9.3.3 Peak Selection After Template Matching
Mostly due to the low SNR in the data, high CCCs are not exclusively obtained for
the macromolecular complex of interest, but in certain cases also for other structures
(‘false positives’). To distinguish between true and false positive peaks of the
cross-correlation volume, a variety of methods is available [7]. A pragmatic way of
assessing the identity of a cross correlation peak is to manually inspect the corresponding position in the queried tomogram. The phenomenal performance of the
human eye is then used to distinguish false from true positives in the context of the
tomogram (Fig. 9.1a). Another option for roughly distinguishing between true and
false positive peaks is to plot a histogram of correlation coefficients: for data with
good quality, a Gaussian-like distribution of high correlation coefficients originating from ‘true positives’ will be mostly separated from an increasing number of low
correlation coefficients originating from ‘false positives’ (Fig. 9.1b). Furthermore,
repeating cross-correlation based pattern recognition with a structurally different
decoy with comparable statistics as the template, most easily obtained by mirroring
the reference structure [10], can help to discriminate true and false positives. For
false positives, cross-correlation coefficients for the right- and wrong-handed reference are typically similar. In contrast, a markedly higher correlation coefficient
can be expected for the right-handed reference structure in case of a true positive
peak (Fig. 9.1c). Expanding the template matching approach to more than one
single template and combining the scores for the different templates can also
enhance the specificity of the approach [7, 11]. Finally, approaches to classification
of subvolumes extracted from the tomogram at the corresponding positions are
powerful to distinguish true and false positive peaks, in particular, if the subvolumes can be classified according to features that are not included in the reference
structure (see below).
9.3.4 Example Dataset
In a representative tomogram of the example dataset depicting ER membraneassociated ribosomes used for illustration, peaks of the cross-correlation volume
resulting from template matching in PyTom with a cryo-EM single particle
reconstruction of the mammalian ribosome (EMD-5592) primarily correspond to
cytosolic and membrane-bound ribosomes. However, also high-contrast features,
such as gold beads used as fiducial markers for tilt image alignment, edges of the
carbon support foil, or the membrane of ER-derived vesicles all yield high
cross-correlation peaks due to their high SNRs and are thus frequently picked up as
false positives (Fig. 9.2).
240
S. Pfeffer and F. Förster
Mostly due to the low SNR in the data, high CCCs are not exclusively obtained for
the macromolecular complex of interest, but in certain cases also for other structures
(‘false positives’). To distinguish between true and false positive peaks of the
cross-correlation volume, a variety of methods is available [7]. A pragmatic way of
assessing the identity of a cross correlation peak is to manually inspect the corresponding position in the queried tomogram. The phenomenal performance of the
human eye is then used to distinguish false from true positives in the context of the
tomogram (Fig. 9.1a). Another option for roughly distinguishing between true and
false positive peaks is to plot a histogram of correlation coefficients: for data with
good quality, a Gaussian-like distribution of high correlation coefficients originating from ‘true positives’ will be mostly separated from an increasing number of low
correlation coefficients originating from ‘false positives’ (Fig. 9.1b). Furthermore,
repeating cross-correlation based pattern recognition with a structurally different
decoy with comparable statistics as the template, most easily obtained by mirroring
the reference structure [10], can help to discriminate true and false positives. For
false positives, cross-correlation coefficients for the right- and wrong-handed reference are typically similar. In contrast, a markedly higher correlation coefficient
can be expected for the right-handed reference structure in case of a true positive
peak (Fig. 9.1c). Expanding the template matching approach to more than one
single template and combining the scores for the different templates can also
enhance the specificity of the approach [7, 11]. Finally, approaches to classification
of subvolumes extracted from the tomogram at the corresponding positions are
powerful to distinguish true and false positive peaks, in particular, if the subvolumes can be classified according to features that are not included in the reference
structure (see below).
9.3.4 Example Dataset
In a representative tomogram of the example dataset depicting ER membraneassociated ribosomes used for illustration, peaks of the cross-correlation volume
resulting from template matching in PyTom with a cryo-EM single particle
reconstruction of the mammalian ribosome (EMD-5592) primarily correspond to
cytosolic and membrane-bound ribosomes. However, also high-contrast features,
such as gold beads used as fiducial markers for tilt image alignment, edges of the
carbon support foil, or the membrane of ER-derived vesicles all yield high
cross-correlation peaks due to their high SNRs and are thus frequently picked up as
false positives (Fig. 9.2).
240
S. Pfeffer and F. Förster