Opus et visualize

OPUS-ET Visualize (in-cell scenes)

Places the refined map at every particle pose inside the tomogram, colored by conformational state, via ChimeraX + ArtiaX.

Agent Rules — read before acting

• Read before editing; verify command behavior before asserting.
• Reconcile pixel sizes explicitly — never assume star coords and tomogram share a pixel size.
• Fair coloring. When showing several candidate volumes/states side by side (e.g. a k-means state gallery), color each distinctly (or all neutrally) — never highlight a subset (e.g. "the selected ones in blue"), which pre-biases the viewer before the evidence is in. Use one consistent palette across the gallery and the latent UMAP so a state is the same color in both. Distinct 20-colors: golden-angle hue spacing gives good neighbour contrast (h=(i*0.618)%1, s≈0.68, v≈0.88).
• ChimeraX tile resets lighting — run lighting soft (and any lighting command) AFTER tile #* columns N, or it won't apply.
• Judge Gate-3 state resolution in 3D at a HIGH percentile contour, not by CC. Render the k-means state gallery with gen_gallery_cxc.py --percentile 98 (per-map 98th-percentile contour). ChimeraX's default/auto contour sits too low and renders high-res detail as low-density "speckle" that reads as junk. Two traps invert the ranking: (a) a sharp high-res map correlates less with the blurry averaged consensus / low-res template, so compare_to_template.py's CC ranking pushes the BEST states to the BOTTOM — treat low consensus-CC as a POSSIBLE high-res signal, not junk, and cross-check in 3D; (b) a cleanly separated, abundant latent-UMAP island is often the best-aligned/sharpest population, not an artifact. (Real case: ribo z8_expanded k17/18/19 were the high-res ribosomes but ranked last by CC and looked grainy at auto-contour.)

Inputs

• WARP/RELION star: rlnCoordinateX/Y/Z, rlnAngleRot/Tilt/Psi, rlnMicrographName.
• Tomogram MRC (cell context) and its pixel size (--tomo-angpix or MRC header).
• Refined/averaged map MRC.
• Optional per-particle labels.pkl from opus-et-analysis k-means (color by state).
• --coords-angpix: pixel size of the star coordinates.

Convention landmine — DEFUSE FIRST (spec §7.1)

ArtiaX vs RELION angle + pixel-size conventions are the top risk. Before rendering thousands, validate ONE particle:

1. particle_transform(row, tomo_angpix) gives [R|t] for particle N.
2. Load the map into ChimeraX and place it with that transform; compare against a manual fitmap of the map into the tomogram at that particle.
3. Only once one particle lands correctly, render the full scene. euler_to_matrix uses the RELION ZYZ Euler_angles2matrix convention.

ArtiaX 0.7.0 scene recipe (verified locally on the Mac)

emit_cxc was rewritten against real ArtiaX 0.7.0 behavior (5 tests cover the pure-Python .cxc generation, and run anywhere via the repo .venv). Old M1-stub commands like artiax open particles and artiax attach ... geomodel do not exist in 0.7.0 — do not use them.

• Renders run LOCALLY on the Mac, not on the cluster. ChimeraX 1.10 + ArtiaX 0.7.0 at /Applications/ChimeraX-1.10.app/Contents/bin/ChimeraX. Pull the maps/poses (.mrc, sel_*.star) down from the cluster first, then render.
• GUI required. ArtiaX's commands only register once its GUI is up. Run scenes as ChimeraX --exit scene.cxc — not --nogui/--offscreen.
• Deterministic model IDs after artiax start -> #1 (#1.1 Tomograms, #1.2 Particle Lists, #1.3 Geometric Models): artiax open tomo <mrc> -> #1.1.1; each plain open <map.mrc> -> #2, #3, ...; each open <star> format relion -> #1.2.1, #1.2.2, ....
• Particle lists load via plain open <star> format relion — there is no artiax open particles in 0.7.0 (generic artiax open errors telling you to use plain open). Two star readers are registered ('RELION STAR file', nickname relion; 'RELION5 STAR file', nickname relion5) — you must pass format relion for old RELION-3.x single-data_-block stars or it errors "Multiple formats ... support .star suffix".
• Coordinate scale (critical). ArtiaX places rlnCoordinate* at 1.0 A/px by default. Rescale with artiax particles #<pl> originScaleFactor <coords_angpix> (e.g. 4.2) so particles register with the tomogram. Symptom if omitted: the particle cloud is ~coords_angpixx too small vs the tomogram. The map needs no scaling — its size comes from its MRC header voxel size.
• Attach + style: artiax attach #<map> toParticleList #<pl> (map model must be a Volume); artiax show #<pl> surface; hide #<pl>.3 models (hide auto markers); color #<pl> <color>.
• Contour: set the map surface to an absolute level (mean + N·sigma) before attach so the contour is identical across all instances. Crop the map to a tight box (strip solvent padding) to cut memory when instancing thousands of copies. emit_cxc exposes this via contour_level/contour_sd.
• Optional translucent tomogram slab for cell context: open a plain second copy of the tomogram, volume #N style image, a faint symmetric transfer function (cryo-ET density is ~0-mean, e.g. volume #N level -0.008,0.7 level -0.002,0.1 level 0.002,0.1 level 0.008,0.7), color light gray, and hide the ArtiaX orthoslice (hide #1.1.1 models). emit_cxc's tomo_transfer param controls the level string.
• emit_cxc params: coords_angpix, contour_level/contour_sd, tomo_transfer, movie_out.
• Judgment note (cross-ref Gate-3 above): the sharpest/high-res state map reads as fragmented "speckle" at a tight contour while blurry low-res maps look deceptively clean — don't let that bias which state you pick.

