Particle Classification

Access: CP::3D Model::Classification (Tab)

In ab initio single-particle 3D reconstruction, raw particle images from similar projection orientation are first grouped into classes (a.k.a. clusters) to enhance the SNR of the averages, which serve as the source for particle angular reconstitution or RCT model reconstruction.  PARTICLE is equipped with unbiased classification algorithms that eliminate the need of particle alignment.

Classification Methods & Parameters

PARTICLE offers a variety of methods for 2D image classification:

• ALNC : as in the conventional practice, registered particle images are subjected to pixel-based PCA for clustering in a reduced eigen-space.

• CLAX1 : utilizes an alignment-free "direct method" to classify particle images.  This method requires large memory (8GB minimum) and hybrid CPU-GPU support.

• CLAX2 : employs Shape Analysis to classify particle images by geometric descriptors.  This method does not require particle pre-alignment.  The CLAX2 algorithm has a higher computational complexity than CLAX1 and consumes more memory space.

And the relevant functional parameters include:

• Weighting : CTF-correction can be incorporated into class averaging.
• Rigor : thoroughness in cluster search.  "3" is the most rigorous level (the default "0" is recommended).
• Members : an estimated particle population in each class (only the order of magnitude matters).
• Curvature : the maximum curvature of a shape contour.  When set to "0" (the default, recommended), the program will optimize this parameter.
• Delta : estimated off-center error (in pixels) in the particle images.
• M-fold : manifold level-sets.  An approximate ratio of 1:30 to the particle frame size (in pixels) is recommended.
• Filtering : Gaussian filtering in pre-processing to reduce noise in the particle images.
• Rank : degree of complexity in class refinement.  When set to "0" (the default, recommended), the program will automatically optimize this parameter.
• Masking Radius : a centered mask for each particle frame.  The image inside the mask will be analyzed for classification.
• Quality Threshold : a quality index (between 0.0 and 1.0) to assess the in-class homogeneity.  "1.0" is the most stringent condition.

Activating the Classify button will initiate the process of particle classification.  Upon completion, two output parameters will be displayed:

• Quality Index : the similarity among class members.  The value can be used to determine a proper value for Quality Threshold.
• Class Total : the number of classes in the result.

Class Editor

Particle class averages can be opened in the PTK-Editor via the montage button (an icon next to the "Main Class" name-box).  The class averages in the editor can then be inspected and modified by the following functional buttons in the group:

• Member : when a single class average is selected (CTRL+LeftClick to highlight in DP) in the PTK-Editor, this function will load its member images with the average at the front.

• Extract : collect all selected class averages (CTRL+LeftClick to highlight in DP) and assemble them into a new sub-class.  The name of the sub-class needs to be defined in the "Sub-Class" name-box. This function can also be used to delete classes by reversing the frame selection.

• Verify : compare a selected set of class averages with the respective seed images side-by-side.

• Merge : select one or a few class averages as the seed and merge all classes according to the seed conformation.  The parameter Rank (an integer, normally a fraction of the estimated class member population) sets the stringency of the merging criterion: the higher the rank, the higher the similarity required for classes to merge.

• Expand : expand the membership of the selected classes.  When there is no selection, all the classes will be processed.

• Q-Sort : sort class averages by their quality index.  The name of the sorted stack needs to be defined in the box above the editor.

Data for Reconstruction

Following particle classification, 3D model reconstruction can proceed with the class averages.  Or, in model refinement, the angular alignment of individual class members can be initialized through their associated class averages.  The source images for model reconstruction and refinement may come from (listed in the pull-down menu)

• Class Averages : export a stack of class-averages for ab initio model reconstruction.
• Single Occupancy : export original particle images from selected classes to a stack for model refinement.  Each particle image has a single entry in the stack.
• Multiple Occupancy : export original particle images from selected classes to a stack for model refinement.  A particle image may have multiple entries in the stack, depending on its memberships associated with the various classes.
• R.C.T. Tilt Pairs : export the tilt-pair particles from selected classes to a pair of stacks.  Then the program will switch to the Reconstruction tab and initialize the RCT setup.

The Export button will output the specified particle stack to the "modeling/" folder in the project space for the subsequent 3D model reconstruction and refinement.