Job: Filament Tracer
Particle picking for filaments.
Description
Pick particles from micrographs of filaments, with or without using templates. The filament tracer uses the same procedure as the template/blob pickers, but during post-processing it attempts to join filament contours together via image skeletonization, similar in concept to the helix tracing functionality from SPRING [1]. In contrast to the template picker, this allows for increased density of picks along the helical axis, the identification of individual filaments, and the pruning of particles that are picked from highly bent filaments. For an example of using the filament tracer in a full helical processing workflow, including subsequent processing using the Inspect Particle Picks job, you may want to view the EMPIAR-10031 case study.
Notes and Limitations
The filament tracer assumes that filaments are roughly cylindrical, and have constant contrast along the helical axis. For some filaments (particularly amyloid fibrils with oblong cross-sections), this assumption may be incorrect. In cases like these, the standard Template Picker or any of the Deep Picking methods may be preferable. Note that all helical reconstruction jobs can work with particles obtained from any of CryoSPARC's particle pickers, as well as imported particles.
Input
Aligned and CTF-estimated micrographs
Optionally, one or more templates (generated in CryoSPARC or elsewhere)
Generally, results improve significantly if multiple templates are used, which cover a diversity of views along the helical axis
If templates are not provided, the Minimum/Maximum filament diameter (A) parameters must be set
Output
Particles
Micrographs
Common Parameters
Note that many parameters with units of length are defined relative to the filament diameter, in order for default values to scale well with differently sized filaments.
Filament diameter (A):
Diameter of the filaments in Angstroms.
Separation distance between segments (diameters):
Distance between picked particles along the helical axis, as a multiple of the filament diameter. Ideally, this should be close to a multiple of the helical rise, in angstroms.
Minimum filament length to consider (diameters):
Reject all picks that are associated with filaments that are less than this length.
Lowpass filter to apply (A):
Lowpass filter corner resolution, to apply to the input micrographs.
Produce diagnostic plots:
Whether or not to produce diagnostic plots showing the cross-correlation figure of merit at various stages in the processing. This is helpful if the default parameters do not produce ideal results, and can help inform which parameters are best to tweak.
Standard deviation of gaussian blur (diameters):
Standard deviation of the gaussian blur on the cross-correlation map prior to filament detection, in multiples of the filament diameter. The gaussian blur is used to remove sharp changes in the cross-correlation, which helps to improve ridge detection. It is recommended to keep this between 0 and 0.3. For larger filaments (diameters over ~200 A), values lower than the default of 0.1 may produce better results.
Hysteresis thresholds:
Low and high threshold values, in percentile, used on the ridge-enhanced cross correlation map. These define percentile values for the hysteresis thresholding algorithm, which is used to detect filament contours. In general, the default values of 93 and 98 may work well for a large variety of datasets.
To increase the number of detected filaments, both thresholds can be decreased
To decrease the number of detected filaments, both thresholds can be increased
Note that if the thresholds are too low, many false crossings may be detected, which will result in many false negatives; alternatively, if the thresholds are too high, the detected contours may be too conservative and hence there also may be many false negatives. For this reason, it is recommended to keep these thresholds at the default value
Radius around crossings to ignore (diameters)
Radius around each filament crossing to avoid picking from, in multiples of the filament diameter. A "crossing" is defined as the point where two detected filament contours overlap. Around these points, a circle of the specified radius will be used to remove all picked particles, as picks with overlapping filaments are usually undesirable. Increasing this value from the default of 1 will result in more conservative picks
Distance to trim from end points (diameters):
The distance to trim from the ends of traced contours, in multiples of the filament diameter.
Trimming may be desirable for both:
removing picks from the very end of traced contours (thus ensuring all picks are in the interior of a filament), and
reducing the number of spurious cross-overs detected (i.e. pruning the skeleton of "hairs", which would otherwise be detected as cross-overs)
Note that this done prior to the removal of filament cross-overs, so it may reduce the number of crossings that are detected and removed; this may or may not be desirable. Good values are between 0 and 2 filament diameters. A value of 0 means that no trimming is done.
