iiinova — where your digital self keeps evolving

.. faith over many years that we'd eventually pull this off, and we are very grateful. Please see seattlehub.org/nextcell for annotated browsable tree plus raw & processed data. We've barely scratched surface of exploring this large tree of regulative development, so please have at it! /finis

.... @choijunhong.bsky.social , @cxqiu.bsky.social. .and thanks to amazing community of scientists & funders made this possible, inc. Seattle Hub for Synthetic Biology @alleninstitute.org @biohub.org @brotmanbaty.bsky.social @uwgenome.bsky.social @hhmi-science.bsky.social ... who had a lot of ...

180

Huge kudos co-authors: Qi Yu, H. Kim, @seidels.bsky.social, J. Acosta-Clark, @bkmartin.bsky.social, K. OConnor, R. Daza, M. Gasperini, @jennynathans.bsky.social, M. Lam, E. Gamo, S.VijayKumar, L. Kuo, JB Lalanne, K. Simeonov, @marionpepper.bsky.social , @coletrapnell.bsky.social , Jesse Gray.. 17/n

You can walk tree in either direction. Descending: blastomere A → one pre-gastrulation founder → a E7.43 progenitor whose 624 sampled descendants exhibit clonally demarcated contributions to three endodermal organs. Zoom in further to homogenous liver subclade. 15/n

Fun result - although tree built from TAPE edits alone , clade co-occurrence recovers germ layer organization & sweeping developmental time turns that into a *dated* hierarchy of couplings 14/n

At tips, 51% of tree siblings share cell type, 9× chance, & extremes track anatomy, e.g. lung & airway 139×, DRG neurons 69× (cell types that commit from spatially restricted pool & expand locally). Recent terminal differentiations are legible among heterotypic tree siblings 13/n

The tree resolves 182 ancestral lineages at E6.0; follow them to E13.5 & just 24 account for half the embryo, where a neutral birth-death process predicts 35. Clonal dominance is set early, then simply inherited. (see paper re: 2nd phase of clonal dominance) 12/n

Then, near silence for 4 days! Followed by fast re-start of keystrokes at ~E6 at ~9 edits/cell/day, fading slowly as TAPEs fill. Avg. 55 / 66 sites written by E13.5. Analogous to 13.5 page book w/ pp2-5 torn out (pre-gastrulation arguably dull anyways :). Not ideal, but legible.

Plot twist: At 2-cell stage there is burst of edits (major ZGA!). Daughter A got 23, B got 19, none shared. Thankfully as we only had one mouse, this hands us an internal replicate. Trees made & measured twice in two half-embryos that shared a zygote. All findings reproduce. 10/n

1100

Annotation didn't start from scratch. Our E8–P0 mouse atlas (by @cxqiu.bsky.social, @bkmartin.bsky.social, Ian Welsh from JAX) covered essentially every cell type at E13.5, so we could immediately label all 1,340,794 tips — 25 major trajectories, 135 cell types. 9/n

180

Phylogenetic tree building methods don't scale well, so @seidels.bsky.social did it in 2 steps. 1) Build a 640,012 cell backbone by NJ on an ordered edit distance (SciPhy) & strict molecular clock dating, then 2) distance-based placement of more cells to get to 1,340,794 tips. 8/n

180

How to read TAPE? Not easy on mRNA-poor nuclei we use for whole-animal scRNA-seq. Fix from Haedong Kim: circular TAPE which, together with sci-RNA-seq3 by Qi Yu and @bkmartin.bsky.social, yielded (11 × 6 TAPE) × 1,753,895 cell-by-gene matrix w/ ~50% of TAPE entries genotyped. 7/n

To get DNA Typewriter in vivo, Qi Yu & @alleninstitute.org / SeaHub's in vivo platform injected PEMax, epegRNAs, blank TAPE into B6 zygotes w/ piggyBac, then let development proceed to E13.5. 100 injected → 78 transferred → 10 embryos → 1 w/ great recording, on 11 x 6-unit TAPEs. 6/n

Building on GESTALT (w/ @aaronmckenna.bsky.social @gregfindlay.bsky.social @james-gagnon.bsky.social @schierlab.bsky.social) we previously developed DNA Typewriter (w/ @choijunhong.bsky.social @chenomics.bsky.social), a sequential molecular recorder that writes insertional symbols to DNA TAPE. 5/n

