Structure-based Antibody Renumbering

SAbR (Structure-based Antibody Renumbering) assigns antibody residue numbers from backbone coordinates. It combines a fine-tuned version of the ProteinMPNN encoder from SoftAlign with the original affine Smith–Waterman alignment and ANARCI numbering rules.

The complete usage guide is available in the SAbR documentation.

Installation

SAbR requires Python 3.11 or newer. Both sabr and softalign modes run on CPU using NumPy and SciPy

pip install sabr-kit

Docker

With Docker installed and running, pull the image and run SAbR from the directory containing antibody.pdb:

docker pull ghcr.io/delalamo/sabr:latest
docker run --rm -v "$PWD:/workspace" ghcr.io/delalamo/sabr:latest \
  -i antibody.pdb -c H -o numbered.pdb

The mounted directory makes the input available inside the container and saves numbered.pdb back to the same directory on your computer. The image runs sabr directly; append any of the command-line options below, or use docker run --rm ghcr.io/delalamo/sabr:latest --help for help. To pin a release, replace latest with its version tag (without the leading v).

Command line

sabr -i antibody.pdb -c H -o numbered.pdb

The complete interface is:

sabr -i INPUT -c CHAIN -o OUTPUT
     [-n imgt|chothia|kabat|martin|aho|wolfguy]
     [-t auto|TYPES]
     [--noise-level 0.0|0.2|0.5|1.0|2.0]
     [-m sabr|softalign]
     [--residue-range START END]
     [--scfv]
     [--no-mmcif]
     [--overwrite] [-v]

Defaults are IMGT numbering, automatic H/K/L selection, noise level 0.0, sabr mode, and the entire selected chain. --chain-type accepts a comma-separated set of candidates. Each candidate is a sequence of H, K, and L domains: H,K tries heavy and kappa single domains, HK,HL tries heavy-kappa and heavy-lambda two-domain chains, and HHK,HHL tries the corresponding three-domain chains. Existing outputs are never replaced unless --overwrite is given. Normal output contains only warnings and errors; -v reports reference scores and pipeline decisions.

Use --mode softalign to select the original SoftAlign encoder weights, reference embeddings, and affine gap penalties together. SoftAlign references do not vary with --noise-level, so that option is ignored in this mode.

Use --scfv to search only the scFv candidate set. It is equivalent to --chain-type HK,HL,KH,LH and still requires the default automatic chain type when used as a flag. Each domain is numbered in the selected scheme, then successive domains receive offsets of 0, 1000, 2000, and so on. For IMGT, this gives domain ranges 1–128, 1001–1128, 2001–2128, and so on. Linker residues continue sequentially from the preceding domain's last number. Multi-domain references use the selected parameter mode, so either form can be combined with --mode softalign.

Input and output may be PDB (.pdb) or mmCIF (.cif or .mmcif). When a requested PDB output needs multi-character insertion codes, SAbR warns and automatically writes it to the corresponding .cif path instead. Pass --no-mmcif to forbid this conversion and fail. Other values that exceed PDB field limits still require an explicitly named mmCIF output. Writes are atomic, so a failed run does not leave a partial output.

CLI conversion guarantees preservation of atomic structure content, not arbitrary non-atomic mmCIF categories. It warns for every mmCIF input.

Python API

from Bio.PDB import PDBParser
from sabr import renumber_structure

structure = PDBParser(PERMISSIVE=False, QUIET=True).get_structure(
    "antibody", "antibody.pdb"
)
numbered = renumber_structure(structure, chain="H")

Strict PDB parsing matches the CLI and rejects malformed duplicate atom records before renumbering.

renumber_structure accepts a Biopython Structure, never mutates its input, and returns a new Biopython Structure. Non-target chains, hetero residues, waters, and residues outside an inclusive residue_range are preserved. SAbR rejects multi-model structures rather than silently modifying only one model. If a partial range would create duplicate residue IDs with unchanged residues, the operation fails with an explanation.

For unusually long loops that need extended insertion codes, use mmCIF output.

IMGT, Kabat, Chothia, Martin, and AHo numbering support up to 5,000 insertion codes per anchor (A–Z, AA–ZZ, then three-letter codes). Wolfguy retains its fixed loop positions: 49 residues per loop, except heavy-chain CDR3 (51). Unsupported lengths raise an error with the scheme, loop, observed length, and supported limit. These are SAbR implementation limits; mmCIF does not remove them or the 1,024-residue structure-selection limit. See long-CDR numbering support for the complete scheme-specific limits.

