henhaobio
Sign in

Journal of Structural BiologyMolecular BiologyPeer-reviewed journal article

Helical reconstruction in RELION

Shaoda He, Sjors H. W. ScheresiD

PDFDOIPreprintCode
In-depth full-text digest available · Jump to digest ↓

Abstract

We describe a new implementation for the reconstruction of helical assemblies in the empirical Bayesian framework of RELION. Our approach calculates optimal linear filters for the 3D reconstruction by embedding helical symmetry operators in Fourier-space, and deals with deviations from perfect helical symmetry through Gaussian-shaped priors on the orientations of individual segments. By incorporating our approach into the standard pipeline for single-particle analysis in RELION, our implementation aims to be easily accessible for non-experienced users. Although our implementation does not solve the problem that grossly incorrect structures can be obtained when the wrong helical symmetry is imposed, we show for four different test cases that it is capable of reconstructing structures to near-atomic resolution.\n\nAbbreviations

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (17 Dec 2016) · CC BY 4.0

RELION 2.0 adds empirical Bayesian helical reconstruction with orientation priors, filament-level half-sets and local twist/rise optimisation, yielding near-atomic maps for four filament datasets.

Fourier-Bessel indexing is error-prone for bent, flattened or disordered filaments, and earlier RELION helical work imposed symmetry with external software, so its data-driven filters ignored the extra averaging and no segment-level orientation priors were used. This work builds IHRSR-style helical processing directly into the RELION empirical Bayesian framework, from filament picking through refinement.

Key results

  1. ShownDe novo initial models converged correctly for all four specimens: sphere lattices built from Fourier-Bessel indexing of 2D class averages for TMV and VipA/VipB, and symmetry-free refinement of a featureless cylinder followed by a symmetry search for MAVS/CARD and MamK.

    Twist/rise initial vs refined: TMV 22.04°/1.45 Å vs 22.03°/1.410 Å; VipA/VipB 30°/21.4 Å vs 29.41°/21.79 Å; MAVS/CARD −101.1°/5.150 Å vs −101.2°/5.071 Å; MamK 18.29°/55.24 Å vs 23.75°/52.11 Å. Cylinder-based asymmetric refinement never gave discernible subunits for TMV or VipA/VipB. · Initial model generation; Discussion

  2. ShownA wrong helical symmetry can still give half-maps that agree to near-atomic nominal resolution, because both halves are biased the same way. Here the missing map features and the shape of the FSC curve, not the FSC value alone, exposed the error.

    VipA/VipB refined as a 6-start C1 helix (start twist 64.9°, rise 3.63 Å; refined 65.07°, 3.76 Å) gave a reported 4.5 Å map without alpha-helices or beta-strands and a spiky FSC. The correct C6 symmetry gave 3.3 Å with the expected features and a smooth FSC. · Pitfalls of imposing incorrect helical symmetry; Discussion

  3. ShownOn TMV, assigning half-sets by whole filament, together with a top-hat translational prior spanning one rise along the helical axis, prevented overfitting from overlapping segments. Either safeguard on its own worked less well.

    TMV with both safeguards: 4.1 Å, FSC near zero beyond 3.8 Å. Both off (isotropic prior, segment-level halves): FSC never reached zero. Filament-level halves with isotropic prior: no overfitting but lower resolution. Anisotropic prior with segment-level halves: slight overfitting. · The half-sets for independent refinements; Discussion

  4. SuggestedIn MamK tests, local twist/rise optimisation from a featureless cylinder reached the correct symmetry only when starting estimates were within about 20% of the true values. Relatively large starting errors could still refine to correct values, but in some cases the resulting map had suboptimal resolution.

    MamK refinements started from a cylinder low-pass filtered at 30 Å needed initial twist within ± 20% (19~25°) and rise within ± 20% (42~62Å). Signs of a wrong symmetry are lower final resolution and twist or rise ending at the edge of the search range. · Local optimisation of twist and rise; Discussion

  5. SuggestedTests supported the practical defaults: larger inter-box distances cost no resolution, Gaussian orientation priors matched or beat top-hat priors, and template-based filament picking matched or beat manual picking.

