A flexible framework for multi-particle refinement in cryo-electron tomography

Abstract

Cryo-electron tomography (cryo-ET) and subtomogram averaging (STA) are increasingly used for macromolecular structure determination in situ. Here, we introduce a set of computational tools and resources designed to enable flexible approaches to STA through increased automation and simplified metadata handling. We create a bidirectional interface between the Dynamo software package and the Warp-Relion-M pipeline, providing a framework for ab initio and geometrical approaches to multiparticle refinement in M. We illustrate the power of working within this framework by applying it to EMPIAR-10164, a publicly available dataset containing immature HIV-1 virus-like particles (VLPs), and a challenging in situ dataset containing chemosensory arrays in bacterial minicells. Additionally, we provide a comprehensive, step-by-step guide to obtaining a 3.4-Å reconstruction from EMPIAR-10164. The guide is hosted on https://teamtomo.org/, a collaborative online platform we establish for sharing knowledge about cryo-ET.

Introduction

Cryo-electron tomography (cryo-ET) is an imaging technique rapidly gaining popularity for the direct visualisation of unique biological objects in 3D. Repeating structural motifs present in cryo-ET data can be reconstructed at higher resolution by subtomogram averaging (STA), allowing for the possibility of studying macromolecular structure in situ [1]. STA has developed alongside single-particle analysis (SPA) in cryo-electron microscopy (cryo-EM), a technique that has benefitted significantly in the last 10 years from advances in both electron detection hardware and image processing software [2,3]⁠.

State-of-the-art STA workflows often co-opt tools and adapt ideas from SPA for tomography and STA [4,5]. An unfortunate consequence of this side-by-side development is a somewhat fragmented software ecosystem with no standardisation of file formats or metadata conventions [6]. Despite the advent of complete or near-complete integrated solutions for cryo-ET and STA [7–9]⁠, optimal methodology for a given dataset often requires the creative combination of different approaches that may not all be present within a single integrated pipeline. For those new to cryo-ET, the burden of interfacing many different software packages in a complex workflow often represents a barrier to the testing of alternative approaches.

Image processing for STA from cryo-ET data is a complex, multistep process that can broadly be divided into 3 blocks (Fig 1). The first, “preprocessing,” generates 3D reconstructions (tomograms) from experimental 2D micrographs. The second, “particle picking,” creates putative positions and orientations for objects of interest within each volume as well as initial reference(s) for subsequent refinement. Finally, the “refinement” block is concerned with the optimisation of reconstruction(s) from imaging data associated with a set of particle positions.

[Figure omitted. See PDF.]

Fig 1. Cryo-ET data processing can be roughly divided into 3 blocks.

Preprocessing (top) transforms raw experimental data, usually in the form of multiframe micrographs, into 3D images called tomograms. Particle picking (middle) derives particle position and orientation information from tomograms as well as an initial reference for subsequent refinement. Refinement (bottom) optimises (multiple) 3D reconstructions from input data. cryo-ET, cryo-electron tomography.

https://doi.org/10.1371/journal.pbio.3001319.g001

The preprocessing block encompasses all steps in the generation of tomograms from experimental data. Experimental cryo-ET data are usually acquired as a set of multiframe 2D micrographs, one per tilt angle in a tilt series. Typically, this block includes per tilt interframe motion estimation and correction, contrast transfer function (CTF) estimation, tilt series alignment, and 3D reconstruction [10]. Tilt series alignment is the least automated of these steps, with significant time often spent optimising alignments in an attempt to produce more accurate reconstructions.

Particle picking is somewhat entangled with the process of initial model generation, with some particle picking methods requiring an initial model and others being reference free. Exhaustive, reference-based template matching approaches are widely accepted as an imperfect solution for particle picking due to the computational overhead and the need for subsequent dataset cleaning [11,12]⁠. Care must also be taken to avoid model bias. Particle picking methods based on deep learning have been proposed, which may help to address some of these limitations [13,14]⁠. Another class of particle picking methods, termed “geometrical particle picking,” derives putative particle positions and orientations from a 3D model of a supporting geometry, usually generated from minimal manual annotations [15]. These methods can be used to impose prior knowledge about how particle poses relate to the supporting geometry during refinement. Employing such a priori knowledge is advantageous, reducing both the computational burden of global searches and the likelihood of ending up in incorrect local minima during refinement. These advantages come at the expense of extra time spent on manual annotation and metadata management.

Refinement in STA has typically meant the optimisation of particle poses, considering particles as rigid bodies within fixed 13D reconstructions [16]. The potential for studying macromolecular structure at intermediate resolution (3 to 5 Å) by STA became reality with the advent of an efficient method for 3D CTF correction [17].⁠ Inspired by progress in SPA, the refinement block may now include reference-based procedures for the a posteriori optimisation of tilt series alignment parameters, frame series alignments, and electron-optical parameters [7–9,18]⁠. Despite these algorithmic advances, there remain relatively few examples of reconstructions from tilt series data in this resolution regime in the Electron Microscopy Data Bank (EMDB), indicating that this work remains nontrivial (S1 Fig). We posit that a possible reason for this discrepancy is the complexity of integrating the various tools available for each step of the workflow in the fragmented cryo-ET software ecosystem.

