Antibody Engineering

Antibody engineering

Sequence-level analysis for antibodies, nanobodies, and other variable-domain formats: numbering, CDR annotation, chemical liabilities, and physicochemical properties. All of it runs in seconds and rules out a surprising fraction of problems before a model or a wet-lab week is spent.

Read [references/numbering-schemes.md](references/numbering-schemes.md) before quoting any residue position, [references/developability.md](references/developability.md) before acting on a liability, [references/humanization-and-design.md](references/humanization-and-design.md) for grafting and humanness, and [references/tools.md](references/tools.md) for the wider ecosystem.

A residue number means nothing without its scheme

"Residue 52" is a different residue in IMGT, Kabat, and Chothia numbering, and the CDRs they define overlap only partially. The same trastuzumab heavy chain:

IMGT    CDRH1 GFNIKDTY (8)   CDRH2 IYPTNGYT (8)            CDRH3 SRWGGDGFYAMDY (13)
Kabat   CDRH1 DTYIH    (5)   CDRH2 RIYPTNGYTRYADSVKG (17)  CDRH3 WGGDGFYAMDY   (11)

Neither is wrong. Use IMGT by default — one definition for both chains, structurally principled gaps, and the germline database is IMGT-numbered — and convert to Kabat when matching legacy literature. State the scheme every time.

python skills/antibody-engineering/scripts/number_antibody.py antibody.fasta
python skills/antibody-engineering/scripts/number_antibody.py antibody.fasta --scheme kabat
python skills/antibody-engineering/scripts/number_antibody.py antibody.fasta \
    --format regions --out regions.tsv

# trastuzumab_VH: chain H, closest germline human_H (human), E=3e-60
#   variable domain spans input residues 1-120
  CDRH1	8	GFNIKDTY
  CDRH2	8	IYPTNGYT
  CDRH3	13	SRWGGDGFYAMDY

Needs pip install anarci plus HMMER (hmmscan on PATH). Note that ANARCI's species call is the closest germline, not an annotation — a humanised antibody reports human because its frameworks are human, which says nothing about its CDRs.

Liabilities, weighted by region

python skills/antibody-engineering/scripts/scan_liabilities.py antibody.fasta \
    --regions regions.tsv --min-severity high

# trastuzumab_VH: 120 residues, 3 finding(s)
  [critical] deamidation (NG) 'NG' at 55 (CDRH2)
             the fastest-deamidating motif; Asn -> iso-Asp/Asp changes charge and can
             abolish binding, and it is the usual cause of potency loss on storage
  [critical] isomerisation (DG) 'DG' at 102 (CDRH3)
             Asp-Gly isomerises to iso-Asp through a succinimide intermediate

Those are trastuzumab's two documented hotspots, found from sequence alone.

Pass --regions. The same NG in framework 3 is usually buried and tolerated; in CDR-H2 it is a redesign candidate. Without region information every finding is reported at the framework baseline, and the script says so.

Motifs covered: N-glycosylation sequons (N-X-[ST], X≠P), deamidation (NG ≫ NS/NT/NN/…), isomerisation (DG ≫ DS/DT/…), acid-labile DP fragmentation, Met and Trp oxidation, unpaired and extra cysteines, N-terminal pyroglutamate, and the RGD/RYD integrin motifs.

A liability is a question, not a veto. Many approved antibodies carry known liabilities and manage them with formulation and release specifications. What settles it is a force-degradation study, not a prediction.

Physicochemical profile

python skills/antibody-engineering/scripts/physchem_profile.py antibody.fasta --combine

# trastuzumab_VH: 120 residues
  molecular weight     13164.7 Da
  isoelectric point    8.17 (EMBOSS pKa set)
  net charge at pH 7.4  +0.89
  extinction (280 nm)  35535 /M/cm (cystine)
  A280 at 1 mg/mL      2.699
  GRAVY                -0.305

• pI drives purification and formulation. Formulate at least a unit away from it — near-zero

net charge means poor colloidal stability, and the script warns when the two are close.

• Net charge at pH 7.4 above roughly +6 associates with fast clearance and polyspecificity in

the published developability sets.

• Extinction coefficient is what turns A280 into a concentration; getting it wrong scales

every downstream number including your affinities.

• Different pKa sets shift pI by a few tenths. The script uses EMBOSS and says so; quote the set.

The order that saves time

1. number_antibody.py --format regions — everything downstream needs regions.
2. scan_liabilities.py --regions — seconds, catches the classics.
3. physchem_profile.py — pI, charge, extinction coefficient.
4. Model the Fv (ABodyBuilder3, IgFold, or boltz) — needed for anything conformational.
5. Structure-based properties: TAP metrics, hydrophobic and charged patches.
6. Humanness, if the molecule is not already human.
7. Test: force degradation, SEC, DSF, HIC, AC-SINS, PSR.

Steps 1–3 cost seconds. Do them before spending a GPU hour.

What sequence cannot tell you

Aggregation, viscosity, polyspecificity, and thermal stability are conformational, and none of them follow from motifs. They need a structure — hydrophobic patch area across the VH/VL interface predicts aggregation and HIC retention far better than GRAVY does — or an experiment. Reporting a clean liability scan as "developable" is the mistake this skill is meant to prevent; say "no sequence liabilities detected, structure-based properties not assessed".

Antibody–antigen complex prediction is also still genuinely hard for every current method, because the interface is a rearranged loop rather than a conserved surface. Check ipTM before believing a predicted complex.

Composing with the rest of the bundle

• glycoengineering — the sequons this skill flags, in depth: occupancy, glycoform engineering,

effector-function consequences.

• esm — language-model scoring for affinity maturation and humanness.
• boltz — antibody–antigen cofolding when you need the complex.
• uniprot-rcsb — antigen sequence and structure; SAbDab entries are PDB entries.
• adaptyv — submit designs and get measured binding and thermostability back.
• open-targets — whether the antigen is validated and accessible to a biologic; its AB

tractability buckets answer exactly that.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: K-Dense-AI
• Source: K-Dense-AI/drug-discovery-agent-skills
• License: MIT
• Homepage: www.k-dense.ai

Install and usage instructions live in the source repository linked above.