For background: in 1983 Sulston & colleagues completed the cell lineage of C. elegans by watching every division under a microscope, & forty years later it is still the only complete animal cell lineage we have. You obviously cannot build a mouse cell lineage this way. 4/n

190

We've also built NextCell (h/t @claude_code ), an interactive browser for the entire annotated phylogeny — seattlehub.org/nextcell Open science, so the trees and data are released there too (or will be very soon, lmk if anything missing). 3/n

190

Thrilled to post thread re: new single-cell lineage of mouse embryo reconstructed w/ DNA Typewriter. One animal, zygote to late organogenesis (E13.5). Tree has 1,340,794 transcriptionally profiled, annotated tips (cells), 1,142,588 dated internal nodes, rooted at zygote 1/n

313458

Apologies, this was an honest oversight!! We will update to add.

110

Ty Daniel! PLEASE tell me what refs we missed, we will update (already saw post from Judith, feel free to reply here or send me DM any others). To second point, entry point = 5 kb, technical path to larger by this general strategy is obvious...

130

This was a very fun project! Big thanks to all the authors: Abby McGee, @carinabiar.bsky.social, @bkmartin.bsky.social , @tony-li-sf.bsky.social , Jean-Benoît Lalanne, and Haedong Kim, as well as everyone who helped along the way! 13/13 (and extra thanks to Abby & Carina for ghostwriting this 🧵!)

041

LAMPRA's slot architecture generalizes: we can swap in different CRE libraries at each slot and expand slot number, element length and combinatorial complexity. This was just a proof of concept — the framework is built for dissecting locus-scale regulatory grammar at scale. 12/n

130

Only ~31% of our training data was needed to get within 5% of max model performance. Readouts from locus-scale libraries like this one may be be a genuinely useful data type for deep learning models of gene regulation, orthogonal to biochemical assays like ATAC-seq and conventional MPRAs. 11/n

140

We also tried predicting activity from sCRL composition. Linear models: r = 0.57. Add interaction terms: r = 0.81-0.82. Tree ensembles (random forest/gradient-boosted trees): r = 0.90. Combinatorial regulatory logic is real, and learnable.

121

In fact, the insulator rules are quite different. Within the most active promoter group, we found that insulator position is very important: activity was ~3-fold higher when the insulator lay upstream of the enhancer than when it intervened between enhancer and promoter. 9/n

110

For these active enhancer elements (e-NMU(iii) and (iv)), position doesn’t seem to matter within the upstream slots. However, that doesn’t hold for all elements – we saw clear activity decreases for insulator positions closer to the promoter. 8/n

110

Not all enhancers behaved the same. Adding more copies of e-NMU(iii) or (iv) in the upstream slots (S1-S4) raised activity, log-additively, while e-NMU(i), (ii), and (v) showed no dose-response at all. Only two of five turned out to be active enhancers in this context. 7/n

121

So we split the data into 4 "promoter groups" by whichever element drives transcription. Each has its own baseline activity and its own dose-response to upstream enhancers — and those two properties are independent of each other. 6/n

121

We found that two of the five "enhancers," e-NMU(iii) and (iv), aren't just enhancers. When placed right next to the minimal promoter in slot 5 (S5), they overpower it, acting as strong, orientation-dependent promoters themselves — up to 110x more active than everything else. 5/n

131

Proof of concept: 5 candidate NMU enhancers + 5 matched shuffled controls + 1 synthetic insulator. 2 orientations, 5 slots. That's 5.2M possible loci — we sampled ~36,000 of them. All for ~$0.30/locus after a one-time synthesis cost. 4/n

130

Long-read (PacBio) sequencing of the loci links each combination of elements to its barcode(s), building a locus-to-barcode dictionary. Short-read sequencing then measures RNA/DNA barcode ratios across the whole pool. One experiment, tens of thousands of synthetic loci. 3/n

150

LAMPRA = Long @$$ Massively Parallel Reporter Assay. Instead of one short CRE next to a promoter, we combinatorially assemble 5 x 1-kb elements to make synthetic cis-regulatory loci (sCRLs), barcode them, and use a standard MPRA readout to measure activity. 2/n

120