Lab Newsletter — August 10, 2026: One Cell, Two Answers

Ask a cell what a gene does and you have always had to choose how to listen. Pooled CRISPR screens are gloriously scalable — perturb every gene in the genome in a single dish — but the readout collapses to a number: did cells carrying this guide grow or die, enrich or drop out? You lose the cell’s shape, its organelles, where a protein went. Arrayed image-based screens keep all of that rich phenotype, but they need one perturbation per well, so genome scale means hundreds of thousands of wells and they simply don’t scale. For years that was the deal. Optical pooled screening (OPS) refuses it — and in 2025 the field carried the idea all the way to the whole genome.

🧬 Read the genotype and the phenotype in the same cell

The trick is disarmingly direct. Grow a pooled library — thousands of genes knocked out across one shared population — image every cell richly, and then, under the same microscope, recover which CRISPR guide each cell received by sequencing its molecular barcode in situ. One cell gives you two answers at once: its perturbation and its phenotype. David Feldman, Paul Blainey and colleagues established the method in “Optical Pooled Screens in Human Cells” (Cell, 2019), using targeted in-situ sequencing to demultiplex a perturbation library after image-based phenotyping — fluorescence microscopy recording both the phenotype and the sequencing reads that name the guide in each cell. (A lovely piece of rigor along the way: a modified lentiviral protocol cut barcode–guide “swapping” from over 28% to under 5%, so the genotype you read is the one the cell actually carries.) By 2023 the throughput had grown past genome scale: a genome-wide OPS of antiviral responses (PNAS, 2023) imaged 10,366,390 cells carrying 80,408 guide RNAs targeting over 20,000 genes, reading them all out with 12 cycles of in-situ sequencing — and pulled out real biology, including that ATP13A1 “is essential for viral sensing.” Why it matters for the lab: this is imaging × perturbation at scale — exactly the regime our self-driving microscope, Agent-Lens and the REEF imaging farm are built for.

🗺️ The genome, mapped by morphology

In 2025 two independent teams turned OPS into genome-scale atlases. The Broad Institute and Calico built PERISCOPE (Nature Methods, 2025) — “perturbation effect readout in situ via single-cell optical phenotyping” — marrying Cell Painting (high-dimensional subcellular imaging) to optical pooled screening. The result is the first unbiased, morphology-based genome-wide perturbation atlas in human cells: three whole-genome CRISPR screens, knockouts of more than 20,000 genes across tens of millions of cells, each scored on hundreds of image-based features — and, crucially, more than 10× cheaper than a comparable single-cell RNA-seq screen, with all data open access. It doesn’t just re-draw known biology; it lit up the poorly known, revealing for instance that TMEM251, tied to a rare lysosomal storage disease, is “required for trafficking enzymes to lysosomes.” As the Broad’s JT Neal put it, it’s a “first-in-class genome-scale resource for linking cell morphology to gene function.” The same year, insitro published a general platform, CellPaint-POSH (Nature Communications, 2025), aimed squarely at OPS’s remaining weakness — most implementations were pathway-specific. Its answer is to stop hand-picking biomarkers: a self-supervised model (CP-DINO) learns representations straight from Cell Painting images, and gene networks emerge without any target-specific readout (AUC ≈ 0.83 against the StringDB interaction network), enabling hypothesis-free discovery across a druggable-genome screen. Tellingly, the model applied zero-shot to PERISCOPE’s data still worked — the learned phenotype generalizes across labs and protocols. Why it matters for the lab: letting a model find the phenotype rather than hand-engineering it is the same thread as our morphological profiling digest — now driving a genome-scale screen, and shipping as open atlases in the BioImage Model Zoo / BioEngine spirit.

🔗 The data engine of a virtual cell — and the honest frontier

Here is why this belongs on the front page. A virtual cell is only as good as the data it learns from, and the hardest data to get is causal: not “what does a healthy cell look like” but “what happens to this cell when you change this gene.” OPS manufactures exactly that — millions of (perturbation → single-cell phenotype) pairs — at a cost that makes genome-scale feasible. It’s the imaging-native complement to yesterday’s genome language models: those read DNA sequence toward function; OPS measures perturbation-to-phenotype in living cells; a Human Cell Simulator is the model that couples the two, and a self-driving lab is what runs the screens that feed it. The honesty clause matters too: morphology is a rich but partial phenotype — it sees shape and localization, not transcriptional state — so OPS and molecular profiling are complementary, not interchangeable; the platforms are still being generalized beyond pathway-specific assays; and a screen is only as trustworthy as the phenotype you can read and believe, the same prove-it discipline we keep insisting on. But the direction is unmistakable. The microscope, which spent a century answering what does this cell look like?, has learned to answer a second question in the same frame — and which gene made it so?

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (X/Twitter sweep was skipped today — our news API is out of credits.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.