NAIAD Dataset Package

This package contains the train/valid/test CSV index files used immediately before NAIAD training, plus every referenced .cif source structure.

Contents

• data/*.csv: model input index files. structure_path has been rewritten to package-relative paths such as structures/101d.cif.
• structures/*.cif: all 3725 unique source structures referenced by the packaged CSV files.
• splits/*.json: released split ID lists used to derive train/valid/test sets.
• scripts/: preprocessing, filtering, CSV-generation, and split/preparation scripts from the training repository.
• configs/: NAIAD training config.

The .cif files were reconstructed locally from the official wwPDB/RCSB mmCIF mirrors using the PDB IDs in manifest.csv. Use scripts/download_structures_from_manifest.py to reproduce this step.

The parser still expects the normal NAIAD chemical component cache (ligands.json.gz and elements.txt) under the repository's data/datasets/rcsb_cif/, or via NAIAD_RCSB_CIF_DIR.

For the reported NAIAD training-data release, use data/train.csv as the training index, data/valid.csv as validation, and data/test.csv as the held-out test index.

To use after extraction, point DF_PATH_TRAIN and DF_PATH_VALID to data/train.csv and data/valid.csv from this package, or copy the package contents under the repository root so the relative structure_path values resolve.

Reproducing The Split CSVs

There are two different workflows in scripts/.

Exact packaged train/valid/test CSVs

Use this route to reproduce the three split CSVs bundled in data/ from the released split ID files and the packaged CIF files:

python scripts/generate_training_csv_from_splits.py \
  --mmcif_dir structures \
  --splits_dir splits \
  --output_dir reproduced_data \
  --split_type design

Expected result when run against this package:

• reproduced_data/train.csv: 3096 rows.
• reproduced_data/valid.csv: 308 rows.
• reproduced_data/test.csv: 321 rows.

The row IDs match the shipped data/train.csv, data/valid.csv, and data/test.csv. This is because the script reads the full split ID lists and keeps only IDs whose CIF file is present under structures/. The regenerated CSVs will contain absolute structure_path values because that script calls os.path.abspath; the shipped CSVs use package-relative paths such as structures/101d.cif for portability.

generate_training_csv_from_splits.py requires only normal tabular Python dependencies (pandas, tqdm) and does not import the NAIAD model/parser code.

Full rescan/filter/preprocess/split from a mmCIF mirror

Use this route only if you want to create a new dataset split from an external mmCIF mirror rather than reproduce the packaged split:

python scripts/prepare_diffusion_dataset_full.py scan \
  --mmcif_dir /path/to/mmcif_files \
  --output_dir new_dataset \
  --num_workers 16 \
  --require_na

python scripts/prepare_diffusion_dataset_full.py preprocess \
  --output_dir new_dataset \
  --num_workers 16

# Optional, if CD-HIT is installed.
python scripts/prepare_diffusion_dataset_full.py cluster \
  --output_dir new_dataset \
  --cdhit_path /path/to/cdhit

python scripts/prepare_diffusion_dataset_full.py split \
  --output_dir new_dataset \
  --valid_fraction 0.1 \
  --test_fraction 0.1 \
  --use_clustering

Equivalent all-in-one form:

python scripts/prepare_diffusion_dataset_full.py all \
  --mmcif_dir /path/to/mmcif_files \
  --output_dir new_dataset \
  --num_workers 16 \
  --require_na \
  --valid_fraction 0.1 \
  --test_fraction 0.1

This full workflow performs scanning, quality filtering, per-structure preprocessing, optional CD-HIT sequence clustering, then writes train.csv, valid.csv, and optionally test.csv. It is not the command used to reproduce the fixed manuscript split files in this package.

prepare_diffusion_dataset_full.py imports NAIAD parser/data modules (cifutils.py, pdbutils.py, na_data_utils.py) and expects the chemical component cache (ligands.json.gz, elements.txt). Run it from a NAIAD source checkout or set PYTHONPATH so those modules are visible, and set NAIAD_RCSB_CIF_DIR if the cache is not under the repository's data/datasets/rcsb_cif/.

Optional preprocessing of an existing CSV

preprocess_dataset.py and preprocess_dataset.sh do not create train/valid/test splits. They consume an existing CSV, such as data/train.csv, and write per-structure sequence and NumPy preprocessing artifacts:

python scripts/preprocess_dataset.py \
  data/train.csv \
  preprocessed/train \
  1 \
  0

The last two arguments are modulo and remainder, used for array-job sharding. For example, an HPC array with 100 tasks would run each shard with modulo=100 and remainder equal to the task index. preprocess_dataset.sh is an example SLURM/Apptainer wrapper for that sharded mode and may need local cluster/container paths edited before use.