README.md

metadata

license: cc-by-4.0
task_categories:
  - tabular-regression
tags:
  - chemistry
  - supramolecular
  - binding-affinity
  - host-guest
  - smiles
  - molecular-structures
pretty_name: SupraBench Binding Affinity (Comprehensive Anchor)
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/records.parquet

SupraBench Binding Affinity — Comprehensive Anchor Dataset

• 📄 Paper: arXiv:2606.13477
• 💻 Code: https://github.com/Tianyi-Billy-Ma/SupraBench

Each row is a host–guest binding-affinity record from SupraBench, enriched with full molecular-structure data for both the host and the guest (SMILES, a 2D depiction, and a 3D conformer) plus the experimental environment (solvent, temperature, pH). It is the structure-complete companion to the SupraBench/bap task split.

Contents

	count
Binding records (rows)	2,609
of which CB[7] supplement	217
host SMILES present	2,386 (91%)
guest SMILES present	2,605 (100%)
host 3D conformer present	2,340
guest 3D conformer present	2,528
environment (solvent) present	2,539

Every record is kept; fields are left empty where the corresponding structure or environment value was unavailable (e.g., exotic synthetic hosts with no public structure). 2D depictions are embedded inline; 3D conformers ship as repo files.

Schema

column	type	notes
task_id	string	record id (t1_<interaction_id>; |-joined ids are averaged measurements)
host_name	string	host molecule name
host_smiles	string	host canonical SMILES (empty if unresolved)
host_image_2d	Image	host 2D depiction (empty if unresolved)
host_3d_basename	string	host 3D file basename -> structures_3d/<base>.{xyz,sdf,pdb}
guest_name	string	guest molecule name
guest_smiles	string	guest canonical SMILES
guest_image_2d	Image	guest 2D depiction
guest_3d_basename	string	guest 3D file basename
binding_affinity	float64	log K_a
solvent	string	measurement solvent
temperature	string	measurement temperature (e.g., 25.0C)
ph	string	measurement pH
source	string	eupmc (SupraBank/Europe PMC pipeline) or cb7_supplement

3D structures

3D conformers (RDKit ETKDGv3 + MMFF94/UFF) are shipped as repo files under structures_3d/, one .xyz, .sdf, and .pdb per molecule, named by InChIKey (1,399 molecules x 3 formats = 4,197 files).

Note. The 3D files are not shown in the dataset viewer (XYZ/SDF/PDB are not a HuggingFace column type). They live under the Files and versions tab. Each record links to them through the host_3d_basename and guest_3d_basename columns, where the file path is structures_3d/<basename>.{xyz,sdf,pdb}.

Download a record's 3D structures programmatically:

from datasets import load_dataset
from huggingface_hub import hf_hub_download

ds = load_dataset("SupraBench/Binding-Affinity", split="train")
row = ds[0]

# host conformer (XYZ); swap the extension for .sdf or .pdb
host_xyz = hf_hub_download(
    repo_id="SupraBench/Binding-Affinity", repo_type="dataset",
    filename=f"structures_3d/{row['host_3d_basename']}.xyz",
)
guest_sdf = hf_hub_download(
    repo_id="SupraBench/Binding-Affinity", repo_type="dataset",
    filename=f"structures_3d/{row['guest_3d_basename']}.sdf",
)

An empty host_3d_basename / guest_3d_basename means no 3D conformer is available for that molecule (see Known limitations and failures.csv).

Resolution & validation

SMILES were resolved through a cascade (local SupraBank/CB7 dictionaries -> InChI -> the record's prompt text -> PubChem by CID/name -> OPSIN -> the SupraBank molecule catalog via exact-name match). Every resolved SMILES passes RDKit parsing; host families (cucurbiturils, sulfonatocalixarenes, calixarenes) additionally pass a chemistry sanity gate (atom/ring/charge counts must match the name). 3D conformers are geometry-validated (no atom clashes / degenerate embeddings). Components that failed the gate, framework materials (zeolites), proteins, and supramolecular assemblies (rotaxanes) are left with empty structure fields; the per-molecule status is in failures.csv.

Known limitations

• Stereoisomer pairs (cis/trans, R/S, L/D, anti/syn) may share a flattened 2D skeleton because the source SMILES lack stereo descriptors.
• \|-joined task_ids are averaged over multiple raw measurements; the reported environment is taken from the first underlying interaction.

Sources & license

Derived from SupraBank (CC-BY-4.0), PubChem, and OPSIN. Released under CC-BY-4.0.