Datasets:

markfaa
/

CryoAtlas

Supporting-file paths below are relative to the repository’s resources/ directory. Values in archive_path and entry_directory retain their documented repository-root and extraction-root meanings.

Entry catalog

metadata/catalog/train.parquet and metadata/catalog/test.parquet contain one row per EMD entry. metadata/catalog/benchmark_targets.parquet is the 618-row subset of the test catalog.

Field	Meaning
`emd_id`, `pdb_id`	EMDB entry and the PDB identifier of the structure provided in the data archive
`split`	Fixed `train` or `test` membership
`benchmark_618`	Whether the entry belongs to the common 618-target benchmark
`resolution_angstrom`	Reported map resolution from `metadata/source_accessions.csv`, in Å
`title`	Deposited description from the source-accession table
`release_component`	Training download part or one of the two test archives
`archive_path`, `download_url`	Repository-relative path and direct URL of the complete containing archive
`entry_directory`	Entry directory inside the extracted training/test release
`has_aligned_cif`	Availability recorded by the released `metadata/Align_cif.json` list
`emdb_url`, `pdb_url`, `source_map_url`	Public accession pages and original map location
`map_release_date`, `structure_determination_method`, `sample_type`	Fields preserved from the source-accession table

Membership is checked against the fixed split lists. Training archive assignments come from the frozen release plan; test assignments are checked against both test-subset EMD lists. Catalog rows contain no server paths, credentials, or upload logs.

Benchmark result table

benchmark/results.parquet has one row per EMD ID × method: 618 targets and five methods, totaling 3,090 rows. It combines the five original CSVs under benchmark/results/. Original CSV values and bytes are preserved.

method identifies ModelAngelo, CryoAtom2, E3-CryoFold, EMProt or AF3.
emd_id, pdb_id, category, total_residues, resolution and n_chains retain the released target descriptors.
Metric names lose only their method prefix: for example, ModelAngelo_gt_score becomes gt_score in a row whose method is ModelAngelo.
tm_score, ca_rmsd, residue coverage, sequence metrics and gt_score concern structure-to-reference agreement.
completeness is the predicted-to-reference model-length ratio. It can exceed 1 and is not a bounded accuracy score.
cc_*, qscore, recall_qscore and map_score describe the released map-agreement measurements.
Ramachandran, rotamer, Cβ, clash, bond/angle, MolProbity and geom_* columns contain the released stereochemical measurements and composite components.
train_pident and train_bin retain the identity annotations shipped in each method CSV. AF3 has no such columns in its source CSV, so the unified table leaves them null. Nulls are not zero measurements.

No scores are recomputed or rescaled. Raw metrics use different units and directions; the table is a convenient view of the released evaluation outputs, not a new scoring protocol. When reproducing a published aggregate, use the paper's specified subset and aggregation rules.

Archive index and checksums

metadata/catalog/archives.json and metadata/catalog/archives.parquet list the same 71 files. bytes is the exact compressed size; size_gb uses decimal GB (10⁹ bytes). sha256 hashes the complete compressed file. entries counts the map–model–sequence entries in a training/test archive; it is null for the mixed benchmark resource package. metadata/checksums/SHA256SUMS gives the same hashes with repository-relative paths.

Density previews

benchmark/examples.parquet contains an HF Image feature with embedded PNG bytes, plus example metadata and archive links. The ten maps are the existing examples in the benchmark resource package: nine benchmark-test entries and the training entry EMD-0719.

Each image displays maximum-intensity projections along array axes 0, 1 and 2 of the stored MRC volume. Each projection is independently displayed between its 2nd and 99.5th percentiles, using a common color map. The axes refer to array storage order, not a standardized physical orientation. These display transformations do not alter the source maps and do not establish map–model alignment or quantify map quality.

The overview figure counts entries in the fixed split lists and displays normalized histograms of reported resolutions from the source-accession table. All 17,809 resolutions are present. The plots and previews are derived browsing aids; the scientific archive files are unchanged.

ModelAngelo training utility

benchmark/training_utility/MA_CryoAtlas.csv and MA_Cryo2StructData.csv preserve the original Kaggle-released evaluation tables byte for byte. Each contains 618 target rows and one AVERAGE row. Both target sets equal the fixed 618-target benchmark.

benchmark/training_utility/results.parquet contains only the 1,236 target rows. The checkpoint column identifies MA_CryoAtlas or MA_Cryo2StructData; the cryo2struct_ prefix is removed from the latter condition's metric names to align columns. Values are neither recomputed nor rescaled. The original CSVs retain their original columns and aggregate rows.

The corresponding predicted CIFs are in the resource archive's training_utility/MA_CryoAtlas/ and training_utility/MA_Cryo2StructData/ folders, 618 per condition. They were generated by the paper's retrained CNN checkpoints and are not serialized checkpoint weights. The official ModelAngelo reference remains part of the five-method benchmark.