Thomas Norman

1/ Most genes and variants matter somewhere but look silent in any one cell model. Can we build conditions that reveal phenotypes? With @kanishkadey.bsky.social, we engineer cell states to expose hidden disease biology and trace how genetic risk reaches disease genes. www.biorxiv.org/content/10.6...

biorxiv.orgHigh-throughput identification of chromatin-opening transcription factorsCell fate transitions involve major changes to the chromatin landscape, some of which are thought to deepen the basin around a newly adopted cell state, stabilizing gene expression programs that might...

19/ Also check out this interesting related preprint from the labs of Jay Shendure (@jshendure.bsky.social) and Silvia Domcke (@sdomcke.bsky.social) on which TFs can open chromatin when newly expressed in mouse ES cells: www.biorxiv.org/content/10.6...

biorxiv.orgEngineered regulatory states resolve context-dependent genetics in macular degenerationThe effects of many genes and disease-associated variants remain hidden in standard cellular models, limiting functional interpretation. Here we engineer regulatory states via CRISPR activation of tra...

20/ Special thanks to co-first authors Eli Metzner and Karthik Guruvayurappan (@somet3000.bsky.social), plus Kaden Southard (@kmsouthard.bsky.social) and Rico Ardy, for their incredible work, and to co-corresponding author Kushal Dey (@kanishkadey.bsky.social)! www.biorxiv.org/content/10.6...

17/ In genetics, we thus find engineered cell states can expose hidden disease variants and show that polygenic risk converges on disease genes in a state-dependent way. We expect the approach and our tools to generalize to other cell types, diseases, and complex traits.

18/ For virtual cells: as Perturb-seq scales to millions of conditions, we need principles for informative experiments. TF interactions are mechanistically interpretable: priming and Mosaic can steer them to new phenotypes. Still unclear what is most informative for other genes.

15/ Our screen offers a way to test this. Each TF is seen with 49 different partners, so we observe the same perturbation across many regulatory states. We can then ask whether polygenic effects converge, and whether that convergence depends on cell state.

16/ To measure convergence state by state, we built Perturb-PoPS, which uses perturbation data to link AMD’s common-variant (GWAS) signals to genes. We found that AMD’s genetic signal converged on inherited retinal disease genes, but precisely which ones varied across states.

13/ To show the power of engineered cell states, we used them to study an important open question in human genetics. Common diseases are often driven by many small-effect variants (polygenic), yet resemble rare diseases caused by single-gene mutations (monogenic).

14/ For AMD, these are inherited retinal diseases that closely resemble it. How do the two connect? The omnigenic model proposes that many small polygenic effects pass through gene networks and converge on a smaller set of “core” genes, such as the monogenic ones.

11/ To exploit priming, we built Mosaic, a transformer that learns from what each TF does alone to predict what a pair will do. It beats simpler models at predicting emergent nonlinear effects. And because its features are interpretable, it can nominate new regulators.

12/ One regulator it nominated, CREB3L3, does little on its own, but a phenotype emerged in a three-TF combination. Added to FOSB + SOX18, it boosted an AMD-associated gene program and switched on BEST1 and CP, two genes that cause inherited retinal diseases.

9/ This suggested a design rule we call chromatin priming: if TF A opens binding sites for TF B, A may prime the cell for B to do something new. The idea echoes priming in development and “chromatin potential” (Buenrostro lab), and can be read from single-TF data alone.

10/ To test it, we used dual-guide Perturb-seq to activate all 1,225 pairs of 50 TFs, one pair per cell. Primed pairs, especially those priming each other (bidirectional), more often had “neomorphic” effects, switching on genes neither TF activates alone, as in FOXA3 + MYRF.

7/ One of the simplest results validated the approach: TF activation opened chromatin at AMD risk variants, sometimes linking them to candidate target genes. Only 5–7% of likely causal variants are open in resting cells, but 28% are across TF states, with TFs chosen blind to AMD.

8/ Looking globally at the ATAC-seq, we found that each TF opens (or closes) chromatin at its own binding motif (diagonal). But many TFs also open the motifs of other TFs (off-diagonal), potentially preparing loci for those partners to bind.

5/ Many phenotypes clearly depend on context. How can we create many contexts in the lab, at scale? We turned to transcription factors (TFs), the natural regulators of cell state. Activating them reshapes a cell’s regulation and, by opening up DNA, can expose disease variants.

6/ We first activated 119 TFs one by one with CRISPRa in RPE-1 cells, a model for age-related macular degeneration (AMD). Our Multiome Perturb-seq assay reads out open chromatin (ATAC-seq) and expression in each cell, capturing what TFs directly do and which variants they expose.

3/ But in our genome-wide screen (Replogle et al. 2022), only about a third of knockdowns strongly changed expression, and other Perturb-seq datasets show similar gaps. In human genetics, over half of disease-associated variants have no detectable effect on expression in GTEx.

4/ Why does this matter? If most perturbations look silent in the cell state we profile, models trained on that data can’t learn what those genes actually do. And in genetics, silent variants can’t easily be linked to their target genes. Profiling more cells alone won’t fix this.

