Tracing low-level structures in cryo-electron tomography

Abstract

Cryo-electron tomography is an imaging technique that provides 3D images (tomograms) in situ of cells with sub-nanometer resolution. Typically, the first step in the analysis is to classify the tomogram voxels into different structures, named semantic segmentation. However, the segmentation results are sets of voxels, hindering further quantitative analysis. In this paper, we define and implement algorithms to convert the semantic segmentation of the main structures in a cellular cryo-electron tomogram (membranes, filaments, cytosolic and membrane-bound macromolecules) into specific skeletons, preserving their topological and geometrical information. Additionally, we have defined a metric for comparing segmentations in cryo-ET coming from different methods more robust than the standard DICE. We also demonstrate how this approach can be used to trace cellular features by analyzing several in situ cellular cryo-electron tomograms.

Fig 1. The workflow.

(A) A scheme of the workflow; solid arrows depict the common steps and the path for filaments, dotted arrows the path for membranes, and dashed ones for macromolecules. (B) A 2D slice of the tomogram provided by [16] authors. Scale bar 200 nm. View of the red box with segmentation for membranes (C) and macromolecules (D). View of the blue box with filaments segmentation (E). Segmentations in (C-E) are not homogeneous, e.g. membranes do not have a constant thickness. (F-H) The results of applying NMS on (C-E) respectively. (I) Instance segmentation from the membranes traced in (F). (J) Positions of the macromolecules obtained from (G). (K) Curves extracted from the graph in (H).

Fig 2. The saliency and NMS.

(A) A 2D slice of a tomogram region with MTs approximately parallel to the X-Y plane, marked by the blue arrows. Scale bar 100 nm. (B) Microtubule segmentation in 3D. (C) Saliency map obtained from the segmentation. This is a 2D slice of the red box marked in B. (D) NMS. (E) NMS with the 3D curves extracted, the color encodes the curve length. (F) Filaments extracted with Amira [3], blue arrows point to fractures artificially created by Amira but not present in our procedure E, each color represents an instance. Colormaps in panels (E) and (F) are different because Paraview and Amira have different ones. Also, Paraview shows distance in Angstroms and Amira in nm.

Fig 3. Dimension of the features.

The analysis of the eigenvectors and eigenvalues determines the dimension of the features. Arrows represent the eigenvectors, and their length represents the eigenvalues.

Fig 4. SEG processing.

(A) Steps to construct the SEG from the NMS point cloud. Balls represent the nodes and lines the edges. (B) Steps to extract the filaments from the SEG. The duplicated node has the same coordinates as the original one (red box); we have shifted it for clarity.

Fig 5. Segmentation process.

(A) A 2D slice of a cellular tomogram. (B) The segmented tomogram, with all the elements together but having different labels. The elements separated: membranes (C), filaments (D), and macromolecules (E). The color code for all panels is: Yellow for membranes, blue for filaments, and red for all types of macromolecules. Scale bars 200 nm.

Fig 6. Membrane segmentation.

(A) Membrane analysis with TracET workflow. Membranes have been classified by the application of a connectivity analysis, and the color code corresponds with the label of every membrane. (B) Membrane segmentation with Membrain. (C) 2D Skeleton from the Membrain segmentation obtained by Skeletonize3D. Scale bars are 200 nm. All panels have zooms to the boxed regions below. Scale bars for zooms are 20 nm. The original tomogram slice can be seen in Fig 5A.

Fig 7. Tracing filaments.

(A) and (D) 2D slices of in situ tomograms with a dense network of filaments. (B) and (E) the SEG obtained from tomograms A and D respectively. (C) and (F) curves extracted from B and E respectively, filtered by length with a threshold of 85 nm. In C the color encodes the length of the curves, a global property, and in F the local curvature, a local property.

Fig 8. Detection of the macromolecular centers.

(A) 2D slice of an in situ tomogram with the segmented ribosomes (red), cytosolic particles (green), and membrane-bound macromolecules (orange) segmented. (B) A zoom of the tomogram with the background context of the particles selected, a cluster of ribosomes is visible. (C) The skeleton with the local minima of the macromolecules in B. (D) The spheres represent the particles in C finally detected by using MeanShift. The radius of the spheres. Scale bar 200 nm.

Fig 9. Robustness of the revised DICE_d.

(A-C) shows an input segmentation in blue, a segmentation used as ground truth in red, and the intersection in green, for dimension d = 2 (membranes), A, d = 1 (filaments), B, and d = 0 (blobs), C. Panel A has zooms to see that displacements and differences in the thickness of the segmentation contribute to reduce the value of the standard DICE. (D-F) compare the standard DICE and revised DICE_d for dimensions d = 2, d = 1, d = 0, increasing the thickness of both input and reference segmentations. (G-I) compare the same metrics but only by increasing the thickness of the input segmentation.