resources/docs/dataset_guide.md

# CryoAtlas dataset guide
### A Large Curated Dataset and Unified Benchmark for Cryo-EM Atomic Model Building

**17,809 entries · 16,657 training · 1,152 test · 618 benchmark targets**

[Project website](https://oulling.github.io/cryoatlas/) · [3D Benchmark Explorer](https://oulling.github.io/cryoatlas/explore/) · [Browse entries](https://huggingface.co/datasets/markfaa/CryoAtlas/viewer/catalog/train) · [Compare results](https://huggingface.co/datasets/markfaa/CryoAtlas/viewer/benchmark_results/test) · [Preview examples](https://huggingface.co/datasets/markfaa/CryoAtlas/viewer/examples/examples) · [Download guide](downloads.md)

CryoAtlas v1.0 pairs experimental cryo-EM density maps with atomic structures and protein sequences, using EMDB and PDB records available up to March 2026. Entries undergo map–model alignment, sequence mapping, and quality filtering to support research on automated protein atomic model building.

The **complete release is available here**: four training parts, the full test set, and benchmark resources. The 70 training and test archives total **2.415 TB** (decimal units); their sizes and SHA-256 hashes have been verified against the frozen v1.0 release.

![CryoAtlas split counts and reported map-resolution distributions](assets/dataset_overview.png)

## Browse before downloading

The Dataset Viewer above displays lightweight tables and image previews. Select a configuration from its dropdown:

| Configuration | Contents | Useful for |
|---|---|---|
| **`catalog`** — default | 16,657 training and 1,152 test rows | Search EMD/PDB IDs, resolution and titles; find the archive containing an entry |
| **`benchmark_targets`** | The 618 common benchmark targets | Inspect the exact evaluation subset |
| **`benchmark_results`** | 3,090 rows: 618 targets × five methods | Compare released metrics using the `emd_id` and `method` columns |
| **`training_utility`** | 1,236 rows: 618 targets × two retraining conditions | Compare released ModelAngelo CNN retraining results |
| **`examples`** | Density projections for 10 example entries | Inspect sample maps without downloading a large archive |
| **`downloads`** | 71 archives, including benchmark resources | Check sizes, entry counts, download URLs and SHA-256 hashes |

The original per-method CSVs remain available in [benchmark/results/](https://huggingface.co/datasets/markfaa/CryoAtlas/tree/main/resources/benchmark/results). Use the `method` column in the unified results table to select a method.

The catalog contains metadata and archive links; map, CIF and FASTA payloads reside in the archives. [Field definitions and provenance](data_dictionary.md) explain the tables and previews.

## Download the data you need

| Component | Entries | Archives | Compressed size | Files |
|---|---:|---:|---:|---|
| Training Part 1 | 4,164 | 17 | 560.0 GB | [Browse](https://huggingface.co/datasets/markfaa/CryoAtlas/tree/main/train/part01) |
| Training Part 2 | 4,164 | 17 | 563.0 GB | [Browse](https://huggingface.co/datasets/markfaa/CryoAtlas/tree/main/train/part02) |
| Training Part 3 | 4,165 | 17 | 564.9 GB | [Browse](https://huggingface.co/datasets/markfaa/CryoAtlas/tree/main/train/part03) |
| Training Part 4 | 4,164 | 17 | 575.8 GB | [Browse](https://huggingface.co/datasets/markfaa/CryoAtlas/tree/main/train/part04) |
| Benchmark test subset | 618 | 1 | 81.9 GB | [Download](https://huggingface.co/datasets/markfaa/CryoAtlas/resolve/main/test/CryoAtlas_v1.0_test_benchmark_618.tar.gz) |
| Remaining test entries | 534 | 1 | 69.0 GB | [Download](https://huggingface.co/datasets/markfaa/CryoAtlas/resolve/main/test/CryoAtlas_v1.0_test_remaining_534.tar.gz) |
| Benchmark resources | Results, predictions and 10 examples | 1 | 1.90 GB | [Download](https://huggingface.co/datasets/markfaa/CryoAtlas/resolve/main/resources/CryoAtlas_v1.0_benchmark_resources.tar.gz) |

The four training parts are disjoint download partitions of one fixed split. **618 + 534 = 1,152 test entries**: the benchmark is a subset of the test set, not an additional split. The 618 targets support paired comparisons after training-overlap filtering and selection of targets with valid outputs from the evaluated learning-based methods. The remaining entries complete the fixed test split.

