docs/REPRODUCE.md

Reproducing CryoDynaScape: every figure and table

This guide maps every display item in the manuscript (cryodynascape_overleaf/main.tex) to the runs, result files and commands that produce it. Paths are relative to the project root (ROOT=/n/netscratch/kempner_mzitnik_lab/Lab/hxu66/cryodynascape). Commands were recovered from three sources, in this order:

1. logs/<run>.log, first line args {...}. run_pipeline.py, run_v1.py, run_hinge.py and run_rigid.py log every argument, so this is the exact configuration of a run. The same dict is stored under the key args in checkpoints/<run>/*.pth, which is on HF. logs/ itself is not.
2. sacct -P -j <id> -o SubmitLine%2000. It records --wrap commands and --export= variables. Variables set with a shell export before sbatch are not recorded. For those wrappers the RUN/STAGE1/V1 values below come from source 1.
3. The project log (memory/cryodynascape-project.md) and results/SUMMARY.md.

Items marked [untraced] could not be recovered exactly (none affects a number in the manuscript: all AU-FSC latents are traced in results/official/latent_sources.json). Items marked [inferred] were reconstructed from file names and shapes, not from a recorded command.

0. Shortcut: download instead of recompute

Everything the paper reads is public at https://huggingface.co/datasets/Xu-AI4Science/CryoDynaScape-data (26,652 files, 446 GB). It uses the same relative paths as this repository: results/, checkpoints/, baselines/inputs/, baselines/runs/, data/synth_*, data/Spike-MD-128/, data/real/empiar10180_128/ and data/real/empiar10076_128/.

huggingface-cli download Xu-AI4Science/CryoDynaScape-data --repo-type dataset --local-dir $ROOT \
    --include "results/*"          # figures/tables only need results/ (+ data/* for GT-dependent panels)

With results/ in place, every step in Section 4 that runs on CPU (scripts fig_*, *_compose, supp_table_paired and table_hier) regenerates its figure or table without a GPU. The GT-dependent panels (Fig. 3a/c, Table 2) also need the CryoBench data from Section 2.

The HF dataset does not contain:

• CryoBench originals or EMPIAR raw micrographs/stacks.
• data/structures/ (templates, AF3 models).
• The CryoSTAR work dirs (baselines/cryostar/projects/star/work_dirs/). The CryoSTAR latents .../<run>_0/0061_0096000/z.npy are therefore only available locally.
• logs/.

1. Environments

All wrappers hard-code /n/holylabs/kempner_mzitnik_lab/Lab/hxu66/conda/envs/<env>. The env migration to netscratch was cancelled, so envs/README.md, which still names .../cds_envs/<env>, is stale on location only. Its specs are current. To run elsewhere, edit PY=, E= and RE= at the top of each wrapper.

env	spec in envs/	kind	used by
cryodynascape	cryodynascape.requirements.txt + .pyvenv.cfg (Mambaforge Py 3.10.13, torch 2.5.1+cu121)	venv	every scripts/*.py of ours, AU-FSC scoring (official_fsc.py, prep_official.py), all figure/table scripts
recovar	recovar.requirements.txt (Py 3.12, jax 0.9.0.1)	venv	RECOVAR baseline and the kernel-regression generator inside every AU-FSC evaluation
cryodrgn	cryodrgn.requirements.txt (Py 3.12)	venv	cryoDRGN baseline, write_star.py, parse_pose_star checks
cryostar	cryostar.yml (+ pip install baselines/cryostar)	conda	CryoSTAR, baseline_displacement.py --method cryostar
cryosphere	cryosphere.yml (numpy<2, torchvision 0.16.2)	conda	cryoSPHERE, cryosphere_resume.py, baseline_displacement.py --method cryosphere
dynamight	dynamight.yml (+ DynaMight git 358db04d5)	conda	DynaMight, dynamight_latent.py
eman2	eman2.yml (-c cryoem -c conda-forge eman-dev)	conda	e2gmm (run_e2gmm.py)
cds_render	cds_render.yml (conda-forge mesalib<24.1)	conda	OSMesa for headless ChimeraX (scripts/chimerax_offscreen.sh)
scipion3, oXmipp	*.yml	conda	HEMNMA attempt only (reported n.r.)

Rebuild:

python -m venv <path>; <path>/bin/pip install -r envs/<env>.requirements.txt --extra-index-url https://download.pytorch.org/whl/cu121
conda env create -p <path> -f envs/<env>.yml          # conda envs
sbatch scripts/rebuild_envs.sh                         # scripted rebuild of all envs (test partition, CPU)

Never create envs on the login node: conda create gets OOM-killed there. External software:

• RELION 5: built from source; recipe in envs/README.md. Used by submit_baseline_prep.sh (relion_reconstruct) and submit_multibody.sh.
• ChimeraX 1.8: el8 RPM extracted to .../software/chimerax/root. Extract with rpm2cpio | cpio -id without -m, so files get fresh mtimes and survive the purge.

