Our paper on helical reconstruction of amyloids in cryoSPARC is out in Acta Cryst. F (open access): https://doi.org/10.1107/S2053230X26003675 and the import-ready .json workflow (v4.7.0) is here: https://github.com/schaefer-jh/CS-amyloids
It came up in Optimization of helical parameters in Helical Refinement some time ago, so here are the practical points in one place.
Nearly all amyloid structures in the EMDB were processed in RELION, though cryoSPARC accounts for a comparable number of depositions overall. We reprocessed nine EMPIAR datasets (tau PHF and MIA, ATTR, β2m V27M, lysozyme D87G, light-chain λ6 - thanks hickschadw) with one set of non-default parameters, validated by Q-score, map–model fit and cross-β-stack separation rather than FSC alone.
What worked
-
8 of 9 datasets gave interpretable, model-buildable maps, most with Q-scores above 0.6; lysozyme D87G reached 2.4 Å in ~12 h on three GPUs (~2000 movies).
-
Filament tracing with a diameter slightly below the measured one and Gaussian blur σ 0.4–0.5. All previously published cryoSPARC amyloid structures used the tracer, not TOPAZ.
-
2D classification in two rounds: K = 100, uncertainty factor 4, 40 EM iterations, hard classification on the last one (ESS → 1).
-
Reciprocal-space symmetry determination: rise straight off the layer lines in Average Power Spectra, twist from the crossover distance. Then constrain the rise search to 4.6–4.9 Å, start symmetry search at 4.5 Å, and read the Helical Symmetry Error Surface (discrete minimum = converged, streaks = unresolved).
-
Masking: cylindrical mask initially, dynamic masking off, then a Gaussian-filtered volume uploaded back to generate the soft mask; z-clip ~0.4 (0.3 for sharpening).
-
Several starting models from different 2D subsets, run in parallel: the most reliable defense against local minima.
What still needs development
-
Ex vivo PHF tau never converged, despite clear β-stack separation in 2D and many attempts with different starting models and symmetry parameters; overfitting signatures in the FSC, negative density between strands. The double-protofilament light-chain polymorph failed the same way. RELION maps of the same data were also somewhat more interpretable by Q-score.
-
No helical initial model from projection overlaps: Cylindrical references are a real bottleneck;
relion_helix_inimodel2dand Helicon’sdenovo3dfill this gap elsewhere. -
No automated clustering of filament 2D classes (cf. FilamentTools, CHEP): polymorph separation is manual class picking, with the false-positive risk that implies.
-
No 3D classification with rise/twist optimization. The biggest limitation for ex vivo mixtures and peptide-level polymorphs, i.e. the intermediates that matter for disease progression and time-course work.
-
Z-shifted half-maps still need manual realignment in ChimeraX and a file swap in the job directory.
Feature requests (mmclean), in priority order: (1) 3D Classification / Heterogeneous Refinement with per-class helical symmetry search; (2) non-cylindrical initial model generation from overlapping 2D class averages; (3) multi-class ab initio helical reconstruction; (4) automatic clustering of filament 2D classes; (5) automatic half-map register-shift correction.
Feedback wanted
These parameters came from a finite set of targets and are a starting point, not a recipe. Where did they break on your sample, and what did you change; especially picking and box sizes for thin or highly curved fibrils? Has anyone gotten ex vivo PHF tau, or any peptide-level polymorph mixture, to converge in cryoSPARC? And how would you rank or extend the feature list above, so the priorities reflect the community rather than just our datasets?
Issues and pull requests on the workflow repo are welcome.