xkronosx/cryoem-acquisition

Cryo-EM Autonomous Acquisition Dataset

Data for training agents that decide which grid square and foil hole to image next during a cryo-EM session, so a sample reaches its target in the fewest acquisitions.

Four EPU sessions derived from public EMPIAR depositions, with per-acquisition outcome labels, pre-acquisition hole features, frozen image embeddings, and orientation-proxy profiles.

candidate holes	27,944
acquired holes	3,936
acquisitions (label rows)	13,526 (10,595 usable)
grid squares	1,795
sessions	4
size	~137 MB

Quick start

pip install huggingface_hub pandas pyarrow numpy
huggingface-cli download xkronosx/cryoem-acquisition --repo-type dataset --local-dir cryoem-acquisition
cd cryoem-acquisition

python -m cryoem_au_data.validate index/combined   # check the contract
python example.py                                  # load, features, labels

Layout

index/
  combined/                       all four sessions concatenated  ← start here
  empiar-11895__epu__grid1/       individual sessions
  empiar-13296__epu__delta9/
  empiar-13296__epu__delta10/
  empiar-13296__epu__deltaRlmE/
  empiar-11895__epu__grid1_uncensored/    censoring-ablation variant
features/
  combined/embeddings_cryo_ief.parquet    8 PCA columns, keyed by hole_uid
  <session>/raw_cryo_ief.parquet          raw 768-d Cryo-IEF vectors
  bases/                                  frozen PCA basis for new sessions
orientation/
  orientation_class2d.parquet             2D-class view proxy (EMPIAR-11895)
cryoem_au_data/                           loader, feature builder, validator
example.py

Each index directory holds four parquet tables:

table	one row per	what it is
target_locations.parquet	candidate hole	every hole EPU's hole-finder proposed — the agent's decision space
holes.parquet	acquired hole	the subset actually imaged
square_geometry.parquet	grid square	shape and extent, for geometric features
labels_v4_outcomes.parquet	acquisition	the outcomes — a hole shot 3× has 3 rows

The contract

UIDs. session:<name>:square:<id>:hole:<id>. The :square: segment is load-bearing: cross-session validation splits on the substring before it, and without it leave-one-session-out silently degrades into leave-one-hole-out.

sample_id is the micrograph stem and the join key to labels, orientation profiles and image features.

Pre- vs post-exposure. target_locations holds only what is knowable before firing; labels_v4_outcomes holds what came back. selected (did a human take this hole) is a legitimate training target but never a model input — in deployment the agent is the selector.

Utility:

w_ctf   = sigmoid((6.0 - ctf_res_adj) / 1.5)
w_ice   = 1.0 / (1.0 + max(0, ice_ring - 1.5))     # missing ice => no penalty
utility = curated_count * w_ctf * w_ice            # 0 where status != "ok"

ctf_res_adj is ctf_res with the defocus trend removed (level preserved). Those three constants were tuned on EMPIAR-11895 and do not transfer — on other sessions the CTF gate saturates or the ice term penalises everything. Re-tuning is a pure relabel; compute_utility() takes them as arguments.

Censoring. Two statuses mark failures that must not be read as zeros:

• ctf_unfit — CTFFIND reports its search limit when a fit fails (a sentinel, not a measurement). Those micrographs still yield ~28 real particles.
• curation_missing — absent from curation is not zero particles.

Known limitations

Stated plainly, because they bound what can be concluded:

• Timestamps exist for only one session (deltaRlmE). The others fall back to a sample_id sort — usable but not true chronological order.
• Label quality varies sharply. Only EMPIAR-11895 has genuine 2D-curated particle counts (count_source == "class2d_curated"); the rest use LoG picks, which carry roughly 7× less learnable signal. deltaRlmE has a between-hole to within-hole variance ratio of ~1.07 — its holes are barely separable from shot noise.
• Transfer between sessions is weak and sometimes negative. Cross-session rank correlation of a feature→utility model is ≈0.04, and training on the 13296 family and testing on 11895 gives ≈−0.30. Treat this as one session plus a family of three, not four exchangeable domains.
• Orientation labels are a 2D-class proxy, available for EMPIAR-11895 only. Class assignment conflates viewing direction with conformation, so coverage conclusions are bounds, not measurements. Real Euler angles from a consensus 3-D refinement would replace it.

Provenance and licence

Derived from public EMPIAR depositions — EMPIAR-11895 (apoferritin; ships a complete RELION project) and EMPIAR-13296 (70S ribosome, three strains). Cite the original depositions in any work using this data. Derived tables released CC-BY-4.0. No raw movies or micrographs are included — this is metadata and labels only.

Building a compatible dataset

If you are producing new EPU / cryoSPARC Live data for these agents, run the validator against your index before handing it over:

python -m cryoem_au_data.validate <your_index_dir>

It checks UID grammar, join integrity, the candidate-vs-acquired ratio, the status vocabulary, timestamp coverage, and whether the utility constants still have dynamic range on your data. Every check corresponds to a failure mode that breaks training silently rather than loudly.