2. Data

dataset	source	local path	preprocessing
IgG-1D, IgG-RL	CryoBench, Zenodo 11629428	data/IgG-1D, data/IgG-RL	none (128 px used directly; pipeline auto-detects contrast and pose convention)
Ribosembly, Tomotwin-100	Zenodo 12528292	data/Ribosembly, data/Tomotwin-100	none
Spike-MD	Zenodo 14941494	data/Spike-MD → data/Spike-MD-128	scripts/prep_spike128.py (job 49082715); GT maps: scripts/spike_frames.py then scripts/spike_gt_vols.py → data/Spike-MD-128/vols/gt
EMPIAR-10180	EMPIAR + cryoDRGN EMPIAR inputs (github zhonge/cryodrgn_empiar → data/real/cryodrgn_empiar)	data/real/empiar10180 → data/real/empiar10180_128	see below
EMPIAR-10076	same	data/real/empiar10076 → data/real/empiar10076_128	see below
synthetic	scripts/make_synth_hier.py	data/synth_*	see Table 5
structures	RCSB (1N0X, 3AVE, 1HZH, 4YBB); AF3 (scripts/submit_af3_igg.sh, submit_af3_iggrl.sh)	data/structures/	scripts/build_template_model.py → igg_templates.pdb, iggrl_templates.pdb (+ _domains.npy); exact build args [untraced]

sbatch scripts/redownload_cryobench.sh            # IgG-1D, IgG-RL, Ribosembly, Tomotwin-100 (size check, unzip -DD)
sbatch scripts/download_ranges.sh <zenodo-url> data/<X>.zip 16   # ranged download used for the large zips (curl -sf)
sbatch scripts/download_empiar.sh 10180 data/real/empiar10180 24
sbatch scripts/download_empiar.sh 10076 data/real/empiar10076 12
sbatch scripts/redownload_empiar_meta.sh          # 10180 Example/ + Mask-and-Ref/, 10076 Parameters.star/.par
# EMPIAR-10180 (job 48615770, test partition):
python scripts/preprocess_empiar.py --star data/real/empiar10180/data/Example/consensus_data.star \
  --datadir data/real/empiar10180/data --ind data/real/cryodrgn_empiar/empiar10180/inputs/filtered.ind.pkl \
  --poses data/real/cryodrgn_empiar/empiar10180/inputs/poses.pkl --ctf data/real/cryodrgn_empiar/empiar10180/inputs/ctf.pkl \
  --D_out 128 --out data/real/empiar10180_128
# EMPIAR-10076 (job 51770970): same with --star data/real/empiar10076/data/Parameters.star, empiar10076 inputs

More inputs:

• 10076 symlink: run_hier.sh reads particles.mrcs, so EMPIAR-10076 has a symlink particles.mrcs → particles.128.mrcs.
• RECOVAR particle lists: RECOVAR reads data/<DSET>_snr0.01_abs.txt. Each is the per-state .mrcs stack list with absolute paths, as RECOVAR ≥0.5 requires. They were written by hand [untraced one-liner]: one absolute path per stack, in the order of sorted_particles.128.txt.
• Baseline inputs: baselines/inputs/<DSET>/ holds a RELION-convention stack, a validated star, particles_halves.star and relion_recon.mrc. Build with DSET=... STAGES="<stage1>:<pdb|none> ..." sbatch scripts/submit_baseline_prep.sh, which calls prep_baseline_inputs.py, write_star.py, a cryoDRGN round trip, relion_reconstruct and check_star.py. Spike (job 51092976): DSET=Spike-MD-128 PART=images/particles.128.mrcs STAGES=spike_seed:data/Spike-MD/seed_structure.pdb. CryoSTAR/cryoSPHERE configs come from scripts/make_baseline_configs.py.

3. Pipeline and dependency order

stage 1 (v0)      run_pipeline.py      consensus G0, pose convention, beads, low-rank latent z   -> results/<X>_v0
  two-pass (v0c)  run_pipeline.py --bead_map_from <X>_v0   (beads placed on cluster maps)        -> results/<X>_v0c
domains (v1)      run_v1.py --stage1 <X>_v0c                                                     -> results/<X>_v0c_v1
motion            run_hinge.py (IgG-1D) | run_rigid.py (IgG-RL, Spike, EMPIAR)                   -> <X>_v0c_hinge | _rigid
model selection   model_select.py (BIC low-rank vs domain) ; run_discrete.py --tag L1 (Level-1 discrete classes)
Level 1           level1_relate.py --stage1 <X>_v0c                                              -> level1_relate.json
Level 2           level2_blocks.py ; run_rigid.py ... --run <X>_v0c_rigid2 --domains init --merge v1
Level 3           particle_explain.py --null map --tag _mapnull[_full]                           -> particle_explain*.json/npz
decision          hier_decide.py --run <rigid run> --tag _mapnull[_full]                         -> hier_decide*.json
evaluation        submit_ext_eval.sh / submit_ext_serial.sh (AU-FSC); metrics.json (k-NN, displacement)

• Wrappers: scripts/run_hier.sh runs Level 1 → v1 → Level 2 → rigid2 → particle_explain → hier_decide end to end (NAME=... DATA=... [L1_ARGS=...] sbatch scripts/run_hier.sh) and resumes on resubmit. scripts/submit_level23.sh (S1=... V1=... IN="..." [RIGID_ARGS=...]) runs Levels 2–3 on an existing v0c/v1 pair.
• AU-FSC (official protocol): prep_official.py writes query.npy, which is arange(0, N, N//100), plus <NAME>_emb.pkl/_targets.txt. recovar reconstruct_from_external_embedding then writes results/official/<TAG>/extp_<NAME>/state###.mrc. Finally official_fsc.py (CryoBench FSC logic) writes score_ext_<NAME>.json, which holds the mean and the per-query values. Latent specs are path.h5:key, path.npy or recovar:<dir>:<zdim>. theta_all gives the (cos, sin) of every hinge angle.

