leonarski_fandClaude Opus 5.5
387a54b33e
The FFT search reaches 500 A (DeAlias unfolds to 1000 A), so a longer axis is never a candidate. The old rescue that widened the bound ran only when fewer than half the validation frames indexed, kept the coarse lattice with the LONGEST axis, and adopted it on frame count. So a hexagonal crystal with a ~1000 A c axis, indexing most frames on a spurious shorter c, never asked it, and a large-cell orthorhombic crystal was given a ~4x-volume triclinic cell where the true cell puts about twice the share of the spots on itself. LongAxisHypothesis replaces it: - asked whenever D * lambda / (pixel * r80) - the longest axis whose rows the detector separates, r80 the pre-scan's spot width - exceeds the standard bound; the coarse search (3.5 A) reaches that length, at most the indexer's 1200 A limit. No frame-count gate. Not asked when a cell is given. - every lattice the coarse schemes return, other than the standing one, is re-indexed constrained by its cell and counted on the pooled validation spots beyond chance; LongAxisChoice takes the largest excess that ValidationEvidencePrefers over the standing lattice, never an exact superlattice of it (the pooled count is biased toward the larger of an exactly related pair; left to the merge's supercell probe). Near ties are decided late on all the frames. Both cases above now index on the true cell at the first pass. The hypothesis costs about 0.02-0.7 s per first pass where it is asked. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HVeDQqN9qRD57vnShzTC37
8.9 KiB
What Rugnux does
The map of a run, in the order it happens — one paragraph per stage, each linking into the data-analysis reference where the algorithm lives. The walk-through is a rotation run with the defaults; stills differences are at the end.
Open the dataset. The geometry, wavelength and goniometer come from the file (What Rugnux reads). A goniometer axis that turns makes it a rotation run; none, or one that stands still from frame to frame, makes it serial stills — nothing is asked of the user.
Pre-scan. A projection of frames (60 by default) spread over the first part of the sweep, the part the geometry pre-pass integrates, finds the beam-stop shadow and masks it (§1.5), masks pixels its frames show to be defective (§1.6), measures the beam centre from the isotropy of the scattered background and compares it with the file's (§1.4), and reads how wide this crystal's spots are, which sets the integration radius (§9.5), and how much longer they are along their radius than across it, which is the beam's bandwidth where the file does not state one (§9.6), and how much wider still they grow away from the beam, which sets the integration footprint (§9.1). In a GPU build the pre-scan runs on the card.
Spots. Every image is decoded — on the GPU straight from the compressed chunk (§0) — and one fused pass computes the azimuthal profile and finds the spots against each image's own per-resolution-ring noise (§2–§3). The ice-ring score is read off the same profile.
Indexing. The spots of a sample of frames are rotated back to a common crystal frame and the FFT search looks for periodicity over thousands of directions; candidate cells are Niggli-reduced, classified by Bravais lattice, refined both constrained and triclinic, and decided on how many validation frames each actually indexes (§4–§7). A de-novo run also tries a second hypothesis with the shortest accepted axis lowered from 10 to 5 Å, for small-molecule cells. The first pass is indexed a second time at the beam centre the pre-scan measured; the file's centre is kept unless it indexes nothing, gives an axis harmonic of the measured centre's lattice, or loses to it when both first passes are merged (§1.4). A failed pass triggers the discrete rescues — the rotation-axis sign, the first-pass depth ladder, the long-axis hypothesis — before anything is given up on.
First integration pass. At the geometry in the file, every frame of the first half of the sweep, at least 120° of it
(--prepass-fraction) is predicted
(§8) and profile-fit integrated
(§9); partials are combined into fulls, scaled and merged
(§10).
Geometry post-refinement. From those reflections the detector distance, beam centre and the cell scale / rotation axis are refined over all those frames at once, each step committed only if it improves a held-out residual (§7.5).
Second pass. The sweep is re-indexed de novo and re-integrated at the refined geometry; this
pass is the canonical output, and a guard compares the two passes and keeps the better one
(reported as PASS= / PASS_DECISION= in the report).
Space group. On the P1 merge of the final pass, the point group is scored operator by operator
on resolution-normalised intensities and the screw axes, glide planes and centring are read from the
systematic absences, with the centre of symmetry from the intensity distribution where the absences
leave it open (§13.1); twinning and translational
pseudo-symmetry are checked beside it (§13.2). CANNOT_DETERMINE and an enantiomorphic pair are real
answers here, not evasions.
Scale and merge. In the determined group: per-frame scales, the cross-validated correction surfaces (decay, absorption, modulation), the error model and ISa, outlier rejection, the CC1/2-based resolution cut, the anisotropy description, French–Wilson amplitudes and the R-free flags (§10, §13.3–§13.5).
Model: validation and molecular replacement. Only with a model (--model or --mr, PDB or
mmCIF). With --model the model is placed as a rigid body and checked against a null; where that does
not place it, or the model is a prediction, it is placed by molecular replacement instead, and --mr
goes to molecular replacement straight away. The search places as many copies as the cell holds, in
every space group the merge cannot tell apart and in the twin hypothesis' subgroup; the placed model
then gives R-work/R-free and σA-weighted maps. The model never changes how the data are processed
(Model and reference,
§14). It runs on every GPU of the machine, and on a Mac its
model density and rotation function on the Apple GPU (GPU data analysis).
Model: refinement. A placed model is then refined against the merge it was placed on (--no-refine
skips it): its chains as rigid bodies and B per residue where that lowers R-free, then every atom's position
and B by a restrained maximum-likelihood refinement whose weights are chosen by R-free in short trials run
side by side; the refined model gives its own maps and the blobs of unexplained density. Like the
validation it decides nothing - the merge, R_WORK/R_FREE and the placed model's files stay as they
were, and the refined results are written beside them under REFINE_
(Model and reference, §15).
On a GPU it takes seconds, with the weight trials spread over every card
(GPU data analysis).
Write. The merged .mtz / .cif / .hkl, the unmerged MTZ and the
results report land next to each other (Output files); with a
model, the maps and the placed model are written too, and beside them the refined model, its maps and
its blobs; where molecular replacement chooses another space group or the twin hypothesis' subgroup, that merge is written beside the answer as
<prefix>_mr.mtz/.cif with the model placed there, <prefix>_mr_model.pdb.
Stills instead. Serial data skip the two-pass machinery: each image is indexed independently
(with the known-cell ffbidx indexer where a cell is given), partiality comes from a per-crystal
orientation-tilt post-refinement rather than a rocking curve, and where the lattice admits a
merohedral indexing ambiguity the crystals are put into one indexing before the first merge: from the
data alone, each crystal into the indexing in which it agrees best with all the others, where the space
group is given (-S); or per image against a reference or a model, which then decides the indexing
the merge is written in. Either way a crystal whose own refined cell already tells the indexings apart
keeps its indexing (Advanced ▸ the indexing ambiguity). The input may be
several files given together - Jungfraujoch masters, or the files an XFEL pipeline writes, read without
a geometry file (Serial XFEL inputs). Before the images are
processed the beam centre is decided on a sample of frames, the file's against the measured ones, and a
given unit cell also decides the detector distance there; the cell is a starting point the frames
refine, and only the frames of its crystal form are merged. With a cell, the geometry refinement places
each detector module as well (§1.4,
§7.7). The report adds
R_split and a CC1/2 from frame half-sets, as CrystFEL computes them (Report).