Warp-RELION-M is a user-friendly solution for cryo-ET preprocessing, STA, and multi-particle refinement, yielding the highest-resolution reconstructions from cryo-ET data seen thus far [9]. The current version of Warp (v1.0.9) is designed to process data collected using SerialEM and does not integrate automated tilt series alignment procedures. The pipeline assumes that an initial model is available for template matching and does not interface directly with tools for geometrical particle picking. Dynamo is a flexible, extensible software environment for STA, providing a variety of powerful tools for the management, annotation, and analysis of cryo-ET data [15,19].

The combination of the tools in these software packages would yield a computational framework capable of more automated, ab initio, and geometrical approaches to multi-particle refinement in M. In this manuscript, we describe our efforts to facilitate working within this framework. In the Materials and methods section, a set of tools is presented, which automate fiducial-based tilt series alignment, simplify metadata handling, and create a bidirectional interface between Dynamo and Warp-RELION-M. In the Results and discussion section, we demonstrate the benefits of our metadata handling tools and illustrate the power of working within this computational framework on 2 datasets. Finally, we provide a comprehensive step-by-step guide to obtaining a 3.4-Å reconstruction of the HIV-1 CA-SP1 hexamer ab initio from 5 tilt series of EMPIAR-10164. This guide is presented as a part of a collaborative, online platform (https://teamtomo.org/) that we establish as a platform for sharing knowledge about cryo-ET data processing.

Materials and methods

Integrating on-the-fly tilt series alignment into the Warp preprocessing pipeline

Tomogram reconstruction within the Warp preprocessing pipeline currently requires the use of an external software package, IMOD [20], for tilt series alignment. In the absence of a fully integrated solution in the currently available version of Warp (1.0.9), we developed autoalign_dynamo, a package for automating fiducial-based tilt series alignment using Dynamo and IMOD. dautoalign4warp is a program for on-the-fly alignment of batches of tilt series from Warp. The program proceeds by dynamic generation of a Dynamo tilt series alignment workflow with appropriate parameters derived from minimal user input. The workflow is executed, generating a set of refined fiducial positions. Refined fiducial positions are converted using dms2mod (provided by autoalign_dynamo) and used to generate tilt series alignment parameters using the IMOD program tiltalign. Key tiltalign parameters fixed within this procedure are the following: solve for 1 tilt axis for the entire tilt series; fix tilt angles at their nominal values; and use robust fitting for parameter estimation. All tiltalign parameters can be seen and modified in the tiltalign_wrapper script. All data are subsequently organised such that alignments can be imported directly into Warp for tomogram reconstruction, requiring no further user input. The base function autoalign, which wraps Dynamo tilt series alignment functionality, is provided for those who wish to apply this procedure outside of the Warp pipeline.

Dynamo tilt series alignment workflows were released recently, and, as such, are not yet described in the literature, although details and the source code are available in the public domain [21]. To facilitate the readers understanding, we provide a brief overview of the alignment algorithm, noting that it was not implemented by us. A binary, synthetic template of a fiducial marker is generated based on user input and used to detect candidate fiducial positions within a tilt series by cross-correlation. A total of 300 subimages are extracted from the tilt series at the peaks of the resulting cross-correlation matrix and averaged to produce a template for detecting fiducial markers in the data. The template is used to detect initial fiducial positions in the tilt series by cross-correlation. Cross-correlation peaks are analysed, and those observations not meeting a minimum degree of rotational symmetry are discarded. Observations are indexed to link observations of the same fiducial marker in multiple micrographs by pairwise cross-correlation of the observations between neighbouring micrographs. The longest “trails” of linked observations are used to generate an initial 3D model of fiducial positions. The 3D model is iteratively reindexed by reprojection of fiducial positions against the tilt series, adding observations that fall within a distance threshold to the set of observations to be included for further refinement. Tilt images lacking at this point are reintegrated by a procedure comparing the reprojection of the current model with the cloud of initial observations found on that micrograph. The number of observations is maximised by the reintegration of missing fiducial markers from other images using the same procedure. The projection model is iteratively refined before the positions of fiducial markers are independently iteratively refined against an average of all observations of that marker. The final set of refined markers is pruned according to the root mean square deviation (RMSD) between measured fiducial position and the reprojected position of the 3D model.

A set of self-contained Python packages for metadata handling

In an effort to interface the primary metadata systems of RELION [22] and Dynamo [19] with the scientific Python [23] ecosystem, we provide 2 Python packages: starfile and dynamotable. These packages provide input/output functionality via a simple application programming interface (API), exposing data as pandas DataFrame objects.

Euler angles are used by many cryo-EM software packages to describe the orientation of a rigid body with respect to a fixed coordinate system. eulerangles is a Python package that provides a simple API for the batch conversion of Euler angles into rotation matrices, rotation matrices into Euler angles, and interconversion of Euler angles defined according to different conventions. Conversions can take place between all possible formulations of Euler angles in a right-handed coordinate system. A simple mechanism is provided for the definition and reuse of conventions from specific software packages. Documentation for eulerangles can be found at https://eulerangles.readthedocs.io/.

Enabling Warp preprocessing for data collected in Tomography 5

In the Warp cryo-ET preprocessing pipeline, only metadata from the SerialEM data collection program in the form of mdoc files are currently supported. Thermo Scientific Tomography 5 (Tomo 5) is the official solution provided with Thermo Scientific microscopes for electron tomography experiments [24]. We provide a small command line tool mdocspoofer for the generation of mdoc files for a directory containing multiframe micrograph files of the form <basename>_<count>[<tilt_angle>]_fractions.mrc, as generated by Tomography 5. This tool enables use of the Warp preprocessing pipeline for users of Tomography 5.

