Subtomogram averaging

Averaging many copies of a repeating particle in 3D raises signal-to-noise and pushes tomography toward sub-nanometer resolution.

Subtomogram averaging (STA) recovers a high-resolution structure from the many noisy copies of a molecule present in a tomogram. Individual particles in a dose-limited reconstruction are buried in noise and distorted by the missing wedge, so no single copy resolves fine detail. But a tomogram may contain hundreds or thousands of instances of the same complex. Each is extracted as a small 3D sub-volume — a subtomogram — and these are aligned to a common reference and averaged.

The whole method rests on one assumption: the copies are identical. They are the same molecular machine repeated throughout the tomogram, each sitting at a different position and turned by a different random angle about its own axis. Once every copy is rotated and translated into the same pose, their true structural signal should line up voxel for voxel; the only thing that does not line up is each copy’s independent noise. STA exploits exactly that, assembling from a low-dose, crowded, in-situ tomogram a structure that would otherwise demand high-dose single-particle imaging.

Why averaging buys resolution

Averaging works because the signal is shared across copies while the noise is independent. Write the value of the $i$ -th aligned subtomogram at some voxel as $x_i = s + n_i$ , where $s$ is the true signal common to every copy and $n_i$ is that copy’s independent, zero-mean noise. Averaging over $N$ particles:

\bar{x} = \frac{1}{N}\sum_{i=1}^{N} x_i = s + \frac{1}{N}\sum_{i=1}^{N} n_i .

Here $\bar{x}$ is the averaged voxel value and $s$ passes through untouched — because it is the same in every copy, averaging a constant leaves it unchanged. The noise term is a sum of $N$ independent random quantities divided by $N$ : if each copy has noise variance $\sigma^2$ , the averaged noise has variance $\sigma^2/N$ and standard deviation $\sigma/\sqrt{N}$ . The signal amplitude is unchanged while the noise standard deviation shrinks by $\sqrt{N}$ , so the signal-to-noise ratio improves by roughly $\sqrt{N}$ (see signal-to-noise ratio).

This $\sqrt{N}$ law sets the cost structure of STA: resolution improves sublinearly in particle count. Going from 100 copies to 400 only doubles the SNR; doubling it again takes 1600. So combining thousands of subtomograms can lift a structure from invisibility to sub-nanometer resolution, but each further notch demands a multiplicative jump in particle number — one reason the yield and purity of particle picking matter so much for the final resolution.

The particles must first be located — see particle picking — and then iteratively aligned: each subtomogram’s orientation and position are refined against the current average, the average is rebuilt, and the cycle repeats until convergence. The first pass usually starts from a coarse, low-resolution reference (even a sphere, or a low-pass-filtered initial average) and tightens the alignment angles and shifts round by round; as the average sharpens, it in turn lets the next round align more accurately. This “better average → better alignment → better average” feedback is what makes STA converge, but it also means a biased starting reference can lock the average onto a wrong structure (reference bias).

Intuition

One blurry copy tells you almost nothing. A thousand blurry copies of the same object, brought into register and stacked, let the random noise cancel while the true structure reinforces itself — the same logic as a long exposure, carried out in three dimensions and over many separate particles.

Averaging more independent copies makes the noise fall as 1/√N while the signal emerges:

Single copy

Average of N

Copies averagedN = 36　· √N = 6.0×

1400

Each subtomogram is the same signal plus independent noise. Averaging N copies leaves the signal intact while the noise standard deviation falls as 1/√N — which is how subtomogram averaging recovers near-atomic structure from copies that are individually very noisy.

How the missing wedge gets filled

The missing wedge is largely overcome in the process. Because particles sit in the cell at random orientations, each one carries its wedge of missing information pointed in a different direction. When subtomograms are rotated into alignment, their wedges scatter across Fourier space, and the union of many partial samplings can fill the gap that no single particle covers.

Concretely: every particle is missing a wedge in Fourier space (the hole left by the limited tilt range, oriented along that particle’s pose in the tomogram). Rotating a subtomogram into the common pose rotates its wedge along with it — so 1000 randomly oriented particles point their wedges in 1000 different directions. In the common frame, a frequency component missed by most particles can still be recovered as long as a few particles happen to be posed so that it falls in their sampled region. The more uniformly the orientations cover the sphere, the more completely Fourier space gets filled. Conversely, if the orientations are preferred (say, membrane proteins all perpendicular to the membrane), certain directions of the wedge are never filled, and the average keeps an anisotropic smear.