Template-free tracing
As a substitute to providing templates, both of these parameters can be set to automatically generate filament templates for picking (in a similar way as to the blob picker).
Minimum filament diameter (A): Minimum filament diameter for reference-free picking
Maximum filament diameter (A): Maximum filament diameter for reference-free picking
Note: If you choose to use template-free tracing by activating the minimum and maximum diameters, better results are often obtained with decreased values for the hysteresis threshold parameters. Thus, when template-free tracing is used, the defaults are lowered by 3 to (90,95).
Interpreting Diagnostic Plots
Diagnostic plots can provide insight as to which inputs and/or parameters should be tweaked to produce ideal results. Below is a description of each of the diagnostic plots, examples of ideal results for each plot, and tips on choosing better parameter values to improve the results at each stage of the processing. For this section, we will use the EMPIAR-10031 (Mitochondrial AntiViral-Signaling protein filaments, or MAVS) as an example dataset. An example micrograph is shown below, as well as an example template generated from manual picking and 2D classification.
Shown above are two example templates. On the left is a low quality template generated using manual picking of 70 particles. On the right is a high quality template generated after one round of filament tracing and 2D classification.
There are four diagnostic plots produced, each of them detailed below.
1. Raw cross correlation (top left)
This is the raw figure-of-merit between the templates and the input micrograph, shown in the top left of the diagnostic plot image. Ideal results have clear filament contours visible as bright streaks across the plot.
The quality of the cross correlation depends solely on the template, input micrographs, and the lowpass filter applied. If streaks appear low-contrast, or are fragmented (have significantly varying intensity along the helical axis), the following may improve results:
Increasing the number of templates used – ideally, templates should cover the full set of views that exist along the helical axis
Increasing the quality of templates used – in most workflows, the first set of templates are generated by using the Manual Picker followed by 2D classification, and these are used to do one round of filament picking, extraction, and 2D classification. Similar to the template picker, improved results may be obtained by using templates generated after one round of filament picking for the next round of filament picking
2. Ridge-enhanced cross correlation (top right)
This is the cross correlation after the gaussian blur, and the ridge filter are applied in order to increase the prominence of the filament contours. This stage of the processing depends only on the Standard deviation of gaussian blur parameter. Ideally, ridges appear as bright connected streaks, and all non-filament regions appear dark. If streaks appear faint or excessively wide, the gaussian blur should be decreased; if streaks appear too narrow or have many "hairs", the gaussian blur should be increased to decrease sensitivity to noise.
3. Thresholded cross correlation (bottom left)
This is a binarized map, obtained after applying the hysteresis threshold to the previous map. This stage of the processing depends only on the Hysteresis thresholds. Ideally, the contours should be thin and surround each ridge, with every "T" or "X" intersection representing a true crossover point.
4. Pruned skeleton (bottom right)
This is the "skeletonized" binary map, obtained by applying a morphological thinning procedure to the previous map. It represents the final locations of the traced contours, after trimming from the skeleton end points and removing cross-overs. It also has filaments of length less than the Minimum filament length to consider removed.
Lastly, we show an example of picks detected on the same micrograph, using two templates of differing quality. On the left are picks obtained using a low quality template generated using manual picking of 70 particles; on the right are picks obtained using a high quality template generated after one round of filament tracing and 2D classification. In many cases, more filaments can be picked using a larger number of higher quality templates, that cover a diversity of views along the helical axis.
Common Next Steps
Inspect Particle Picks
After filament tracing, it is strongly recommended to run an Inspect Picks job to remove false positives. In addition to the NCC and Power thresholds, picks can also be pruned based on their estimated local curvature, and entire filaments can be pruned based on their estimated sinuosity
Citations
[1] Stefan T. Huber, Tanja Kuhm, Carsten Sachse. Automated tracing of helical assemblies from electron cryo-micrographs. Journal of Structural Biology, Volume 202, Issue 1, 2018, Pages 1-12, ISSN 1047-8477. [ https://doi.org/10.1016/j.jsb.2017.11.013 ]
Last updated
