DTI benchmarks with 3D structures

The BindingDB, BioSNAP and Human drug–target interaction benchmarks, with the same files and splits as DrugBAN (DrugBAN/datasets). Two kinds of structures are added: one RDKit 3D conformer for every compound, and one AlphaFold 3 structure for every protein. The 11 longest proteins (over 4,000 residues) could not be predicted, and one metal complex has no conformer, so their pairs were removed from every csv file; everything else matches DrugBAN. So every pair that is left can be featurised from both structure files.

A training notebook that uses these files (EGNN encoders + DrugBAN's bilinear attention) is at github.com/XingqiaoLin/final-project-9.

Licenses. The interaction csv files and the RDKit conformers are MIT licensed. The AlphaFold 3 structures in main/protein_af3 (including protein_ca.npz, which is derived from them) are not MIT licensed: they are provided under and subject to the AlphaFold 3 Output Terms of Use. See main/protein_af3/OUTPUT_TERMS_OF_USE.md and main/protein_af3/LEGALLY_BINDING_TERMS_OF_USE.txt. novel target benchmark/ has its own sources and licence notes in its README.

Layout

main/
  bindingdb/   full.csv  random/{train,val,test}.csv  cluster/{source_train,target_train,target_test}.csv
  biosnap/     full.csv  random/{train,val,test}.csv  cluster/{source_train,target_train,target_test}.csv
  human/       full.csv  random/{train,val,test}.csv  cold/{train,val,test}.csv
  drug_sdf/
    drugs.csv                 one row per compound: did, SMILES, how its conformer was made
    drug_conformers.sdf.gz    every conformer in one SDF file; each molecule's name is its did
  protein_af3/
    proteins.csv              one row per protein: pid, sequence, whether it has a structure, mean pLDDT, pTM
    protein_ca.npz            Cα coordinates + pLDDT of every structure (small; enough for residue-level models)
    af3_structures.tar.gz     full AF3 output: af3/<pid>/<pid>_model.cif, <pid>_summary_confidences.json
    OUTPUT_TERMS_OF_USE.md  LEGALLY_BINDING_TERMS_OF_USE.txt
novel target benchmark/       a separate test set of proteins unlike any in main/ (see its README)

Each split csv has the columns SMILES, Protein, Y, where Y is 1 for an interaction and 0 otherwise. The cluster splits are DrugBAN's cross-domain splits: source_train is the source domain, and target_train and target_test are the target domain.

Ids. Compounds and proteins are identified by did = md5(SMILES)[:16] and pid = md5(sequence)[:16], computed on the exact SMILES and Protein strings in the split files. So a pair's structures can be found directly from its row:

import hashlib
md5 = lambda s: hashlib.md5(s.encode()).hexdigest()[:16]
did, pid = md5(row.SMILES), md5(row.Protein)

Statistics (full.csv)

Dataset	Pairs	Unique proteins	Unique compounds	Splits provided
BindingDB	49,179	2,619	14,643	random, cluster
BioSNAP	27,409	2,175	4,504	random, cluster
Human	6,696	1,995	2,710	random, cold

Human full.csv contains 731 repeated (SMILES, Protein) rows. After removing them there are 5,965 pairs, which is what the train, val and test files of the random split add up to. BioSNAP has 7 repeated rows (27,402 after removing them).

Rows removed from DrugBAN's files. Every row whose protein has no AlphaFold 3 structure (11 proteins, see below) or whose compound has no RDKit conformer (1 compound, see below) was removed from full.csv and from every split file. Nothing else was changed: no rows were moved between splits, and the columns are the same.

Dataset	Proteins removed	Rows removed from full.csv	Rows removed per split file
BindingDB	4 of 2,623	20 of 49,199 (0.04%)	random: train 16, val 0, test 4; cluster: source_train 10, target_train 1, target_test 0
BioSNAP	6 of 2,181	55 of 27,464 (0.20%)	random: train 35, val 7, test 13; cluster: source_train 2, target_train 19, target_test 6
Human	6 of 2,001	32 of 6,728 (0.48%)	random: train 21, val 2, test 9; cold: train 18, val 0, test 0

Some removed proteins appear in more than one dataset, so the per-dataset counts add up to more than 11. The 5 extra BioSNAP rows are the ones of the compound without a conformer. The unfiltered files are in the DrugBAN repository.

