Toward detecting and identifying macromolecules in a cellular context: template matching applied to electron tomograms

Abstract

Electron tomography is the only technique available that allows us to visualize the three-dimensional structure of unfixed and unstained cells currently with a resolution of 6-8 nm, but with the prospect to reach 2-4 nm. This raises the possibility of detecting and identifying specific macromolecular complexes within their cellular context by virtue of their structural signature. Templates derived from the high-resolution structure of the molecule under scrutiny are used to search the reconstructed volume. Here we outline and test a computationally feasible two-step procedure: In a first step, mean-curvature motion is used for segmentation, yielding subvolumes that contain with a high probability macromolecules in the expected size range. Subsequently, the particles contained in the subvolumes are identified by cross-correlation, using a set of three-dimensional templates. With simulated and real tomographic data we demonstrate that such an approach is feasible and we explore the detection limits. Even structurally similar particles, such as the thermosome, GroEL, and the 20S proteasome can be identified with high fidelity. This opens up exciting prospects for mapping the territorial distribution of macromolecules and for analyzing molecular interactions in situ.

Figure 1

Two x-y slices of a tomographic reconstruction of a whole ice-embedded Pyrodictium abyssi cell. The plasma membrane and intracellular vesicles are clearly recognizable. The vesicles are surrounded by dark protein masses, probably macromolecular assemblies involved in exo- or endocytosis. The resolution of the tomographic data set, obtained with a CM 120 Biofilter, is about 8 nm.

Figure 2

The three macromolecular assemblies used as test particles: (A) the thermosome (℘ ≈16 nm) with an 8-fold rotational symmetry, (B) GroEL (℘ ≈15 nm) with a 7-fold rotational symmetry, and (C) the 20S proteasome (℘ ≈12 nm) with a 7-fold rotational symmetry. The particles were filtered to 2-, 4-, and 8-nm resolution (from left to right).

Figure 3

Artificial electron-tomographic volumes of (A) thermosome, (B) 20S proteasome, and (C) GroEL macromolecules. The particles were randomly distributed and oriented within the volumes. To simulate electron-tomographic data acquisition, the volumes were projected perpendicular to a virtual tilt axis from −60° to + 60° with 5° increments. The projections were shifted in the x-y plane, convoluted with a realistic contrast transfer function, backprojected and obscured by “colored noise” with a SNR of 0.5.

Figure 4

(A) Slice from the tomographic reconstruction of ice-embedded thermosome particles. The particles are preferentially arranged in a top-view orientation at the water-air interface before freezing. Two high-contrast gold particles, used for aligning the projections, are visible in the right half of the left image. (B) A projection of the MCM-processed version of A after 10 iterations. The peaks detected by the peak search algorithm described are shown as white crosses. The particles close to the border of the volume were excluded from detection.

Figure 5

Performance of MCM on the detection of macromolecules in simulated electron-tomographic volumes. The percentage of particles located correctly, of existing particles not detected (false negatives) and nonexisting particles detected (false positives) was measured at different SNRs.

Figure 6

Distribution of the CCC peak-heights for the reconstructed 20S proteasomes (A) and thermosome (B) particles. The reconstructed volumes were independently correlated with a 20S proteasome, a GroEL, and a thermosome template. The correlation peaks are distinctively higher if particle volume and template correspond to each other and thus the correct particle can be discriminated. Upon visual inspection, the particles corresponding to the left tails of the distribution appear to have structural defects. All particles have been identified correctly.

Figure 7

Identification results for simulated volumes at different resolutions. (A) The percentage of particles detected correctly is shown. The detection criterion was the following. The correlation peak of particle volume i and template i was calculated. The result was divided by the correlation peak of particle volume i and template j. If the result was > 1, the identification was assumed to be correct. (B) The average of this ratio over all particles is plotted. Because of the difference in diameter, the 20S proteasome can be easily discriminated from the two other particles. To distinguish the thermosome and GroEL, a good resolution is obligatory as the low-resolution information of the two particles is basically identical (same size and shape), whereas the high-resolution data differ because of the distinct symmetry (8-fold vs. 7-fold). The opposite is true for the discrimination of GroEL and the 20S proteasome: the two particles are identical in symmetry, but differ in size, therefore a resolution of 8 nm is sufficient for a successful identification.