What cryo-EM is

Freeze molecules in vitreous ice and image them with electrons — what cryo-EM is, and its two main roads.

Cryo-electron microscopy (cryo-EM) does one thing, plainly put: it freezes biomolecules into a layer of glass-like ice and photographs them with a beam of electrons. It does not stain, dry, or crystallize the sample — it tries to keep the molecule as it was in water, close to its native state. The 2017 Nobel Prize in Chemistry went to this method.

To see why this is hard, line up three scales. A typical protein complex is a few nanometres across; to resolve atomic detail you need to see spacings near 0.3 nm; and a frozen sample is often tens to a few hundred nanometres thick. So you are reading out a nanometre-scale shape inside a slab of ice a hundred times thicker than your target — a slab that is easily destroyed. Two things make it possible: the short wavelength of electrons, and freezing the sample into a glass.

Intuition

Why freeze at all? The microscope column is a hard vacuum, and the electron beam does radiation damage; a hydrated room-temperature sample would dehydrate and be destroyed. Flash-freezing it in milliseconds into vitreous ice (glass-like, no crystal lattice, no ice crystals) both locks in the near-native conformation and lets it survive the vacuum and part of the electron dose. Another way to see it: at room temperature water molecules are always moving — like motion blur in a long exposure. Flash-freezing fixes that single frame, so what you photograph is this instant, not a time-average.

One beam, one projection

Electrons have a far shorter wavelength than visible light, so an electron microscope reaches the nanometre and near-atomic scale. A light microscope is capped by the wavelength of visible light (about 400–700 nm) and cannot separate two points much closer than that; an electron accelerated through 300 kV has a wavelength near 2 pm (0.002 nm), so its in-principle resolution limit sits far below atomic spacings. What actually caps resolution is no longer the wavelength but beam damage, ice thickness, and lens aberrations.

The beam passes through the thin icy sample; structures of different density scatter electrons differently, forming a 2-D projection on the detector. The word projection is the point: the detector records the scattering accumulated by electrons travelling along the optical axis through the whole sample — it flattens the 3-D structure along one direction into a 2-D image. Depth information is lost in that step.

Worse, the signal is faint. Every electron that contributes to an image also breaks chemical bonds, so the total dose must be kept very low (typically tens of electrons per Å²), which makes a single projection extremely noisy — the structure is barely visible by eye. How you recover 3-D structure from these thin, noisy, flattened projections is what splits cryo-EM into two roads.

Two roads: single-particle averaging vs tomography

Two roads of cryo-EM: single-particle vs Cryo-ET

Acquire → Reconstruct

Fourier coverage

Limited tilt → missing wedge → vertical smear

Tomography (Cryo-ET)

For a unique, in-situ structure, averaging is impossible. You tilt the same region through ±60°, angle by angle, then reconstruct. Fourier space is missing a wedge, so the reconstruction is stretched along the beam axis — the missing wedge this site sets out to fix.

Both roads attack the same mathematical problem: a single projection collapses 3-D into 2-D and throws away depth. The only way to put depth back is to record several projections from different directions and stitch them in Fourier space — this is the central-slice theorem: the Fourier transform of one projection equals a slice through the origin of the Fourier transform of the 3-D object. The more directions you collect, the more of Fourier space you fill, and the more complete the reconstruction. The two roads differ only in where those different directions come from.

Single-particle analysis (SPA): the sample holds thousands of copies of one molecule, in random orientations. Align and average them and each orientation fills in a slice of Fourier space — full coverage, near-atomic resolution. One concrete payoff is noise: averaging $N$ aligned copies suppresses random noise by roughly $\sqrt{N}$ , so even when a single projection is dominated by noise, averaging tens of thousands of them lets the signal surface. The cost: you need many separable, conformationally consistent copies — which means purifying the molecule out of the cell, and so discarding its native surroundings.
Cryo-electron tomography (cryo-ET): for a unique, in-situ structure (the inside of a whole cell, say), averaging is impossible. You tilt the same region through a series of angles, one image each, and reconstruct a 3-D volume. Two hard constraints follow: the dose budget is fixed and split across dozens of tilt images, so each is noisier than an SPA exposure; and the stage cannot tilt to 90° (it usually stops at ±60°–70°), so the directions never reached leave a whole gap in Fourier space — the missing wedge.

Deep

Write the above as formulas. Let the 3-D object be $f(\mathbf{r})$ , and let the projection along direction $\theta$ be $p_\theta(\mathbf{x})$ — the line integral of $f$ along the optical axis. The central-slice theorem states: the 2-D Fourier transform $\hat p_\theta$ of $p_\theta$ equals $\hat f$ , the 3-D Fourier transform of $f$ , evaluated on the plane through the origin with normal $\theta$ . Here $\mathbf{r}$ is a 3-D coordinate, $\mathbf{x}$ a detector-plane coordinate, and $\hat f$ describes the structure’s content at each spatial frequency (higher frequency = finer detail).

So acquisition is filling $\hat f$ one slice at a time. SPA’s orientations cover the whole sphere and can in principle fill the entire Fourier ball; cryo-ET’s tilts sweep only a limited arc, leaving a pair of cone-shaped gaps along the tilt axis — in a 2-D cross-section these look like two wedges, hence missing wedge. Frequencies never sampled are set to zero in the reconstruction, which shows up as stretching, blur, and artefacts along the missing direction. Filling that gap back in is exactly where statistical machine learning enters.

This site is about the latter — cryo-ET, and how we use statistical machine learning to fill that missing wedge back in (see Cryo-ET Reconstruction). This is not inventing detail from nothing: it takes the set of all 3-D structures consistent with the measured projections and uses a statistical model to pick one plausible answer, or to give a family of possible answers. Both roads share the same imaging physics — vitrification, the dose budget, low-SNR projections — and the next page lays out exactly how a single dataset is acquired.

← Electron Tomography