anonymousapple/Assay-aware-BindingDB

Assay-aware BindingDB

Assay-aware BindingDB is a collection of protein–ligand binding records organized by experimental assay type. Each row represents a BindingDB reactant set and includes its measured affinity, source publication, original experimental context, and an assay-specific structured description.

The complete dataset remains available as the full split. Four assay configurations provide direct access to ITC, SPR, FPA, or RBA records, and 40 training-compatible configurations provide predefined training, validation, and test partitions for seeds 0 through 9.

Configurations and splits

The default configuration contains all four assays in one full split. The itc, spr, rba, and fpa configurations each expose one complete assay as full without applying training eligibility filters.

Configurations named <assay>_seed_<seed> reproduce the data used in the downstream binding affinity prediction experiments for seeds 0–9 and expose train, validation, and test splits.

Eligible records require:

• A precomputed Boltz-2 affinity representation.
• A Qwen3 assay-context embedding.
• Exactly one positive numeric Kd, Ki, or IC50 value.
• No < or > qualifier.

Unique PMIDs are shuffled with NumPy RandomState(seed). Using integer truncation, 20% are assigned to test, 10% to validation, and the remainder to train, preventing PMID leakage.

Data schema

Field	Type	Description
reactant_set_id	integer	BindingDB reactant-set identifier and primary record identifier.
pmid	integer	PubMed identifier for the source publication.
protein	string	Protein or biological target name.
ligand.smiles	string	Ligand structure represented as SMILES.
affinity_data.type	string	Measurement type, such as Kd, Ki, or IC50.
affinity_data.value	float	Numeric affinity value.
affinity_data.relation	string	Reported comparison operator, such as =, <, or >.
affinity_data.unit	string	Unit associated with the affinity value.
DESCRIPTION	string	BindingDB assay description.
assay_type	string	One of itc, spr, fpa, or rba.
search_path	list of strings	Locations searched when extracting experimental context.
supplementary_source	list of strings	Supplementary sources used during extraction.
references_previous	string, nullable	Relevant preceding references captured from the publication.
original_paragraph	JSON string, nullable	Source passages serialized as JSON.
structured_description	JSON string, nullable	Assay-aware structured extraction serialized as JSON.
source_filename	string	Name of the source JSON file.
source_record_key	string	Original record key in the source file.

original_paragraph and structured_description are JSON-encoded strings rather than nested Arrow objects because their internal structures vary among publications and assay types. They can be decoded into Python objects when nested data is needed. Both fields are nullable.

Examples

Install the Hugging Face Datasets library before running the examples:

pip install datasets

Load the full dataset

from datasets import load_dataset

dataset = load_dataset(
    "anonymousapple/Assay-aware-BindingDB",
    split="full",
)

print(dataset)
print(f"Number of records: {len(dataset):,}")

Load one complete assay

Pass the assay configuration name as the second argument:

from datasets import load_dataset

dataset = load_dataset(
    "anonymousapple/Assay-aware-BindingDB",
    "itc",
    split="full",
)

print(f"ITC records: {len(dataset)}")

Read an individual record

record = dataset[0]

print("Reactant set:", record["reactant_set_id"])
print("PMID:", record["pmid"])
print("Assay:", record["assay_type"])
print("Protein:", record["protein"])
print("Ligand SMILES:", record["ligand"]["smiles"])

affinity = record["affinity_data"]
print(
    "Affinity:",
    affinity["type"],
    affinity["relation"],
    affinity["value"],
    affinity["unit"],
)

Read multiple records

for record in dataset.select(range(5)):
    print(
        record["reactant_set_id"],
        record["protein"],
        record["assay_type"],
    )

Read structured descriptions

Use json.loads() to decode the JSON string and json.dumps() with indentation to display it in a readable structure:

import json

record = dataset[0]
value = record["structured_description"]

if value is not None:
    structured_description = json.loads(value)
    print(
        json.dumps(
            structured_description,
            indent=2,
            ensure_ascii=False,
        )
    )

After decoding, nested values can be accessed normally:

if record["structured_description"] is not None:
    structured_description = json.loads(
        record["structured_description"]
    )
    print(json.dumps(structured_description, indent=2, ensure_ascii=False))

The same approach works for original_paragraph:

value = record["original_paragraph"]

if value is not None:
    original_paragraph = json.loads(value)
    print(json.dumps(original_paragraph, indent=2, ensure_ascii=False))

Load a training-compatible seeded split

Combine the assay and split seed in the configuration name, then select a normal Hugging Face split:

from datasets import load_dataset

itc_train_seed_1 = load_dataset(
    "anonymousapple/Assay-aware-BindingDB",
    "itc_seed_1",
    split="train",
)

itc_validation_seed_1 = load_dataset(
    "anonymousapple/Assay-aware-BindingDB",
    "itc_seed_1",
    split="validation",
)

itc_test_seed_1 = load_dataset(
    "anonymousapple/Assay-aware-BindingDB",
    "itc_seed_1",
    split="test",
)