Interfacing Dynamo and the Warp-RELION-M pipeline

The recommended procedure for particle picking in the Warp-RELION-M pipeline is exhaustive, reference-based template matching. Dynamo offers an interactive environment for picking particles based on supporting geometries in cryo-electron tomograms [15]. We provide warp2catalogue, a program that sets up a database called a Dynamo catalogue for tomograms reconstructed in Warp. The catalogue is set up such that all visualisation operations make use of a deconvolved reconstruction, filtered for improved visualisation, while subsequent particle extraction uses the corresponding unfiltered volume, simplifying the experience for the user.

For rigid body particle pose optimisation, STA in RELION is integrated into the existing pipeline. Alternative approaches to particle picking, pose optimisation, and classification may provide advantages over this workflow. We provide dynamo2m, a set of tools that create a bidirectional interface between Dynamo and the Warp-RELION-M pipeline. dynamo2warp allows for particle position and orientation data in Dynamo to be used for particle reconstruction in Warp, rigid body refinement in RELION, and multi-particle refinement in M. warp2dynamo provides a route to using particles reconstructed in Warp within Dynamo. An additional utility, relion_star_downgrade, is provided to enable reconstructing particles refined using RELION version 3.1 or higher in Warp 1.0.9.

Results and discussion

Tools for simplified metadata handling

The ability to test alternative approaches and rapidly iterate is key to optimising complex data analysis workflows like cryo-ET and STA. Exploratory data analysis such as checking the internal consistency of particle positions within a supramolecular assembly after a refinement is often key to understanding the limitations of a given approach. Implementing custom input/output functionality when attempting to work with a variety of nonstandard metadata presents a barrier to entry for those wishing to interactively explore their data or implement custom analysis routines. The Python packages starfile and dynamotable provide simple interfaces between the metadata systems of RELION, Dynamo, and a wealth of data analysis infrastructure, visualisation tools, and educational resources in the scientific Python ecosystem [25,26]⁠.

Converting between different Euler angle conventions is often a pain point for those wishing to interface different cryo-EM software packages due to the abundance of ambiguities in their interpretation [27]. eulerangles simplifies generating rotation matrices from Euler angles for custom analyses and facilitates building interfaces between pieces of software that interpret Euler angles according to differing conventions. Our vectorised implementation is approximately 10× faster than the popular scipy utility scipy.spatial.transform.Rotation, taking 381 ms versus 4.38 s for the conversion of a million Euler triplets, the eulerangles package thus being more suitable for working with large sets of Euler angles in Python.

As an example of the benefits derived from working within this ecosystem, we combine starfile and eulerangles with the packages mrcfile [28] (for reading and writing MRC format image files) and napari [29]⁠ (a fast, powerful multidimensional data visualisation library) to visualise tomograms, particle positions, and their oriented Z-vectors in 3D (Fig 2). Using these packages, creating such a visualisation is achieved in less than 25 lines of (human readable) code (S2 Fig).

[Figure omitted. See PDF.]

Fig 2. Modular tools designed to integrate with the scientific Python ecosystem simplify the creation of custom visualisations in napari.

(a) Particle positions and orientations, easily accessed and manipulated using starfile and eulerangles, are rendered in 3D with the imaging data from which they were derived. (b) A flowchart of the process for creating the visualisation in (a) demonstrates the benefits provided by an interface with the scientific Python ecosystem.

https://doi.org/10.1371/journal.pbio.3001319.g002

A flexible framework for ab initio and geometrical approaches to subtomogram averaging

Each STA dataset is different, and analysis presents its own unique challenges. The tools provided in autoalign_dynamo and dynamo2m were designed to simplify cryo-ET data processing where possible and empower researchers with access to a vast array of alternative tooling where it may provide benefit. Combined, these tools enable working within a powerful, flexible framework for more automated, ab initio, and geometrical approaches to multi-particle refinement, linking Warp-RELION-M with Dynamo (Fig 3). In this section, we discuss and illustrate the benefits provided by our tools.

[Figure omitted. See PDF.]

Fig 3. autoalign_dynamo and dynamo2m enable a powerful, flexible framework for ab initio and geometrical approaches to multi-particle refinement of cryo-ET data.

The cryo-ET preprocessing pipeline in Warp (steps marked by a W, top) is extended to include on-the-fly tilt series alignment in Dynamo (marked by a grey cog). Users benefit from increased automation. warp2catalogue augments the particle picking block in Warp-RELION-M to facilitate the use of geometrical particle picking tools in Dynamo, enabling ab initio reconstruction leveraging prior knowledge of particle orientation during alignment and averaging. In the refinement block, dynamo2m enables users to move freely between Dynamo and Warp-RELION-M, choosing the tool most well suited to the problems posed by their data. Combined, these tools enable ab initio and geometrical approaches to multi-particle refinement of cryo-ET data. cryo-ET, cryo-electron tomography; CTF, contrast transfer function.