Proper STA weights each contribution by its own missing-wedge mask so that the average is not biased by the uneven sampling. Intuitively: at a given frequency component, only the particles that actually sampled it count toward the denominator there; the ones that missed it sit out that part of the average — otherwise the average would be diluted by a pile of zeros.

Deep

This is clearest in Fourier space. Let the Fourier transform of the $i$ -th particle’s subtomogram be $X_i(\mathbf{k})$ , and represent its missing wedge with a binary (or soft) mask $M_i(\mathbf{k})\in[0,1]$ — equal to $1$ where the frequency was sampled, $0$ in the wedge hole. The weighted average is computed independently at each frequency $\mathbf{k}$ :

\bar{X}(\mathbf{k}) = \frac{\sum_{i=1}^{N} M_i(\mathbf{k})\, X_i(\mathbf{k})}{\sum_{i=1}^{N} M_i(\mathbf{k})} .

The numerator sums only over the particles that sampled $\mathbf{k}$ ; the denominator is the effective number of samples $\sum_i M_i(\mathbf{k})$ at that frequency. Here $\bar{X}(\mathbf{k})$ is the averaged spectrum, $X_i(\mathbf{k})$ is the spectrum of the $i$ -th aligned subtomogram, and $M_i(\mathbf{k})$ marks whether it has data at that frequency. If some $\mathbf{k}$ is missed by every particle, the denominator is zero and that point stays empty — which is exactly why the orientations have to be diverse. Plotting the effective sample count $\sum_i M_i(\mathbf{k})$ over the sphere gives the sampling-density map routinely used to diagnose whether coverage is even. In practice one also multiplies in weights for alignment accuracy and per-copy dose/CTF weighting, but the skeleton is this mask-normalized Fourier-space average.

Measuring resolution with FSC

Resolution is assessed by Fourier shell correlation (FSC): the dataset is split in half, two independent averages are computed, and their correlation is measured as a function of spatial frequency. With the Fourier transforms of the two half-set averages denoted $F_1(\mathbf{k})$ and $F_2(\mathbf{k})$ , the correlation is taken over each frequency shell (the spherical shell of constant $|\mathbf{k}|$ ):

\mathrm{FSC}(k) = \frac{\sum_{|\mathbf{k}|=k} F_1(\mathbf{k})\,\overline{F_2(\mathbf{k})}}{\sqrt{\sum_{|\mathbf{k}|=k} |F_1(\mathbf{k})|^2 \;\sum_{|\mathbf{k}|=k} |F_2(\mathbf{k})|^2}} .

Here $F_1,F_2$ are the spectra of the two independent half-set averages, $\overline{F_2}$ is the complex conjugate, and the sums run over the whole shell of radius $k$ . The FSC is a correlation coefficient between 0 and 1: at low frequency the two halves agree and $\mathrm{FSC}\approx 1$ ; in the noise-dominated high frequencies the halves are uncorrelated and the FSC falls toward 0. The frequency at which the FSC drops below a fixed threshold defines the reported resolution. Keeping the two halves independent is the point — if both halves are aligned against a shared reference, the noise gets correlated artificially, inflating the FSC and overstating resolution, an artifact known as overfitting or “noise alignment”.

Established pipelines for STA include RELION and dedicated tomography packages; the same low-SNR conditions that motivate averaging also motivate learned restoration such as CryoWGEN. The two paths are complementary: STA fills holes and beats down noise in Fourier space through sheer redundancy, given that you have already located and aligned thousands of identical copies; learned methods restore the missing wedge and noise directly on a single tomogram, making upstream particle picking, segmentation, and alignment more dependable — a cleaner input means more accurate poses, less reference bias, and ultimately a better FSC resolution. Placing this in the four-method taxonomy: MAP (CryoGEN-I) gives a point estimate, WAE/OT (CryoGEN-II) gives a stable single answer, and EVIA Monte-Carlo (CryoWGEN-I) and EVIA Langevin (CryoWGEN-II) give a family of posterior solutions; what they improve is the quality of the volume that enters STA, not the per-particle averaging itself.

← Structural Analysis