Scientific behavior

• The default sabr mode preserves the trained SAbR encoder weights, references, and gap penalties unchanged.
• The optional softalign mode uses softalign_encoder.npz, softalign_embeddings.npz, and the exact penalties in softalign_gap.npz as one parameter set.
• Alignment uses the original affine Smith–Waterman recurrence, with soft assignments computed by an explicit NumPy reverse pass.
• Deterministic CDR gap distribution and DE-loop correction are always applied. Between IMGT anchors 79 and 85, DE-loop residues fill 80 first, then 84 back through 81; additional residues are inserted after 82.
• Automatic chain selection aligns against H, K, and L references and uses the highest score, with deterministic H/K/L tie order.
• chain_type candidate order is deterministic and resolves score ties.
• scFv mode searches only the HK,HL,KH,LH candidate list in that order.
• Multi-domain candidates do not apply gap-open or gap-extension costs to query linker residues aligned at any boundary between domain references.
• Multi-domain candidates receive normal affine gap-open and gap-extension costs for unaligned query and reference termini when their selection scores are compared; the underlying alignments and raw alignment scores are unchanged.

A structural gap is detected when the C–N distance between consecutive residues exceeds 2.66 Å. SAbR handles gaps automatically by waiving reference-deletion opening costs at each detected boundary; extension costs remain unchanged. For CDR and DE-loop numbering, SAbR keeps the structural anchors and counts unresolved residues declared in PDB REMARK 465 or mmCIF _pdbx_unobs_or_zero_occ_residues metadata between them. These residues occupy sequence positions through scheme conversion, including insertion codes, but no atoms are created. Author-number jumps alone are never treated as missing sequence. Regions with missing anchors or breaks not explained by metadata retain the existing alignment fallback; numbering in those regions can still be inaccurate.

The CLI reads this metadata automatically with Biopython's MMCIF2Dict. For in-memory mmCIF inputs, use the standard MMCIFParser and supply the metadata dictionary explicitly (using the default author chain/residue IDs):

from Bio.PDB import MMCIFParser
from Bio.PDB.MMCIF2Dict import MMCIF2Dict
from sabr import renumber_structure

structure = MMCIFParser(QUIET=True).get_structure("input", "input.cif")
numbered = renumber_structure(
    structure, chain="H", mmcif_dict=MMCIF2Dict("input.cif")
)

PDB inputs from Bio.PDB.PDBParser already retain REMARK 465 in their headers. Explicit mmcif_dict data takes precedence over header records; it must describe the input structure's author IDs. Neither input is modified. The returned structure drops missing-residue header records for the renumbered chain because their original author IDs are no longer valid.

T-cell receptors are not an officially supported SAbR target. For experimental use, pass the actual TCR chain type (A, B, G, or D) with --chain-type or chain_type. SAbR aligns only against the K reference, which includes IMGT position 10 as TCRs do, while retaining the TCR type for ANARCI conversion. This workflow is limited to IMGT and AHo numbering; the other bundled schemes are antibody-specific.

Development

pip install -c constraints.txt -e '.[test]'
pytest
pre-commit run --all-files

constraints.txt records the exact canonical development and CI environment. Package metadata remains ranged for normal installation.

The committed tests are self-contained and never download data. They verify the fixed asset hashes, encoder and alignment baselines, all numbering schemes, H/K/L selection, regional corrections, structure-object behavior, and CLI failure handling.

Independent antibody structure tests use PDB fixtures and SAbDab IMGT labels in tests/data/antibody. They check residue numbering, domain boundaries, coordinate and input-numbering invariance, and real scFv behavior.

License and attribution

SAbR is distributed under the repository license. The vendored ANARCI numbering code retains its original license in src/sabr/_anarci/LICENSE.

If you use SAbR, please cite:

@article{delAlamo2026,
  title = {Structure-based Antibody Renumbering},
  url = {http://dx.doi.org/10.64898/2026.09.16.751788},
  DOI = {10.64898/2026.09.16.751788},
  publisher = {openRxiv},
  author = {del Alamo,  Diego},
  year = {2026},
  month = Sept 
}