Alterlab Bindingdb

BindingDB Database

Overview

BindingDB (https://www.bindingdb.org/) is the primary public database of measured drug-protein binding affinities. It contains roughly 3.2 million binding data records for ~1.4 million compounds tested against ~11,400 protein targets, curated from scientific literature and patent literature. BindingDB stores quantitative binding measurements (Ki, Kd, IC50, EC50) essential for drug discovery, pharmacology, and computational chemistry research.

Key resources:

• BindingDB website: https://www.bindingdb.org/
• REST API base: https://bindingdb.org/rest/ (no key; default response is XML, append response=application/json)
• Downloads page: https://www.bindingdb.org/rwd/bind/chemsearch/marvin/Download.jsp (the full TSV is the dated BindingDB_All_<YYYYMM>_tsv.zip, ~560 MB zipped, refreshed monthly)

When to Use This Skill

Use BindingDB when:

• Target-based drug discovery: What known compounds bind to a target protein? What are their affinities?
• SAR analysis: How do structural modifications affect binding affinity for a series of analogs?
• Lead compound profiling: What targets does a compound bind (selectivity/polypharmacology)?
• Benchmark datasets: Obtain curated protein-ligand affinity data for ML model training
• Repurposing analysis: Does an approved drug bind to an unintended target?
• Competitive analysis: What is the best reported affinity for a target class?
• Fragment screening: Find validated binding data for fragments against a target

Core Capabilities

1. BindingDB REST API

Base URL: https://bindingdb.org/rest

import requests

BASE_URL = "https://bindingdb.org/rest"

def bindingdb_query(method, params):
    """Query the BindingDB REST API."""
    url = f"{BASE_URL}/{method}"
    response = requests.get(url, params=params, headers={"Accept": "application/json"})
    response.raise_for_status()
    return response.json()

2. Query by Target (UniProt ID)

The getLigandsByUniprot endpoint takes a single uniprot parameter formatted as <UniProt accession>;<cutoff in nM>, e.g. P00519;10000.

def get_ligands_for_target(uniprot_id, cutoff=10000):
    """
    Get all ligands with measured affinity for a UniProt target.

    Args:
        uniprot_id: UniProt accession (e.g., "P00519" for ABL1)
        cutoff: Maximum affinity value to return (in nM)
    """
    params = {
        "uniprot": f"{uniprot_id};{cutoff}",
        "response": "application/json",
    }
    return bindingdb_query("getLigandsByUniprot", params)

# Example: Get all compounds binding ABL1 (imatinib target) at <=100 nM
ligands = get_ligands_for_target("P00519", cutoff=100)

3. Query by SMILES (structural similarity)

def search_by_smiles(smiles, cutoff=0.85):
    """
    Search BindingDB by SMILES string (structural-similarity search).

    Args:
        smiles: SMILES string of the compound
        cutoff: Tanimoto similarity threshold (0.0-1.0; e.g. 0.85)
    """
    params = {
        "smiles": smiles,
        "cutoff": cutoff,
        "response": "application/json",
    }
    return bindingdb_query("getTargetByCompound", params)

# Example: structural-similarity search for imatinib's binding targets
result = search_by_smiles("Cc1ccc(NC(=O)c2ccc(CN3CCN(C)CC3)cc2)cc1Nc1nccc(-c2cccnc2)n1")

To query by a PubChem CID, first convert the CID to a SMILES string via PubChem PUG-REST, then pass it to search_by_smiles. There is no by-name or by-CID BindingDB REST endpoint.

4. Download-Based Analysis (Recommended for Large Queries)

For comprehensive analyses, download BindingDB data directly:

