Strategies and Potentials Towards High-Resolution Subtomogram Averaging

Le résumé fourni par la source

Cryo-electron tomography (cryo-ET) has emerged as a leading technique for visualizing large, transient, dynamic, and heterogenous macromolecules in their native or near-native states. From three-dimensional reconstructions (e.g., tomograms) of these complex systems, abundant molecules can be further analyzed for high-resolution structural information using sub-tomogram averaging (STA) [2] or contextual detail using 3D-rendering approaches [3, 4]. Once sub-tomogram averages are generated, individual molecules can be mapped back into their original positions to reveal their spatial organization and provide important biological context. While recent advances have pushed the achievable resolution from STA workflows to below 4 Å [4-8], the process is usually complicated and involves cross-program integration. Here, we explored several STA strategies including the early steps of reconstruction generation, denoising, and particle picking. We present a ∼3 Å structure of DNA protection during starvation (DPS) protein (223 kDa), a popular cryo-EM standard [9, 10] whose overall molecular weight and size is similar to biological complexes in situ. DPS protein was induced via arabinose from pBAD24-DPS in a mini cell producing Escherichia coli strain WM3516, and purified as described previously [11]. 4 µl of the purified protein in a Tris buffer (pH of 7.5) was deposited onto R1.2/1.3 200 mesh Cu grid (Quantifoil), followed by plunge-freezing using a Vitrobot Mark IV (ThermoFisher Scientific). Tilt series were collected in SerialEM (v.4.1) [12] using a dose symmetric scheme [13] in groups of 2 from -60∘ to 60∘⁠, 3∘ increment, with a total dose of 80 e-/Å2. Frame movies per tilt were collected on a Krios G4 (Thermo Fisher Scientific) equipped with an E-CFEG operating at 300 kV, Selectris X imaging filter using a 10-eV slit, and a Falcon 4i direct electron detector operating in EER mode. The magnification (EFTEM mode) was 105,000x with an unbinned pixel size of 1.22 Å. A nominal defocus range of -1.5 to -3 µm in 0.325 µm steps was used. In total 63 tilt series were collected, of those, 43 were used for subsequent data processing. Raw EER fractions were motion corrected using MotionCor2 [14] (1.6.4). Briefly, a total of 72 hardware raw EER frames were grouped (fraction of 9), up sampled at 2x (0.61 Å/pixel), and corrected in 5x5 patches on which the local motion was measured. Motion corrected frames per tilt were then assembled via alignframes (IMOD [15] 4.11.24) to obtain sorted tilt series. We first used automated packages, such as AreTomo [16], to quickly assess data quality. The AreTomo approach implemented back-projection to align the tilt series and reconstruct them into tomograms at a binned pixel size of 7.32 Å (bin 6). In parallel, Etomo/IMOD was used to generate tomograms using the weighted-back projection reconstruction process [15]. Both tomograms were used for subsequent STA analyses. Particle picking is an essential step in STA. We explored several popular picking strategies including template matching, geometric picking, and automated neural network segmentation approach using simulated tomograms via TomoSIM [17] and Dragonfly. Previously, we had demonstrated the efficacy of combining PEET[18] and RELION 4.0 [8] for STA of a viral matrix protein lattice structure which reached 4.6 Å [19]. Here, we integrated Dynamo [20] and RELION 4.0 [8] to utilize GPU-empowered alignment, CTF refinement, frame-alignment, and 3D classification. At various stages of the process, we used UCSF Chimera [21] and ChimeraX [22] combined with plugins PlaceObject [23] and ArtiaX [24] to map back particles to the original tomograms so that particle orientation could be examined and masks generated for STA alignment. Finally, we built a ∼3 Å model of DPS using less than 10,000 unique particles (Figure 1). In this study, we investigated recent hardware (Titan Krios G4, SelectrisX, Falcon 4i) and software advances to prepare a workflow (Figure 2) that demonstrates the integration of these developments and the resulting outcomes. There are many options for obtaining high-resolution structures using subtomogram averaging approaches [5, 7, 25, 26]. In each case, the STA workflow will need to be tailored based on the biological targets, programs used, and computational resources available. Subtomogram averaging and model building of DPS. (A-B) A representative tomogram slice of DPS (bin 6x, pixel size of 7.32 Å) and the corresponding deep-learning based segmentation for particle picking via TomoSIM and Dragonfly. (C) Multiple Z slices of the final averaged structure. (D) Model built into the density map with one monomer highlighted in red box. (E) Model in (D) colored by Q-score[1] (residue) according to the legend shown. (F-G) Zoomed-in views of a monomer (F) and a representative alpha-helix (G) fitted with the model. Subtomogram averaging workflow. 43 complete tilt series were used for tilt series alignment and reconstruction. A total of 31291 particles were picked and went into particle volume extraction and subtomogram averaging. 19936 particles and their refined positions were imported into RELION 4.0. In the end, a final total of 9180 particles contributed to the final average of ∼3 Å (2.95 Å, FSC cutoff at 0.143)

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.