Subtomogram averaging

Research Area: A.1-Electron Microscopy
Status: In progress  
Project leaders:
  • Fernando Amat
  • Mark Horowitz
Collaborators:
  • Puey Ounjai
  • Daphne Koller
  • Kenneth H. Downing
  • Luis R. Comolli
  • Farshid Moussavi
  • John Smit
Proposed start date: 2008-12-03 Proposed end date: 2011-06-02
Description:

Subtomogram averaging is a registration algorithm in order to obtain higher resolution structures by averaging thousands of aligned subvolumes containing the same structural unit (Fig. 1). For example, we can obtain multiple tomograms of bacteria and select subvolumes containing flagellar motors from each tomogram. If the hypthesis that flagellar motor structure is the same in all the samples, we can register all those subvolumes and average them in order to obtain a higher resolution structure than from a single tomogram. The basic idea is very similar to single particle approaches: by combining two similar images that have been blurred in different directions by the missing wedge, we can obtain a better reconstruction of the original image. However it has the advantage that we can apply to all sorts of tomographic samples and not only to purified structures. As of 2011, resolutions close to 20 ̊A have been reported using subtomogram averaging, and continuing progress in the field makes higher resolutions expected. The end goal is to achieve near-atomic resolution of biological complexes close to their native state. A review by Bartesaghi and Subramaniam contains an excellent review on this topic.

 

Figure 1: schematic to understand the main principle behind subtomogram averaging. Each subtomogram represents the same structure affected by a different missing wedge and noise. If we can register all the same structures and average them, we can recover a much cleaner version of the underlying signal.

We borrow concepts from sparse signal representation and compress sensing in order to improve robustness during the registration process in low SNR environments such cryo-electron tomography. In particular, instead of registering subvolumes using the standard metric in the field known as Constrained Cross-Correlation (CCC) we define a new metric named Thresholded Constrained Cross-Correlation. This metric borrows takes advantage of two facts: first, most biological structures can be represented sparsely in Fourier domain. Moreover, the number of voxels N in subtomograms is much larger than the number of pixels in two-dimensional images, which is known to improve sparse representations. Second, in subtomogram averaging, instead of a single image we have multiple copies of the same object, so we can obtain more reliable noise statistics that facilitates the threshold selection. We have applied this approach to find new structures and we have released open-source code with its implementation.