Compound structures (main/drug_sdf)

• drug_conformers.sdf.gz holds one 3D conformer per compound, with explicit hydrogens, in a single SDF file. Each molecule's name (_Name) is its did, and its smiles property is the original SMILES.
• drugs.csv lists all 21,833 compounds of DrugBAN's original files: did, n_pairs, datasets, has_conformer, method, n_heavy_atoms, smiles. method says how the conformer was made (see below). n_pairs and datasets count the original files; 17 compounds no longer appear in any csv (16 were only paired with removed proteins, and 1 is the compound without a conformer).
• How the conformers were made (RDKit 2026.03.1):
  • The SMILES is reduced to its largest fragment, which removes salts and solvents.
  • It is embedded with ETKDGv3 (randomSeed=42) and then optimized with MMFF94 for up to 2000 iterations. UFF is used where MMFF parameters are missing.
  • If the embedding fails, it is retried with random initial coordinates, and then without enforcing chirality.
• 21,832 of the 21,833 compounds succeeded. The exception is 15952d0a50480e89 (nitroprusside, O=N[Fe--](C#N)(C#N)(C#N)(C#N)C#N), whose 5 BioSNAP rows were removed from the csv files. It is still listed in drugs.csv with has_conformer = False.

Protein structures (main/protein_af3)

4,517 of the 4,528 unique proteins have a structure. The 11 without one are the longest proteins (4,128 to 7,073 residues, median 4,548): their predictions did not fit in the 48 GB of GPU memory, which is enough for about 3,600 residues, and spilling them to host memory was too slow to be worth it. The longest protein with a structure has 3,661 residues. These 11 proteins are still listed in proteins.csv with has_structure = False, but their pairs are no longer in the csv files.

How the structures were predicted:

• Model: AlphaFold 3 v3.0.1 with the official weights. Each protein is predicted as a monomer, and there is one prediction per protein (model seed 1, one diffusion sample).
• MSA: HMMER 3.4 jackhmmer searched UniRef90 (2022_05) only. The AF3 default settings were used (1 iteration, E-value 1e-4, at most 10,000 sequences). There is no paired MSA, and the MGnify, BFD and UniProt searches were not run.
• Templates: hmmsearch was run against PDB seqres (2022-09-28), with a template date cutoff of 2021-09-30 and at most 4 templates.
• Confidence: the median over proteins of the mean pLDDT is 84.6, and the median pTM is 0.78.

The files:

• proteins.csv: pid, length, n_pairs, datasets, has_structure, mean_plddt, ptm, seq. seq is the sequence exactly as it appears in the Protein column of the split files.

• af3_structures.tar.gz: the full mmCIF model (per-atom pLDDT in the B-factor column) and AF3's summary confidences (pTM, ranking score) for each protein.

• protein_ca.npz: only the Cα atom of each residue, for all proteins back to back. It is 32 MB instead of 453 MB and loads in a second.

  Array	Shape	Content
  pid	(P,)	protein id
  offset	(P+1,)	protein i is rows offset[i] : offset[i+1] of the arrays below
  aa	(R,)	one-letter residue type; X for non-standard residues (e.g. selenocysteine)
  ca_xyz	(R, 3)	Cα coordinates in Å
  plddt	(R,)	per-residue pLDDT of the Cα atom, 0–100

  Every protein in the npz has exactly one row per residue of its sequence.

If you use these structures, please cite Abramson, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493–500 (2024).

Novel target benchmark (novel target benchmark/)

A separate test set for generalization to unseen proteins: protein–ligand complexes from FoldBench whose proteins are less than 50% identical to every protein in main, with measured negatives from ChEMBL and BindingDB. Its README describes how it was built, its columns and its licence. These proteins do not have AF3 structures in this dataset.

