Particle picking & template matching

Locating copies of a target molecule in a noisy tomogram is the step that precedes subtomogram averaging.

Particle picking is the task of finding the positions — and often the rough orientations — of a molecule of interest within a tomogram. It precedes subtomogram averaging, which needs the coordinates of every copy before it can extract and average them. Picking is hard precisely because the data is so noisy: the targets sit at very low signal-to-noise ratio, are distorted by the missing wedge, and may be densely packed among other cellular material. A single tomogram often contains hundreds to thousands of copies awaiting localization, so picking must be both accurate and scalable across the whole volume.

Template matching: sliding a stencil through the volume

The classical approach is template matching. A 3D template — a known or low-resolution model of the target — is correlated against the tomogram at every position and orientation. The cross-correlation between template $t$ and volume $v$,

peaks where the local density resembles the template. Here $\mathbf{x}$ is the position being tested in the tomogram, $\mathbf{u}$ runs over every voxel the template covers, $t(\mathbf{u})$ is the template density at offset $\mathbf{u}$, and $v(\mathbf{x}+\mathbf{u})$ is the tomogram density at the corresponding spot. The sum is a voxel-by-voxel multiply-and-add: when the template’s shape matches the local density everywhere — same sign, same height — the products are all positive and the sum is large; when the shapes disagree, positives and negatives cancel and the sum falls toward zero. Searching over a grid of orientations — typically thousands of angle combinations — and taking local maxima of the correlation map yields candidate positions and angles.

Raw correlation values cannot simply be compared, because dense regions of the tomogram naturally produce a larger sum and drown out weaker signals whose shape actually matches. So in practice one uses normalized cross-correlation: at each position the local patch and the template are each mean-subtracted and divided by their standard deviation, which puts the score between $-1$ and $1$, with $1$ a perfect linear match independent of absolute density. There is a second subtlety. The tomogram is missing a missing wedge of Fourier information, while the template is complete — compared directly, template density in directions the data never sampled would spuriously lower the correlation. The standard fix is to mask the template with the same missing wedge as the data, so both are compared on the same incomplete Fourier support and the comparison is fair.

Once the candidate peaks are found, the list almost always contains false positives and must be pruned. Common criteria include peak height (too low usually means a chance coincidence of noise), peak shape — sharp and isolated versus diffuse (diffuse peaks often correspond to membranes or ice contamination rather than a compact particle), a minimum spacing between neighboring peaks (to stop one particle being picked twice, usually set by the particle diameter), and whether the location is plausible (a ribosome should not land in the middle of a lipid bilayer). Template matching is noise-sensitive, but it remains a workhorse: it needs no human annotations, runs from even a crude initial model, and supplies the initial orientations that seed averaging — something a pure detector cannot give.

Neural picking: from correlation to learning

Neural picking has become a strong complement. A convolutional network (such as a 3D U-Net) or a transformer-based network is trained — from a modest set of annotations, or via self-supervised pretraining — to flag voxels belonging to the target. Its fundamental difference from template matching is this: template matching only asks “how much does the local density resemble this one stencil,” whereas a network has seen the target across many orientations and crowded backgrounds during training, so what it learns is a family of features more tolerant of noise and deformation than any single rigid stencil.

The cost is annotation: a supervised network needs experts to mark true particles by hand in several tomograms first, and labeling cellular data is both slow and subjective. This is why self- and semi-supervised pretraining matter — learn a general representation on a large pool of unlabeled tomograms, then fine-tune with a handful of labels. Whichever route, one thing holds for both methods: the cleaner the underlying tomogram, the more reliable picking becomes. Template-matching peaks rise more clearly from the noise floor, and networks separate target from background more easily. This is one reason learned restoration methods such as CryoGEN are relevant upstream of this step — suppress missing-wedge artifacts and noise first, and downstream picking, alignment, and averaging all benefit.

Picking also works hand in hand with segmentation: segmenting large-scale structures like membranes and organelles first lets you restrict picking to plausible regions (search for ribosomes only in the cytosol, say), which both cuts false positives and saves compute. In in-situ imaging this “frame the scene, then pick the molecules” pairing is especially useful. Finally, the picked coordinates and angles flow directly into subtomogram extraction and on to subtomogram averaging — so the accuracy of picking is written directly into the resolution of the final structure.

← Structural Analysis