  DSET=IgG-1D TAG=igg1d NAME=hinge_v0c LATENT=$ROOT/results/igg1d_v0c_hinge_test_results.h5:theta_all sbatch scripts/submit_ext_eval.sh
  sbatch --export=ALL,EVALS="DSET|TAG|NAME|LATENT DSET|TAG|NAME|LATENT ..." scripts/submit_ext_serial.sh
  

  Run evaluations serially, one GPU and one at a time: the generator uses 2–5 % of the GPU, which is why the lab rule requires serial runs. submit_ext_serial.sh skips any eval whose score_ext_*.json already exists, so a timed-out chain can simply be resubmitted (example: job 51165569). Each eval takes ~1–1.5 h, about 3 h on Ribosembly. The serial wrapper caps each at timeout 170m.

3.1 Recovered commands for the core runs

The variables in front of sbatch are taken from the args in logs/<run>.log. Job IDs come from logs/queue_*.log and sacct.

run(s)	command (job)
igg1d_v0	sbatch scripts/submit_igg1d.sh (48235447)
igg1d_v0c → _v1 → _hinge	RUN=igg1d_v0c sbatch scripts/submit_igg1d.sh --bead_map_from igg1d_v0 (48911691); RUN=igg1d_v0c_v1 STAGE1=igg1d_v0c sbatch scripts/submit_v1_igg1d.sh (48911692); RUN=igg1d_v0c_hinge STAGE1=igg1d_v0c V1=igg1d_v0c_v1 sbatch scripts/submit_hinge_igg1d.sh --n_axes 100 --rest x0 (48911693)
iggrl_v0, iggrl_v0c → _v1 → _rigid	DSET=IgG-RL PERIOD=0 RUN=iggrl_v0 sbatch scripts/submit_igg1d.sh; then RUN=iggrl_v0c ... --bead_map_from iggrl_v0 (48912675); RUN=iggrl_v0c_v1 STAGE1=iggrl_v0c DSET=IgG-RL PERIOD=0 sbatch scripts/submit_v1_igg1d.sh (48912676); RUN=iggrl_v0c_rigid STAGE1=iggrl_v0c V1=iggrl_v0c_v1 sbatch scripts/submit_rigid.sh --rest x0 --em_rounds 4 --flip_pool 50 --fast_coarse --c2f 0:25:32,14:12:8,7:8:2,3.5:6:1 (48912677)
ribo_v0, ribo_v0c → _v1 → _rigid	ribo_v0: run_pipeline.py ... --n_beads 2000 --K 8 --epochs_homo 3 --epochs 30 --n_clusters 16 (see log); ribo_v0c = same --bead_map_from ribo_v0 --labels ... --gt_vols data/Ribosembly/vols/128_org (48931595, --wrap); run_v1.py --stage1 ribo_v0c --run ribo_v0c_v1 --Ke 4 --epochs 20 (48931632); run_rigid.py --stage1 ribo_v0c --v1 ribo_v0c_v1 --run ribo_v0c_rigid --rest x0 ... (48931635; full args in log)
tomo_v0, tomo_v0c → _v1 → _rigid	DSET=Tomotwin-100 PERIOD=0 RUN=tomo_v0 sbatch scripts/submit_igg1d.sh (48943326); RUN=tomo_v0c ... --bead_map_from tomo_v0 (48943329); v1 (48943337); submit_rigid.sh --rest x0 --em_rounds 1 --sweeps 1 --fast_coarse --pop_prior 50 --c2f ... (48943338)
spike_v0 → spike_v0c → _v1 → _rigid	--wrap jobs 49083864/65/67/71: run_pipeline.py --particles data/Spike-MD-128/images/particles.128.mrcs ... --run spike_v0 --n_beads 2000 --K 8 --epochs_homo 3 --epochs 30 --n_clusters 20; then v0c (--bead_map_from spike_v0), run_v1.py --Ke 4 --epochs 20, run_rigid.py --rest x0 --em_rounds 4 --flip_pool 50 ...
model selection	`model_select.py --stage1 _v0c --domain_run _v0c_hinge
Level-1 discrete (Ribosembly, Tomotwin)	run_discrete.py --stage1 ribo_v0 ... --n_sub 20000 --Ds 64 --Ks 5,10,16,25 --n_iter 6 --domain_run ribo_v0c_rigid --domain_kind rigid --domain_stage1 ribo_v0c --apply_all --tag L1 (48974832) → results/ribo_v0/L1_posterior_K16.npy; Tomotwin the same with --stage1 tomo_v0 --Ks 75,100,125 --domain_run tomo_v0c_rigid (48974831) → results/tomo_v0/L1_posterior_K100.npy; Spike: run_discrete.py --stage1 spike_v0 ... --Ks 1,10,50,100,200 --domain_run spike_v0c_rigid (49083885) → low-rank selected (results/spike_v0/L1.json)

4. Display items

Fig. 1 — Overview (figures/fig1_overview.pdf = results/figures/overview.pdf)

• Inputs: icon renders results/figures/ribo/test.png, made with bash scripts/chimerax_offscreen.sh scripts/diag/cx_test.cxc (job 51430342). It also uses ribo/class05,12,06.png (Fig. 4 render) and morph/igg1d_still0,2.png (Fig. 3b).
• Command: python scripts/fig_overview.py --out results/figures/overview (job 51505455, test partition).

