---
{
  "pretty_name": "MolAS Active Learning — AI4DD 2026 benchmark and research artifacts",
  "language": [
    "en"
  ],
  "license": "other",
  "license_name": "cc-by-4.0-author-content-and-upstream-terms",
  "license_link": "LICENSE",
  "size_categories": [
    "10K<n<100K"
  ],
  "task_categories": [
    "tabular-regression"
  ],
  "tags": [
    "chemistry",
    "biology",
    "drug-discovery",
    "molecular-docking",
    "algorithm-selection",
    "active-learning",
    "runtime-aware",
    "benchmark",
    "ai4dd"
  ],
  "configs": [
    {
      "config_name": "benchmark_cells",
      "data_files": [
        {
          "split": "full",
          "path": "data/viewer/benchmark_cells.csv"
        }
      ],
      "default": true
    },
    {
      "config_name": "rmsd",
      "data_files": [
        {
          "split": "full",
          "path": "data/raw/moad_rmsd_trimmed.csv"
        }
      ]
    },
    {
      "config_name": "pb_ratio",
      "data_files": [
        {
          "split": "full",
          "path": "data/raw/moad_pb_ratio_trimmed.csv"
        }
      ]
    },
    {
      "config_name": "runtime",
      "data_files": [
        {
          "split": "full",
          "path": "data/raw/moad_runtime_trimmed.csv"
        }
      ]
    },
    {
      "config_name": "acquisition_scores",
      "data_files": [
        {
          "split": "full",
          "path": "data/targets/full_strict_score_matrix.csv"
        }
      ]
    },
    {
      "config_name": "native_evaluation_scores",
      "data_files": [
        {
          "split": "full",
          "path": "data/targets/moad_score_native_reconstructed.csv"
        }
      ]
    },
    {
      "config_name": "folds",
      "data_files": [
        {
          "split": "full",
          "path": "data/splits/moad_5fold_split_seed42.csv"
        }
      ]
    },
    {
      "config_name": "physical_descriptors",
      "data_files": [
        {
          "split": "full",
          "path": "data/descriptors/moad_physical_descriptors_filtered.csv"
        }
      ]
    },
    {
      "config_name": "protein_graph_index",
      "data_files": [
        {
          "split": "full",
          "path": "data/viewer/protein_graph_index.csv"
        }
      ]
    }
  ]
}
---

<h1 align="center">🧬 MolAS Active Learning — Dataset</h1>

<p align="center"><strong>Active Learning for Molecular-Docking Algorithm Selection<br>under Heterogeneous Runtime</strong><br>
Accepted at the NeurIPS 2026 Workshop on AI for Drug Discovery (AI4DD)<br>
Sean Wan · Jiabao Brad Wang · Mustafa Misir — Duke Kunshan University</p>

[🤗 Dataset](https://huggingface.co/datasets/SeanWan05/MolAS-Active-Learning) · [💻 Code](https://github.com/SeanWan514/MolAS_Active_Learning) · [📄 Paper](paper/AI4DD52_camera_ready.pdf) · [📝 OpenReview](https://openreview.net/forum?id=tSz3wBLABm) · [🗃️ Data dictionary](docs/data_card.md) · [🔎 Verification](provenance/final_verification.json)

GitHub repository: https://github.com/SeanWan514/MolAS_Active_Learning

HuggingFace repository: https://huggingface.co/datasets/SeanWan05/MolAS-Active-Learning

## 🔬 Dataset summary

This dataset preserves the inputs and saved experimental artifacts for retrospective, cell-wise active learning in molecular-docking algorithm selection. It contains **3,179 protein–ligand complexes**, **seven docking engines**, fixed five-fold memberships, original molecular features, and acquisition/downstream outputs. The dataset also includes **all 3,179 original protein graph files**, which are distributed separately from the GitHub code release.

The default `benchmark_cells` configuration contains **22,253 rows**, one per complex–algorithm cell. It is a deterministic reshaping of the unchanged original tables, with original outer-fold assignments. No docking, feature inference, acquisition campaign, or selector training was executed to prepare this release.

