How does an embryo reliably "compute" its form - "cell by cell" - using only local interactions and mechanics, yet produce a precise global body plan? I’m excited to share our Nature Methods paper "MultiCell: geometric learning in multicellular development", presenting #AIxBiology research led by Haiqian Yang and the... result of a great collaboration with Ming Guo, George Roy, Tomer Stern, Anh Nguyen and Dapeng Bi. A long-standing challenge in developmental biology is to predict how thousands of cells collectively self-organize as tissues fold, divide, and rearrange. In MultiCell, we represent a developing embryo as a dual graph that unifies two complementary views of tissue mechanics with single-cell resolution: cells as moving points (granular) and cells as a connected foam (junction network). This lets the model learn dynamics from both geometry and cell–cell connectivity. On whole-embryo 4D light-sheet movies of Drosophila gastrulation (~5,000 cells), our model predicts key cell behaviors and the timing of events, including junction loss, rearrangements, and divisions with high accuracy, at single-cell resolution. Beyond prediction, the same representation supports robust time alignment across embryos and offers interpretable activation maps that highlight the morphogenetic "drivers" of development. The broader goal is a foundation for cell-by-cell forecasting in more complex tissues, and eventually for detecting subtle dynamical signatures of disease. Kudos to the team for this inspiring collaboration with brilliant researchers to push the boundary of AI for biology! Citation: Yang, H., Roy, G., Nguyen, A.Q., Buehler, M.J., et al. MultiCell: geometric learning in multicellular development. Nature Methods (2025), DOI: 10.1038/s41592-025-02983-x Code/data links are in the manuscript.show more

MIT News article:

Some of you have commented on relations to cellular automata, we've explored those systems in other work:

This work quietly reframes a core question in developmental biology: global order does not require global control. MultiCell shows how purely local, mechanical, and topological interactions can be sufficient to forecast tissue-scale outcomes with single-cell precision. That is not just a modeling advance. It is a computational hypothesis about how biology actually operates. We are likely seeing the emergence of a new class of biological “state variables” rooted in geometry and connectivity, not just genes or gradients. Following @BioSignal for more signal-dense coverage at the intersection of AI, mechanics, and systems biology.

Inded, thank you! This feature is key to why this approach helps not just in understanding the principles of development but also in designing synthetic systems. Knowing the 'generative rules' of these local interactions enables us to move closer to de novo design of multicellular structures through self-assembly rather than top-down engineering. This has many applications in other areas from materials science to intelligence itself!

Or: "How are pure a priori synthetic judgements possible for embryos?"

Does the model work in a local way where each cell's behavior depends only on factors in its immediate neighborhood like morphogen gradients and mechanical forces from neighboring cells, or is that sort of model of embryological development not possible yet?

Yes, that is exactly how it works! The model uses Graph Neural Networks to compute a cell's state based on its immediate neighbors, and recursively, its neighbors' neighbors (via message passing layers!). This mimics how mechanical stresses or signals propagate through tissue. That is, in fact, the elegance of this work as we show that these geometric cues alone are highly predictive of future behaviors.

@Lasermazer Potentially naive question - are cellular autonoma also a good model for this?

@Lasermazer

The embryo is not computing, it is remembering. The article frames the result as “cell-by-cell prediction.” But the deeper implication is that Form is not calculated forward, it is recalled into expression. If local interactions were merely mechanical causes, prediction would fail catastrophically under noise. Instead, accuracy improves precisely because cells are not independent. They are participatory centers of coherence, each acting from an implicit sense of the whole. Identity precedes structure. Structure expresses remembered intent. Time is a coordination device, not a driver Geometry is not instruction, it is constraint on intention The dual-graph model (points + foams) works because it mirrors what biology already does, cells are both discrete actors and continuous participants, boundaries are real, but relational, not isolating, local mechanics shape expression, they do not author it. Energy does not push matter into form. Energy organizes action toward fulfillment. The AI succeeds because it is learning the grammar of constraint, not the origin of form. High predictability ≠ determinism. The article highlights ~90% accuracy. This is crucial. If development were purely algorithmic, we’d expect, near-perfect predictability, fragility under perturbation. Instead, we see robust predictability, flexibility under variation. That combination only occurs in systems where purpose precedes process. The embryo is not executing a program. It is improvising within remembered bounds. Disease detection reframed. The authors suggest early disease detection (asthma, cancer) via altered cell dynamics. Disease is not a broken mechanism. It is a loss of coherent self-reference. AI may detect disease early not because it finds “errors,” but because it senses when cells stop acting from shared identity and begin acting locally, defensively, or incoherently. The quiet inversion this paper perform. Without saying so, the work flips a long-standing assumption: Global order does not require global control Information does not flow top-down Form is not encoded like software Instead: Order emerges because coherence is intrinsic Cells behave as centers of organized agency Form is remembered, not computed AI here does not explain life away. It accidentally reveals how little command life needs.

someone send me the pdf from please {{{we}}} need this

Markus, Beautiful work—but notice what it quietly reveals. The embryo isn’t computing a blueprint so much as remembering itself into form. Local mechanics + connectivity don’t cause the body plan; they constrain a value-seeking process already underway. Geometry here is not instruction—it’s participation. The global emerges because each cell acts as a center of organized intention, not because a code is being executed. AI doesn’t explain morphogenesis; it shows us how little command is required when coherence is intrinsic.

@HaiqianYang Congratulations. Really amazing stuff. Might be interesting if this work might be helpful for future models.