https://doi.org/10.1371/journal.pbio.3001319.g003

Automation of tilt series alignment in Warp-RELION-M with autoalign_dynamo

Optimisation of data collection strategies has significantly increased the throughput of cryo-ET data collection in recent years, with datasets now routinely exceeding 100 tilt series. As dataset sizes increase, so does the need for accurate, automated solutions to each and every step of complex data analysis workflows. Accurate tilt series alignment is often achieved using semiautomated procedures that quickly become tedious when faced with a large dataset. Providing the automated, robust fiducial-based tilt series alignment procedures from Dynamo as an on-the-fly solution tightly integrated into the Warp preprocessing pipeline greatly simplifies the generation of accurate 3D reconstructions for downstream data analysis. Algorithmically, Dynamo tilt series alignment differs from batchruntomo in IMOD in a few significant ways [30]. In IMOD, the fiducial center position is determined on a Sobel-filtered image. Dynamo iteratively refines the center position of fiducials in each image by alignment to a per-fiducial reference image generated from images of that fiducial in the whole tilt series. Additionally, the Dynamo procedure explicitly attempts to maximise the number of observations in a tilt series by reintegrating missing fiducials, leveraging an existing projection model for detection of missing fiducials and subsequent improvement of the model. In contrast, IMOD does not attempt to reintegrate markers not included in the initial seed model unless explicitly added by the user. A quantitative comparison between different automated tilt series alignment algorithms could form the basis for interesting future work.

We provide no integrated solution for aligning tilt series lacking exogenous fiducial markers, such as those from samples prepared by focused ion beam (FIB) milling, although procedures exist in other software packages [8]⁠. As new tilt series alignment algorithms are developed, their integration into on-the-fly data processing workflows will be important for optimal use of both microscope and researcher time.

Enabling geometrical approaches to subtomogram averaging in Warp-RELION-M

warp2catalogue and dynamo2warp enable the use of geometrical approaches to STA to users of the Warp-RELION-M pipeline (Fig 4). The catalogue system in Dynamo and the interactive 3D modelling tools in the dtmslice viewer make it easy to generate models of supporting geometries in sets of tomograms and use them to derive particle positions and orientations for subsequent STA experiments [15]⁠. Examples of easy-to-annotate supporting geometries from which putative particle poses can be easily derived are vesicles (Fig 4A), arbitrarily shaped membrane surfaces (Fig 4B), filaments, and crystals. warp2catalogue simplifies the setup of a Dynamo catalogue for Warp-RELION-M users: The user annotates volumes filtered to aid visualisation, whereas particle extraction operations performed from the catalogue use the corresponding unfiltered data. This automation simplifies the use of an optimal workflow without unnecessary cognitive burden for the user.

[Figure omitted. See PDF.]

Fig 4. Geometrical approaches to STA.

Particle positions and orientations are generated from minimal manual annotations. In (a), a sphere is defined by a center and edge point, and positions with an orientation normal to the surface are seeded on the sphere. In (b), an arbitrary surface is defined by annotation of points on that surface in a number of tomographic slices. A supporting mesh is derived from these points on which a set of particle positions and orientations are seeded, again with an orientation normal to the surface. (c) Different positions in a chemosensory array (EMD-10160; see also below) are annotated in blue and orange. A set of particle poses from an STA experiment (blue) is used to derive the positions and orientations of the orange particles in a “subboxing” procedure. STA, subtomogram averaging.

https://doi.org/10.1371/journal.pbio.3001319.g004

Dynamo also provides dpktbl.subbox.tableOnTable for “subboxing,” deriving particle positions and orientations, which are geometrically related to an existing set of positions and orientations (Fig 4C). This procedure is often useful in cryo-ET for focussing analysis on subunits of a large complex after an initial consensus refinement [31]⁠. dynamo2warp provides a simple mechanism for the integration of these powerful tools into the Warp-RELION-M pipeline.

Providing access to alternative refinement and classification procedures

RELION implements a Bayesian approach to the refinement of one or many 3D volumes from cryo-EM data called maximum a posteriori refinement [22]⁠. Originally designed for single-particle cryo-EM data and more recently extended to work with cryo-ET data [32]⁠, RELION provides solutions for the “refinement” block of the cryo-ET pipeline. The package features an easy-to-use graphical user interface (GUI) and is already in use in a large number of cryo-EM labs. Rather than a data processing pipeline, Dynamo is a collection of powerful tools for working with STA data. With an emphasis on customisation, Dynamo often exposes the user to large numbers of parameters in fully featured GUIs.

By interfacing Dynamo with the Warp-RELION-M pipeline in the “refinement” block, we provide Warp-RELION-M users access to highly customisable STA and multireference classification procedures, principal component analysis (PCA)-based classification, and myriad data visualisation tools and analysis tools. Dynamo users gain a route to making use of the CTF estimation and particle reconstruction tools in Warp, STA and classification in RELION, and the multi-particle refinement procedure implemented in M.

As an example of a possible benefit of increased flexibility in STA workflows, we compare the options for masking during iterative refinement in Dynamo and RELION. In a 3D classification experiment using relion_refine, one mask is used for alignment, classification, and Fourier shell correlation (FSC) for the determination of regularisation parameters between iterations. In a Dynamo multireference classification project, a user may choose to provide separate masks for these 3 procedures. As an example of the possible benefits, this allows for the optimisation of particle alignment parameters during a focused classification experiment. In addition, creative combination of masks can allow for exclusion of a membrane from alignment for a target that would otherwise require using masks with extremely soft edges to avoid FSC artefacts. Such a soft mask would necessarily include the membrane or exclude membrane-proximal protein density.

