liuganghuggingface/moldata-bde_db2

BDE-db2 and external bond-energy tests

Bond-specific energy prediction data with molecular structures, bond indices, and radical fragments. The main training labels are calculated; an experimental external test is identified separately.

Part of the Molecular Datasets collection.

Get started

from datasets import load_dataset
dataset = load_dataset("liuganghuggingface/moldata-bde_db2", "bde_rdf_with_multi_halo_cfc_model_3")

Where the data comes from

• Original data: https://github.com/patonlab/bde-db2
• Original study: BDE-db2 and external bond-energy tests publication

Tables

Config	Splits	Rows	Targets
bde_rdf_with_multi_halo_cfc_model_3	train, validation, test	834,066	bde Bond dissociation enthalpy (regression; units not specified), bdfe Bond dissociation free energy (regression; units not specified)
test_set_exp_bdes	test	82	exp_kcal_mol Experimental bond dissociation enthalpy (regression; kcal/mol)
test_set_natural_products	test	81	bde Bond dissociation enthalpy (regression; units not specified), bdfe Bond dissociation free energy (regression; units not specified)
test_set_pfas	test	1,162	bde Bond dissociation enthalpy (regression; units not specified), bdfe Bond dissociation free energy (regression; units not specified)

data/ contains rows with readable molecular structures; identical observations are collapsed. observations/ contains all source rows.

Source label columns are text, including values such as <10. SMILES_original retains the input representation; SMILES_canonical provides a canonical representation. structure_key groups molecular components, and structure_status records parsing outcomes.

Data quality

• computed: Labels are computed (for example DFT or molecular dynamics).
• all structures readable: All 835,391 SMILES were read with RDKit.
• 1.8% undefined stereo: 1.8% of molecules have a stereocenter whose configuration is not given, so mirror-image forms cannot be told apart.
• 24 of 7,000 stereoisomer sets differ: Found 7,000 groups of molecules that differ only in stereochemistry. In 24, the values are 10-fold apart (log-scale labels), more than one standard deviation apart, or in different classes. That can be a real effect or a mistake in the source; a model that ignores stereochemistry treats these molecules as the same.
• R² limit 0.65: Across 191,754 cases of a molecule measured more than once, the spread between repeats is 0.58 times the spread of all values, so even a perfect model reaches an R² of about 0.65. Part of the spread can come from conditions the source does not record.
• 100% measured per task: In a typical task, 100% of molecules have a value. A missing value means not measured, not inactive.
• 1,818 extreme values: 1,818 of 1,670,700 values lie more than 6 robust standard deviations from the median of their target (on a log scale when values span 1,000-fold or more). They are often typing or unit mistakes, and sometimes real; look at them before training.
• 57 molecules per task: A typical task has 57 labeled molecules. Results on fewer than 1,000 molecules vary a lot between splits.
• source split: train / validation / test: The source divides the data into train (808,628 rows), validation (12,852 rows), test (13,911 rows). Use this division to compare with published results.

Dataset notes

• Training labels are calculated; external test labels are experimental.
• Bond identity and radical fragments are part of each example. Different bonds in one molecule are distinct prediction targets.
• The model-3 expanded training set overlaps earlier training versions.
• Dataset reuse terms are linked from the original source.

License

Dataset-wide license: unspecified. Original source.

Citation

Cite the original study and data source:

• BDE-db2 and external bond-energy tests publication