Data Pipeline

Use uqdd/data/data_papyrus.py to prepare Papyrus++ datasets.

Key arguments

• --activity: xc50 | kx
• --descriptor-protein: ankh-large | ankh-small | unirep
• --descriptor-chemical: ecfp2048 | ecfp1024
• --split-type: random | scaffold | time
• --n-targets: -1 for all targets, or a positive integer
• --file-ext: pkl | parquet | csv
• --sanitize: clean invalid entries
• --verbose: print progress

Example

python uqdd/data/data_papyrus.py \
  --activity xc50 \
  --descriptor-protein ankh-large \
  --descriptor-chemical ecfp2048 \
  --split-type time \
  --n-targets -1 \
  --file-ext pkl \
  --sanitize \
  --verbose

Outputs

• Preprocessed files stored under data/, organized by activity/split.
• Descriptor choices should match dimensions in uqdd/config/desc_dim.json.

Auxiliary scripts

• uqdd/data/papyrus_prepare_example.sh: template for preparing a small example subset.
• uqdd/data/papyrus_extensive.sh: comprehensive data preparation workflow for large runs.

Integration notes

• Downstream training scripts (uqdd/models/model_parser.py) expect consistent file extensions and split naming.
• Time-based splits better reflect prospective evaluation, while scaffold splits assess generalization across chemotypes.