Usage

M1: scripts/gen_artiax_scene.py is a library, not a CLI (no __main__). The conductor (or you) imports it and calls its functions directly — there is no python scripts/gen_artiax_scene.py ... invocation yet.

import pickle
import gen_artiax_scene as gs

labels = pickle.load(open("analyze.39/kmeans12/labels.pkl", "rb"))
df = gs.attach_labels(
    gs.reconcile_coords(gs.read_particles("picks.star"),
                         coords_angpix=A, tomo_angpix=A),
    labels,
)
stars = {s: gs.write_relion_star(df, f"state{s}.star", state=s)
         for s in sorted(set(df["state"]))}
gs.emit_cxc(tomogram="TS_026.mrc", map_path="ref.mrc",
            state_stars=stars, out_cxc="scene.cxc")

Produces one RELION star per state plus scene.cxc.

M2 target — not yet wired. The plan calls for a real main()/argparse CLI wrapping the same calls (see the M1 plan's "Notes for M2–M5"):

python scripts/gen_artiax_scene.py \
    --star picks.star --tomogram TS_026.mrc --map ref.mrc \
    --labels analyze.39/kmeans12/labels.pkl \
    --coords-angpix <A> --tomo-angpix <A> \
    --out-cxc scene.cxc --out-star-prefix state

Open the resulting scene in ChimeraX: ChimeraX --exit scene.cxc (GUI required — ArtiaX commands don't register under --nogui/--offscreen; see the ArtiaX 0.7.0 recipe above).

Raw-density marker reveals — show the density, don't replace it

The finale (above) places the refined map at each pose — idealized, and it hides the raw tomogram. To instead reveal the raw reconstruction with the picks only marked (validation, or a rare-species "here it really is" shot):

• Per-particle zoomed gallery — scripts/particle_gallery.py (20 tests). One small crop per pick, taken at the pick's own Z-plane, marker drawn so the interior stays visible. Reuses tm_picks_overlay's pick reader + Å coordinate reconciliation. Best for a sparse species where a whole-slab overlay loses it.

  python scripts/particle_gallery.py --tomogram TS_028_bin2.mrc \
      --picks sel_fas30.star --tomo TS_028 --coords-angpix 4.2 \
      --style ring --num 24 --half 17 --species FAS -o fas_raw_gallery.png \
      --bild fas_markers.bild        # optional: 3D markers for the scan below
  

• Whole-slab context is already tm_picks_overlay.py (mean-projected slabs, ring markers, all/top-N). Use it for "where in the cell", the gallery for "what the density looks like".
• Marker style is the reveal/spot trade-off (--style): ring (open circle, interior untouched — best for revealing density), transparent (tinted disc, density shows through), solid (opaque — unmissable but hides the particle). For a moving Z-scan, solid reads best (a thin ring flickers against the grain).
• Scannable Z-scan reveal (plain ChimeraX, not ArtiaX). ArtiaX's orthoslice can't be driven by command, so scan a plain volume image-plane with the picks as fixed BILD markers (--bild above); each barrel appears under its marker as the plane crosses its Z:

  open TS_028_bin2.mrc ; open fas_markers.bild
  volume #1 style image ; volume #1 color white
  volume #1 level -0.016,0 level 0.016,1 ; volume #1 planes z,62
  set bgColor black ; camera ortho ; view orient
  movie record ; perframe "volume #1 planes z,$1" range 62,176 frames 90 ; wait 90
  perframe stop ; movie encode fas_scan.mp4 framerate 24 quality medium
  