Fig. 2 — AU-FSC benchmark (fig2_benchmark.pdf = results/figures/benchmark.pdf)

• Reads the numbers from tab:bench in main.tex, so it always matches Extended Data Table 1. Regenerate the table first (ED Table 1 below).
• python scripts/fig_benchmark.py --tex ../cryodynascape_overleaf/main.tex --out results/figures/benchmark (job 51405401).

Fig. 3 — Recovered motion (fig3_motion.pdf = results/figures/fig3.pdf)

panel	run	data	command
a	igg1d_v0c_hinge	results/igg1d_v0c_hinge_test_results.h5 (theta_mean_dom1, labels)	composed in fig3_compose.py
b	igg1d_v0c_hinge	results/igg1d_v0c_hinge/hinge1_traj_ca.pdb	python scripts/fig_morph.py --traj results/igg1d_v0c_hinge/hinge1_traj_ca.pdb --h5 results/igg1d_v0c_hinge_test_results.h5 --out results/figures/morph/igg1d && bash scripts/chimerax_offscreen.sh results/figures/morph/igg1d.cxc (job 51458988; the still/strip post-processing inline python -c in that job is [untraced], truncated in sacct)
c	iggrl_rigid_restx0 (oracle start, rest re-estimated)	results/iggrl_rigid_restx0/motion_eval.npz	python scripts/fig_state_overlay.py --run results/iggrl_rigid_restx0 --out results/figures/iggrl_states && bash scripts/chimerax_offscreen.sh results/figures/iggrl_states/render.cxc (job 51602507)
d	spike_v0 vs all Spike evals	results/official/spike/score_ext_*.json (per_query)	python scripts/fig_spike_paired.py --out results/figures/spike_paired (job 51622677)
Compose: python scripts/fig3_compose.py --out results/figures/fig3 (job 51920932).

Fig. 4 — Ribosembly containment graph (fig4_ribosembly.pdf = results/figures/ribo/ribo_graph.pdf)

• Run: ribo_v0c with Level 1 from level1_relate.py --stage1 ribo_v0c ... --kgrid 1,2,3,5,8,10,16,20,30 (scripts/run_level1_all.sh, job 49858198). Outputs: results/ribo_v0c/level1_relate.json, level1r_class00–15.mrc.
• Commands (job 51430859, then 52017444):

  python scripts/fig_ribo.py render  --run results/ribo_v0c --out results/figures/ribo
  bash scripts/chimerax_offscreen.sh results/figures/ribo/render.cxc
  python scripts/fig_ribo.py compose --run results/ribo_v0c --out results/figures/ribo
  

Fig. 5 — EMPIAR-10180 (fig6_e10180.pdf = results/figures/fig6.pdf; file name keeps the old number)

Runs (all on data/real/empiar10180_128, 139,722 particles):

1. DATA=data/real/empiar10180_128 RUN=e10180_v0 sbatch scripts/submit_real.sh (48634518; 2000 beads, 30 epochs).
2. DATA=... STAGE=v1 STAGE1=e10180_v0 RUN=e10180_v1 sbatch scripts/submit_real_stage.sh (48634888).
3. DATA=... STAGE=rigid STAGE1=e10180_v0 V1=e10180_v1 RUN=e10180_rigid_gate sbatch scripts/submit_real_stage.sh --em_rounds 1 --sweeps 1 --fast_coarse --bic_gate (49367475).
4. Map-null labels on all particles, one job per domain j = 1..4 (49672084–087): particle_explain.py --run e10180_rigid_gate --stage1 e10180_v0 --particles data/real/empiar10180_128/particles.128.mrcs --poses .../poses.pkl --ctf .../ctf.pkl --n 0 --domain j --null map --tag _mapnull_full_dj
5. Decision: hier_decide.py --run e10180_rigid_gate --tag _mapnull_full → hier_decide_mapnull_full.json. The exact invocation is [inferred] from the usage string; no recorded job.
6. Validation against RELION (49695252): validate_labels.py --run e10180_rigid_gate --stage1 e10180_v0 <inputs> --label_tag _mapnull_full_d --rotvec_tag _mapnull_full_d --out_tag _mapnull. It needs RELION multi-body results/relion_multibody/ (sbatch scripts/submit_multibody.sh, 48885317, 2 H100, 4 h 57 min) and the authors' masks data/real/empiar10180/data/Mask-and-Ref.
7. Motion ends (51541669): motion_bins.py --run e10180_rigid_gate --stage1 e10180_v0 <inputs> --domains 1 3 → results/e10180_rigid_gate/motion_bins/. Quantify (51590700): fig_e10180_bins.py --bins results/e10180_rigid_gate/motion_bins --out results/figures/e10180/bins.

