wjulius/HumanBindome

Website • GitHub • MCP • Preprint (tba)

Towards a protein binder candidate for every human protein

Designed Target-Binder Interface Benchmark

Dataset Description

This dataset contains designed target-binder protein complexes generated with BindCraft. Each row in metadata.csv corresponds to one designed two-chain target-binder complex and links the design metadata, benchmark split assignment, interface-cluster assignment, ProteinMPNN binder variant information, and associated structure files. For each design, the release includes the bound target-binder complex structure as well as predicted apo structures for the target and binder.

The dataset is intended for machine-learning studies of designed protein-protein interactions, including model training, validation, and benchmark evaluation.

Files

The dataset is indexed by a single metadata.csv file. Benchmark partitions are encoded in the split column. Structure paths are provided directly in the metadata table.

The released complex structures are full target-binder complex models in a superimposed coordinate frame. The relative path to each structure is given in the target_binder_complex_pdb_path column. Predicted apo structures for the target domain and binder are provided under target_domain_pdb_path and binder_alone_pdb_path, and the full-length apo target structure under target_structure_pdb_path.

Repository organization:

.
├── README.md
├── metadata.csv
└── structures/
    ├── [xx]/                       # first two characters of target_name
    │   ├── [target_name_1].tar
    │   └── [target_name_2].tar
    └── ...

Each archive is an uncompressed .tar containing the structure files for one target:

[target_name].tar
├── [target_name].pdb                                  # full-length predicted target structure
└── [target_name]_[domain]/
    ├── [target_name]_[domain].pdb              # predicted apo target domain
    └── [design_name]/
        ├── [design_name].pdb                          # superimposed target-binder complex
        └── [design_name]_binder_alone.pdb             # predicted apo binder

Metadata Columns

Column	Description
design_name	Unique identifier for the designed target-domain–binder complex.
target_domain	Target domain or subdomain used for design.
target_name	Name or identifier of the full target.
design_seed	Integer design seed used to generate the binder.
mpnn_variant	ProteinMPNN variant index for the same design seed.
split	Split assignment: train, validation, test, or deleaked.
paired_interface_cluster_id	Paired target-side/binder-side interface cluster identifier used for leakage-controlled splitting.
cluster_member_type	Cluster membership annotation for the validation and test splits. cluster_rep denotes the randomly selected representative used for one-per-cluster validation/test evaluation; cluster_member denotes other members of the same cluster.
structure_archive	Relative path to the target-level tar archive containing the structure files.
target_binder_complex_pdb_path	Path inside the archive to the superimposed target-domain–binder complex PDB file.
binder_alone_pdb_path	Path inside the archive to the binder-only PDB file.
target_domain_pdb_path	Path inside the archive to the apo target-domain PDB file (one per target domain, shared across its designs).
target_structure_pdb_path	Path inside the archive to the full-length apo target structure PDB file (one per target).
binder_sequence	Amino-acid sequence (one-letter) of the designed binder, parsed from the apo binder PDB.

Dataset Splits

The dataset provides predefined train, validation, and test partitions. Complexes removed during leakage control are annotated as deleaked and are not part of the usable benchmark.

Split	Complexes	Share
train	205,553	67.14%
validation	21,205	6.93%
test	21,955	7.17%
deleaked	57,424	18.76%
total	306,137	100.00%
usable benchmark	248,713	81.24%

The usable benchmark consists of the train, validation, and test partitions only.

Split Construction

Leakage was defined at the level of interface similarity, using both structural (Foldseek) and sequence (MMseqs) interface similarity during the subsequent deleaking procedure, rather than by target identity, binder identity or design metadata alone. This is important because designed complexes can share related binding epitopes or recurrent interface geometries even when their identifiers or sequences differ.

The split was constructed using an interface-level procedure adapted from PINDER. Interface residues were defined as residues with at least one heavy backbone atom within 10 Å of the opposite chain. Target and binder chains were aligned all-vs-all with Foldseek. Alignments were retained for clustering only when the aligned region covered at least 50% of an interface residue set.

The resulting Foldseek interface-similarity graph was thresholded using the PINDER clustering threshold of 0.70, and weighted asynchronous label propagation was used to assign chain-interface cluster labels. Each complex was assigned to a paired-interface cluster defined by its target-side and binder-side interface cluster identifiers. Train, validation, and test assignment was performed at the paired-interface-cluster level.

After holdout clusters were selected, PINDER’s depth-2 transitive deleaking procedure was applied. Candidate training complexes connected to validation or test complexes through one or two edges in either the Foldseek interface-similarity graph or a separately constructed interface-filtered MMseqs sequence-similarity graph were assigned to the deleaked partition and excluded from the usable benchmark.

Redundancy Audit

As an independent audit, we assessed whether ProteinMPNN sister variants derived from the same binder backbone on the same target domain were assigned to the same paired-interface cluster. ProteinMPNN grouping was not used as input to the clustering algorithm.

Among 120,609 backbones with ProteinMPNN sister variants, 120,082 were contained within a single paired-interface cluster, whereas 527 were split across multiple clusters. 99.56% of multi-member backbone groups were captured by the paired-interface clustering procedure.

After deleaking, only 23 of the 88,720 multi-member backbone groups in the usable benchmark (0.026%) have sister variants split between train and the held-out (validation or test) split.

Intended Use

The predefined split should be used for benchmark comparisons at the interface level. The validation split is intended for model selection and hyperparameter tuning. The test split should be reserved for final evaluation.

Citation

If you use this dataset, please cite this repository

Acknowledgments

We thank the authors of the following methods, datasets and tools, on which Bindome is built:

• BindCraft
• ColabDesign
• AlphaFold2
• AlphaFold-Multimer
• ProteinMPNN
• Foldseek
• PINDER