Molecular-sociology / cellular-context render (hero aesthetic) — in emit_cxc

This IS gen_artiax_scene.emit_cxc — placing each species' refined map at every pose in the tomogram. The hero look (molecules as a 3D cloud spilling past a floating slab, not a flat decorated micrograph) is opt-in on the same call:

emit_cxc(tomogram="TS_028_bin4.mrc", map_path=None,
         state_stars={0: "ribo_TS028.star", 1: "fas_TS028.star"},
         state_maps={0: "ribo.mrc", 1: "fas.mrc"},
         state_contours={0: 0.010, 1: 0.0072},
         state_colors={0: "cornflower blue", 1: "gold"},
         coords_angpix=4.2, tilt_x=-55,          # tilt so the slab is edge-on
         silhouettes=True,                        # 3D pop; safe at any instance count
         still_out="finale_still.png",
         movie_out="finale.mp4", movie_rock=30,   # rock +-30, NOT a 360 (thin slab)
         movie_step=3, out_cxc="finale.cxc")

silhouettes needs the tilt (a face-on view flattens it and silhouettes outline only near-plane particles). Do NOT pass shadows=True for a thousands-of- instances scene — the shadow map covers every placed copy and even a single still hangs (learned the hard way at 3,387 ribosomes). Shadows are fine for the single-map showcases (M-spins) and small scenes (≲2,000 instances, e.g. TS_029); the big in-cell finale uses silhouettes + depth-cue only. movie_step coarsens surfaces for a tractable movie (use 3, not 2, for thousands of instances); movie_rock avoids the edge-on midpoint of a 360. Defaults (all off) reproduce the plain turntable. Pick the tomogram by cross-referencing organelle content (clean *_13.48Apx.png central slices) against per-tomogram particle counts (grep -oE "TS_0[0-9]+" sel.star | sort | uniq -c, ÷5 for the star's 5 mentions/row). Rotating single-map showcases (the M result) are a plain-ChimeraX open map ; volume level <mean+4sd> ; surface dust #1 size 120 ; lighting shadows/silhouettes ; turn y 3 120 turntable. See demo/render_commands.md §B/§B3, demo/finale/build_ts029_cell_scene.py.

ChimeraX/ArtiaX gotchas (learned the hard way)

• Headless --offscreen/--nogui rendering often fails. ChimeraX offscreen rendering needs a working OSMesa / virtual-display GL context that many machines (especially cluster nodes) lack, so the render comes back blank or errors. GUI mode is the reliable path for all renders, not just ArtiaX — if a headless render fails, run it in a local GUI session. (ArtiaX is GUI-only regardless; its commands don't even register without the GUI.)
• view does NOT reset orientation — it re-fits zoom but keeps the current rotation. view; turn x -38 after a prior turn x -55 silently stacks to −93° (edge-on). Re-issue the full setup, or view orient, between angles.
• Shadows scale badly with ArtiaX instance count. lighting shadows on thousands of placed copies makes even a single still hang (the shadow map covers every instance) — a 3,387-ribosome still never finished. Use silhouettes only for the big in-cell scene; keep shadows for single-map showcases and scenes ≲2,000 instances. For a thousands-of-instances movie also use volume … step 3 (not 2) so it stays ~2 min, not ~6.
• volume … planes z,N breaks the ArtiaX orthoslice (goes black) — it only works on a plain (non-ArtiaX) volume. That's why the Z-scan uses plain ChimeraX + BILD markers rather than ArtiaX.
• turn y 360° on a thin slab goes edge-on mid-spin and hides in-plane features (an organelle arc). For a slab, rock ±~30° instead of a full turntable.
• Silhouettes at a face-on view only outline near-plane particles (the opaque slice occludes the rest) → an uneven, messy subset. They render cleanly once the slab is tilted edge-on so molecules float clear of it.
• Python split on the Mac render box: the ChimeraX-bundled python has matplotlib but NOT mrcfile; the repo .venv (3.14) has both. Extract crops with a python that has mrcfile, render figures with either. Verify an mp4 without ffmpeg via qlmanage -t -s 1100 -o <dir> movie.mp4 (QuickLook thumbnail).

Status

M1: scene generation + single-particle validation. ArtiaX command syntax is verified against ArtiaX 0.7.0 / ChimeraX 1.10 locally on the Mac (see recipe above); the cluster is only the source of the .mrc/.star inputs, not the render/verification host. Added this session: particle_gallery.py (raw-density marker galleries + BILD markers, 20 tests), the cellular-context render aesthetic, and the scannable Z-scan reveal.

Files in this skill

scripts/
  gen_artiax_scene.py    # in-cell finale — emit_cxc() places a refined map at every pose
                         #   (library; driven by demo/finale/build_insitu_scene.py)
  particle_gallery.py    # CLI — per-particle zoomed raw-density gallery + BILD 3D markers (reveal mode)
  tm_picks_overlay.py    # CLI — pick markers on mean-projected raw slabs (whole-slab context)
  slice_preview.py       # CLI — central-slice tomogram previews (Gate-1 alignment QC)
tests/                   # pytest — test_scene_pose + test_scene_coords (gen_artiax_scene),
                         #   test_particle_gallery, test_tm_picks_overlay, test_slice_preview

No references/ — the ArtiaX recipe + gotchas are inline above.