SABLE - Structural Antibody Benchmark for deep-Learning Evaluation - is a curated database and benchmark for antibody-structure machine learning. SABLE provides processed antibody variable-domain structures, sequence and CDR annotations, antigen metadata, train/test split labels, redundancy annotations, downloadable data files, and software tools for data loading and training.

The goal of SABLE is to make antibody-structure data easy to inspect, download, and use in reproducible machine-learning experiments. Antibody datasets are especially sensitive to data leakage: related antibodies can share nearly identical complementarity-determining regions (CDRs), high variable-domain sequence identity, or homologous antigen targets. SABLE therefore treats splits, redundancy grouping, and leakage-control metadata as core parts of the dataset rather than as optional annotations.

By providing a fixed, leakage-aware train/test split, SABLE aims to support fairer comparison between deep-learning methods for antibody structure prediction, antibody-antigen modeling, representation learning, docking, paratope and epitope prediction, and related tasks. We encourage users to report the exact SABLE version and training subset used so that results can be compared consistently across models.

Dataset sources

Experimental structures

Experimental SABLE entries are derived from antibody structures curated in SAbDab, which in turn is built from structures deposited in the Protein Data Bank (PDB). SABLE processes these source structures into standardized antibody entries, extracts heavy- and light-chain variable domains, includes matched antigen chains when available, records missing-residue information, and stores per-entry metadata such as source PDB ID, release date, experimental method, resolution, antigen annotation, chain identifiers, framework regions, and CDR sequences.

A row with synthetic=False is an experimental structure derived from SAbDab/PDB.

Synthetic/modelled structures

SABLE also contains synthetic entries. In the metadata, these entries are marked with synthetic=True and have method = AF3 MODELING. These entries are not experimental PDB structures. They are modeled antibody-antigen complex structures generated from paired antibody-antigen sequences obtained from the NaturalAntibody patent-derived antibody sequence data. The synthetic structures were modeled with AlphaFold3 and filtered to retain only high-confidence models.

Synthetic entries are included to expand the distribution of the data. They should be interpreted differently from experimental structures

Redundancy definition

SABLE redundancy is designed to reduce near-duplicate antibody examples in training and to let users explicitly control how much redundancy is used. Experimental training entries are grouped into redundancy clusters according to antibody CDR sequence identity. In the SABLE construction procedure, entries sharing the six CDR sequences are assigned to the same redundancy group, and a representative is selected using structural-quality criteria such as experimental resolution, missing residues in CDRs, and total missing residues.

The redundancy_cluster column identifies the cluster, and the non_redundant column marks the selected representative. This design supports several training protocols:

• Non-redundant experimental training: use one representative per redundancy cluster.
• Expanded experimental training: use all experimental training entries, including redundant cluster members.

A row with non_redundant=True is the selected representative of its redundancy cluster. A row with non_redundant=False is a redundant member assigned to a representative cluster.

Splits

SABLE uses frozen split labels so that models can be evaluated and compared consistently.

• train: entries intended for model training.
• test: held-out experimental entries intended for final benchmark evaluation.

Users should not tune model choices, hyperparameters, or early-stopping decisions on the test set. When reporting results, specify the SABLE release version and whether training used experimental-only data, synthetic data, non-redundant representatives only, or the expanded redundant training set.

For compact training, users can select experimental training entries with split=train, synthetic=False, and non_redundant=True.

Leakage-aware benchmark design

The SABLE benchmark was designed to reduce common sources of leakage in antibody machine learning, including similar HCDR3 loops, high variable-domain sequence identity, and homologous antigen sequences. Candidate held-out structures are selected using a temporal cutoff and then compared to the training set using antibody-level and antigen-level similarity filters.