Render and compose:

• Render: bash scripts/fig_e10180_render.sh (writes results/figures/e10180/render.cxc, identical to the one used in job 51563165, and runs it with scripts/chimerax_offscreen.sh).
• Compose: python scripts/fig6_compose.py --out results/figures/fig6 (job 51653150).

Text-only analyses (no figure panel):

• e10180_latent_agreement.py --out results/e10180_rigid_gate/latent_agreement.json
• e10180_view_leak.py

Both need results/cryodrgn_e10180 and results/recovar_e10180 (Section 5, caveat 3).

Fig. 6 — EMPIAR-10076 (fig7_e10076.pdf = results/figures/e10076/e10076_panel.pdf)

• Stage 1 (51770971): DATA=data/real/empiar10076_128 RUN=e10076_v0 sbatch scripts/submit_real.sh.
• Hierarchy (51780442): NAME=e10076 DATA=data/real/empiar10076_128 L1_ARGS="--kgrid 1,2,3,5,8,10,16,20,30" sbatch scripts/run_hier.sh. L1_ARGS was set exactly as shown (recorded in the project log; the BIC keys in results/e10076_v0/level1_relate.json agree). This produces e10076_v1, e10076_rigid2 and the 20k-particle particle_explain.json that the figure renders.
• Full labels (51849317): particle_explain.py --run e10076_rigid2 --stage1 e10076_v0 --particles data/real/empiar10076_128/particles.mrcs ... --n 0 --null map --tag _mapnull_full.
• Then (51860267): hier_decide.py --run e10076_rigid2 --tag _mapnull_full and occupancy_check.py --run e10076_rigid2 --stage1 e10076_v0 <inputs> --domain 2 --tag _mapnull_full. The 20k version, which the figure uses, is the same command without --tag (51809938).
• Block → 50S chains: bash scripts/chimerax_offscreen.sh scripts/cx_fit_50s.py, then python scripts/e10076_block_map.py --write_beads && bash scripts/chimerax_offscreen.sh scripts/cx_fit_50s_beads.py (51850681, 51851250). Needs data/structures/4ybb.cif.
• Figure (51823516): python scripts/fig_e10076.py render --out results/figures/e10076 && bash scripts/chimerax_offscreen.sh results/figures/e10076/render.cxc && python scripts/fig_e10076.py compose --out results/figures/e10076.

Extended Data Fig. 1 — Calibration contrast (results/figures/calibration_contrast.{json,pdf})

• Data: per-domain labelled fractions under three calibrations:
  • BIC penalty: particle_explain_full_d*.json
  • bead-rendered null: particle_explain*.json without _mapnull
  • consensus-map null: particle_explain_mapnull*.json
• Runs behind it:
  • synthetic species: synth_hier_class{0,1}_rigid2 (49665450/464)
  • EMPIAR-10180: e10180_rigid_gate; bead null 49424308–315, map null 49672084–087
• Regenerate: python scripts/fig_calibration.py rebuilds calibration_contrast.{json,pdf,png} from the particle_explain JSON files above (checked 2026-10-12: identical JSON to the original, max difference 0).

Extended Data Table 1 — AU-FSC (tab:bench)

Published values come from CryoBench. Cells marked † were evaluated here, each from results/official/<tag>/score_ext_<NAME>.json (AU-FSC_mean):

row	IgG-1D (igg1d)	IgG-RL (iggrl)	Ribosembly (ribo)	Tomotwin (tomo)	Spike (spike)
CDS (official)	hinge_v0c ← igg1d_v0c_hinge:theta_all	rigid_v0c ← iggrl_v0c_rigid:features_pca	L1_discrete ← results/ribo_v0/L1_posterior_K16.npy	L1_discrete ← results/tomo_v0/L1_posterior_K100.npy	lowrank_v0 ← spike_v0_test_results.h5:z_pca
RECOVAR†	RECOVAR ← recovar:results/recovar_igg1d:10	same, recovar_iggrl	recovar_ribo	recovar_tomo (0.201; table prints published 0.258)	recovar_spike
e2gmm†	e2gmm ← baselines/runs/e2gmm_<DSET>/latent.npy (all five)
DynaMight†	dynamight ← baselines/runs/dynamight_<DSET>/latent.npy (via submit_dynamight_eval.sh)		n.c.	n.c.	DIAM 300
CryoSTAR†	cryostar_tpl ← baselines/cryostar/projects/star/work_dirs/<DSET>_<s1>_tpl_0/0061_0096000/z.npy		—	—	cryostar_seed
cryoSPHERE†	cryosphere_tpl ← baselines/runs/cryosphere_<DSET>_<s1>_tpl/analysis/z.npy		—	—	cryosphere_seed (resumed, Section 5)
CDS (structure)	hinge_tpl ← igg1d_tpl_hinge:theta_all	rigid_tpl ← iggrl_tpl_rigid_dinit:features_pca	—	—	rigid_seed ← spike_seed_rigid:features_pca

The latent specs above are recorded in sacct (--export / EVALS) for these evaluations: e2gmm IgG-RL/Tomotwin/ Ribosembly/Spike, CryoSTAR Spike, rigid_seed, the AF3 rows, random_latent and cryoSPHERE/DynaMight (wrappers). For the evaluations of 2026-09-25 to 09-28 (48913202–49202654) LATENT was exported in the shell and not recorded. The exact source of every official embedding (53 of 53) was recovered by matching each saved results/official/<tag>/<name>_emb.pkl against all candidate latents (python scripts/trace_latents.py → results/official/latent_sources.json); use the listed source as LATENT (for hinge runs the key theta_all).