## ✨ Included artifacts

| Artifact | Contents |
|---|---|
| Raw outcomes | Aligned RMSD, PoseBusters-validity ratio, and runtime tables: 3,179 × 7 each |
| Targets | Continuous acquisition scores; the explicitly reconstructed native MolAS evaluation scores; preserved earlier target serialization |
| Splits and descriptors | Original five outer folds and four physical descriptor columns |
| Small features | Original 3,179 × 256 projected cache, its verified row IDs, and 3,179 × 384 ligand embeddings |
| Protein graphs | **3,179 original PyTorch/PyG graph files; 14,985,036,224 bytes** |
| Acquisition artifacts | **90 run archives** covering six policies × five folds × three acquisition seeds; **27,000 original CSV identities** |
| Downstream artifacts | **4,508 original CSV payloads** in six ZIPs and 1,771 readable summary mirrors; acquisition seed 0 |
| Research context | Paper PDF/source/figures, source-code mirror, data dictionary, historical provenance, and portable inspection utilities |

The GitHub repository is the primary code reference. The source mirror here is pinned to the GitHub commit recorded in `provenance/hf_source_snapshot.json`. The acquisition archive uses lossless storage deduplication: earlier cumulative CSVs are exact prefixes of final ledgers, with original byte identities recorded and verified. Extraction utilities recover the original CSV streams; a duplicate multi-gigabyte transport ZIP is unnecessary.

## 🚀 Inspect the dataset

For the loading/download examples, install `datasets` and `huggingface_hub` (`python3 -m pip install -r requirements-hub.txt`). The standalone file/data validators require only Python’s standard library.

The Dataset Viewer is explicitly configured for nine independent CSV tables. The binary graphs/features and compressed result archives are download assets and are not automatically deserialized by the Viewer. Every configuration has a split named **`full`**.

```python
from datasets import load_dataset

repo_id = "SeanWan05/MolAS-Active-Learning"
cells = load_dataset(repo_id, "benchmark_cells", split="full")
runtime = load_dataset(repo_id, "runtime", split="full")
graphs = load_dataset(repo_id, "protein_graph_index", split="full")
```

Available configurations: `benchmark_cells` (default), `rmsd`, `pb_ratio`, `runtime`, `acquisition_scores`, `native_evaluation_scores`, `folds`, `physical_descriptors`, and `protein_graph_index`. Their paths and row counts are in `provenance/hf_viewer_configurations.json`.

### Default cellwise schema

| Column | Meaning |
|---|---|
| `PDB_CCD_ID` | Original complex identifier; preserve as a string |
| `algorithm` | One of the seven engine IDs listed below |
| `outer_test_fold` | Original test-fold membership, integers 0–4 |
| `rmsd_angstrom` | Original docking RMSD, interpreted in Å by the paper |
| `pb_ratio` | Original PoseBusters validity ratio; strict validity requires exact equality to 1 |
| `runtime_seconds` | Original runtime, treated as seconds by the experiment code |
| `acquisition_strict_score` | Original continuous acquisition target |
| `native_evaluation_score` | Native evaluation target from the explicitly named, independently checked reconstruction |

Engine order in the original wide matrices: `surfdock`, `unimol`, `diffdock`, `diffdock_l`, `gnina`, `autodock`, `smina`.

## 🧬 Download and use the structural inputs

```python
from huggingface_hub import snapshot_download

folder = snapshot_download(
    repo_id="SeanWan05/MolAS-Active-Learning",
    repo_type="dataset",
    local_dir="MolAS_HF",
)
```

For a versioned analysis, pass the actual published dataset commit to `revision=`. The graph index provides each graph's relative path, size, and SHA-256 for selective downloading. Original graph files are under `data/external/protein_graphs_esmc_600m/`; ligand and projected features are under `data/features/`. Graphs are serialized molecular data, not trained selector weights. Use the preserved MolAS loader and its PyTorch/PyG dependencies rather than treating these binary files as ordinary Dataset Viewer rows.

## 🔎 Local verification and archive access

From the downloaded folder, Python 3.11+ can validate all file hashes and data consistency without executing scientific experiments:

```bash
python3 scripts/verify_hf_package.py
python3 scripts/validate_data.py --require-complete
python3 code/downstream/verify_saved_outputs.py --require-complete --original-archive-manifest provenance/downstream_original_archive_recovery.json
```

`SHA256SUMS` and `provenance/repository_files.json` cover the upload payloads, including all protein graphs. The self-generated manifests and final report are excluded from their own recursive hashes. Verification checks available bytes and recorded provenance; it does not establish a fresh end-to-end reproduction of the historical scientific experiments.

