The Microscope That Had to Be Frozen

Consider an instrument with a self-defeating design. An electron microscope can in principle resolve individual atoms; the wavelength of an accelerated electron is far shorter than anything visible light can offer, and the physics has never been the obstacle. The obstacle is that electrons deposit energy in whatever they pass through, and biological material is held together by bonds that a fast electron breaks casually. By the time enough electrons have gone through a protein to form a decent image of it, the protein is no longer there. What has been photographed is the wreckage.

There was a second problem, in some ways more absurd. An electron microscope works in a high vacuum, and a protein without water around it collapses. So the technique that could theoretically see atoms was, for biology, restricted to specimens that had been dried, stained with heavy metal salts, and thereby killed twice over. For decades the honest description of biological electron microscopy was that it produced pictures of the shape of the dried residue a molecule leaves behind.

The instrument that destroys its subject

Radiation damage sets a hard budget. Beyond a certain accumulated dose the fine structural detail is gone, and that limit is reached long before a single molecule has scattered enough electrons to produce an interpretable picture of itself. The arithmetic is unforgiving: the image you are permitted to take is far too noisy to read.

The workaround the field settled on in the 1950s and 1960s was negative staining. Surround the specimen with a heavy-metal salt, dry it, and image the cast. The stain scatters electrons strongly and tolerates the beam, so the picture is robust — but it is a picture of a mould, at whatever resolution the stain's own granularity permits, and it says nothing about the interior of the molecule. Useful for shape and size. Useless for chemistry.

The alternative was to accept the noise and defeat it statistically. If you have many identical objects, each imaged at a dose too low to damage it and therefore too low to see it, you can add the images together. Random noise cancels; the common signal accumulates. Nothing about this is obvious in practice, because before you can average images you have to know how they are related to each other — how each particle is rotated and positioned — and that has to be worked out from the same hopelessly noisy pictures.

The first demonstration that this could reach real structural detail came in 1975, when Richard Henderson and Nigel Unwin produced a three-dimensional model of bacteriorhodopsin in purple membrane from low-dose electron images and diffraction patterns 1. It resolved the arrangement of the helices crossing the membrane and it did so on a membrane protein, at a time when membrane proteins were essentially unavailable to crystallography. It was also, importantly, a hint rather than a general solution: it depended on the protein being arranged in a naturally occurring two-dimensional lattice, which is a kind of crystal by another name.

Freezing water without letting it become ice

The vacuum problem was solved by a route that sounds like it should not work. Freezing a specimen keeps its water in place, but ordinary freezing forms crystalline ice, and ice crystals wreck a specimen mechanically while also diffracting the beam and obliterating the picture. Water, however, does not have to crystallise. Cooled fast enough, its molecules are immobilised before they can arrange themselves, and it sets as a glass — amorphous, transparent to electrons, holding whatever is suspended in it exactly where it was.

Jacques Dubochet's group at the European Molecular Biology Laboratory worked out how to do this reliably: a thin film of sample spread across a holey support, blotted to a film only tens of nanometres thick, then plunged into liquid ethane, which conducts heat away fast enough to beat crystallisation. The 1988 review setting out the method and its physical basis remains the reference account 2.

Two things follow from this that are worth separating. The first is that the specimen is now hydrated and unstained — the electrons interact with the protein itself, not with a metal cast of it. The second, and larger, is that the molecules do not need to be in a crystal. They are suspended in the glassy film in whatever orientations they happened to adopt, scattered at random. If the computational problem of sorting out those orientations can be solved, then the requirement that has governed structural biology since 1912 — grow a crystal first — quietly disappears.

Joachim Frank had been developing exactly that computational machinery since the 1970s: classify the particle images, align them, determine each one's orientation, and back-project the lot into a three-dimensional density. Single-particle analysis worked, and through the 1990s and 2000s it produced a steady output of structures — at resolutions where you could see the shape of a molecule and the disposition of its domains, but rarely the chain itself. The field acquired the affectionate and slightly wounding nickname blobology.

The camera, and the revolution it caused

What changed around 2012 and 2013 was not the microscope. It was the thing at the bottom of it recording the picture.

