RELION single-particle refinement
Use for a RELION single-particle project, especially extracted particles → homogeneous selected
particle subset → gold-standard refinement → half-map validation and postprocessing. The bundled
runner starts from CTF-annotated extracted particles and an initial 3D reference. It does not
replace motion correction, picking, 2D/3D selection, or a biological interpretation of map quality.
For tomography, helical reconstruction, Blush, or heterogeneous-state modeling, use the appropriate
upstream workflow rather than forcing those data into this bounded SPA runner.
Preserve acquisition and coordinate conventions
Read references/acquisition-and-restarts.md when starting
from movies or resuming jobs. Confirm pixel size in Å/pixel, voltage in kV, spherical
aberration in mm, defocus in Å, amplitude contrast as a fraction, and the symmetry justified
by the specimen. Do not “correct” a suspicious value by guessing its units.
data_optics describes acquisition/image groups; data_particles references them through
_rlnOpticsGroup. Particle filenames use one-based index@stack.mrcs; leading zeros such as
00000001@stack.mrcs are valid. Relative paths resolve
from the RELION project directory, not the STAR file's directory. Keep optics groups when merging
or subsetting STAR files. _rlnOriginXAngst/_rlnOriginYAngst are Å translations, not pixels.
Run from this skill directory with paths to the real project:
python scripts/spa_workflow.py validate-star project/particles.star --project project
This opens referenced stacks and checks optics membership, finite acquisition/CTF values, indices,
box sizes, duplicate particle references and existing half-set assignments. Use --metadata-only
only when stacks are genuinely unavailable; the JSON records stack_checks_performed: false.
It does not scan every particle pixel for corruption or establish correct image normalization.
Physical-range warnings are review prompts, not proof that unusual microscope settings are wrong.
Refine a selected particle population
Before running, inspect representative particles and class averages, defocus distributions, CTF
fits, particle orientation distribution, and the initial reference. Ensure the map and particle
boxes/pixel sizes agree after any downsampling. The runner deliberately supports one effective
box/pixel size across optics groups; handle heterogeneous sampling with an explicit upstream
resampling workflow. Use conventionally extracted, normalized particles that have not already
been phase-flipped or Wiener-filtered; this runner does not configure those special input cases.
python scripts/spa_workflow.py refine \
--star project/particles.star --reference project/initial.mrc \
--project project --diameter 180 --symmetry C1 \
--initial-lowpass 40 --mpi-ranks 3 --threads 2 --output project/RefinePilot
The diameter and low-pass filter above are illustrative Å values. Use specimen-appropriate
values. Refinement executes mpirun -np 3 relion_refine_mpi with --auto_refine,
--split_random_halves, --ctf, and a low-pass starting reference. Gold-standard splitting
requires MPI; the plain sequential relion_refine executable cannot perform this split.
Use odd ranks ≥3 (master plus balanced half-set workers), with a matching MPI installation.
The CPU command is useful for a bounded pilot; choose a documented GPU/MPI launch for full data.
The runner keeps the command, native version and log in a new output directory, records an
explicit random seed (default 1), surfaces runtime warnings, stops on process failure, and
requires converged unfiltered half maps before reporting success. It does not automatically retry
expensive jobs or silently discard failed-job artifacts. Keep _optimiser.star, model/sampling
STAR files, and referenced particle paths for restart. Use the original job's optimiser rather
than starting a new random split from a partially processed table.
Inspect independent half maps
Use the two independently refined unfiltered half maps, never two copies of the combined,
sharpened map. Matching headers cannot establish statistical independence; the independent
particle assignments and refinement history provide that evidence. Inspect directional
anisotropy, preferred orientation and local resolution as well as a global FSC curve.
python scripts/spa_workflow.py fsc \
project/RefinePilot/run_half1_class001_unfil.mrc \
project/RefinePilot/run_half2_class001_unfil.mrc --output diagnostic-fsc.tsv
This checks map dimensions, finite values, pixel size, origin, axis order and duplicate maps, then
writes an unmasked diagnostic FSC. The reported 0.143 crossing uses linear interpolation;
null means no downward crossing was detected, not infinite resolution. Nyquist resolution is
2 × pixel size. This diagnostic is limited to even cubic maps ≤256³; use RELION's native
relion_image_handler --fsc for larger maps. It does not substitute for mask-corrected FSC.
The helper requires real-space maps with canonical axes, zero MRC start indices and orthogonal
cell angles. Convert other grids explicitly with provenance; merely editing headers can misalign
density. Matching headers and FSC cannot determine absolute handedness.
Postprocess with a soft mask
Construct the solvent mask from an appropriately low-pass-filtered density, with an expanded
boundary and a smooth edge. Inspect all slices; a tight mask can inflate correlation. Avoid a
mask derived from high-frequency noise shared between half maps.
python scripts/spa_workflow.py postprocess \
--half1 project/RefinePilot/run_half1_class001_unfil.mrc \
--half2 project/RefinePilot/run_half2_class001_unfil.mrc \
--mask project/soft_mask.mrc --output project/PostProcessPilot
The helper checks a nonconstant mask in [0,1], soft-edge voxels and matching map grids, then runs
relion_postprocess with explicit half maps, mask and pixel size. RELION performs its own
mask/randomization correction and writes postprocess.star. The bounded command leaves the
B-factor at zero (no automatic B-factor estimation); add automatic/manual sharpening only after choosing a defensible fit
range and inspecting map quality. A valid range and some fractional mask voxels do not prove the
mask is scientifically appropriate. Inspect the phase-randomized masked FSC near the reported
resolution: residual correlation calls for a smoother/wider mask and another postprocessing run.
See references/runtime-and-validation.md for the tested
native utilities and the distinction between pipeline execution and reconstruction validation.
Primary references