Application to the HIV-1 CA-SP1 hexamer (EMPIAR-10164)

We illustrate the power of working within this framework on an example dataset by reprocessing a 5 tilt-series subset of EMPIAR-10164 containing immature HIV-1 dMACANC virus-like particles (VLPs). This dataset was contributed to the community by the Briggs group, and this subset has previously been used to benchmark NovaCTF [17]⁠ and Warp [10]⁠, resulting in 3.9 Å and 3.8 Å reconstructions, respectively.

Initial motion correction and initial CTF estimation were performed in Warp with respective spatiotemporal model resolutions of (1, 1, 8) and (2, 2, 1). Tilt series were automatically aligned using autoalign_dynamo before CTF estimation and tomogram reconstruction at 10 Å/px in Warp. VLPs were annotated in Dynamo, and initial estimates of positions and orientations were generated with an interparticle distance of 45 Å, oversampled relative to the approximately 75 Å lattice spacing seen in the tomograms. A total of 500 particles were extracted and averaged in Dynamo to produce a coarse template. The same data were subject to STA against this template in Dynamo without imposing symmetry during refinement. The resulting average contained a hexagonal lattice. To obtain an initial model, the 6-fold axis of the lattice was aligned to the z-axis, and the resulting volume was symmetrised 6-fold in Dynamo. A total of 28,516 particles were extracted from annotated VLPs and aligned against the initial model for one iteration with a limited angular search range in Dynamo. The resulting particle positions and orientations, visualised in Dynamo, formed regular lattices on the surfaces of each VLP with some less regular areas. Using simple MATLAB scripts, only the 19,810 particles with 3 or more neighbours at the expected interparticle distance were retained for subsequent analysis. Using dynamo2m, metadata were converted to enable working in Warp-RELION-M. Particle reconstructions were carried out in Warp and 3D refinements in RELION 3.1, starting with local angular searches of ±15°. Resolution estimates were measured by FSC between independent reconstructions from random half-sets of the data using relion_postprocess. Particles were (i) reconstructed at 5 Å/px and refined (unmasked) to 10 Å resolution; (ii) reconstructed at 2.5 Å/px and refined (unmasked) to 5 Å resolution; and (iii) reconstructed at 1.7 Å/px in Warp and refined (masked to include only the central hexamer) to 3.8 Å resolution. This intermediate result is indicative of the accuracy of fiducial-based alignment workflows in Dynamo and alignments in RELION. Seven rounds of multi-particle refinement were performed in M at 1.6 Å/px in a masked region containing only the central hexamer. The first 4 rounds of multi-particle refinement, optimising only for deformation parameters, yielded a 3.6-Å reconstruction. Three further rounds including optimisation of electron-optical aberration–related parameters and tilt frame alignments refined to a resolution of 3.4 Å (Fig 5).

[Figure omitted. See PDF.]

Fig 5. A (nominally) 3.4-Å reconstruction of the HIV-1 CA-SP1 hexamer from 19,810 particles in 5 tilt series of EMPIAR-10164.

The whole hexamer with one monomer highlighted in yellow (a), one monomer (b), and the region of the map around helix 292–307 PDB 5L93 (fitted) (c) are shown. The gold-standard FSC curve (d) used to estimate the nominal resolution value is shown as well as the phase-randomised masked FSC used to validate the mask used for FSC calculation. The data used to produce this figure can be found at http://doi.org/10.5281/zenodo.4783129. FSC, Fourier shell correlation.

https://doi.org/10.1371/journal.pbio.3001319.g005

It should be noted that working within this framework readily yielded a 3.4-Å reconstruction from 19,810 particles in 5 tomograms, without use of an external reference and leveraging the geometry of the system evidenced in the tomograms during alignments. The combination of Dynamo tools and the Warp-RELION-M pipeline, enabled by autoalign_dynamo and dynamo2m, rendered this ab initio, geometrical approach a smooth, efficient process. An ab initio approach is not strictly necessary for this dataset, as demonstrated by emClarity [7]⁠ and Warp-RELION-M [9]⁠. We, however, note that obtaining an initial model is often a challenging step in the first stages of a project. Thus, access to ab initio and geometrical approaches is provided as an alternative, should the integrated template matching procedure prove inadequate. The final resolution of the reconstruction serves only to demonstrate the ability of M to further optimise image alignment and electron-optical parameters.

Application to the Escherichia coli chemosensory array in situ

We are currently using an optimised minicell system [33]⁠ to investigate structural bases of the chemotactic response in motile bacteria. To illustrate the advantages of working within this framework, we choose to present intermediate results from a set of 109 minicell tomograms (Fig 6).

[Figure omitted. See PDF.]

Fig 6. A (nominally) 11-Å reconstruction of the in situ chemosensory array from 17,484 particles in 48 tomograms of E. coli minicells.

