Extensive Angular Sampling Enables the Sensitive Localization of Macromolecules in Electron Tomograms

Abstract

Cryo-electron tomography provides 3D images of macromolecules in their cellular context. To detect macromolecules in tomograms, template matching (TM) is often used, which uses 3D models that are often reliable for substantial parts of the macromolecules. However, the extent of rotational searches in particle detection has not been investigated due to computational limitations. Here, we provide a GPU implementation of TM as part of the PyTOM software package, which drastically speeds up the orientational search and allows for sampling beyond the Crowther criterion within a feasible timeframe. We quantify the improvements in sensitivity and false-discovery rate for the examples of ribosome identification and detection. Sampling at the Crowther criterion, which was effectively impossible with CPU implementations due to the extensive computation times, allows for automated extraction with high sensitivity. Consequently, we also show that an extensive angular sample renders 3D TM sensitive to the local alignment of tilt series and damage induced by focused ion beam milling. With this new release of PyTOM, we focused on integration with other software packages that support more refined subtomogram-averaging workflows. The automated classification of ribosomes by TM with appropriate angular sampling on locally corrected tomograms has a sufficiently low false-discovery rate, allowing for it to be directly used for high-resolution averaging and adequate sensitivity to reveal polysome organization.

Keywords: GPU acceleration; electron cryo-tomography; particle localization and identification; template matching; volume registration.

Figure 1

PyTOM’s particle detection can be integrated into a workflow for tilt-series processing. Top left: a tomogram, represented as a slice of a tomogram from the ER microsome dataset (denoised for visualization, displayed between ±2σ) is the input for TM in PyTOM. Top center: The particle candidates are indicated by blue circles in a slice of the tomogram. Top right: A model of two Gaussians is fitted to the LCC_max histogram to estimate the true positives in the dataset and to determine a cutoff to classify the candidates. Bottom center: These automatically annotated particles are shown as surface rendering in 3D. Bottom right: The extracted particles are the input for subtomogram averaging in M 1.0.9 and RELION 3.1.4. Bottom left: A zoom-in highlights the analysis of the spatial distribution of neighboring particles in PyTOM.

Figure 2

Increased rotational sampling improves particle detection. (A) Illustration of the relation between particle diameter and angular increment in Equation (1). The orange outline shows the original orientation of the template, while the dashed line shows the template after rotation. A point on the template is shown, which transforms by one pixel for the rotation $\mathsf{\Delta }\alpha$. (B) LCC_max values of the 300 highest scoring particles for rotation sampling at 13° (blue line), 7° (orange line) and 3° (green line). (C) Evaluation of performance of LCC_max as a classifier for different rotation samplings. Histogram of the LCC_max values on the left, with the Gaussian fitted to the true positives in the candidates (red line) and the bimodal model of the background and positives (blue line); the dashed line illustrates the LCC_max cutoff estimated from the RUC. On the right, the corresponding ROC curves are shown (red dots) with the RUC point (green dot) that provides a good balance between sensitivity and FDR.

Figure 3

Local alignment improves particle detection in tilt-series with large deformation. (A) ${\mathrm{LCC}}_{\max }$ values from TM were ordered from high to low for globally (blue) and locally (orange) aligned tomograms. Left tomogram has residual error mean and standard deviations reported from the IMOD fiducial model as 0.780 and 0.547, middle tomogram has 1.210 and 0.638, and right tomogram has 2.206 and 1.459, indicating the decreasing fit of a rigid-body model from left to right. (B) Slice of the tomograms along the y-axis (or xz-plane) shown between ±3σ.

Figure 4

TM can indicate sample damage in plasma FIB milling. (A) Representative slice of tomograms from EMPIAR-11306 [29], denoised for visualization purposes and displayed between ±2σ (Position 106 and Position 142 of dataset). (B) A scatter plot of the coordinates projected along the y-axis and colored by their LCC_max values (blue = low, yellow = high). (C) Scores plotted by z-height in the ice layer; coordinates were first rotated to the plane with the least variation, then the median z-height was subtracted from all coordinates to set the center of the ice layer to z = 0. Red dashed lines were added as a visual aid to indicate the drop-off in LCC_max values when moving from the center to the edges of the ice layer.

Figure 5

Detected ribosomes can be used for high-resolution averaging. (A) Refined ribosome from the 12,343 particles extracted from PyTOM. (B) FSC between half maps of the refined structure within a masked area for the set of particles automatically classified in PyTOM (orange) and after additional 3D classification of the subtomograms to remove false positives in the particle set (green). (C) Local resolution of the structure in (A); the values on the color bar are in Å units.

Figure 6

Ribosome neighbor density. (A) The subtomogram average of the particles filtered to (14 Å)⁻¹ resolution shows a ribosome associated with the ER membrane, with two neighboring densities on the left and right, which are visible at lower density thresholds. (B) Neighbor density for the particles in the dataset calculated for 4 neighbors around each ribosome. The probability is projected onto the xy-plane (left) and yz-plane (right) corresponding with the view of the averages in (A), where the grayscales indicate the probability, with black corresponding to the highest probability of finding neighbors.