Use `scripts/read_acquisition_artifact.py` to inspect one original acquisition CSV. Use `scripts/materialize_acquisition.py --destination /path/to/new/folder` to estimate expansion; add `--write` to extract. Seed 0 requires about 3.22 GB expanded, and all three acquisition seeds can be selected explicitly. Original downstream CSVs are in the six ZIPs under `results/downstream/`.

## 📁 Dataset layout

```text
MolAS_HF/
├── data/                  # raw outcomes, targets, folds, descriptors, and features
│   ├── external/protein_graphs_esmc_600m/  # all 3,179 original graphs
│   └── viewer/            # deterministic cellwise table and graph index
├── results/               # original acquisition and downstream archives/summaries
├── code/                  # pinned source mirror; GitHub remains the code reference
├── paper/                 # unchanged paper, source, bibliography, and figures
├── docs/                  # data dictionary, interpretation, and source attribution
├── provenance/            # source identities, component checks, and HF manifest
├── scripts/               # inspection, validation, and explicit upload helper
├── LICENSE                # component-specific bundle terms
├── LICENSE_DATA.md         # CC BY 4.0 grant for author-owned data contributions
├── LICENSE_CODE           # MIT license for original software
└── UPLOAD.md              # instructions for publishing this prepared folder
```

## ⚖️ Licensing and attribution

Author-owned data contributions, original result outputs, and curation metadata are provided under **CC BY 4.0**, to the extent the authors hold the relevant rights. Original software is **MIT**. Supplied third-party data, model-derived features/graphs, software, and manuscript assets retain their separately applicable terms. The Hub metadata therefore uses `license: other` and points to the explicit [component-specific LICENSE](LICENSE), rather than assigning a blanket CC BY or MIT license to every bundled artifact. See [LICENSE_DATA.md](LICENSE_DATA.md), [LICENSE_CODE](LICENSE_CODE), and [third-party provenance](docs/third_party.md).

Upstream sources include [MolAS](https://github.com/BradWangW/MolAS), its [processed-data record](https://zenodo.org/records/17760688), MOAD/PDB-derived complexes, ESM-C protein features, and the recovered ChemBERTa ligand-feature workflow. The author confirmed redistribution terms and delegated selection of the dataset's author-owned-content license; the release record is in `provenance/hf_release_authorization.json`.

## 📌 Intended use and interpretation limits

- This is a retrospective reference benchmark. The full tables are offline oracles; acquisition policies may access only previously acquired training-pool outcomes/runtimes. Preserve original held-out folds and warm-start cost accounting.
- **Do not randomly split the cellwise rows.** The seven rows of one complex share the same original outer-test fold. Derive each training pool from the original complex-level memberships. No protein-family/scaffold-disjoint split is newly asserted.
- Acquisition and native evaluation targets are distinct. Native MolAS evaluation zeroes RMSD >5 Å or <0; the continuous acquisition target does not use that truncation. The original server-side native-target CSV bytes were not recovered; the clearly named reconstruction matches all 22,253 saved engine scores within 1.11e-16.
- Extreme finite RMSD values are preserved, including 211 above 1,000 Å. Their failure/sentinel semantics are not established. They are not evidence of valid physical poses.
- Runtime is hardware/workflow dependent; exact historical timing boundaries and all engine-version details are not established by the numeric tables alone.
- The projected feature cache uses the original fixed full-set normalization/projection. It is not a newly fitted fold-local representation. Runtime-surrogate fitting remains a separate observed-data operation.
- Acquisition artifacts cover three seeds; downstream saved results use acquisition seed 0. Recovered launcher defaults do not establish every historical selector RNG/configuration.
- Per-run training configurations and some final figure-export details remain unpinned. No new experiments, significance tests, generalization studies, or prospective docking validations are claimed by this data release.

## 📚 Citation

Please cite the active-learning paper and the upstream MolAS/data/feature sources when using their respective components. The manuscript remains the supplied camera-ready PDF; its original availability statement is preserved in this frozen copy.

```bibtex
@inproceedings{wan2026molas_active_learning,
  title = {Active Learning for Molecular-Docking Algorithm Selection under Heterogeneous Runtime},
  author = {Wan, Sean and Wang, Jiabao Brad and Misir, Mustafa},
  booktitle = {NeurIPS Workshop on AI for Drug Discovery},
  year = {2026},
  url = {https://openreview.net/forum?id=tSz3wBLABm}
}
```

Code: https://github.com/SeanWan514/MolAS_Active_Learning · Data snapshot version: **1.0.0**. When citing or reusing the data, record the actual dataset identifier and immutable Hub commit alongside the code version.