PicoBlog

Q1 2024 Update: Please read my Winter Update for a short note regarding this post. The methods-specific comparisons in this essay are accurate, and the questions towards the end are still highly relevant to the direction morphology-driven research efforts are heading. TLDR, the methods are so similar because the researcher who developed POS in Paul Blainey’s lab moved to Insitro to duplicate the method!

Over the past week and a half, the Broad’s Cell Painting group and Insitro published very similar approaches for large scale, hypothesis-free, pooled optical screening. Cell Painting (Carpenter-Singh lab) is embedded in the PDD landscape as a publication powerhouse, with Dr. Anne Carpenter serving as a board member of Recursion Pharmaceuticals, and Dr. Shantanu Singh advising Dewpoint Therapeutics. Insitro, led by Daphne Koller from Stanford, is a well-funded AI/ML-focused PDD player who has published minimally until now. These papers are a demonstration of the continued value of profiling-style cell-based imaging assays in general, and invite process-optimizations from the TechBio industry that would enable biological insights of increasing depth in phenotypic drug discovery in particular.

I have focused on wet-lab methodology because I like to think of ways to accelerate lower throughput methods via denser plate formats and intentional lab automation.

CAS9 cell lines and sgRNA transduction: Insitro (POSH) purchased a CAS9 plasmid from Horizon Discovery and screened a single cell type (A549) while Cell Painting (PERISCOPE) screened both A549 and HeLa cells, with the A549 CAS9 constructs developed in house, and the HeLa CAS9 over-expressed cells gifted by Professor Iain Cheeseman. PERISCOPE developed their sgRNA libraries in house, while POSH ordered libraries from Twist Biosciences. Both cloned their libraries into CROP-seq vectors manually, with POSH using 4 sgRNAs/ gene, and PERISCOPE using 10 sgRNAs/ gene for their pilot screen, and moving to 4/gene for their subsequent experiments.

The cell line divergence may highlight a difference between academia and industry, in that Insitro was able to purchase numerous CRISPR-related reagents for their study from external CRO vendors, while the Broad sought to leverage internal teams (Cell Painting & MIT) for necessary materials. Also, until now, it seemed as though Insitro was focusing on iPSC models for their programs. Library design is also a key difference between the papers, with PERISCOPE screening against roughly 10X the amount of genes (whole genome) as POSH, whose library was more narrowly focused on genes that were known to induce morphological changes, were druggable, and whose products were targets of annotated tool compounds.

Imaging and In-Situ-Sequencing: A process-related similarity between the two papers is that they both used Cellvis 6-well plates as the experimental footprint (compared to 12/24/96/384/1536 well plates), and imaged for morphology and sequencing readouts with the same microscope (Nikon Ti-2 Eclipse with 20x objectives)! Additionally, Insitro notes that running RT after Cell Painting causes degradation of RNA, and the same order of RT->CP->ISS is followed in PERISCOPE as well. A key difference between the reports is that the Broad developed novel Cell Painting probes whose fluorophores could be chemically cleaved off via TCEP before in-situ-sequencing, to get over the spectral overlap expected to occur by combining these methods, while Insitro developed a proprietary RNA-FISH mitochondrial stain (Mitoprobe) that could be washed off during ISS.

It is surprising how similar the plate footprint and acquisition parameters are between both, and that the key difference here is in how both teams used different approaches to develop unique mitochondrial imaging probes to overcome the predicted challenge of fluorescence overlap between Cell Painting stains and ISS probes. The following intro figures from both papers illustrate the similarities in approaches: POSH on top, PERISCOPE on the bottom

Analysis and conclusions: The products for both teams, so to speak, would be wet-lab/ dry-lab pipelines that allow one to map relationships (ie. determine pathway interactions) between morphological changes and corresponding single-gene knockouts, in a given cell type. Both teams ran CellProfiler with unique tweaks, however a major difference is that Insitro used deep learning for identifying amplified barcodes, segmenting cells, and extracting features (specifically adapting the DINO transformer by Caron et al.) while the Broad didn’t, and instead used traditional CellProfiler to determine whether the imaging features that reported on gene knockouts corresponded to the whole cell or single/ multi compartment. The biological questions asked were also different, with Insitro’s figures looking more like UMAPs with representative images, with metrics for model accuracy. Cell Painting actually made some claims about gene dependencies (Figure 6). I could go into more detail here if it is of interest, but 1) I question the translatability of these claims (ie. image similarity of a single cell can inform on the relatedness of barcoded knockouts) 2) I’m way more interested in thinking about broader (I also suppose deeper) applications in small molecule-focused phenotypic drug discovery, rather than asking specific biological questions with such approaches.

When we say that this technology WORKS, what that means is that you can link cell morphology to single gene knockouts, at single cell resolution. Refer to Chris Bock & Jonathan Weissman’s article, High-content CRISPR Screening, to learn more about successful attempts to link CRISPR knockouts to other single cell readouts.

Current Realities and Abstract Challenges: While the Broad chose to focus on making biological pathway connections, I liked Insitro’s emphasis on Deep Learning methods for all aspects of their screen. If one were to run broadly-combinatorial screens (described in Future Opportunities), Deep Learning methods would likely be necessary to deconvolute the synergistic or inhibitory effects of compounds and knockouts at the single cell level. However I will present two general challenges: 1) How we deal with the fact that the connected plated cells, while individually experience different knockouts, have proliferative agendas that are coordinated at the population level (picture below from POSH to show physically connect cells)? And how do we know that morphology is a direct consequence of a given knockout, and not a due to physical cell-to-cell contact? Perhaps non-dividing cells, like iPSCs, would be better suited for future screens, however there may be transfection-related challenges to overcome. 2) In phenotypic drug discovery, the goal is for imaging assays to have some kind of predictive value of a compound’s safety and efficacy in animals or humans, and these experiments do not do that.

Future Opportunities: What these assays do enable, today, is the ability for teams to run compounds in plates, without any ISS probes, and match morphological similarity to barcoded gene knockouts (present in the training data), enabling out of the box target deconvolution. More specifically, if you chose to run all programs with the same cell line (why would you do that?), then you may choose to run the broadest sgRNA library possible on day 1, to link all future compound perturbations to going forward. From a process standpoint, some companies may choose to choose internally-engineered model cell lines and run the same 6-well approach, or may choose to make the plate footprint more dense and broaden their knockout library (may also consider combinatorial knockouts, or mixing knockouts with tool compounds- a similar approach is suggested in the Bock & Weissman paper linked earlier) in the same wells for early MOA insights. Perhaps reaching a threshold of replicates would improve predictive accuracy of such approaches. Or, if your company is already deep into multiple campaigns, choose a narrower sgRNA library that represents pathways you expect your compounds to be acting by, and use that for determining MOA.

If these methods can scale reproducibly, there is no reason why a single company can’t run this assay as part of their ‘platform-as-a-service’ model for other PDD companies, or develop a subscription-based barcoded-morphology library for other companies to reference on an ongoing basis. While I proposed that every PDD company could run this at the start or end of a campaign, teams may not all be technically equipped/ capable of performing this assay. The ongoing success of such approaches depend on people! Simply, the informational richness afforded by this assay may be WORTH sharing.

ncG1vNJzZmiakZi4tbvToZylmZJjwLau0q2YnKNemLyue89op6irmGLDtHnPnqmiq5OkvaY%3D