import pandas as pd

def load_bindingdb(filepath="BindingDB_All.tsv"):
    """
    Load BindingDB TSV file (the unzipped BindingDB_All_<YYYYMM>.tsv).
    Download the dated BindingDB_All_<YYYYMM>_tsv.zip from:
    https://www.bindingdb.org/rwd/bind/chemsearch/marvin/Download.jsp
    """
    # Key columns
    usecols = [
        "BindingDB Reactant_set_id",
        "Ligand SMILES",
        "Ligand InChI",
        "Ligand InChI Key",
        "BindingDB Target Chain  Sequence",
        "PDB ID(s) for Ligand-Target Complex",
        "UniProt (SwissProt) Entry Name of Target Chain",
        "UniProt (SwissProt) Primary ID of Target Chain",
        "UniProt (TrEMBL) Primary ID of Target Chain",
        "Ki (nM)",
        "IC50 (nM)",
        "Kd (nM)",
        "EC50 (nM)",
        "kon (M-1-s-1)",
        "koff (s-1)",
        "Target Name",
        "Target Source Organism According to Curator or DataSource",
        "Number of Protein Chains in Target (>1 implies a multichain complex)",
        "PubChem CID",
        "PubChem SID",
        "ChEMBL ID of Ligand",
        "DrugBank ID of Ligand",
    ]

    # Exact TSV headers drift between monthly releases (some contain double
    # spaces), so intersect with the actual header rather than hard-failing.
    header = pd.read_csv(filepath, sep="\t", nrows=0).columns
    keep = [c for c in usecols if c in header]
    df = pd.read_csv(filepath, sep="\t", usecols=keep,
                     low_memory=False,

    # Convert affinity columns to numeric
    for col in ["Ki (nM)", "IC50 (nM)", "Kd (nM)", "EC50 (nM)"]:
        if col in df.columns:
            df[col] = pd.to_numeric(df[col], errors='coerce')

    return df

def query_target_affinity(df, uniprot_id, affinity_types=None, max_nm=10000):
    """Query loaded BindingDB for a specific target."""
    if affinity_types is None:
        affinity_types = ["Ki (nM)", "IC50 (nM)", "Kd (nM)"]

    # Filter by UniProt ID
    mask = df["UniProt (SwissProt) Primary ID of Target Chain"] == uniprot_id
    target_df = df[mask].copy()

    # Filter by affinity cutoff
    has_affinity = pd.Series(False, index=target_df.index)
    for col in affinity_types:
        if col in target_df.columns:
            has_affinity |= target_df[col] <= max_nm

    result = target_df[has_affinity][["Ligand SMILES"] + affinity_types +
                                      ["PubChem CID", "ChEMBL ID of Ligand"]].dropna(how='all')
    return result.sort_values(affinity_types[0])

5. SAR Analysis

import pandas as pd

def sar_analysis(df, target_uniprot, affinity_col="IC50 (nM)"):
    """
    Structure-activity relationship analysis for a target.
    Retrieves all compounds with affinity data and ranks by potency.
    """
    target_data = query_target_affinity(df, target_uniprot, [affinity_col])

    if target_data.empty:
        return target_data

    # Add pIC50 (negative log of IC50 in molar)
    if affinity_col in target_data.columns:
        target_data = target_data[target_data[affinity_col].notna()].copy()
        target_data["pAffinity"] = -((target_data[affinity_col] * 1e-9).apply(
            lambda x: __import__('math').log10(x)
        ))
        target_data = target_data.sort_values("pAffinity", ascending=False)

    return target_data

# Most potent compounds against EGFR (P00533)
# sar = sar_analysis(df, "P00533", "IC50 (nM)")
# print(sar.head(20))

6. Polypharmacology Profile

def polypharmacology_profile(df, ligand_smiles, affinity_cutoff_nM=1000):
    """
    Find all targets a compound binds to, by exact SMILES match.
    For tolerance to tautomers/salts/charge states, match on the InChIKey
    skeleton (first 14 chars of "Ligand InChI Key") instead of raw SMILES.
    """
    # Search by ligand SMILES (exact string match)
    mask = df["Ligand SMILES"] == ligand_smiles

    ligand_data = df[mask].copy()

    # Filter by affinity
    aff_cols = ["Ki (nM)", "IC50 (nM)", "Kd (nM)"]
    has_aff = pd.Series(False, index=ligand_data.index)
    for col in aff_cols:
        if col in ligand_data.columns:
            has_aff |= ligand_data[col] <= affinity_cutoff_nM

    result = ligand_data[has_aff][
        ["Target Name", "UniProt (SwissProt) Primary ID of Target Chain"] + aff_cols
    ].dropna(how='all')

