Welcome to the exhibit. Please step closer to the glass, metaphorically speaking, because the object on display is only about 27 nanometers wide: the MS2 bacteriophage, a tiny RNA virus that infects E. coli. It wears an icosahedral capsid, which is a fancy way of saying nature gave it 20 triangular faces and the geometric confidence of a Dungeons & Dragons die. Icosahedral shells are common in viruses because they pack genetic material efficiently, like a very committed carry-on suitcase with protein panels instead of zippers (Capsid, Wikipedia; MS2, Wikipedia).

The new paper by Mall and colleagues asks a deceptively simple question: what happens to that shell when the virus dries out in air? Not “after drying,” not “in a frozen pose,” but while it is passing through the weird, hostile airport security line of aerosolization, vacuum, and X-ray beams (Mall et al., 2026).

The Tiny Shell That Flinched

If you look closely, the surprise is not that the capsid changes. Soft biological stuff changes all the time. The surprise is how it changes.

The old expectation was something like a neat mechanical transition: a symmetric shell buckles in a low-energy way, maybe protecting the RNA inside from further dehydration. The paper reports direct experimental evidence for this kind of drying-linked buckling, but with a twist: the capsid does not politely preserve its icosahedral symmetry. It buckles incoherently. Translation: different regions seem to cave or compact at different times, like a plastic water bottle losing pressure in one awkward dent instead of shrinking with architectural dignity.

The team found a previously unreported compact conformation and intermediate shapes between hydrated and dehydrated states. Notice how that matters: intermediates are the good museum labels. They tell you not just what something is, but how it got there.

The Camera Is Basically Lightning With Homework

The experiment used X-ray single-particle imaging at the European XFEL. The researchers aerosolized MS2 particles with electrospray, focused them into an X-ray beam, and collected diffraction snapshots at up to 3520 frames per second. That sounds smooth until you remember the particles are tiny, randomly oriented, and usually miss the beam entirely. Science, as usual, is mostly trying to photograph a flea in a snowstorm while the camera is exploding politely.

They identified 287,168 potential diffraction hits, with each useful frame containing only a few thousand scattered photons on average. Instead of taking one clean picture, they gathered a crowd of noisy shadows and reconstructed the story from the crowd.

This is where machine learning walks into the exhibit wearing sensible shoes.

The AI Sorting Hat For Viral Origami

The researchers used 2D classification, then trained a beta-variational autoencoder, or beta-VAE, on thousands of averaged diffraction patterns. A VAE learns a compressed “latent space,” which is basically a map of hidden variation. If regular image sorting says “this one goes in pile A,” a VAE says “this one lives over here between slightly squashed, suspiciously round, and what-happened-to-your-left-side.” Neural networks: occasionally useful, despite being math lasagna.

The model helped organize MS2 shapes into a structural landscape. One path showed larger, water-wrapped particles moving toward smaller, drier capsids. Another showed shape changes at roughly constant size, revealing asymmetry and local deformation. The authors are careful here: the particles were not timestamped individually, so the drying timeline is inferred from the experimental setup, particle sizes, and known hydrated structure. Good. We like our claims with seatbelts.

The 19-Residue Plot Twist

Now come around to the molecular display case. The proposed trigger involves the FG loop, a flexible segment of the MS2 coat protein around residues 66-82. In the hydrated capsid, water molecules help stabilize parts of this loop. During drying, simulations suggest some loops become destabilized, especially around the 3-fold and 5-fold pores. The capsid then compacts locally.

That is wonderfully specific. Not “the virus shrinks because dry bad,” but “this little protein loop may lose stabilizing water and help the shell buckle.” Biology often hides the big drama in a short floppy segment, like a blockbuster movie hinging on one intern with a keycard.

Why This Belongs In The AI/ML Gallery

This paper is not just about one bacteriophage having a bad day in dry air. It shows how AI can help structural biology handle messy heterogeneity: many particles, many orientations, weak signals, and no single perfect average. Recent work is pushing the same direction, including scalable online reconstruction for X-ray single-particle imaging (Shenoy et al., 2025), self-supervised reconstruction from diffraction data (Chen et al., arXiv:2311.16652), and studies of gas background and sample delivery that determine whether the signal survives the trip through the machine (You et al., 2025; Yenupuri et al., 2024).

If this approach scales, researchers could watch more biomolecules move through realistic, non-frozen conditions: viruses drying in aerosols, vaccine particles surviving spray-drying, proteins changing shape before traditional methods can pin them down. The caveats stay on the wall label: this is MS2, not every airborne virus; the resolution is coarse compared with atomic structures; and morphology does not automatically tell us infectivity. Still, the exhibit earns your attention.

Because if a virus shell can protect its genome by denting in just the right places, and machine learning can help us see the dent, then the next room of structural biology may look a lot less like a still-life gallery and more like a slow-motion action sequence.

References

Mall, A., Munke, A., Mazumder, P., et al. “High-throughput in situ single particle X-ray imaging of dehydrating viral capsids.” Light: Science & Applications 15, 280 (2026). DOI: 10.1038/s41377-026-02262-0. PMID: 42337245.

Shenoy, J., Levy, A., Ayyer, K., et al. “Scalable 3D reconstruction for X-ray single particle imaging with online machine learning.” Nature Communications 16, 6812 (2025). DOI: 10.1038/s41467-025-62226-7.

Chen, Z., Wang, C., Gao, M., et al. “Augmenting X-ray single particle imaging reconstruction with self-supervised machine learning.” arXiv: 2311.16652 (2023). DOI: 10.48550/arXiv.2311.16652.

You, T., Bielecki, J., & Maia, F. R. N. C. “Impact of gas background on XFEL single-particle imaging.” Scientific Reports 15, 29559 (2025). DOI: 10.1038/s41598-025-15092-8.

Yenupuri, T. V., Rafie-Zinedine, S., Worbs, L., et al. “Helium-electrospray improves sample delivery in X-ray single-particle imaging experiments.” Scientific Reports 14, 4401 (2024). DOI: 10.1038/s41598-024-54605-9.

Ekeberg, T., Assalauova, D., Bielecki, J., et al. “Observation of a single protein by ultrafast X-ray diffraction.” Light: Science & Applications 13, 15 (2024). DOI: 10.1038/s41377-023-01352-7.

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.