Fundamental limits

The dose–resolution trade-off, sample thickness, and the missing wedge unique to tomography — the hurdles cryo-EM cannot get around.

It is not that cryo-EM simply lacks resolution — it is that a few physical hurdles cannot be gotten around. Seeing them clearly is what makes every method later on this site make sense.

Intuition

Compress all three hurdles into one sentence: you can only pour so many electrons onto the sample, the sample itself blurs, and you cannot turn it all the way around. The first caps how sharp any single image can be, the second caps how thick the sample can be, the third caps how many directions you can look from. Stacked together, they mean what you actually collect is a set of observations that are few, noisy, and missing angles — not the structure itself. Every method later on this site is, at bottom, fighting those three words: few, noisy, incomplete.

1. The dose budget: the clearer you look, the more you break

Electrons image the sample and destroy it at the same time: every electron that passes through breaks chemical bonds and creates radicals. And high-resolution information dies first — fine structure (side chains, secondary structure, lattice fringes) is the most radiation-sensitive and is wiped out while the accumulated dose is still low, whereas coarse outlines survive longer. That forces a brutal trade-off:

want a sharp (high-resolution) view → you need more electrons to lift the signal out of the noise → but the sample is wrecked first;
want to preserve the structure → you must go low-dose → and a single image has miserable signal-to-noise.

Put numbers on it. A single cryo projection typically uses only a few electrons per square ångström, and a whole tilt series is often kept within a few tens of $e^-/\text{Å}^2$ total (beyond that, high-resolution information is destroyed before it can be imaged). At electron counts this sparse, the number of electrons a pixel receives is itself a small random number governed by Poisson statistics: if the expected count is $N$ , the fluctuation is about $\sqrt{N}$ , so the relative noise is about $1/\sqrt{N}$ . Small $N$ means large relative noise — a hard physical floor that no better camera can lift.

This is why a single cryo image is almost “all noise”: by eye you can barely tell there is any structure in there at all.

Depth

Think through how the dose is split, and you see why single-particle and tomography have such different fates. Given a total dose budget $D$ (capped by radiation damage), you can spend it two ways:

Single-particle: put all of $D$ into one image, getting a 2D projection with tolerable SNR — but each molecule is captured in only one orientation. You then find thousands of identical copies of the same molecule in the sample, align them, and average; the SNR improves as $\sqrt{n}$ , where $n$ is the number of copies — average ten thousand copies and the effective SNR rises about 100-fold. The price: you can only solve structures that exist in many identical copies.
Tomography (cryo-ET): you need many tilt angles of the same one-of-a-kind object (say a stretch of membrane inside a cell), so you must split the same budget $D$ across dozens of images. Each gets a fraction of the single-particle dose, so the per-image SNR is so low it is nearly invisible — and the object is unique, with no copies to average.

So tomography works at intrinsically worse SNR and has no “average over copies” escape hatch. Recovering the signal from that noise then rests on statistical modelling of the structure: quantify how much information survives in the language of signal-to-noise, then fill the rest with a prior.

Single-particle pulls the SNR up by averaging thousands of copies; tomography has no such crowd and must cope at low SNR directly — which is exactly where statistical modelling earns its keep. And because the dose has to be split, the one exposure at each tilt is itself broken into a movie of dozens of frames, motion-corrected and summed back afterwards (dose fractionation) — a little post-processing to squeeze every one of those scarce electrons.

2. The sample must be thin

Electrons have limited penetration: the thicker the sample, the more likely a single electron is scattered more than once on the way through (multiple scattering). Single scattering carries the phase information we want; multiple scattering scrambles it together, so the image blurs and contrast degrades. That caps the imageable thickness at a few hundred nanometres.

Intuition

Picture the beam as light through frosted glass. Thin glass: you can read the text behind it (single scattering, clean information). Thick glass: just a uniform glow (multiple scattering, information mashed together). The ice slab is that glass — too thick, and no microscope can recover it.

