Methods

Sequence inspection and calculated descriptors

Input text is retained separately from canonical sequences in the analysis provenance. A single FASTA header, whitespace and letter case can be normalized; unsupported residues and malformed records fail explicitly. No residue is substituted. Numbering must map every input residue back to its exact amino acid and reproduce the stated CDR coordinates before a construct is returned.

Mass is the sum of IUPAC average amino-acid masses minus 18.0153 Da for each peptide bond. We report mature, reduced, unmodified polypeptide mass. Glycans, terminal processing and disulfide connectivity are not inferred. The construct mass is the stoichiometric sum of chain masses, not a mass-spectrometry observation.

Net charge is calculated at pH 7.4 using Henderson–Hasselbalch fractions, Bjellqvist pKa parameters and sequence-dependent terminal parameters. The theoretical pI is the zero-charge root on pH 0–14; no root produces an abstention. Mean hydropathy uses the Kyte–Doolittle scale. Composition and cysteine counts are exact sequence counts. Extinction coefficients assume reduced cysteines. All values are deterministic descriptors under stated assumptions; no experimental interval or solubility, stability, aggregation or yield prediction is supplied. Independent calculation reference: Biopython 1.86.

The numerical implementation was compared with independently generated Biopython outputs on a predefined 64-case panel: four public-example domains, five constant segments, 20 single-residue edge cases and 35 seeded synthetic chains. All 64 cases met prespecified tolerances: mass 10⁻⁶ Da, charge 10⁻¹² e, pI 10⁻⁴ pH units and mean hydropathy 10⁻¹². This is software agreement, with no inference about performance on biological outcomes.

A separate independently specified Python composition panel checks 152 configurations in four families: Tandem scFv (32), CrossMab KiH (6), DVD-Ig (18), and IgG–scFv (96). All exact chain hashes, lengths, copy counts and parent-domain preservation checks passed. This supports sequence-construction correctness in those configurations; it does not measure assembly yield or validate every library configuration.

Potential chemical motifs and context

The read-only sequence inspector uses explicitly stated motif definitions informed by the Liability Antibody Profiler paper (Satława et al., 2024). It displays exact chain positions, inherited binder versus constant/linker/boundary origin, and CDR or framework context when numbering is available. Unassigned context remains unassigned. A motif outside a published definition’s CDR scope is displayed without transferring the CDR-specific risk interpretation.

This is not a reproduction of LAP’s complete risk adjustment: no germline prevalence, therapeutic prevalence or surface accessibility is calculated. Methionine and tryptophan occurrences are potential sequence flags, not observations of oxidation. The added inspector does not change engineering tiers; existing introduced-junction rules remain separate from inherited sequence annotations.

Module J · junction inspection

Engineered junctions compared against natural flexible linkers. This is a sequence observation, not an engineering tier or a prediction of a bispecific outcome. It does not affect candidate ordering.

Natural-boundary reference

We use gaps of 3–100 residues between immediately adjacent, non-overlapping Pfam Domain annotations in reviewed UniProt proteins. Full-protein sequence checksums must match the Pfam release. Unclassified boundaries, overlapping domains, non-standard residues and ambiguous cross-stratum windows are excluded. Each retained window has 15 residues on either side of its boundary. The reference is natural interdomain boundaries, not bispecific outcomes.

Disorder proxy: metapredict distribution 3.1.1, V3 network, checkpoint 3209c42d91a77dc5828618c7835ebc5dd912f56d43d0111d3c3c36138d9739a8. Residues at or above 0.5 use the documented disorder threshold. A gap is flexible-proxy if at least 80% of its residues meet that threshold; structured-proxy if at most 20% do. The 80% homogeneity rule is a stated method choice, not a calibrated biological threshold. Low predicted disorder does not establish mechanical rigidity. metapredict documentation.

Before proxy inference, 25,000 proteins were selected by a reproducible accession hash from 74,944 eligible proteins. After inference and deduplication, the same hash-sampling rule and 25,000-window cap applied to both strata. All 3029 flexible windows survived that cap; the structured population was sampled from 65169 to 25000. The strata must not be interpreted as uncapped population proportions.

Reference dependence and resolution

The flexible reference contains 3,029 windows, 1,554 distinct gaps, 1,390 source proteins and 916 protein clusters at 50% identity. There are 1,475 gaps with both boundaries and 79 with one retained boundary. Paired boundaries share a gap; multiple gaps can share a protein; homologous proteins can share a cluster. The windows are not independent samples.

Full source proteins were clustered with MMseqs2 at 50% aligned-residue identity, requiring 80% bidirectional sequence coverage, using a single-step connected-component graph and sensitivity 7.5. Connected components include transitive relationships; every pair within a cluster need not share 50% identity. Software version 21fbeafa3eaed712b82b5743b78f1a36690e7d51. This full-length coverage rule can miss shared local or domain homology. Cluster counts summarize redundancy and do not establish independence or a formal effective sample size. MMseqs2 method and parameter definitions.

The reported percentile remains a descriptive window-weighted empirical rank. Its arithmetic step size is not independent biological resolution. Independent information is bounded by the underlying gaps and further limited by protein and homology dependence; 3,029 windows do not provide 3,029 independent observations. No confidence interval or inferential precision is calculated from the window count. Any future uncertainty analysis must resample whole protein clusters and retain their gaps and paired boundaries together.

Fallback (a): only the flexible stratum supplies percentiles. Gly-Ser junctions use their matching boundary direction, never the pooled corpus: 1513 domain-to-linker windows and 1516 linker-to-domain windows. Other boundary chemistries abstain when no matching comparison is available.

Sources: Pfam 38.2, release 2026-01 (month precision), based on UniProt 2025_03; UniProtKB 2026_03, release 02-Sep-2026. Extraction commit d31e7823289f7db3321ce65568f623b6e3f428be; window artifact SHA-256 fbd72c7a36033224250b8db217a54270c51e0a50a373e1b4183536a75f54b33e. Source checksums, proxy parameters and counts are recorded in each run manifest.

Masked likelihood and its limits

ESM-2 esm2_t33_650M_UR50D, without fine-tuning, masks each residue in turn. Pseudo-perplexity is the exponential of the mean negative log-probability assigned to the actual masked residues. Masks are batched; BOS/EOS tokens are not scored. Higher values mean the sequence is less expected by this language model. The three largest negative log-probabilities identify contributing residues, with chain and sequential coordinates. All residues in the window provide context; a window entirely inside a preserved parent domain is never scored.

Reference and candidate windows require the identical checkpoint and inference policy. Cache keys include the window hash and that policy. A percentile is withheld until all 3,029 flexible windows have real scores. It is the fraction of matching-direction natural reference windows with strictly lower pseudo-perplexity; ties are not broken. The empirical reference distribution accompanies the percentile. Predictive performance for expression, aggregation or pairing has not been established for any threshold. The percentile is reported for inspection only and does not affect ordering. ESM-2 method.

Comparison with Phase 1 liabilities

Convergence requires an exact overlap on the same chain and sequential residue coordinate with an actual Phase 1 liability. In the uncalibrated raw-coordinate audit of public-example Tandem scFv BSD2-000181, none of the 18 reported contributors across six boundary observations overlaps an actual Phase 1 liability. W110, M107 and the other contributors lie in preserved parent segments, which Phase 1 excludes from junction penalties. Methionine or tryptophan identity alone is not convergence. ESM surprisal is not an oxidation detector; even an overlap would not validate chemical reactivity or a bispecific outcome. This audit does not require or report a reference percentile; calibration remains withheld until the complete reference batch is verified.