The full reconstruction with a pixel size of 7.5 Å/px is shown in (a) with the membrane density superposed at a lower isosurface threshold. The quality of the density is visible in a slice through the map reconstructed with a pixel size of 2.24 Å/px and filtered to 11 Å (b, left). Holes are clearly visible in the densities of receptor dimers at this resolution (b, right). The gold-standard FSC curve (c) used to estimate the nominal resolution value is shown alongside the phase-randomised masked FSC curve used to validate the mask used for FSC calculation. The data used to produce this figure can be found at http://doi.org/10.5281/zenodo.4783151. FSC, Fourier shell correlation.

https://doi.org/10.1371/journal.pbio.3001319.g006

The chemotaxis-mediating molecular machinery forms large supramolecular arrays, a highly cooperative network of core signalling units (CSUs), in the bacterial inner membrane at the cell poles. The elongated nature of the CSU, the crowded cellular milieu, and increased sample thickness complicate template matching approaches with this system, increasing the false-positive rate and impeding appropriate peak extraction. Instead, we opt for a seed-oversampling approach in which initial positions and orientations are distributed on Dynamo surface models, oversampling with respect to the expected interparticle distance of approximately 120 Å. Particle positions were generated from the surface model with an interparticle distance of 30 Å using Dynamo. Particles were extracted and subject to one round of STA in Dynamo with local angular searches of ±30°. Constraining the angles of particles relative to initial estimates from a supporting geometry ensures that particles within each array have orientational parameters coherent with our understanding of the system. Particles were extracted at 7.5 Å/px in Warp and imported into Dynamo using warp2dynamo.

A consensus refinement of these particles, centred on the 3-fold symmetry axis of a p6-symmetric assembly, was performed in Dynamo, allowing local deviations (±30°) in out-of-plane orientation, complete search of in-plane orientation, and enforcing C3 symmetry during refinement. In the resulting average, the expected 3-fold symmetric arrangement of P4 domains in the structure was not readily observed. The resulting reconstruction had a nominal resolution of 18 Å, measured by gold-standard FSC, with the FSC curve indicating the presence of significant heterogeneity. The Dynamo tool dpktbl.subbox.tableOnTable was used to derive the positions and orientations of particles centred on 6 intertrimer-of-dimer axes of the array. Duplicate particles were removed using dpktbl.exclusionPerVolume prior to a consensus refinement of these particles enforcing C2 symmetry. The resulting reconstruction had a nominal resolution of 19 Å. Attempts to disentangle structural heterogeneity using multireference approaches to 3D classification in a variety of masks using both Dynamo and RELION failed to yield sensible results at this stage.

Multi-particle refinement in M yielded improvements when performed in a masked region encompassing a large array region of 400 Å diameter, resulting in a reconstruction with a nominal resolution of 11 Å. No improvement was seen when refinement was performed on smaller regions encompassing 1 to 4 CSUs, presumably due to insufficient signal for accurate alignments. Work is underway to improve upon this intermediate result by classification and further refinement. However, these initial results are already quite promising, showing features such as a hole in the center of the 4-helical bundle of a receptor dimer, not yet seen in situ.

A comprehensive guide to obtaining a 3.4-Å reconstruction of the HIV-1 CA-SP1 hexamer

Information about best practices and approaches to problems presented by cryo-ET data analysis is constantly evolving and currently divided between the literature, mailing lists, and documentation specific to various pieces of software. We provide, in the form of a rich, living document, a step-by-step walk-through for the procedure used to obtain the 3.4-Å reconstruction of the HIV-1 CA-SP1 hexamer (https://teamtomo.org/walkthroughs/EMPIAR-10164/introduction.html). This guide takes a user through downloading a 5 tilt series subset of EMPIAR-10164, preprocessing multiframe micrographs in Warp, aligning tilt series automatically using autoalign_dynamo, generating tomograms in Warp, setting up a Dynamo catalogue for geometrical particle picking using warp2catalogue, picking particles based on supporting geometries in Dynamo, initial model generation in Dynamo, isolating an optimal subset of particles using custom scripts, reconstruction of particles in Warp via dynamo2m, rigid body STA in RELION, and multi-particle refinement in M. Working through this guide is designed to expose a user to the theory and practice of obtaining a reconstruction from cryo-ET data. They will learn to use a variety of computational tools to address problems they will likely encounter when analysing their own data. The guide is provided as part of https://teamtomo.org/, an open collaborative platform we established for sharing knowledge in the growing cryo-ET community. Those wishing to share their expertise for the benefit of the wider community are invited to contribute.

Conclusions

The tools and resources presented in this manuscript increase automation within the Warp preprocessing pipeline, simplify metadata handling, and interface two complementary cryo-ET data processing packages. The interface provided by dynamo2m enables working within a powerful, flexible framework for geometrical and ab initio approaches to STA, which can benefit from multi-particle refinement in M. Our tools increase the functionality available to users of both Dynamo and Warp-RELION-M, lowering the barrier to employing more customisable workflows for cryo-ET data processing, which are often key to success. Finally, we have attempted to make structural cryo-ET more accessible by providing a comprehensive, step-by-step walk-through to obtaining state-of-the-art STA results. It is our hope that this online guide can serve as a starting point for newcomers to the growing cryo-ET community, enabling them to more quickly understand and use cryo-ET to solve their specific biological questions. Furthermore, we anticipate that the existence of a collaborative, community-driven resource for sharing cryo-ET knowledge and tools will encourage open science and contribute to the development of cryo-ET as a valuable tool for structural cell biology.

Supporting information

S1 Fig. The number of STA reconstructions with different reported resolution values in the EMDB.

The data used to produce this figure can be found at https://www.ebi.ac.uk/pdbe/emdb/statistics_main.html. EMDB, Electron Microscopy Data Bank; STA, subtomogram averaging.

https://doi.org/10.1371/journal.pbio.3001319.s001

(TIF)

S2 Fig. The source code for generating the scene depicted in Fig 2, combining starfile and eulerangles with existing tools in the scientific Python ecosystem.

https://doi.org/10.1371/journal.pbio.3001319.s002

(TIF)

Acknowledgments

We acknowledge Diamond Light Source for access and support of the cryo-EM facilities at the UK’s national Electron Bio-imaging Centre (eBIC), funded by the Wellcome Trust, MRC, and BBRSC. Cryo-ET data acquisition has been supported by iNEXT, grant number 653706 (PID:2626 to IG), funded by the EU Horizon 2020 programme.

