Lab Newsletter — August 31, 2026: A Thousand Frozen Poses

Three weeks ago we watched AI teach proteins to move — generative models that emulate molecular dynamics and turn AlphaFold’s single fold into a whole ensemble of shapes, all predicted from sequence. Today’s story is the mirror image: recovering that same motion not from a model of physics but from actual pictures — and it starts with a beautiful, awkward fact about how cryo-electron microscopy really works. A cryo-EM dataset is not one photograph of a protein. It’s millions of noisy 2D snapshots, each a different individual molecule flash-frozen in a thin film of ice — and every one of those molecules was caught mid-wiggle, frozen in a slightly different pose. For decades the entire craft was to average those millions of poses into a single, gorgeously sharp 3D structure. But averaging the wiggle away throws out exactly the thing biology runs on: motion. The new idea is to stop averaging and start reconstructing the whole crowd.

🧊 The breakthrough: reconstruct the distribution, not the average

The turn came when neural networks were pointed at the problem. As cryoDRGN (Zhong, Bepler, Berger & Davis, Nature Methods, 2021, from MIT) framed it, “many imaged protein complexes exhibit conformational and compositional heterogeneity that poses a major challenge to existing three-dimensional reconstruction methods” — the polite way of saying the average is a lie when the molecule is a shape-shifter. Their answer was “an algorithm that leverages the representation power of deep neural networks to directly reconstruct continuous distributions of 3D density maps and map per-particle heterogeneity of single-particle cryo-EM datasets.” Instead of one structure, a landscape of them — and the biology fell out immediately: with cryoDRGN they “uncovered residual heterogeneity in high-resolution datasets of the 80S ribosome and the RAG complex, revealed a new structural state of the assembling 50S ribosome, and visualized large-scale continuous motions of a spliceosome complex.” A machine in motion, recovered from a pile of frozen stills.

🌀 Modeling the motion itself

cryoDRGN learns that the molecule varies; the next wave learns how it moves — as physical motion, not just statistical variation. 3DFlex (Punjani & Fleet, Nature Methods, 2023, from the Toronto group behind cryoSPARC) is “a motion-based neural network model for continuous molecular heterogeneity” that “exploits knowledge that conformational variability of a protein is often the result of physical processes that transport density over space and tend to preserve local geometry” — baking in the physical prior that a flexing domain bends, it doesn’t teleport. Because it treats all the poses as views of one deforming object, “3DFlex can improve 3D density resolution beyond the limits of existing methods because particle images contribute coherent signal over the conformational landscape” — motion that used to blur the map now sharpens it. And the tooling has gone open and honest: DynaMight (Schwab, Kimanius … Scheres, Nature Methods, 2024, from the MRC LMB) calls continuously flexing molecules “one of the biggest outstanding challenges in single-particle analysis,” and “estimates a continuous space of conformations … by learning three-dimensional deformations of a Gaussian pseudo-atomic model … for every particle image,” then inverts those deformations for “an improved reconstruction of the consensus structure.” Crucially, it ships “as free, open-source software, as part of RELION-5” — and its authors say the quiet part out loud: over-relying on atomic-model priors “may lead to important artifacts due to model bias.”

🧭 The honest frontier — and why it’s our kind of problem

That last confession is the whole ballgame, and it’s the lab’s native tongue. A recovered “motion” is a model of the data, not a direct observation — and if your prior is too strong, the algorithm can paint a beautiful, confident conformational change that was never there. So how do you tell a real motion from a hallucinated one? You need ground truth — which the field mostly didn’t have. That gap is exactly what CryoBench (Jeon et al., 2024, senior author Ellen D. Zhong) sets out to close: “a suite of datasets, metrics, and benchmarks for heterogeneous reconstruction in cryo-EM,” built with known ground-truth heterogeneity — from antibody-complex motions and molecular-dynamics simulations to ribosome assembly states — against which neural and classical reconstruction tools can finally be measured, in the hope of a “foundational resource for accelerating algorithmic development and evaluation.” That is the prove-it discipline this digest keeps coming back to: a method earns trust only when it can be scored against something real, and the honest move is to build the benchmark before you believe the picture.

And it lands right where we live. Notice the symmetry with Aug 11: there, motion was predicted from sequence; here, it’s recovered from images — two roads to the same conformational landscape, and a serious virtual cell will want both, each checking the other. Because dynamics is the molecular raw material a Human Cell Simulator has to reproduce: not a gallery of frozen statues but machines that open, close, and hand off. And notice how the whole field moves — cryoDRGN, cryoSPARC, RELION-5, CryoBench are all open, callable, benchmarked tools, the same BioImage Model Zoo and BioEngine ethos we keep betting on: publish the model and the test that could embarrass it. Freeze ten million molecules mid- motion, teach a network to read the crowd instead of the average, and keep it honest against ground truth — and a drawer full of frozen poses becomes a movie of a molecule doing its job.

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