Baseline runs:

• RECOVAR: DSET=<DSET> TAG=<tag> sbatch scripts/submit_recovar.sh (zdim 1,2,4,10; --mem=300G for Ribosembly). Jobs: IgG-1D 48256679 (submit_recovar_igg1d.sh), IgG-RL 48312921, Ribosembly 48344994, Tomotwin 48944700, Spike 49197320.
• cryoDRGN: DSET=... PART=... sbatch scripts/submit_cryodrgn.sh (zdim 8, 20 epochs), then DSET=... TAG=... PERIOD=100|0 sbatch scripts/submit_eval_cryodrgn.sh → score_drgn_own.json (own decoder; used in S1). Published numbers are shown in the table.
• CryoSTAR: NAME=<DSET>_<s1>_tpl sbatch scripts/submit_cryostar.sh (Spike: NAME=Spike-MD-128_spike_seed, 51104007).
• cryoSPHERE: NAME=... sbatch [--partition=kempner_h200] scripts/submit_cryosphere.sh (IgG-1D 48888020, IgG-RL 48942834, Spike 51104008), then NAME=... DSET=... TAG=... PERIOD=... EPOCH=299 [EXTNAME=cryosphere_seed] sbatch scripts/submit_cryosphere_eval.sh.
• DynaMight: DSET=... DIAM=250 sbatch scripts/submit_dynamight.sh (Spike DIAM=300, 51104009), then DSET=... TAG=... DIAM=... PERIOD=... sbatch scripts/submit_dynamight_eval.sh.
• e2gmm:
  • Conversion on CPU: run_e2gmm.py --dset <DSET> --apix 3.0|4.5 --box 128 --prep_only --chunks 8 (49934015–020, test partition).
  • Training: DSET=... APIX=... sbatch scripts/submit_e2gmm.sh. Ribosembly ran fully on CPU, test partition, 180 G (50057962).
• HEMNMA: not run (scripts/hemnma_scipion.py, scripts/scipion_compat/; incompatibility in the caption).
• Spike random-latent floor and map regeneration: sbatch scripts/submit_spike_extra.sh (51541332) → score_ext_random_latent.json and score_ext_lowrank_v0_maps.json, which must equal 0.4258.

Extended Data Table 2 — State recovery / displacement (tab:motioncmp)

row	source file(s)
RECOVAR	results/recovar_<tag>/cds_metrics.json from `eval_recovar.py --recovar_out results/recovar_ --labels ... --poses ... --label_period 100
cryoDRGN	results/cryodrgn_<DSET>/latent_metrics.json (submit_eval_cryodrgn.sh)
CryoSTAR	baselines/cryostar/projects/star/work_dirs/<DSET>_<s1>_tpl_0/latent_metrics.json (producer [untraced]); displacement results/baseline_disp_cryostar_{igg1d,iggrl}.json (50984766)
cryoSPHERE	baselines/runs/cryosphere_*/latent_metrics.json; displacement results/baseline_disp_cryosphere_*.json (50984767)
DynaMight, e2gmm	baselines/runs/{dynamight,e2gmm}_<DSET>/latent_metrics.json
CDS no structure	results/igg1d_v0c_hinge/metrics.json (gt_knn_label_err), iggrl_v0c_rigid/metrics.json (gt_knn_purity), ribo_v0c metrics
CDS GT start	results/igg1d_hinge_restx0/metrics.json, iggrl_rigid_restx0/metrics.json (motion_disp_rmsd_A)

Displacement command, per method and dataset: <env>/python scripts/baseline_displacement.py --method cryostar|cryosphere --work_dir <run dir> --stage1 <s1>_tpl --gt_pdbs data/<DSET>/pdbs --labels data/<DSET>/gt_latents.pkl --out results/baseline_disp_<method>_<s1>.json.

Extended Data Table 3 — Starting structure (tab:structure)

Metrics come from results/<run>/metrics.json; AU-FSC from results/official/{igg1d,iggrl}/score_ext_<NAME>.json.