In the current SABLE split, test-set entries satisfy the following benchmark constraints:

• Temporal separation: test entries were released after 30 September 2021 (AlphaFold3 cutoff).
• Structural quality: test entries have resolution ≤ 3.0 Å and no missing residues in CDRs.
• HCDR3 dissimilarity: each test entry has HCDR3 edit distance ≥ 3 from all training examples.
• Variable-domain dissimilarity: each test entry has variable-domain sequence identity ≤ 90% to all training examples.
• Antigen leakage filtering: protein or peptide antigens were filtered against training antigens using BLASTP. A candidate was removed if it had a training antigen hit with identity ≥ 30%, coverage ≥ 70% of the shorter antigen, and e-value ≤ 1e-5, or if it had a high-identity antigen hit with identity ≥ 70%.
• Manual inspection: final test-set entries were inspected manually to check that their metadata and processed structures are suitable for benchmark evaluation.

These filters reduce, but cannot eliminate, all possible forms of biological similarity. For example, remote structural similarity between antigens may remain even when sequence similarity is low. SABLE therefore provides the split criteria and nearest-neighbor metadata so that users can interpret benchmark results transparently.

Interpreting missing values

Blank or missing values usually mean that the field is not available or not applicable. Examples:

• pdb_id, release_date, resolution, r_free, and r_factor are usually missing for synthetic entries.
• L_sequence and light-chain region fields are missing for nanobody/VHH or heavy-only entries.
• antigen fields are missing when has_antigen=False.
• affinity fields are missing when no binding measurement is available.

Region sequence and index convention

SABLE stores full heavy/light variable-domain sequences and separate FR/CDR region sequences according to the IMGT numbering scheme. Region index columns such as H_CDR3_indices are zero-based, half-open intervals of the form (start, end), meaning the region is located at sequence[start:end]. Placeholder intervals indicate that a region is absent or not applicable. Missing residues in antibody sequences are represented with lower-case amino-acid letters.

Typical uses

SABLE can be used for:

• antibody structure prediction and refinement.
• antibody-antigen interaction modeling.
• paratope and epitope prediction.
• antibody representation learning.
• docking and interface learning.
• sequence-to-structure and structure-to-sequence modeling.
• generative antibody design.
• benchmarking the effects of synthetic data and training-set redundancy.

Counts in this release build

• Total entries: 20112
• Source PDB IDs: 9615
• Train entries: 19785
• Test entries: 327
• Experimental train entries: 16511
• Synthetic entries: 3274
• Experimental non-redundant train representatives: 5668
• Located structure files: 20112
• Missing structure files: 0

SABLE GitHub repository

This Zenodo record consists of flat files:

• sable_v0.1.0_metadata.csv: human-readable per-entry metadata.
• sable_v0.1.0_metadata.parquet: same rows and columns as the CSV, provided for faster typed loading.
• sable_v0.1.0_metadata_schema.json: JSON schema for metadata.csv and metadata.parquet.
• sable_v0.1.0_column_descriptions.md: human-readable metadata column descriptions.
• sable_v0.1.0_splits.zip: frozen train/test split files and redundancy mappings.
• sable_v0.1.0_sequences.zip: heavy-chain, light-chain and paired variable-domain FASTA files.
• sable_v0.1.0_structures_pdb_*.tar.gz: processed PDB files under structures/pdb/.
• sable_v0.1.0_manifest.json: machine-readable release manifest.
• sable_v0.1.0_checksums.sha256: checksums for the flat Zenodo files.
• sable_v0.1.0_LICENSE-DATA.txt: data-license.
• sable_v0.1.0_VALIDATION_REPORT.md: validation summary generated during release construction.
• sable_v0.1.0_README.md

The SABLE GitHub repository https://github.com/dina-lab3D/SABLE provides a Python software package for downloading, validating, exploring, and using the dataset in machine-learning workflows. The package downloads the flat Zenodo release files, verifies their checksums, reconstructs the local release tree (according to sable_v0.1.0_manifest.json), and provides utilities for working with the metadata, structures, splits, and PyTorch-compatible dataset classes.

After reconstruction, the local dataset directory has the following structure:

sable_v0.1.0/
  metadata/
    metadata.csv
    metadata.parquet
    metadata_schema.json
    column_descriptions.md
  splits/
  sequences/
  structures/pdb/
  manifest.json
  checksums.sha256
  README.md
  LICENSE-DATA.txt