←
Jonathan Weissman's Lab209 posts
Jonathan Weissman's Lab avatar

Jonathan Weissman's Lab

@JswLab

Jonathan S. Weissman's lab at the @WhiteheadInst/@MIT. Account run by lab members.

◷ Joined October 2019▧ 40 Photos and videos
gokul gowri avatar
⟳ Jonathan Weissman's Lab retweeted
gokul gowri@gokl_g·Nov 6

the transcriptional state of a cell is not generally sufficient to predict its past or future. in our new preprint we show that conditioning on lineage context (sisters, cousins, etc) can help a lot. this one was v fun :) arxiv.org/abs/2610.40447

New preprint on generative modeling for mouse development! Rather than single cell trajectories, we set out to model embryos as trees. Led by @gokl_g, @njwfish, @WilliamNColgan, w/ Philippe Rigollet, @stats_stephen, Zack Smith, @marinkazitnik, @lukekoblan and a wonderful team

0:05
Jonathan Weissman's Lab avatar
Jonathan Weissman's Lab@JswLab·Nov 5

There's a lot more in the paper, and we also have a website w/ informal writeup of the method, code, data, and an interactive viewer for reconstructed dynamics. Looking forward to feedback! Paper: arxiv.org/abs/2610.40447 Notes and interactive viewer: lineageinpainting.info

Jonathan Weissman's Lab avatar
Jonathan Weissman's Lab@JswLab·Nov 5

Overall, we're excited about the direction of modeling multicellular dynamics using lineage (and other kinds of) context, rather than one cell at a time. We hope that similar techniques can be applied to other systems: regeneration, inflammation, cancer and more.

Jonathan Weissman's Lab avatar
Jonathan Weissman's Lab@JswLab·Nov 5

Even this crude approach can already provide useful predictions. We tested them experimentally for Tbx5 by profiling crispant embryos with scRNA-seq. The model's predictions about heart effects largely aligned with the experiment, though not perfectly (see e.g. AV canal).

Post attachment
Jonathan Weissman's Lab avatar
Jonathan Weissman's Lab@JswLab·Nov 5

The model is fit only to normal development, but intuitively it should be informative about perturbations: knocking out a regulator expressed in a progenitor likely affects its descendants. So we simply shift cells to the nearest "gene-off" state and run the model forward.

Post attachment
Jonathan Weissman's Lab avatar
Jonathan Weissman's Lab@JswLab·Nov 5

Once lineages converge, to what extent do cells remember where they came from? In 1976, Lewis Wolpert called this non-equivalence. Our model suggests that neurons retain more gene expression memory than the endothelium. Hox genes persist in every case we studied.

Post attachment
Jonathan Weissman's Lab avatar
Jonathan Weissman's Lab@JswLab·Nov 5

Fit to our recent atlas of 1.4M lineage-traced cells (E7.5–E10), we find that lineage inpainting disagrees with expression-only models mostly when multiple transcriptional states converge, as in endothelium and some neurons.

Post attachment
Jonathan Weissman's Lab avatar
Jonathan Weissman's Lab@JswLab·Nov 5

In many simulations where the true dynamics are known, LI recovers non-least-action paths that tree-less and cell-level methods miss. It also recovers dynamics that can be determined only from lineage context, which single-cell trajectory models fundamentally cannot capture.

Jonathan Weissman's Lab avatar
Jonathan Weissman's Lab@JswLab·Nov 5

But there's a wrinkle - while we need complete trees to fit the sampler, we only ever see the leaves. To solve this, we use a second network to infer ancestors that distributionally match cells from an earlier embryo. The sampler then learns from these complete trees.

Post attachment
Jonathan Weissman's Lab avatar
Jonathan Weissman's Lab@JswLab·Nov 5

To denoise whole (sub)trees simultaneously, we develop a tree-structured transformer, where each token represents a node of the tree. Inspired by ALiBi, we encode the tree structure through a bias term with learned linear weight per head.

Post attachment
Jonathan Weissman's Lab avatar
Jonathan Weissman's Lab@JswLab·Nov 5

So we model cells in their lineage context, rather than one at a time. We introduce lineage inpainting: given a tree with states observed at one time, sample its ancestors and descendants to fill in missing states, using stochastic interpolants to transport noise to states.

Post attachment
Jonathan Weissman's Lab avatar
Jonathan Weissman's Lab@JswLab·Nov 5

Our view: the lineage tree provides a context in which to interpret cell state. Two cells can be in the same state but have distinct contexts, suggesting possibly distinct dynamics.

Post attachment
Jonathan Weissman's Lab avatar
Jonathan Weissman's Lab@JswLab·Nov 5

Recent experimental techniques from us and others have enabled the reconstruction of lineage trees, providing critical -- yet still indirect -- information about dynamics: cells are profiled only at the leaves. How should we use this information?

Post attachment
Jonathan Weissman's Lab avatar
Jonathan Weissman's Lab@JswLab·Nov 5

scRNA-seq has led to remarkably detailed landscapes of cell states during development. Because it is destructive, we must infer dynamics from snapshots. But many dynamics can give rise to the same snapshots.

Post attachment