Many interesting objects (a whole eukaryotic cell) are several micrometres thick, far past this ceiling. To do in-situ tomography (imaging molecular machines directly inside a cell), you usually first use a focused ion beam (FIB) to shave the frozen cell, like a plane, into a lamella roughly 100–200 nm thick, then image that. This turns “the sample is too thick” from a dead end into a process step — but the price is that you only ever see the thin slice that was milled out.

3. Unique to tomography: the missing wedge

Key

The stage cannot tilt the full ±90°: at steep tilts the effective thickness the beam must cross grows as $1/\cos\theta$ (already doubled at $60^\circ$ , where signal degrades sharply), and the holder mechanically gets in the way. So the tilt usually stops near ±60°, leaving an unsampled pair of wedges in Fourier space — the missing wedge. It stretches and smears the reconstruction along one axis.

Quantify it. The angles you can collect cover only $\pm\theta_{\max}$ ; the rest is a gap. The half-angle of the missing wedge is

\alpha = 90^\circ - \theta_{\max},

where $\theta_{\max}$ is the largest tilt you can actually reach, and $\alpha$ is the half-angle of the unsampled wedge on each side in Fourier space. At $\theta_{\max}=60^\circ$ , $\alpha=30^\circ$ — meaning about a third of all directions are never measured. Why the gap is exactly a pair of wedges, and why it stretches the structure along Z, follows from the central-slice theorem; see the missing wedge page for the geometry.

The key point: the missing wedge is not noise — it is a whole block of information that was never measured. Noise can be beaten down by taking and averaging more shots; missing data cannot be denoised or filtered into existence — it simply is not in the data. To fill it, you must bring in a prior about what a plausible structure looks like (membranes are continuous and smooth, density is non-negative, similar molecular machines should look similar) and solve it as an inverse problem. That prior is the common starting point for every reconstruction method later on this site.

And one more: CTF modulation

Even setting the missing wedge aside, the microscope does not hand you a faithful projection of the structure. To get contrast, images are usually taken underfocus, which introduces a contrast transfer function (CTF): in Fourier space it acts like a set of oscillating rings that attenuate some spatial frequencies, flip the sign of others, and zero out specific frequencies entirely at each ring of zeros. In other words, the observation is the true signal modulated by this CTF (and then buried in noise). The reconstruction has to deconvolve it, or the structure comes out with spurious fringes and inverted contrast. For the shape of the CTF and where its zeros fall, see the contrast transfer function page.

Turning limits into a problem

Gather the hurdles together — low SNR, CTF modulation, the missing wedge — and they are all the same sentence: infer the one clean structure from incomplete, noisy, modulated observations. Written out, the observation $y$ relates to the unknown structure $x$ roughly as

y = \underbrace{C}_{\text{CTF}}\;\underbrace{W}_{\text{wedge-limited projection}}\,x \;+\; \underbrace{n}_{\text{noise}},

where $x$ is the 3D density we want, $W$ is the incomplete measurement operator that projects only from the angles in $\pm\theta_{\max}$ , $C$ is the CTF modulation, and $n$ is the noise from low dose. $W$ is not invertible (a block is missing) and $n$ is large, so inverting it directly is ill-posed: infinitely many $x$ satisfy this equation.

To pick the sensible one out of that infinity, you add a prior $p(x)$ , and the problem becomes a Bayesian inverse problem: infer $x$ under $p(x\mid y)\propto p(y\mid x)\,p(x)$ . The rest of this site takes it apart with signal processing (CTF, filtering, sampling), probability and statistics (MAP and EM, Bayesian inference), optimal transport, and the other tools — landing finally on our own Cryo-ET reconstruction methods. What separates those methods is precisely how they answer this ill-posed inverse problem: a point estimate (CryoGEN-I), a stable single answer (CryoGEN-II), or a whole family of posterior samples (CryoWGEN).

← Electron Tomography