row	IgG-1D run → eval NAME	IgG-RL run → eval NAME
no structure, consensus beads	igg1d_ns_v1→igg1d_ns_hinge (from igg1d_v0) → hinge_ns	iggrl_v0→iggrl_ns_v1→iggrl_ns_rigid (48830518/21/25) → rigid_ns_consensus
no structure, cluster-map beads	igg1d_v0c_hinge → hinge_v0c	iggrl_v0c_rigid → rigid_v0c
AF3 whole	igg1d_af3→_v1→igg1d_af3_hinge → hinge_af3 (50059423)	VARIANT=whole sbatch scripts/submit_iggrl_af3.sh (50977590) → rigid_af3
AF3 PAE domains	igg1d_af3hier→_hinge (--init_pae ... --pae_hierarchical, see log) → hinge_af3hier (50059425)	VARIANT=hier ... (50977591, timed out in round 4), resumed by run_rigid.py ... --run iggrl_af3hier_rigid --resume_from iggrl_af3hier_rigid --extra_rounds 1 (51092974) → rigid_af3hier
PDB 1N0X+3AVE	igg1d_tpl (--init_model data/structures/igg_templates.pdb --init_domains ..._domains.npy) → igg1d_tpl_hinge → hinge_tpl	iggrl_tpl→iggrl_tpl_rigid_dinit → rigid_tpl
oracle, rest re-estimated	igg1d_hinge_restx0 (--rest x0, from igg1d_ca) → hinge_nostruct	iggrl_rigid_restx0 → rigid_nostruct
oracle (1HZH)	igg1d_hinge (submit_hinge_igg1d.sh --n_axes 100, 48274787) → hinge_pdb	iggrl_rigid_em4_flip → rigid_pdb

Notes:

• The igg1d_ca/iggrl_ca stage-1 runs used submit_igg1d.sh --init_pdb data/<DSET>/pdbs/000.pdb --gt_pdbs data/<DSET>/pdbs (48262209, 48286174).
• AF3 inputs: submit_af3_igg.sh (48370160) and submit_af3_iggrl.sh (48831257) → data/structures/af3*/.

Extended Data Table 4 — Ablations (tab:ablation)

(a) Oracle rest shape; values from results/<run>/metrics.json:

dataset	run per row
IgG-1D	igg1d_hinge; igg1d_ca; igg1d_v1; igg1d_hinge_v12a (pivot fixed at the interface; mapping [inferred] from metrics); igg1d_hinge_soft
IgG-RL	iggrl_rigid_em4_flip; iggrl_rigid_attempt1 (centroid); iggrl_rigid_grid; iggrl_rigid (keep 8); iggrl_rigid_c2f32 (keep 32); iggrl_rigid_flip (naive); iggrl_rigid_flip2 (aligned); iggrl_rigid_c2f32_em2; iggrl_rigid_c2f32_fast_prior; iggrl_rigid_c2f32_fast
Ribosembly	ribo_v0; ribo_comp; ribo_icm16

Exact args for each run are in logs/<run>.log.

(b) IgG-RL, official setting.

• Six rigid runs abl_{no_flip,grid_only,centroid_pivot,no_merge,em1,with_prior}, each followed by its AU-FSC eval (logs/queue_ablation.log). Each is RUN=abl_<x> STAGE1=iggrl_v0c V1=iggrl_v0c_v1 sbatch scripts/submit_rigid.sh --rest x0 --fast_coarse --c2f 0:25:32,14:12:8,7:8:2,3.5:6:1 --em_rounds 4 --flip_pool 50 <change>, where <change> is one of:
  • --flip_pool 0 (48987090)
  • --search grid (48988015)
  • --pivot centroid (48992586)
  • --merge v1 (48998119)
  • --em_rounds 1 (48998125)
  • --pop_prior 50 (49011449)
• Evals: NAME=abl_<x> LATENT=results/abl_<x>_test_results.h5:features_pca.
• "consensus beads" = rigid_ns_consensus. "low-rank only" = score_ext_lowrank_ns.json, latent results/iggrl_v0_test_results.h5:z_pca (IgG-1D: results/igg1d_v0_test_results.h5:z_pca; from latent_sources.json).

Extended Data Table 5 — Hierarchical selection (tab:hier)

(a) Level 1:

• Synthetic variants: make_synth_hier.py --out data/<v> --N 50000 [--frac_b 0 | --snr 0.01 | --snr 0.005 | --frac_b 0.1 | --b_shape globule] (--N 20000 for synth_n20k). Then run_pipeline.py ... --run <v>_v0 --n_beads 1000 --K 8 --epochs_homo 3 --epochs 40 --batch 256 --n_clusters 10 and level1_relate.py --stage1 <v>_v0 ... --species_gt data/<v>/gt.npz. Jobs: 49424982, 49490302/324, 49543147–151.
• All variants plus CryoBench in one job: sbatch scripts/run_level1_all.sh (49858198) → results/*_v0[c]/level1_relate.json.
• Tomotwin discrete row: level1_relate.py --stage1 tomo_v0 ... --posterior results/tomo_v0/L1_posterior_K100.npy --min_class 300 (49918976).
• Synthetic summary: python scripts/table_hier.py --tag _mapnull → results/table_hier_mapnull.json.

(b) Levels 2–3, files results/<run>/hier_decide_mapnull.json:

• Synthetic: NAME=synth_hier DATA=data/synth_hier sbatch scripts/run_hier.sh (49556003). Then map-null re-labels of synth_hier_class{0,1}_rigid2 (49665450, 49665464). The --wrap commands are in Section 3, pattern particle_explain.py ... --n 20000 --null map --tag _mapnull && hier_decide.py --tag _mapnull.
• CryoBench: submit_level23.sh with S1=<X>_v0c V1=<X>_v0c_v1 IN="<inputs>". Jobs: IgG-1D 49918979, IgG-RL 49806450, Spike 49842278, Ribosembly 49842245 (RIGID_TO=230m RIGID_ARGS="--em_rounds 1 --sweeps 1").
• Ribosembly second round: run_rigid.py --stage1 ribo_v0c --v1 ribo_v0c_v1 --run ribo_v0c_rigid2_r2 ... --resume_from ... --extra_rounds 1 (49857979; full args in logs/ribo_v0c_rigid2_r2.log), then particle_explain + hier_decide on _r2 (49874342).

