What shapes template-matching performance in cryogenic electron tomography in situ?

Abstract

The detection of specific biological macromolecules in cryogenic electron tomography data is frequently approached by applying cross-correlation-based 3D template matching. To reduce computational cost and noise, high binning is used to aggregate voxels before template matching. This remains a prevalent practice in both practical applications and methods development. Here, the relation between template size, shape and angular sampling is systematically evaluated to identify ribosomes in a ground-truth annotated data set. It is shown that at the commonly used binning, a detailed subtomogram average, a sphere and a heart emoji result in near-identical performance. These findings indicate that with current template-matching practices macromolecules can only be detected with high precision if their shape and size are sufficiently different from the background. Using theoretical considerations, the experimental results are rationalized and it is discussed why primarily low-frequency information remains at high binning and that template matching fails to be accurate because similarly shaped and sized macromolecules have similar low-frequency spectra. These challenges are discussed and potential enhancements for future template-matching methodologies are proposed.

Keywords: computer vision; cryo-electron tomography; particle picking; template matching.

Figure 1

Template classes used to match the ribosome in the previously annotated tomogram from de Teresa-Trueba et al. (2023 ▸). Different shapes with different radii sampled at a 13.8 Å voxel size, to match the voxel size of the tomogram, were used as templates for template matching with PyTom. Specifically, the map of the 80S ribosome (a) (EMDB entry EMD-3228), a sphere (b) and a rotationally symmetrized heart emoji (c) were used. The used radii range from 1 to 19 voxels in one-voxel increments. Only three representative radii are shown.

Figure 2

Template-matching performance using three distinct template classes scaled to different radii (see Fig. 1 ▸). Radius scaling was performed by resampling each template to 10 × (radius)⁻¹ times the sampling rate of the tomogram, starting from an initial template with an assigned radius of 10 and the same sampling rate as the tomogram (see Section 2). (a) Ribosome-picking recall by the number of picked particles. We used linear sum assignment to achieve an optimal one-to-one mapping between ground-truth and picked particles. A particle is considered to be correctly picked if it is within a five-voxel distance of its assigned ground-truth particle. Consequently, all particles without assignment to ground-truth particles were considered to be false positives. All 4000 picked particles were considered in the following figures. (b) As in (a) but showing precision instead of recall.

Figure 3

Proportion of true positives out of all picked particles (precision) by radius of templates. Template-matching results obtained with PyTom are shown in the left column and those obtained with pyTME are on the right. In each row, the sampled number of angles is shown: namely, 980, 1944 and 15 192. 4000 picks were considered when determining the precision.

Figure 4

Template-matching performance on the FAS complex using three distinct templates. Picked particles were one-to-one assigned to the union of ground-truth FAS and ribosome coordinates using linear sum assignment. Each particle is assigned to no more than one class and is considered to correctly pick that class if it is within a five-voxel distance of its assigned ground-truth particle.

Figure 5

Fourier magnitude spectrum averages by ‘distance’ and template. ‘Distance’ was computed as the Euclidean distance from the zero-frequency component and was rounded to the nearest integer. Templates used for template matching at a radius of 10 voxels are shown (Fig. 1 ▸). ‘Sphere (theory)’ refers to the theoretical derivations made in Section 3.4 with R = 10 (Friedman, 1997 ▸). Magnitude spectrum averages were linearly scaled to the interval [0, 1] to facilitate curve-shape comparison.