Semi-automated selection of cryo-EM particles in RELION-1.3

Abstract

The selection of particles suitable for high-resolution cryo-EM structure determination from noisy micrographs may represent a tedious and time-consuming step. Here, a semi-automated particle selection procedure is presented that has been implemented within the open-source software RELION. At the heart of the procedure lies a fully CTF-corrected template-based picking algorithm, which is supplemented by a fast sorting algorithm and reference-free 2D class averaging to remove false positives. With only limited user-interaction, the proposed procedure yields results that are comparable to manual particle selection. Together with an improved graphical user interface, these developments further contribute to turning RELION from a stand-alone refinement program into a convenient image processing pipeline for the entire single-particle approach.

Keywords: Automated particle picking; Electron cryo-microscopy; Single-particle analysis.

Fig.1

Schematic representation of the data model. (A) Representation of a micrograph, with coordinate vectors $r=(r_x\text{,}{r}_y)$ inside the micrograph, and coordinate vectors $q=(q_x\text{,}{q}_y)$ inside each particle image. Vectors $t_i=(t_x\text{,}{t}_y)$ place the ith particle inside the micrograph with an unknown in-plane rotation ${\varphi }_i$ with respect to a common frame of reference. Inset (B), mask ${\mathbf{M}}_{\mathbf{o}}$ which is used for normalisation of the particle images: average and standard deviation of the background pixels are calculated in the white area of this mask. Inset (C), mask ${\mathbf{M}}_{\mathbf{i}}$ which is used for the particle sorting algorithm: all statistics on the difference images between each particle and its corresponding template are calculated in the white area of this mask.

Fig.2

Improved GUI and workflow. (A) Screenshot of the new GUI in RELION-1.3. (B) Proposed workflow for semi-automated particle selection in RELION-1.3. After CTFs have been estimated for all micrographs, the particle selection procedure consists of five steps (numbered 1–5), as explained in more detail in Section 4.

Fig.3

Particle selection for the KLH data. (A) The ten reference-free class averages (ordered from larger to smaller classes) that were calculated from the manually selected particles. The two classes indicated with an asterisk were selected as templates for the auto-picking. (B) Curves of precision, recall and false discovery rate against the pick threshold. A picking threshold of 0.3 was chosen. (C) The 15 particles with the highest average Z-scores after sorting. (D) The 15 largest classes (ordered from larger to smaller) after 2D class averaging of the auto-picked particles. Particles assigned to the classes indicated with an asterisk were selected for subsequent 3D refinement. (E) 3D map after refinement of the semi-automatically selected particles from the combined near-to-focus (NTF) and far-from-focus (FFF) KLH data sets.

Fig.4

Particle selection for the $\beta$-galactosidase data. (A) The 25 class averages that were calculated from the manually selected particles (ordered from larger to smaller class). The 10 class averages indicated with an asterisk were selected as templates for the auto-picking. (B) Curves of precision, recall and false discovery rate against the pick threshold. A picking threshold of 0.4 was chosen. (C) Map obtained after 3D refinement with the manually picked particles. (D) Map obtained after 3D refinement with the semi-automatically selected particles.

Fig.5

The “Einstein-from-noise” pitfall. (A) Class averages for the 15 largest classes (ordered from larger to smaller) after sorting and 2D class averaging of the auto-picked particles that were picked with a threshold of 0.1. Class averages indicated with an asteriks were identified as artificial classes caused by template bias (see Section 5). (B) Examples of particle images assigned to one of the artificial classes: the third class in A. No clear particles are visible. The lower-right image shows the average of all assigned particles in this class without any CTF correction. (C) Examples of particle images assigned to a good class: the first class in A. Particles are clearly visible. The lower-right image shows the average of all assigned particles in this class without any CTF correction. (D) 3D map obtained from 17,082 particles that were assigned to artificial classes.