Citation: Burt A, Gaifas L, Dendooven T, Gutsche I (2021) A flexible framework for multi-particle refinement in cryo-electron tomography. PLoS Biol 19(8): e3001319. https://doi.org/10.1371/journal.pbio.3001319

References

1. Schur FK. Toward high-resolution in situ structural biology with cryo-electron tomography and subtomogram averaging. Curr Opin Struct Biol [Internet]. 2019;58:1–9. Available from: http://www.ncbi.nlm.nih.gov/pubmed/31005754. pmid:31005754

2. Bai X, McMullan G, Scheres SHW. How cryo-EM is revolutionizing structural biology. Trends Biochem Sci [Internet]. 2015 Jan;40(1):49–57. Available from: http://www.ncbi.nlm.nih.gov/pubmed/25544475. pmid:25544475

3. Nakane T, Kotecha A, Sente A, McMullan G, Masiulis S, Brown PMGE, et al. Single-particle cryo-EM at atomic resolution. Nature [Internet]. 2020;587(7832):152–6. Available from: http://www.ncbi.nlm.nih.gov/pubmed/33087931. pmid:33087931

4. Turoňová B, Sikora M, Schürmann C, Hagen WJH, Welsch S, Blanc FEC, et al. In situ structural analysis of SARS-CoV-2 spike reveals flexibility mediated by three hinges. Science [Internet]. 2020;370(6513):203–8. Available from: http://www.ncbi.nlm.nih.gov/pubmed/32817270. pmid:32817270

5. Sanchez RM, Zhang Y, Chen W, Dietrich L, Kudryashev M. Subnanometer-resolution structure determination in situ by hybrid subtomogram averaging—single particle cryo-EM. Nat Commun [Internet]. 2020;11(1):3709. Available from: http://www.ncbi.nlm.nih.gov/pubmed/32709843. pmid:32709843

6. Leigh KE, Navarro PP, Scaramuzza S, Chen W, Zhang Y, Castaño-Díez D, et al. Subtomogram averaging from cryo-electron tomograms. Methods Cell Biol [Internet]. 2019;152:217–59. Available from: http://www.ncbi.nlm.nih.gov/pubmed/31326022. pmid:31326022

7. Himes BA, Zhang P. emClarity: software for high-resolution cryo-electron tomography and subtomogram averaging. Nat Methods [Internet]. 2018;15(11):955–61. Available from: http://www.ncbi.nlm.nih.gov/pubmed/30349041. pmid:30349041

8. Chen M, Bell JM, Shi X, Sun SY, Wang Z, Ludtke SJ. A complete data processing workflow for cryo-ET and subtomogram averaging. Nat Methods [Internet]. 2019;16(11):1161–8. Available from: http://www.ncbi.nlm.nih.gov/pubmed/31611690. pmid:31611690

9. Tegunov D, Xue L, Dienemann C, Cramer P, Mahamid J. Multi-particle cryo-EM refinement with M visualizes ribosome-antibiotic complex at 3.7 Å inside cells. Nat Methods [Internet]. 2021; Available from: https://www.nature.com/articles/s41592-020-01054-7.

10. Tegunov D, Cramer P. Real-time cryo-electron microscopy data preprocessing with Warp. Nat Methods [Internet]. 2019;16(11):1146–52. Available from: http://www.ncbi.nlm.nih.gov/pubmed/31591575. pmid:31591575

11. Schaffer M, Pfeffer S, Mahamid J, Kleindiek S, Laugks T, Albert S, et al. A cryo-FIB lift-out technique enables molecular-resolution cryo-ET within native Caenorhabditis elegans tissue. Nat Methods [Internet]. 2019;16(8):757–62. Available from: http://www.ncbi.nlm.nih.gov/pubmed/31363205. pmid:31363205

12. Basanta B, Chowdhury S, Lander GC, Grotjahn DA. A guided approach for subtomogram averaging of challenging macromolecular assemblies. J Struct Biol X [Internet]. 2020;4:100041. Available from: http://www.ncbi.nlm.nih.gov/pubmed/33319208. pmid:33319208

13. Chen M, Dai W, Sun SY, Jonasch D, He CY, Schmid MF, et al. Convolutional neural networks for automated annotation of cellular cryo-electron tomograms. Nat Methods [Internet]. 2017 Oct;14(10):983–5. Available from: http://www.ncbi.nlm.nih.gov/pubmed/28846087. pmid:28846087

14. Moebel E, Martinez-Sanchez A, Larivière D, Fourmentin E, Ortitz J, Baumeister W, et al. Deep Learning Improves Macromolecules Localization and Identification in 3D Cellular Cryo-Electron Tomograms. bioRxiv. 2021.

15. Castaño-Díez D, Kudryashev M, Stahlberg H. Dynamo Catalogue: Geometrical tools and data management for particle picking in subtomogram averaging of cryo-electron tomograms. J Struct Biol [Internet]. 2017;197(2):135–44. Available from: http://www.ncbi.nlm.nih.gov/pubmed/27288866. pmid:27288866

16. Wan W, Briggs JAG. Cryo-Electron Tomography and Subtomogram Averaging. Methods Enzymol [Internet]. 2016;579:329–67. Available from: http://www.ncbi.nlm.nih.gov/pubmed/27572733.

