subtomogram average of the mammalian ribosome bound to the protein-conducting
channel in the native ER-membrane is obtained from these tomograms.
9.2 Increasing the Resolution of Volumetric
Data by Averaging
The interpretation of tomographic data from frozen hydrated samples is a formidable challenge due to their low signal-to-noise ratio (SNR). The low SNR is a direct
consequence of the high dose sensitivity of vitrified specimens, which must be
hence imaged with very few electrons [1]. Typically, the cumulative electron dose
used for acquisition of a tomogram should not exceed 100 e
− /Å
2 to avoid specimen
alteration and excessive structural damage. This electron dose ultimately limits the
resolution of the raw tomogram. The SNR strongly decreases as a function of
spatial frequency until it is too low to distinguish signal from the background.
Depending on the sample and equipment used for imaging this frequency is
somewhere in the range of 3–10 nm [2, 3]. However, by iterative alignment and
averaging of subtomograms, each depicting the same macromolecular complex, the
inherently low SNR can be increased and the higher resolution signal hidden in the
raw tomogram can be retrieved. Assuming a perfect spatial alignment of subtomograms, the SNR scales linearly with the square of the number of subtomograms
used for averaging. With an increasing number of structure factors rising above the
significance (noise) level, the resolution of the subtomogram average improves.
9.3 Particle Localization
Although the SNR is typically low in CET, it is possible to locate macromolecular
structures of 500 kDa or more with acceptable specificity. This section introduces
different concepts for particle localization, with an emphasis on cross-correlation
based pattern recognition algorithms.
9.3.1 Template-Free Approaches
Template-free approaches for detection of macromolecules in tomograms rely on
generic features of complexes. For example, the strong variation of the intensity at
the boundaries can be used for detection in mean curvature motion (MCM) based
particle localization [4]. More elaborate approaches first filter the tomogram
according to objects of target sizes by convolution with a Gaussian and
238
S. Pfeffer and F. Förster
channel in the native ER-membrane is obtained from these tomograms.
9.2 Increasing the Resolution of Volumetric
Data by Averaging
The interpretation of tomographic data from frozen hydrated samples is a formidable challenge due to their low signal-to-noise ratio (SNR). The low SNR is a direct
consequence of the high dose sensitivity of vitrified specimens, which must be
hence imaged with very few electrons [1]. Typically, the cumulative electron dose
used for acquisition of a tomogram should not exceed 100 e
− /Å
2 to avoid specimen
alteration and excessive structural damage. This electron dose ultimately limits the
resolution of the raw tomogram. The SNR strongly decreases as a function of
spatial frequency until it is too low to distinguish signal from the background.
Depending on the sample and equipment used for imaging this frequency is
somewhere in the range of 3–10 nm [2, 3]. However, by iterative alignment and
averaging of subtomograms, each depicting the same macromolecular complex, the
inherently low SNR can be increased and the higher resolution signal hidden in the
raw tomogram can be retrieved. Assuming a perfect spatial alignment of subtomograms, the SNR scales linearly with the square of the number of subtomograms
used for averaging. With an increasing number of structure factors rising above the
significance (noise) level, the resolution of the subtomogram average improves.
9.3 Particle Localization
Although the SNR is typically low in CET, it is possible to locate macromolecular
structures of 500 kDa or more with acceptable specificity. This section introduces
different concepts for particle localization, with an emphasis on cross-correlation
based pattern recognition algorithms.
9.3.1 Template-Free Approaches
Template-free approaches for detection of macromolecules in tomograms rely on
generic features of complexes. For example, the strong variation of the intensity at
the boundaries can be used for detection in mean curvature motion (MCM) based
particle localization [4]. More elaborate approaches first filter the tomogram
according to objects of target sizes by convolution with a Gaussian and
238
S. Pfeffer and F. Förster