2/ Some background. Perturb-seq pairs CRISPR with single-cell sequencing: each cell gets one genetic perturbation, and we read out its effect on the whole transcriptome. The scale of these experiments is growing exponentially, in part to train AI “virtual cell” models.

My lab at MSKCC in New York is hiring for two positions. Join us at the frontier of functional genomics, studying fibroblast state transitions, combinatorial genetics, and ECM in disease. Please share with anyone who might be a good fit! (Mustache not required.)

msk.wd108.myworkdayjobs.comResearch Technician, Norman LabAbout Us: The people of Memorial Sloan Kettering Cancer Center (MSK) are united by a singular mission: ending cancer for life. Our specialized care teams provide personalized, compassionate, expert ca...

We're also hiring a research technician. This is an opportunity to learn CRISPR screening and single-cell genomics directly from the people who developed these approaches. Previous techs have gone on to great biotech roles and graduate programs. msk.wd108.myworkdayjobs.com/MSKCC_Career...

thenormanlab.comthenormanlab.com

Happy to answer questions if you're interested. More about the lab and our research here: thenormanlab.com

msk.wd1.myworkdayjobs.commsk.wd1.myworkdayjobs.com

We're looking for a postdoc with expertise in ECM biology, fibrosis, or mechanobiology who wants to apply functional genomics tools to their questions OR someone with a technology development background who wants to help build new methods. msk.wd1.myworkdayjobs.com/MSKCC_Career...

New preprint on technologies to scale up CRISPR screens. We use them to map 665,856 pairwise genetic perturbations and outline a path to comprehensive interaction mapping in human cells. We also introduce an approach for cloning lentiviral libraries with billions of elements.

bsky.app/profile/norm...

Thomas Norman@normanlab.bsky.social

In my excitement I forgot to actually link the preprint! www.biorxiv.org/content/10.6...

biorxiv.orgbiorxiv.org

These technologies open several future directions in large-scale genetics. The lab is actively recruiting postdoctoral fellows and technicians. If you’re interested in working at the frontiers of experimental genomics, please get in touch!

In my excitement I forgot to actually link the preprint! www.biorxiv.org/content/10.6...

biorxiv.orgbiorxiv.org

One key implication: we think comprehensive genetic interaction mapping among all human genes may be within reach. Scaling to ~10,000 expressed genes would require ~2.5 billion lineages—about 100× beyond what we show here.

This project has been ongoing nearly since the start of my lab. Huge thanks to lead authors Anran (Angel) Tang and Rico Ardy for years of persistence, and to Rafaela Mendes for foundational early work during the height of COVID.

Using PORTAL with an AP-1 reporter, we measured 665,856 pairwise perturbations across 612 genes and 46 million clonal lineages. To our knowledge, this is the largest exhaustively measured GI map in human cells, and the first at this scale with a non-fitness phenotype.

Why AP-1? Our results show that it responds to diverse biological processes. So AP-1 activity provides a readout with fitness-like breadth without requiring large changes in cell growth or death, enabling interaction mapping at much larger scale.

Each clonal lineage thus becomes an independent replicate. (We show that single-cell resolution is also possible via combinatorial indexing.) For example, here we see knockdown of KDM5C increasing AP-1 reporter activity across hundreds of lineages.

Finally, we developed a dual-sgRNA PORTAL vector using compact “mini” Pol III promoters. These pieces together enable systematic genetic interaction mapping, our main application. (P.S. The same cassette can also be used for dual-sgRNA CROP-seq experiments.)

To address the second bottleneck, we developed PORTAL (Perturbation Output via Reporter Transcriptional Activity in Lineages). Instead of reading perturbation identity from genomic DNA, PORTAL encodes perturbation effects in expressed transcripts.

The PORTAL vector actually produces two transcripts: • a pathway-responsive reporter • a constitutive identity transcript for normalization Both carry the sgRNA and a clonal barcode in their 3′ ends, linking phenotype to lineage.

The result is extremely even, high-complexity libraries. For example, we cloned a library of 812 × 812 guide pairs × 120,000 clonal barcodes (~80 billion elements), recovering 99.99% of guide pairs with near-Gaussian representation.

CAP cloning should be broadly useful wherever transformation is limiting, including lineage tracing, MPRAs, combinatorial protein or antibody engineering, toxic ORF libraries, synthetic biology circuits, and, as we show, new scales of functional genomic screens…

In traditional cloning, you construct plasmids in vitro and then transform them into bacteria to amplify. That transformation step is where complexity dies, as anyone who has failed to clone a library has experienced firsthand.

CAP cloning instead assembles and amplifies linear DNA in vitro, then uses TelN protelomerase to generate exonuclease-resistant, covalently closed molecules that package directly into lentivirus. (Shoutout to Touchlight who pioneered this chemistry.)

Two bottlenecks limit scaling. First, even constructing very large perturbation libraries is hard. Second, most screens read out only a single molecule per cell—the sgRNA—which is an inefficient use of cells. We address both with new technologies.

The first and likely most broadly useful advance is CAP cloning (Covalently closed Assembly Products), a new approach for cloning ultracomplex lentiviral libraries by avoiding bacterial transformation.