17. Turoňová B, Schur FKM, Wan W, Briggs JAG. Efficient 3D-CTF correction for cryo-electron tomography using NovaCTF improves subtomogram averaging resolution to 3.4Å. J Struct Biol [Internet]. 2017;199(3):187–95. Available from: http://www.ncbi.nlm.nih.gov/pubmed/28743638. pmid:28743638

18. Bartesaghi A, Lecumberry F, Sapiro G, Subramaniam S. Protein secondary structure determination by constrained single-particle cryo-electron tomography. Structure [Internet]. 2012 Dec 5;20(12):2003–13. Available from: http://www.ncbi.nlm.nih.gov/pubmed/23217682. pmid:23217682

19. Castaño-Díez D, Kudryashev M, Arheit M, Stahlberg H. Dynamo: a flexible, user-friendly development tool for subtomogram averaging of cryo-EM data in high-performance computing environments. J Struct Biol [Internet]. 2012 May;178(2):139–51. Available from: http://www.ncbi.nlm.nih.gov/pubmed/22245546. pmid:22245546

20. Kremer JR, Mastronarde DN, McIntosh JR. Computer visualization of three-dimensional image data using IMOD. J Struct Biol [Internet]. 1996;116(1):71–6. Available from: http://www.ncbi.nlm.nih.gov/pubmed/8742726. pmid:8742726

21. Castaño-Díez D. Dynamo: Walkthrough on GUI based tilt series alignment [Internet]. 2019. Available from: https://wiki.dynamo.biozentrum.unibas.ch/w/index.php/Walkthrough_on_GUI_based_tilt_series_alignment.

22. Scheres SHW. RELION: implementation of a Bayesian approach to cryo-EM structure determination. J Struct Biol [Internet]. 2012 Dec;180(3):519–30. Available from: http://www.ncbi.nlm.nih.gov/pubmed/23000701. pmid:23000701

23. Virtanen P, Gommers R, Oliphant TE, Haberland M, Reddy T, Cournapeau D, et al. SciPy 1.0: fundamental algorithms for scientific computing in Python. Nat Methods [Internet]. 2020;17(3):261–72. Available from: http://www.ncbi.nlm.nih.gov/pubmed/32015543. pmid:32015543

24. Thermo Fisher Scientific. Tomography. Software version 5.X. Release Notes [Internet]. 2020. Available from: https://assets.thermofisher.com/TFS-Assets/MSD/Product-Updates/105074-TOMO-Release-notes.pdf.

25. Harris CR, Millman KJ, van der Walt SJ, Gommers R, Virtanen P, Cournapeau D, et al. Array programming with NumPy. Nature [Internet]. 2020;585(7825):357–62. Available from: http://www.ncbi.nlm.nih.gov/pubmed/32939066. pmid:32939066

26. Paszke A, Gross S, Massa F, Lerer A, Bradbury J, Chanan G, et al. PyTorch: An Imperative Style, High-Performance Deep Learning Library. 2019 Dec 3. Available from: http://arxiv.org/abs/1912.01703.

27. Urzhumtseva L, Urzhumtsev A. py_convrot : rotation conventions, to understand and to apply. J Appl Crystallogr [Internet]. 2019 Aug 1;52(4):869–81. Available from: http://scripts.iucr.org/cgi-bin/paper?S1600576719007313.

28. Burnley T, Palmer CM, Winn M. Recent developments in the CCP-EM software suite. Acta Crystallogr D Struct Biol [Internet]. 2017 Jun 1;73(Pt 6):469–77. Available from: http://www.ncbi.nlm.nih.gov/pubmed/28580908. pmid:28580908

29. Sofroniew N, Lambert T, Evans K, Winston P, Nunez-Iglesias J, Bokota G. napari/napari: 0.4.3rc1 [Internet]. 2021. Available from: https://zenodo.org/record/4437554#.YA6FldYo83E.

30. Mastronarde D, Held S. Automated tilt series alignment and tomographic reconstruction in IMOD. J Struct Biol [Internet]. 2017;197(2):102–13. Available from: http://www.ncbi.nlm.nih.gov/pubmed/27444392. pmid:27444392

31. Ertel KJ, Benefield D, Castaño-Diez D, Pennington JG, Horswill M, den Boon JA, et al. Cryo-electron tomography reveals novel features of a viral RNA replication compartment. Elife [Internet]. 2017 Jun 27;6. Available from: https://elifesciences.org/articles/25940. pmid:28653620

32. Bharat TAM, Russo CJ, Löwe J, Passmore LA, Scheres SHW. Advances in Single-Particle Electron Cryomicroscopy Structure Determination applied to Sub-tomogram Averaging. Structure [Internet]. 2015 Sep 1;23(9):1743–53. Available from: http://www.ncbi.nlm.nih.gov/pubmed/26256537. pmid:26256537

33. Burt A, Cassidy CK, Ames P, Bacia-Verloop M, Baulard M, Huard K, et al. Complete structure of the chemosensory array core signalling unit in an E. coli minicell strain. Nat Commun [Internet]. 2020 Dec 6 [cited 2020 Jun 27];11(1):743. Available from: http://www.ncbi.nlm.nih.gov/pubmed/32029744. pmid:32029744