    # Rank by the tightest measured constant per row (NaNs ignored)
    result = result.assign(best_nM=result[aff_cols].min(axis=1))
    return result.sort_values("best_nM")

Query Workflows

Workflow 1: Find Best Inhibitors for a Target

import pandas as pd

def find_best_inhibitors(uniprot_id, affinity_type="IC50 (nM)", top_n=20):
    """Find the most potent inhibitors for a target in BindingDB."""
    df = load_bindingdb("BindingDB_All.tsv")  # Load once and reuse
    result = query_target_affinity(df, uniprot_id, [affinity_type])

    if result.empty:
        print(f"No data found for {uniprot_id}")
        return result

    result = result.sort_values(affinity_type).head(top_n)
    print(f"Top {top_n} inhibitors for {uniprot_id} by {affinity_type}:")
    for _, row in result.iterrows():
        print(f"  {row['PubChem CID']}: {row[affinity_type]:.1f} nM | SMILES: {row['Ligand SMILES'][:40]}...")
    return result

Workflow 2: Selectivity Profiling

1. Get all affinity data for your compound across all targets
2. Compare affinity ratios between on-target and off-targets
3. Identify selectivity cliffs (structural changes that improve selectivity)
4. Cross-reference with ChEMBL for additional selectivity data

Workflow 3: Machine Learning Dataset Preparation

def prepare_ml_dataset(df, uniprot_ids, affinity_col="IC50 (nM)",
                        max_affinity_nM=100000, min_count=50):
    """Prepare BindingDB data for ML model training."""
    records = []
    for uid in uniprot_ids:
        target_df = query_target_affinity(df, uid, [affinity_col], max_affinity_nM)
        if len(target_df) >= min_count:
            target_df = target_df.copy()
            target_df["target"] = uid
            records.append(target_df)

    if not records:
        return pd.DataFrame()

    combined = pd.concat(records)
    # Add pAffinity (normalized)
    combined["pAffinity"] = -((combined[affinity_col] * 1e-9).apply(
        lambda x: __import__('math').log10(max(x, 1e-12))
    ))
    return combined[["Ligand SMILES", "target", "pAffinity", affinity_col]].dropna()

Key Data Fields

Affinity Interpretation

Best Practices

• Use Ki for direct binding: Ki reflects true binding affinity independent of enzymatic mechanism
• IC50 context-dependency: IC50 values depend on substrate concentration (Cheng-Prusoff equation)
• Watch for qualifier prefixes: TSV affinity cells often carry >, <, or >= prefixes (e.g. >10000) for censored measurements. pd.to_numeric(errors='coerce') silently turns these into NaN — strip the prefix first (e.g. df[col].astype(str).str.lstrip("<>= ")) and decide explicitly whether to keep, drop, or treat censored values as inequalities before modeling
• Normalize units: BindingDB reports in nM; verify units when comparing across studies
• Filter by target organism: Use Target Source Organism to ensure human protein data
• Handle missing values: Not all compounds have all measurement types
• Cross-reference with ChEMBL: ChEMBL has more curated activity data for medicinal chemistry

Additional Resources

• BindingDB website: https://www.bindingdb.org/
• Data downloads: https://www.bindingdb.org/rwd/bind/chemsearch/marvin/Download.jsp
• REST API documentation: https://www.bindingdb.org/rwd/bind/BindingDBRESTfulAPI.jsp (REST base: https://bindingdb.org/rest)
• Citation: Gilson MK et al. "BindingDB in 2015." Nucleic Acids Research 2016;44(D1):D1045-53. PMID: 26481362, doi:10.1093/nar/gkv1072
• Related resources: ChEMBL (https://www.ebi.ac.uk/chembl/), PubChem BioAssay

Scripts

scripts/query_bindingdb.py — runnable helper for the BindingDB REST API (no key):

python scripts/query_bindingdb.py uniprot P00519 --cutoff 10000
python scripts/query_bindingdb.py pdb 1Q0L,3ANM --cutoff 100 --identity 92
python scripts/query_bindingdb.py compound "<SMILES>" --cutoff 0.85