Data#

This section summarizes the datasets used by SynPlanner and how to obtain them.

Overview of datasets#

SynPlanner operates reaction and molecule data stored in different formats.

Data type

Description

Reactions

Reactions can be loaded and stored as the list of reaction smiles in the file (.smi) or RDF File (.rdf)

Molecules

Molecules can be loaded and stored as the list of molecule smiles in the file (.smi) or SDF File (.sdf)

Reaction rules

Reaction rules stored as TSV (.tsv, preferred) or pickled list (.pickle, legacy)

Retrosynthetic models

Retrosynthetic models (neural networks) can be loaded and stored as serialized PyTorch models (.ckpt)

Retrosynthetic routes

Retrosynthetic routes can be visualized and stored as HTML files (.html) and can be stored as JSON files (.json)

Note

Reaction and molecule file formats are parsed and recognized automatically by SynPlanner from file extensions. Be sure to store the data with the correct extension.

Data repository structure#

Data is hosted on HuggingFace in a component-based structure:

SynPlanner-data/
├── policy/                    # Reaction rules + policy network weights
│   └── {architecture}/
│       └── {rules_version}/
│           ├── reaction_rules.tsv
│           ├── pipeline.yaml
│           └── {weights_version}/
│               ├── ranking_policy.ckpt
│               └── filtering_policy.ckpt
├── value/                     # Value network weights
│   └── {architecture}/
│       └── {version}/
│           ├── value_network.ckpt
│           └── meta.yaml
├── building_blocks/           # Building block sets
│   └── {name}/
│       ├── building_blocks.tsv
│       └── meta.yaml
├── reaction_data/             # Raw → standardized → filtered pipeline
│   └── {source}/
│       ├── raw/
│       ├── standardized/{YYYY-MM-DD}/
│       └── filtered/{YYYY-MM-DD}/
├── training_data/             # Per-network training inputs
│   ├── ranking_policy/{YYYY-MM-DD}/
│   ├── filtering_policy/{YYYY-MM-DD}/
│   └── value_network/{YYYY-MM-DD}/
├── presets/                   # Ready-to-use preset definitions
│   └── {name}.yaml
└── benchmarks/                # Benchmark target sets (downloaded separately)
    └── sascore/

Versioning: model components (policy/, value/) are versioned by directory name (architecture family, rules version, weights version). Pipeline data (reaction_data/, training_data/) is versioned by processing date (YYYY-MM-DD).

Data sources and bundles#

policy/supervised_gcn/v1/ — reaction rules and policy weights
reaction_rules.tsv — 24k reaction rules in SMARTS format (TSV)
v1/ranking_policy.ckpt — ranking policy network trained on filtered USPTO and corresponding rules
v1/filtering_policy.ckpt — filtering policy network trained on ChEMBL molecules and corresponding rules
pipeline.yaml — full reproducibility manifest (standardization, filtration, extraction, training configs)
value/supervised_gcn/v1/ — value network weights
value_network.ckpt — value network trained from planning simulations on ChEMBL targets
building_blocks/emolecules-salt-ln/ — building blocks
building_blocks.tsv — 186k standardized building blocks (eMolecules + Sigma Aldrich)
reaction_data/uspto/ — reaction data pipeline
raw/uspto_full_mapped.smi.zip — original USPTO dataset (1.48M reactions, compressed)
standardized/2024-12-31/ — standardized reactions + config + errors
filtered/2024-12-31/ — filtered reactions + config + errors
training_data/ — per-network training inputs (date-versioned)
filtering_policy/2024-12-31/molecules_for_training.smi.zip — ChEMBL molecules for filtering policy training
value_network/2024-12-31/targets_for_training.smi.zip — ChEMBL targets for value network tuning
benchmarks/sascore/ — SynPlanner original benchmarks (downloaded separately)
7 target subsets split by SAScore range (1.5–8.5)

Download data#

Data download#

Use the built-in downloader to fetch pre-trained models, reaction rules, and building blocks from HuggingFace.

Details#

For a full list of datasets and descriptions, see Data.

Download from Hugging Face (browse)#

On this page