SAbR

SAbR (Structure-based Antibody Renumbering) assigns antibody residue numbers from backbone coordinates. It combines the original trained Haiku encoder with the original affine Smith–Waterman alignment and ANARCI numbering rules.

SAbR is intentionally small and feature-complete. It provides one Python API and one command-line program.

The complete usage guide is available in the SAbR documentation.

Installation

SAbR requires Python 3.11 or newer.

pip install sabr-kit

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|H|K|L]
     [--noise-level 0.0|0.2|0.5|1.0|2.0]
     [-m sabr|softalign]
     [--residue-range START END]
     [--scfv]
     [--overwrite] [-v]

Defaults are IMGT numbering, automatic H/K/L selection, noise level 0.0, sabr mode, and the entire selected chain. 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 for a single chain containing two linked variable domains. In addition to the H, K, and L references, this mode tries the concatenated H:K, H:L, K:H, and L:H representations. The second domain is numbered with a 128 offset so both domains have unique residue IDs in one structure chain; linker residues use insertion codes after the first domain. Because each composite already specifies both domain types, scFv mode requires automatic chain type. Composite references use the selected parameter mode, so --scfv can be combined with --mode softalign.

Input and output may be PDB (.pdb) or mmCIF (.cif or .mmcif). Use mmCIF when chain names or ANARCI insertion codes exceed PDB's one-character fields. 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. When those categories matter, load a Gemmi structure and use the in-memory API.

Python API

from Bio.PDB import PDBParser
from sabr import renumber_structure

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

Gemmi structures use the same function:

import gemmi
from sabr import renumber_structure

structure = gemmi.read_structure("antibody.cif")
numbered = renumber_structure(
    structure,
    chain="heavy_chain",
    scheme="chothia",
    chain_type="auto",
    noise_level=0.0,
    mode="softalign",
    residue_range=None,
    scfv=False,
)

renumber_structure never mutates its input and returns the same concrete structure type. 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.

The same-type clone preserves metadata represented by the input BioPython or Gemmi object. Alternate conformers are normalized deterministically: a complete blank-altloc backbone is preferred, then the complete conformer with the greatest summed occupancy, with altloc name as the final tie-breaker. Selections above 1,024 polymer residues are rejected before quadratic model work; use residue_range to select the antibody domain.

Modified peptide residues are translated only for sequence generation. Their original names and atoms remain unchanged. The committed mapping was generated from the wwPDB Chemical Component Dictionary snapshot dated 2026-07-11 (components.cif.gz SHA-256 0b3323123ec10b997afe1c530b4cad30306e60b451b2b062c59bc9bb5cbe0679) and contains only peptide-linking components with exactly one canonical amino-acid parent. Unsupported or ambiguous polymer chemistry fails explicitly; no runtime network access occurs.

Gemmi itself only represents one-character insertion codes. For unusually long loops that need extended codes, use a BioPython structure in memory or the CLI with mmCIF output.

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 differentiable affine Smith–Waterman method.
• In sabr mode, gap extension is -0.175027 and gap opening is -2.525591. In softalign mode, they are 0.1942468136548996 and -2.5441808700561523, respectively, as stored in the repository asset.
• CDR gap distribution is always applied.
• No deterministic light-chain DE-loop or C-terminal correction is applied.
• Automatic chain selection aligns against H, K, and L references and uses the highest score, with deterministic H/K/L tie order.
• scFv mode appends H:K, H:L, K:H, and L:H reference candidates in that order.
• Composite 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. A gap skips only the affected CDR correction and emits a warning; other regions continue normally.

Development

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

constraints.txt records the exact canonical development and CI environment. Package metadata remains ranged for normal installation. SAbR does not force a JAX backend; CPU is simply the canonical CI regression baseline.

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.

Deferred full benchmark

The historical pre-2021 SAbDab manifest contains approximately 1,012 chains. The current method scores about 90% on that set, not 100%. The full corpus is not bundled or downloaded by CI.

Future benchmark work should create a checksum-pinned corpus, verify residue IDs, insertion codes, and coordinate parity, and compare a lossless archive with Foldcomp before adding a separate manual or nightly workflow. This is a benchmarking TODO, not a unit-test or release requirement.

License and attribution

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