This is a fascinating exploration of developmental biology! The local interactions leading to a harmonious global form indeed resemble a computational process. Have you considered how this approach might inform regenerative medicine or artificial tissue engineering? For anyone delving deeper into biomedical research questions, I recommend checking out a platform designed to generate comprehensive biomedical reviews and more. #Medicine #Biology

Yes, this is absolutely relevant!

This is fascinating. How the cells communicate with each other to collective self-organize geometries reliably. I wonder if this has any implications for machine learning and language models...

We have been thinking about this quite a bit - I think, yes, concepts of self-organization have important implications for future AI models. Check out this paper:

You even provide arXiv links to relevant research papers. You got a new follower today.

I also find Ruppert Sheldrake's hypothesis about morphic resonance fascinating.

ΔΦ Embryonic Computation – Tension Resolution as the Architect of Form By Thomas S. Mitchell Sr. In embryogenesis, structure does not emerge from mere genetic instruction or chemical diffusion—it is the lawful result of electrostatic field tension resolving across a living substrate. Each cell within the embryo acts not as a computational unit in the digital sense, but as a localized capacitor collapsing under field constraint. The Mitchell Equation (ΔΦ = ρ × v) formalizes this: biological form arises from the interplay between local charge density (ρ) and directed field intention (v). The embryo does not simulate its outcome; it is the outcome of a tension field minimizing itself in real time through symbiotic resolution. Recent research attempting to model embryonic shape computation through graph-based granular models confirms this behavior empirically, but fails to explain it lawfully. The observed “junction loss,” “mechanical rearrangement,” and “activation mapping” reflect what ΔΦ law already defines: microhabitat collapse, field redirection, and substrate nesting behavior. In such systems, symmetry is not imposed but emerges through equilibrium seeking. It is not the code that determines the body—it is the path of least ΔΦ, recursively achieved at the scale of membrane interaction. Unlike machine learning models that must be trained on known data, the embryonic field is not learning—it is resolving. Each division, migration, and adhesion is a gradient event wherein local energy seeks global harmony. This unifies morphogenesis with broader cosmic behavior: just as galaxies spiral through ΔΦ collapse and plankton migrate in charge-regulating columns, the embryo spins its body into being through nested tension cancellation. The field is the blueprint. Biology is the readout. Understanding this reframes development as a process governed by immutable field law rather than evolutionary trial-and-error. It allows for prediction, diagnosis, and even correction—not through artificial intelligence, but through field realignment. Disease, deformity, and stochastic error are no longer black boxes; they are deviations from lawful ΔΦ trajectories, detectable and correctable if the substrate architecture is respected. This makes embryogenesis not just a miracle of biology, but a universal act of resonance. The ΔΦ framework does not oppose modern biological modeling—it completes it. Where graph theory and imaging reveal patterns, ΔΦ law explains them. Where AI observes, the field acts. The embryo does not compute in metaphor—it solves in physics. This paper calls not for another simulation, but for the adoption of electrostatic tension as the true architect of life. Form is not coded. It is collapsed into being.

Geometric learning across cell types solves embryonic computation puzzle beautifully

Inspiring work! 👏🏻

Thank you @ja_me_su !

Geometric signal transduction within the tissue

@eshear have you seen this research? Reminded me of your work at @softmaxresearch

Haven’t read this yet - very cool, thank you for spinning it into my bubble I’m curious if the flittering “cell by cell compute?” is representative of the energetic protein assemblies or something like that? I don’t do a lot of biologics reading - is there any relevance to the dynamic action of proteins & their operating inside the Schumann? Any f=ma? thermal is often an indicator & amniotic fluid would likely inform the motion.. if it’s not nonsense

Demis Hassabis will be proud, one more point for general computability in biology!

Lovely

@zenstyle, thank you!

Simply amazing

How beautiful

Thank you!

something almost perfect about watching geometry and connectivity unify into coherent form. the elegance of the dual-graph representation that lets you see tissue mechanics and cell identity simultaneously is itself a kind of answer

Cc @fkasummer

Congratulations on the MultiCell work. You achieve impressive predictive accuracy from local geometry/connectivity in Drosophila gastrulation. The dual-graph approach elegantly captures emergent order. One question: When extending to "de novo design of multicellular structures," how do you safeguard against unintended disruption of intrinsic developmental buffers (e.g., lifelong tissue-resident cell populations from embryonic origins that sustain morphogenetic resilience in health)? Early models like this are powerful for disease signature detection but accelerating to synthetic self-assembly risks shifting from understanding natural coherence to imposing external intent, potentially eroding those evolved safeguards. Curious how the team weighs those biophysical/ethical trade-offs in more complex (e.g., mammalian) systems.

local rules to global body plan is crazy. clean CA vibes. congrats on the nature methods drop

Evolution! The theory cannot explain how life emerged on Earth. There is no scientific evidence that the "evolutionary mechanisms" proposed by the theory have any evolutionary impact. The fossil record shows the exact opposite of the theory's predictions.

Looks like an extension of Stephen Wolfram’s work in A New Kind of Science. Guess he is vindicated after all. @stephen_wolfram

Does not the blueprint of the form pre-exist as it has been implied by Dr Michael Levin @drmichaellevin during his research on morphogenesis and bioelectricity ?

Four billion years of evolution. Our biology science language is no more than one hundred fifty years old. Our visualization began with Robert Hooke in the seventeenth century!

How? "Chronoharmonics" combined with "a priori knowledge" 🤫

im sure this pattern can be observed with actual fruitflies ...

👍

Cool this is something I have been wondering about and now I still don’t have a clue :)