In situ structural biology

Cryo-ET resolves macromolecular structures directly inside cells, in their native cellular context.

In situ structural biology studies macromolecules where they actually work — inside the cell — rather than after purification. Electron tomography is its central tool: a vitrified cell, or a thin lamella milled from one, is imaged as a tilt series and reconstructed into a 3D tomogram that captures complexes in their native surroundings, embedded in membranes, bound to partners, and arranged as the living cell left them.

Zoom across scales — from a whole cell to molecules in the tomogram:

In-situ Cryo-ET spans three scales: vitrify a whole cell, cut a ~200 nm slab from it with a focused ion beam (FIB), then resolve macromolecules in their native locations inside the tomogram — without ever purifying them out.

Reaching an interpretable tomogram spans several orders of magnitude in scale. A eukaryotic cell is roughly 10–20 micrometers across, while a ribosome we want to resolve is about 25 nanometers — nearly three orders of magnitude apart. A whole cell is too thick for the electron beam, so cryo-focused-ion-beam (cryo-FIB) milling carves away most of the material above and below a target region in the vitrified state, leaving a slab roughly one to two hundred nanometers thick; the internal detail of the target complexes then emerges at the nanometer scale. This progression — cell, lamella, tomogram — is the backbone of the in situ workflow.

This native context is the defining advantage. Purification can strip away interaction partners, membrane environments, and the spatial organization of a complex; in situ imaging preserves all of it. The approach gives rise to visual proteomics — mapping the identities and locations of many molecular species within a single cellular volume — and to what is sometimes called molecular sociology, the study of how molecules are distributed, oriented, and associated with one another in space. Beyond individual structures, tomograms reveal the crowding of the cytoplasm, where macromolecules pack at concentrations that purified samples never reproduce.

Intuition

A purified structure is a portrait taken in a studio. An in situ structure is the same molecule photographed in the crowd it lives in — less posed and noisier, but showing who it stands next to and where it sits, information no studio portrait can carry.

The cost is signal. A cell is far thicker and denser than a thin layer of isolated particles, so even a milled lamella scatters more, and every target sits against a dense background of other molecules. The signal-to-noise ratio is markedly lower than for purified specimens, and the missing wedge still distorts the reconstruction. Identifying and resolving a complex therefore leans on the full downstream pipeline: segmentation to map cellular architecture, particle picking to locate copies, and subtomogram averaging to reach high resolution by combining many instances.

The same low-SNR, crowded conditions motivate learned restoration. Learned missing-wedge restoration methods such as CryoGEN and CryoWGEN aim to make in situ tomograms more interpretable, supporting the picking, segmentation, and averaging on which in situ structural biology depends.

Finding the target: CLEM and FIB milling

A whole frozen cell offers no obvious landmark for a complex that occupies a vanishingly small fraction of its volume, so targeting is a problem in its own right. Correlative light and electron microscopy (CLEM) addresses it by first imaging the cell with cryo-fluorescence: a labeled structure lights up at a known position, and that fluorescence map is registered to the electron view so the milling and imaging can be aimed at the right region. The thin lamella itself is produced by cryo-FIB milling, where a focused gallium ion beam ablates material above and below a chosen plane until only a slab of one to a few hundred nanometers remains — thin enough for the electron beam, yet still embedded in vitreous ice. Correlation can be carried through milling so the fluorescent target is not accidentally removed.

The lamella thickness is not arbitrary; it is wedged between two demands that pull against each other. More signal argues for a thicker slab, since a thicker slab holds more copies of the target in the field of view. But electrons scatter inelastically inside the specimen, and the longer their path, the more of them are scattered out of the imaging optics and into the background — so a thicker slab also raises the noise.

Depth

To make the thickness trade-off concrete, the governing scale is the inelastic mean free path $\Lambda$ — the average distance an electron travels before it is inelastically scattered once. At 300 kV in vitreous ice, $\Lambda$ is about 300–400 nm. The fraction of electrons that cross a slab of thickness $t$ having scattered zero or one times — that is, usefully for imaging — falls off roughly as

I/I_0 \approx e^{-t/\Lambda}

where $I_0$ is the incident electron count, $I$ is the count still inside the useful energy window, $t$ is the lamella thickness, and $\Lambda$ is the inelastic mean free path.

Taking $\Lambda \approx 350$ nm: at $t = 100$ nm about $e^{-0.29}\approx 75\%$ of electrons are usable; at $t = 200$ nm this drops to about $e^{-0.57}\approx 57\%$ ; and by $t = 500$ nm only about $24\%$ survive. That exponential is why lamellae land in the 100–200 nm range — in a thicker slab nearly every electron is multiply scattered, adding blur without adding signal.

This constraint compounds with the tilt series: at tilt angle $\theta$ the effective path through the lamella grows as $t/\cos\theta$ , already doubled at $60^\circ$ . So high-tilt images are not only limited by missing-wedge geometry but are themselves the thickest and noisiest — a double source of the tomogram’s anisotropy.

Why crowding makes restoration matter

Inside a cell the macromolecular concentration is hundreds of milligrams per milliliter, so every target sits against a dense, structured background rather than empty buffer. That background is signal from other molecules, not just noise, and it cannot be subtracted away the way a thin layer of ice can. Combined with the dose limits of a thick specimen and the missing wedge, the result is tomograms in which individual complexes are barely distinguishable by eye.

The reason averaging is unavoidable is a simple counting argument. The SNR of a single particle in an in situ tomogram can be of order 0.1 — in one copy, the molecule’s shape is buried in noise. Stack $N$ aligned identical copies and the coherent signal adds linearly while the incoherent noise adds only as $\sqrt{N}$ , so the SNR improves as $\sqrt{N}$ . Raising the SNR by one order of magnitude takes about 100 copies; another order of magnitude takes tens of thousands. That is what subtomogram averaging does — but it carries a precondition: the copies must first be found and oriented in a volume where they are hard to see.

Learned restoration is valuable here precisely because that precondition is itself bottlenecked by low SNR: picking and segmentation run on the un-averaged, anisotropic, crowded volume. Methods such as CryoWGEN that restore the missing wedge raise the contrast of crowded tomograms enough for particle picking and segmentation to gain a foothold, handing the downstream average the copies it needs.

← Electron Tomography