Loading

Download only what you need, e.g. one split plus the compact structure files (about 50 MB):

import gzip, hashlib
import numpy as np, pandas as pd
from huggingface_hub import snapshot_download
from rdkit import Chem

root = snapshot_download('Xingqiao-2002/DTI', repo_type='dataset', allow_patterns=[
    'main/biosnap/random/*', 'main/drug_sdf/drug_conformers.sdf.gz', 'main/protein_af3/protein_ca.npz'])
md5 = lambda s: hashlib.md5(s.encode()).hexdigest()[:16]

train = pd.read_csv(f'{root}/main/biosnap/random/train.csv')                     # SMILES, Protein, Y

with gzip.open(f'{root}/main/drug_sdf/drug_conformers.sdf.gz') as fh:            # did -> RDKit molecule
    mols = {m.GetProp('_Name'): m for m in Chem.ForwardSDMolSupplier(fh)}

z = np.load(f'{root}/main/protein_af3/protein_ca.npz')                           # pid -> Cα coordinates, pLDDT
row = {pid: i for i, pid in enumerate(z['pid'])}
def protein(pid):
    s, e = z['offset'][row[pid]], z['offset'][row[pid] + 1]
    return ''.join(z['aa'][s:e]), z['ca_xyz'][s:e], z['plddt'][s:e]

pair = train.iloc[0]
mol, (aa, xyz, plddt) = mols[md5(pair.SMILES)], protein(md5(pair.Protein))

For the full-atom structures, download and unpack the tarball:

import tarfile
path = snapshot_download('Xingqiao-2002/DTI', repo_type='dataset', allow_patterns=['main/protein_af3/af3_structures.tar.gz'])
tarfile.open(f'{path}/main/protein_af3/af3_structures.tar.gz').extractall('protein_af3')   # protein_af3/af3/<pid>/

Sources

Interaction data. These come from DrugBAN (Bai et al., Nature Machine Intelligence 2023), which collected them from these sources:

1. Liu et al. (2007). BindingDB: a web-accessible database of experimentally determined protein–ligand binding affinities. Nucleic Acids Research 35, D198–D201.
2. Huang et al. (2021). MolTrans: Molecular Interaction Transformer for drug–target interaction prediction. Bioinformatics 37(6), 830–836. (BioSNAP)
3. Chen et al. (2020). TransformerCPI: improving compound–protein interaction prediction by sequence-based deep learning with self-attention mechanism and label reversal experiments. Bioinformatics 36(16), 4406–4414. (Human)
4. Bai, P., Miljković, F., John, B. & Lu, H. (2023). Interpretable bilinear attention network with domain adaptation improves drug–target prediction. Nature Machine Intelligence 5, 126–136. (DrugBAN; source of all files and splits)

Protein structures. These were predicted with AlphaFold 3. Its MSA and template searches used UniRef90 and the PDB:

5. Abramson, J. et al. (2024). Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493–500. https://doi.org/10.1038/s41586-024-07487-w
6. Suzek, B. E. et al. (2015). UniRef clusters: a comprehensive and scalable alternative for improving sequence similarity searches. Bioinformatics 31(6), 926–932.
7. Berman, H. M. et al. (2000). The Protein Data Bank. Nucleic Acids Research 28(1), 235–242.
8. Eddy, S. R. (2011). Accelerated profile HMM searches. PLoS Computational Biology 7(10), e1002195. (HMMER)

Compound structures. These were generated with RDKit:

9. RDKit: Open-source cheminformatics, version 2026.03.1. https://www.rdkit.org
10. Riniker, S. & Landrum, G. A. (2015). Better informed distance geometry: using what we know to improve conformation generation. Journal of Chemical Information and Modeling 55(12), 2562–2574. (ETKDG)
11. Halgren, T. A. (1996). Merck molecular force field. I. Basis, form, scope, parameterization, and performance of MMFF94. Journal of Computational Chemistry 17(5–6), 490–519.

Novel target benchmark:

12. Xu, S. et al. (2025). Benchmarking all-atom biomolecular structure prediction with FoldBench. Nature Communications. https://doi.org/10.1038/s41467-025-67127-3
13. Steinegger, M. & Söding, J. (2017). MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. Nature Biotechnology 35, 1026–1028.
14. Burley, S. K. et al. (2023). RCSB Protein Data Bank (RCSB.org): delivery of experimentally-determined PDB structures alongside one million computed structure models of proteins from artificial intelligence/machine learning. Nucleic Acids Research 51(D1), D488–D508.

Acknowledgements

We thank Google DeepMind and Isomorphic Labs for making AlphaFold 3 and its model parameters available for academic use. The protein structures in this dataset were generated with AlphaFold 3 (Abramson et al., 2024), and they are subject to the AlphaFold 3 Output Terms of Use. We also thank the authors of DrugBAN, BindingDB, BIOSNAP and TransformerCPI for releasing the benchmark data, the UniProt Consortium and the Worldwide Protein Data Bank for the sequence and structure databases, and the RDKit developers.