    Inter-box distances of 2-30% gave very similar map quality for TMV and MAVS/CARD. Gaussian priors beat top-hat priors (spanning −3 to +3 SD) on VipA/VipB but made no notable difference on MamK. Auto-picking selected many more MAVS/CARD particles and gave a higher-resolution map than the EMPIAR-10031 manual picks. · Particle picking and 2D classification; The inter-box distance; Angular and translational priors

Methods, data and reproducibility

RELION's empirical Bayesian expectation-maximisation refinement is extended to helical segments with Fourier-space symmetrisation over H new asymmetric units per box, real-space symmetrisation of central sections with a grid search over twist and rise, helical 2D/3D masks, anisotropic translation and bimodal in-plane rotation priors, and filament-aware picking.

  • Tobacco mosaic virus (TMV) cryo-EM data: straight, regular helices interspersed with symmetry-breaking discs; manually picked, movie frames polished · tobacco mosaic virus · EMPIAR-10020
  • Contractile sheath filaments (VipA/VipB) of the type-VI secretion system: relatively straight, manually picked; the entry contains no movies · bacterium (species not stated) · EMPIAR-10019
  • Flexible, prion-like MAVS/CARD filaments: template-picked and compared with deposited manual coordinates; movie frames polished · EMPIAR-10031
  • In-house cryo-EM data on flexible filaments of the bacterial actin homolog MamK: template-picked; used for convergence and prior-shape tests · bacterium (species not stated)
Methods & tools
RELION 2.0, relion_refine (--helical_sigma_distance, --ignore_helical_symmetry), relion_helix_toolbox (--simulate_helix, --cylinder, --search, --impose, --pdb_helix), IHRSR (iterative helical real-space reconstruction), Fourier-Bessel indexing, Expectation maximisation, RELION template-based auto-picking, RELION particle polishing, RELION post-processing (FSC with phase randomisation, B-factor sharpening), EMPIAR, EMDB
Software & environment
RELION version 2.0, open source and free for academic and non-academic users; the helical code is GPU-accelerated together with the standard single-particle code.
Assay conditions & compute
2D classification or 3D auto-refinement took a few hours on 64-128 CPU cores for TMV and VipA/VipB, and up to an entire day on similar resources for MAVS/CARD and MamK; one Nvidia GTX1080 does the work of approximately 100 CPU cores.
Reproducibility gaps
The MamK data are in-house with no accession given; the VipA/VipB EMPIAR entry has no movies; MamK movies were not re-polished for this paper; Table 1 and the supplementary tables and figures are not included in the provided text.
Limitations (6)
  • Twist and rise refinement is a local optimiser that converges to the nearest minimum of a complex energy landscape. It cannot replace Fourier-Bessel indexing, cylinder-based asymmetric refinement or tomography for establishing helical parameters. · stated by the authors
  • FSC only measures consistency between half-maps, so a wrong symmetry imposed on both halves can produce a near-atomic estimate (4.5 Å for VipA/VipB). Symmetry validation stays uncertain unless side chains, separated beta-strands and a sound atomic model can be resolved. · stated by the authors
  • Starting from a featureless cylinder suits simple helices but failed for TMV and VipA/VipB, and its early iterations are slow. The alternative, sphere models built from indexing, depends on user expertise in reading diffraction patterns. · stated by the authors
  • 3D classification did not find a better subset in any test case. Heterogeneity that varies independently from segment to segment along a filament fits poorly into 3D helical classes, and refining twist and rise per class did not improve these maps. · stated by the authors
  • No prior is placed on the rot angle between neighbouring segments, and the width of the tilt prior is fixed by the user. Helices with seams (such as microtubules) and helical processing of sub-tomograms are not yet supported. · stated by the authors
  • The claim that results are comparable to other programs is not backed by side-by-side resolutions in the provided text (Table 1 contents are absent). Each parameter test (inter-box distance, prior shape, convergence range) used only one or two datasets. · noted in this summary

Read this if you process helical or filamentous cryo-EM data and want guidance on RELION helical priors, inter-box distance, estimating initial twist and rise, or avoiding wrong-symmetry overfitting.

Citations
566
FWCI
18.89
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISLi '25Scheres '20Randolph '19Cook '19Zivanov '20Li '25
71% semantic matchBiophysics
Daoyi Li, Xiaoqi Zhang, Wen Jiang·Journal of Structural Biology·Oct 21, 2025· 6 citations