For a first look, use the online previews or the **1.90 GB benchmark resource archive**. To locate a particular entry or download a selected training part, use the [download helper](../scripts/download.py):

```bash
pip install -U huggingface_hub
hf download markfaa/CryoAtlas resources/scripts/download.py --repo-type dataset --local-dir .

# Show the archive and size for one entry; do not download the archive yet.
python resources/scripts/download.py --emd EMD-0719 --dry-run

# Download Training Part 1 and verify each archive.
python resources/scripts/download.py --part 1 --destination CryoAtlas-downloads
```

`--emd` selects the **complete archive containing that entry**. More options, direct downloads, extraction instructions and checksum commands are in the [download guide](downloads.md).

## Repository layout

```text
README.md                 Short introduction and main download links
train/part01/ ... part04/  68 training archives
test/                     Two test archives
resources/                Benchmark resource package and supporting files
    benchmark/            Evaluation tables and image previews
    metadata/             Source accessions, splits, inventories, catalogs and checksums
    docs/                 Detailed documentation and figures
    scripts/              Selective download helper
```

Repository-side supporting paths are grouped under `resources/`. Paths *inside extracted archives* retain their original v1.0 layout.

## What is inside an entry?

Training and test archives extract into `CryoAtlas_v1.0/train/EMD-.../` and `CryoAtlas_v1.0/test/EMD-.../`. Each archive is independently extractable; extract the desired archives into the same parent directory.

| File | Contents |
|---|---|
| `emd_<id>_fullmap.map` | Experimental density map in MRC/CCP4 format |
| `<pdb>.cif` | Atomic structure with deposited metadata |
| `<pdb>.fasta` | Original protein sequence records |
| `<pdb>_aligned.fasta` | Aligned sequences, including gap characters where applicable |
| `<pdb>_trimmed.fasta` | Terminal-trimmed sequences |
| `<pdb>_aligned.cif`, where supplied | Structure aligned to the density-map frame |

Use the aligned CIF for map-frame comparisons when it is provided. The per-archive inventories record the exact files and hashes for each entry.

## Benchmark resources and examples

Visit the [CryoAtlas project website](https://oulling.github.io/cryoatlas/) for the paper overview and the [3D Benchmark Explorer](https://oulling.github.io/cryoatlas/explore/) for interactive structure and density visualization.

The resource archive extracts into **`CryoAtlas_v1.0_resources/`**. It contains:

- `Benchmark_outcome/`: the five original evaluation CSVs, with 618 target rows each.
- `Benchmark_rawdata/`: predicted CIF files from AF3, CryoAtom2, E3-CryoFold, EMProt and ModelAngelo, with 618 targets per method.
- `metadata/`: source accessions, fixed split lists, alignment availability and resource-file checksums.
- `training_utility/`: `MA_CryoAtlas.csv` and `MA_Cryo2StructData.csv` (618 target rows plus one AVERAGE row each), with the corresponding `MA_CryoAtlas/` and `MA_Cryo2StructData/` predicted CIF folders (618 per condition). These are existing paper results from the original Kaggle resource release.
- `sample/Dataset/`: 10 self-contained examples with maps, structures, sequences and available predictions.

The same five CSVs are available directly in this repository under `resources/benchmark/results/`. The unified Viewer table adds a method column and removes method prefixes from column names; it preserves the released measurements. AF3 is a sequence-based reference; the other four methods use density maps.

<a id="examples"></a>
### Example gallery

Nine examples belong to the 618-target benchmark. **EMD-0719 belongs to training**, includes an aligned CIF, and has no baseline predictions in the example package. These examples also occur in the full release and do not add new entries.

[Open the image Viewer](https://huggingface.co/datasets/markfaa/CryoAtlas/viewer/examples/examples)

<details>
<summary>Show all 10 density previews</summary>

![Maximum-intensity projections of the 10 released examples](assets/example_gallery.png)

</details>

Previews show maximum-intensity projections along the three stored array axes. Each view uses its own 2nd–99.5th percentile display range. They are visual summaries, not aligned structure overlays or quantitative map-quality measurements. Original maps are available in the resource archive and full release.

## Provenance, scope and license

[Source accessions](../metadata/source_accessions.csv) identify EMDB/PDB records and original map URLs. [Fixed split lists](https://huggingface.co/datasets/markfaa/CryoAtlas/tree/main/resources/metadata/splits), per-part EMD lists, file inventories and checksums support reproducible retrieval. The overview and catalog use the published metadata; no scientific data or benchmark measurements were modified during repository organization.

CryoAtlas supports protein atomic model building from cryo-EM maps. The benchmark results describe the released methods and evaluation subset; the 10 examples are for inspection, not an independent evaluation set.

The dataset and benchmark resources are distributed under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/). Source attribution and deposited metadata are retained.