the published methods are manual, take two weeks or more, depend on homebrew in situ sequencing chemistry, and often recover the guide in fewer than half the cells. And even then, you only get morphology, with no RNA or protein.

With Myllia, we wanted to show something different: a screen that captures guide identity, transcriptome, protein, and morphology from the same cell, on a single automated instrument, at a scale that's genuinely useful for target discovery rather than a proof of concept.

Traditional pooled CRISPR screening assesses which genes are necessary for cells to survive and grow. This is useful in certain areas of biology, e.g., oncology, where you want to trigger cell death, but less so in other areas.

Over the past few years, we have focused on single-cell CRISPR screens, which record transcriptomic snapshots in cells perturbed with CRISPR. While this is useful in certain areas of biology, it does not capture key features such as the shape and granularity of cells, nor does it allow us to look at phenotypes that can be visualized by antibody staining and microscopy.

Optical pooled screening offers a solution to this problem. We can knock out genes at will in a Petri dish and observe what happens in each knockout by microscopy. 

In a traditional pooled screen, you transduce a population of cells with a pooled CRISPR library, apply some selection (a drug, a growth condition, a sort), then sequence the survivors to see which guides were enriched. You find out which genes shift that one endpoint, but everything happening inside the cell is invisible, and you're limited to whatever your selection measures.

Optical pooled screening keeps the pooled format but reads each cell by microscopy. You image the cells, then sequence the guide in place to assign every cell its perturbation. So instead of one enrichment value per gene, you get rich single-cell phenotypes (where a protein sits, how the cell is shaped) tied directly to the genetic change, across millions of cells at once.

Several optical pooled screening approaches were published before. And while all of them were exciting and significant improvements have been made over the past two years, all of them were difficult to set up.

The main constraint was the difficulty associated with in situ sequencing, which was rather manual and time-consuming. As a consequence, this technology has not been widely adopted, despite the fact that it addressed an unmet need.

DISS, as implemented on the AVITI24 instrument, closes this gap and allows researchers to easily access optical pooled screening. The one thing researchers have to keep in mind is that, because of the nature of the barcoded antibody technology that is used to detect subcellular localization of proteins, the resolution of the imaged phenotype is still rather coarse.

Consequently, you have to either dedicate more cells to an experiment to increase resolution or accept the fact that you will only uncover strong regulators of your phenotype of interest. 

DISS sequences native RNA directly inside intact, fixed cells, with no library prep. You use a single-sided probe that hybridizes to the known guide-RNA scaffold and then extends to read the sequence downstream, so we capture unmodified sgRNAs (single guide RNAs) without engineered constructs, dual-flank probe designs, or large probe panels.

Two things follow from that. First, guide detection becomes reliable, whereas older in situ methods often recovered the guide and threw away most of the experiment.

Second, because the same chemistry reads the 3-prime transcriptome, you get genome-scale RNA in addition to the guide, on the very same instrument that's doing the imaging.

That's what turns optical pooled screening from a morphology-plus-barcode assay into a true multimodal screen, and it's why the whole protocol can run as one automated workflow instead of a multi-week manual process.

RNA, protein, and cellular morphology are somewhat connected, i.e., one would expect an RNA to become a protein to exert an effect on cell morphology. However, this is not always true. In fact, cell morphology or, more specifically, a change in subcellular localization of a biomarker may not be connected to a change in RNA or protein levels.

Consequently, capturing all of these features at the same time gives us a broader picture of cell state and thus allows us to better infer what is going on in cells.

Because those layers don't move in lockstep, and you lose the mechanism if you measure them separately. A perturbation can shift a transcript expression's location within minutes, while protein changes over hours, and morphology captures the structural consequences further downstream.

In our lung cancer research, knocking out the IL-1 receptor and then adding IL-1 beta showed loss of p38 and HSP27 phosphorylation within 30 minutes at the protein level, while the RNA confirmed broad suppression of NF-kB target genes such as CXCL8 and CCL2.

Either readout on its own tells a partial story. Measured together from the same cell, you can trace how one genetic change propagates through signaling, transcription, and structure, and you avoid the usual problem of stitching together separate experiments that were never looking at the same cells in the first place.

There is still a trade-off between the scale of these screens and the biological depth of the read-out, inasmuch as the simplest screens are the most scalable.

However, approaches such as the one we present here promise to close this gap and offer a compromise: medium scale (at least thousands of genes profiled) with a deep phenotypic read-out.

This is exciting because it will enable more unbiased screening campaigns linked to more complex readouts, thereby opening the door to novel discoveries.

That trade-off came from the different tools, not from the biology. If depth means manual imaging and homebrew in situ sequencing, you cap how many cells and conditions you can realistically run.

If scale means a pooled screen with a single enrichment readout, you give up everything happening inside the cell. Automating the whole thing on one instrument changes things: guide identity, transcriptome, protein, and morphology all come off the same cells in the same run, so adding depth no longer costs you throughput.

You also collapse what used to be several separate assays into one, which cuts both cost and time to insight.

To my knowledge, this is the first CRISPR screen to be conducted on this instrument. Hence, our questions were more of a technical nature than of a biological nature.

On a technical level, we learned that the instrument is able to capture guide RNAs in a very solid fashion—this is key to establishing the link between cells and the genes that were perturbed.

Secondly, we got an appreciation of the number of cells that are needed to recapitulate strong, intermediate, or weak phenotypes. This is important to conceive future studies aimed at discovering novel biology.

And finally, we appreciated how the unbiased study of morphology can be utilized to classify genes according to gene function. 

They also showed that it holds up at the scale and resolution discovery actually demands. Myllia profiled roughly 440,000 cells across a screen of 195 genes, reading p65 localization (whether NF-kB had moved into the nucleus) alongside cell-painting features.

The screen recovered the canonical NF-kB machinery, which is the control you need before you trust anything else, and it also flagged regulatory roles for chromatin-modifying complexes that aren't the obvious candidates in this pathway.

Hitting the known biology validates the assay, and surfacing the less-expected regulators shows it can generate hypotheses rather than just confirm them. Doing both in a single automated run, across hundreds of thousands of cells, at high guide detection efficiency, is the bar optical pooled screening has to clear to work as a primary discovery engine.

There are entire areas of biology that strongly depend on imaging as a phenotype. For instance, in neuroscience, scientists want to study the shape and connectivity of neurons as a proxy for cell function.

Alternatively, scientists wish to study the occurrence of certain protein aggregates that represent biomarkers for disease. Being able to conduct screens will allow us to elucidate novel drug targets that have the potential to reverse disease phenotypes (such as the ones mentioned above) and thus could give rise to novel drug discovery campaigns down the line.

The introduction to this interview includes text that has been created with the assistance of generative AI and has undergone editorial review before publishing. Technology Networks' AI policy can be found here.