electron dense material such as gold labels. Another possibility is the use of correlative imaging where a fluorescence signal from light microscopy is overlaid with
the electron tomogram to pin-point features or macromolecular assemblies of
interest [52]. Localization accuracy is currently in the range of 50–100 nm [53, 54]
and correlative imaging techniques using super-resolution light microscopy with
potentially higher localization accuracy are also under development [55]. However,
both techniques will only give access to a small subset of assemblies and can
introduce perturbations into the cellular system.
A more general and possibly more attractive way of detection is through computational methods where some model of the assembly is used as a template to find
copies of that assembly in the tomographic reconstructions. The first feasibility test
for this sort of approach was done using correlation-based template matching with
tomograms of reconstituted systems containing purified thermosomes, 20S proteasomes, and GroEL respectively [56]. While the results were encouraging with
very high detection fidelity, the conditions of the specimen were idealized and not
very close to the situation in cells, which are highly crowded with many different
constituents and interacting partners. Slightly more realistic follow-up tests with
liposomes filled with 20S proteasomes, thermosomes, or both [44] showed somewhat less convincing results, but were still encouraging.
This type of correlation-based template matching uses a ‘matched filter’ in terms
of detection theory [57]. A matched filter can be shown to minimize the probability of
identification errors, as long as the template and the target are nearly identical and the
noise is independent and identically distributed, additive, and Gaussian [58, 59].
These conditions are not very well met for reconstructions done by electron
tomography. The noise in these reconstructions is spatially correlated and the tails of
the noise distribution are often quite heavy, especially in stained samples [60]. These
issues generate a noise distribution that is distinctly non-Gaussian. In addition, the
potential mix of conformations, potential inaccuracies in the magnification estimate,
and/or the presence of stain make it difficult to obtain accurate templates. As a
consequence, correlation-based template matching tends to generate false hits in areas
of high density such as membranes or vesicles when used with cellular tomograms
[61–63]. The method generally performs better when the signal-to-noise ratio is high
[64] and the use of the new Volta phase plate technology [15, 65] has led to very
encouraging results when correlation-based template matching was used in full
cellular environments for assemblies such as proteasomes [66] or ribosomes [67].
Several recent studies address the shortcomings of the matched filter approach
by introducing new ways of scoring or defining templates. One set of studies uses a
measure in real space, somewhat analogous to the R-free value in X-ray crystallography [68, 69]. In this approach a mask is used to outline the area of the template
without modification of the template itself. The mask is then split into a working
and a testing area. The working area is used for the calculation of the correlation
during the template-matching search. The information from both areas is used to
calculate a score (M-free) that measures how strongly the template influences the
correlation signal in each position. The M-free score gives a significantly better
distinction between true and false hits in the test cases presented [69]. While the
12 Segmentation of Features in Electron Tomographic Reconstructions
307
the electron tomogram to pin-point features or macromolecular assemblies of
interest [52]. Localization accuracy is currently in the range of 50–100 nm [53, 54]
and correlative imaging techniques using super-resolution light microscopy with
potentially higher localization accuracy are also under development [55]. However,
both techniques will only give access to a small subset of assemblies and can
introduce perturbations into the cellular system.
A more general and possibly more attractive way of detection is through computational methods where some model of the assembly is used as a template to find
copies of that assembly in the tomographic reconstructions. The first feasibility test
for this sort of approach was done using correlation-based template matching with
tomograms of reconstituted systems containing purified thermosomes, 20S proteasomes, and GroEL respectively [56]. While the results were encouraging with
very high detection fidelity, the conditions of the specimen were idealized and not
very close to the situation in cells, which are highly crowded with many different
constituents and interacting partners. Slightly more realistic follow-up tests with
liposomes filled with 20S proteasomes, thermosomes, or both [44] showed somewhat less convincing results, but were still encouraging.
This type of correlation-based template matching uses a ‘matched filter’ in terms
of detection theory [57]. A matched filter can be shown to minimize the probability of
identification errors, as long as the template and the target are nearly identical and the
noise is independent and identically distributed, additive, and Gaussian [58, 59].
These conditions are not very well met for reconstructions done by electron
tomography. The noise in these reconstructions is spatially correlated and the tails of
the noise distribution are often quite heavy, especially in stained samples [60]. These
issues generate a noise distribution that is distinctly non-Gaussian. In addition, the
potential mix of conformations, potential inaccuracies in the magnification estimate,
and/or the presence of stain make it difficult to obtain accurate templates. As a
consequence, correlation-based template matching tends to generate false hits in areas
of high density such as membranes or vesicles when used with cellular tomograms
[61–63]. The method generally performs better when the signal-to-noise ratio is high
[64] and the use of the new Volta phase plate technology [15, 65] has led to very
encouraging results when correlation-based template matching was used in full
cellular environments for assemblies such as proteasomes [66] or ribosomes [67].
Several recent studies address the shortcomings of the matched filter approach
by introducing new ways of scoring or defining templates. One set of studies uses a
measure in real space, somewhat analogous to the R-free value in X-ray crystallography [68, 69]. In this approach a mask is used to outline the area of the template
without modification of the template itself. The mask is then split into a working
and a testing area. The working area is used for the calculation of the correlation
during the template-matching search. The information from both areas is used to
calculate a score (M-free) that measures how strongly the template influences the
correlation signal in each position. The M-free score gives a significantly better
distinction between true and false hits in the test cases presented [69]. While the
12 Segmentation of Features in Electron Tomographic Reconstructions
307