Until then, electrons were detected indirectly. They struck a scintillator, which converted them into light, which was piped to a conventional sensor. Every conversion in that chain blurred and lost signal, which was catastrophic in a technique already operating at the absolute floor of available dose. Direct electron detectors dispensed with the intermediaries: the electrons hit the semiconductor sensor itself, and were registered individually as counted events.

The gain in sensitivity would have been significant on its own. The second consequence was greater. Because these detectors could read out very fast, an exposure could be recorded as a short movie rather than a single frame — and inspecting those movies revealed that the specimen had been moving all along, the beam itself causing the film to shift and buckle during the exposure. Every previous image had been blurred by motion nobody could see. Once the frames could be aligned to each other computationally before summing, that blur could be removed 3.

Structures that had been stuck at low resolution for years were re-solved with atomic detail within months. Werner Kühlbrandt gave the phenomenon its name in a 2014 commentary, and the phrase resolution revolution stuck because for once the hyperbole was proportionate 4. The Nobel Prize in Chemistry followed in 2017, awarded to Dubochet, Frank and Henderson for developing cryo-electron microscopy for the high-resolution structure determination of biomolecules in solution — a citation that neatly divides the three problems between the three recipients.

The receptors finally became visible

For peptide science specifically, one class of target mattered more than any other, and it had been the least cooperative in all of structural biology.

G protein-coupled receptors sit in the cell membrane, and most peptide hormones act on them. They are flexible by design, since signalling requires them to change shape. They are unstable outside a lipid environment. They exist in equilibrium between states, which is precisely the property a crystal cannot accommodate. Crystallography did eventually reach some of them, but only through heroic engineering — thermostabilising mutations, fusion partners inserted into flexible loops, deletion of the very regions that moved — and each of those interventions raised a fair question about whether the structure obtained was the structure of the receptor or of the construct built to make it crystallisable.

The larger prize was the activated complex: the receptor with its agonist bound and its G protein attached, which is the assembly that actually transmits a signal. That object is large, asymmetric, flexible and firmly uncrystallisable. From 2017 onward, cryo-EM began delivering it. Among the first was the activated GLP-1 receptor bound to its G protein, published that year 5, and the peptide-hormone receptors that had been inferred from decades of binding assays and mutagenesis began appearing as structures — the peptide threaded into a large extracellular domain and down into the transmembrane core, the receptor's helices splayed on the intracellular side to grip the G protein.

The consequence for receptor pharmacology is that arguments about how a peptide agonist engages its receptor, or why one analogue favours one downstream pathway over another, moved from indirect inference to something that could be looked at. Whether looking at it is enough to design the next analogue is a separate and much less settled question.

What is still out of reach

The technique has a lower size limit, and it is inconvenient. Determining a particle's orientation requires enough contrast in its image to align it against others, and contrast scales with mass. Very small proteins produce images too faint to orient reliably, which is why the method's routine territory begins somewhere in the low hundreds of kilodaltons and becomes difficult well above the size of anything that could be called a peptide. A short peptide bound to a large receptor is entirely visible; the same peptide on its own is not a cryo-EM target in any sense.

• Small and flexible objects remain hard, because alignment needs mass and needs a consistent shape to align.
• Preferred orientation persists — particles that all lie the same way in the film leave whole directions of the reconstruction poorly sampled.
• Averaging assumes the particles are identical, so genuine conformational heterogeneity is smeared out unless it can be sorted computationally.
• The map is a static average of a molecule that was moving, and it carries no direct information about the timescales on which it moved.

Those limits are being worked at rather than lived with. Sorting mixed conformations out of a dataset has become a research field of its own, and imaging molecules inside intact cellular material rather than in purified suspension is now a serious pursuit rather than an aspiration.

But the arc has already completed in the way these stories do. A structure that would once have consumed a laboratory for a decade is now a booking on an instrument, a few days of data collection, and a processing pipeline. The photographs are still individually almost empty — each particle still recorded at a dose that reveals nearly nothing, exactly as the radiation problem demands. There are simply so many of them, added together so well, that the emptiness no longer matters.