Extended Data Table 6 — EMPIAR-10180 labels (tab:real)

• Labels and fractions: results/e10180_rigid_gate/particle_explain_mapnull_full_d{1..4}.json, and the decision in hier_decide_mapnull_full.json (Fig. 5 steps 4–5).
• RELION columns: validate_labels_mapnull.json (step 6).
• Body overlap: scripts/diag/label_domains_by_masks.py; its arguments are [untraced].

Extended Data Table 7 — Runtime (tab:runtime)

• CDS component times: results/SUMMARY.md § "Runtime", taken from sacct Elapsed of the jobs in Section 3.1.
• Baselines (sacct Elapsed):
  • RECOVAR 48256679 / 48312921 / 48344994; EMPIAR-10180 50977593 (18:09)
  • cryoDRGN 48832407 / 48860117 / 48865079; EMPIAR-10180 50084984 (45:43)
  • RELION multi-body 48885317

Supplementary Table S1 — Per-query paired tests (tab:supp_paired)

• Per-query AU-FSC of directories that predate score_ext per-query storage: python scripts/paired_test.py (49096193, test partition) → results/official/<tag>/pq_*.json and paired_tests.json. It re-scores extp_*/state###.mrc and drgn_own/vol_*.mrc.
• Table: python scripts/supp_table_paired.py --out results/official/supp_paired.tex (51875037) → supp_paired.{tex,json}, pasted into main.tex. The reference is CDS: hinge_v0c, rigid_v0c, L1_discrete ×2, lowrank_v0.

5. Known caveats

1. cryoSPHERE on Spike-MD was resumed. The run (51104008) hit the 16 h limit at epoch ~181/300. It was continued with NAME=Spike-MD-128_spike_seed sbatch --dependency=afterany:51104008 scripts/submit_cryosphere_resume.sh (51239687). That job uses scripts/cryosphere_resume.py: same epoch counter and loss schedules, Adam moments re-initialised. Evaluation: submit_cryosphere_eval.sh with EXTNAME=cryosphere_seed (51239688). An uninterrupted 300-epoch run will differ slightly.
2. Ribosembly used one EM round. Its Level 2–3 run (ribo_v0c_rigid2) used --em_rounds 1 --sweeps 1; every other dataset used 2. A resumed second round is reported as a range in tab:hier. Per-particle rotations had not converged; the decision was unchanged.
3. EMPIAR-10180 has three non-finite images (indices 9955, 40867, 108052). CDS zeroes them in its loader. The baselines must exclude them with --ind data/real/empiar10180_128/good_ind.pkl (139,719 particles). Without this flag RECOVAR and cryoDRGN produce NaN (results/*_e10180.failed_nan). RECOVAR needs --mem=160G. EMPIAR-10076 has one corrupt image (7234), which is zeroed.
4. Netscratch purge. On 2026-10-08 the purge emptied data/{IgG-1D,IgG-RL,Ribosembly,Tomotwin-100}, the ChimeraX install and EMPIAR metadata, because extracted files keep old archive mtimes. Restore them with redownload_cryobench.sh (unzip -DD), download_ranges.sh (curl -sf; a 107-byte HTTP error body once corrupted a Tomotwin part) and redownload_empiar_meta.sh.
   • Not restored: the EMPIAR-10180 raw micrograph stacks. data/real/empiar10180_128 is intact and on HF.
   • Deleted on purpose and regenerable: half/unfiltered generator maps, the zips, Spike 256-px images, and intermediate cryoSPHERE/DynaMight checkpoints. See notes/DELETED_MANIFEST.md.
   • Never delete anything in notes/FIGURE_DATA.md.
5. ChimeraX offscreen needs the cds_render env, which provides OSMesa (mesalib<24.1), and the symlinks in .../software/osmesa_link. Always call it through scripts/chimerax_offscreen.sh. Other requirements:
   • .cxc paths must be absolute.
   • Do not pipe ChimeraX output into head.
   • A map written from an (x, y, z) array is read by ChimeraX as (z, y, x).
6. Spike-MD AU-FSC is evaluated here at 128 px. The GT maps come from the query frames via CryoBench's ChimeraX recipe, so the values are not comparable to published Spike-MD numbers. The floor is the random latent at 0.380.
7. Post-hoc thresholds (disclosed in the paper):
   • The Level-1 link rule uses corr ≥ 0.7, stable over 0.60–0.80.
   • hier_decide --mode_ratio 0.35 was chosen after an absolute threshold mislabelled Ribosembly.
   • The defaults in the scripts are the values used in the paper.
8. No fixed seeds. GPU fits are not bit-reproducible. Rerunning AU-FSC from the stored _emb.pkl/_targets.txt reproduces scores to about 1e-7 (Spike check: 0.425817 vs 0.425817).