[0, 1]; higher means greater model confidence in the generated chain arrangement; no binding cutoff

Interface predicted TM-score. It describes model confidence in the relative chain arrangement of a generated pose, not binding probability, affinity, or specificity.

Output of AlphaFold-Multimer / Boltz-2 / Protenix, aggregated from predicted inter-chain positional-error information into one model-confidence score.

A non-binder or incorrect biological pairing can still receive high ipTM, while a real binder can receive a low score. Construct boundaries, model version, and seed can change the result.

Useful for same-protocol pose triage. Multi-model or multi-seed agreement can expose sensitivity but cannot prove that an agreed pose is biologically correct.

Rank generated-pose confidence only under the same protocol. Do not rank binding, affinity, specificity, or candidate advancement from ipTM alone; require orthogonal experiments.

[0, 1]

Arithmetic mean of Boltz-2 and Protenix ipTM. It summarizes model-reported pose confidence, not binding probability.

mean(boltz_iptm, protenix_iptm). Future versions will weight by per-model reliability.

Averaging can hide disagreement, and correlated models can agree on an incorrect or non-binding pose.

A pose-triage summary only. Pair with explicit model/seed agreement and inspect structures; require experiments for biological claims.

Use only to compare pose confidence under a fixed protocol, not to rank affinity, specificity, or probability of binding.

[0, 1] when reported; interpretation requires a versioned method

A model-to-model agreement feature. Scalar ipTM similarity, contact-set overlap, and coordinate convergence are different definitions and must not be mixed.

Read the stored method/version with the value. If provenance is absent, treat the number as legacy/undefined rather than assuming an ipTM-difference or Jaccard formula.

Models can share a failure, and similar scalar confidence does not imply similar coordinates. Low agreement does not reveal which model is closer to reality.

Useful for flagging model sensitivity, not as evidence of binding or correctness.

Use as a manual-review flag alongside coordinate/contact convergence. Do not automatically drop or advance candidates from this scalar.

[0, 100]; <50 disordered; 50-70 low; 70-90 confident; >90 very high

Per-residue confidence in atomic positions. AlphaFold's own granular signal.

Model output, distilled from training-set positional accuracy.

Mean pLDDT masks region-specific issues. CDR-H3 natively 60-80 is normal; framework at 70 indicates structural problems. Doesn't capture rigidity vs flexibility.

Well-validated against PDB. For antibody framework, expect >90 as baseline; <80 is a red flag.

Report per-region (FR / CDR-H1/H2/H3 / LCDR) when possible. Mean pLDDT alone is a weak signal for antibodies.

[0, 30+] Å; lower means the model reports less inter-chain positional uncertainty

Predicted aligned error over selected inter-chain pairs. It describes uncertainty in generated geometry, not whether the molecules bind.

AlphaFold / Boltz output. Mean PAE over inter-chain residue pairs at the interface.

A low value can occur for a confident but biologically incorrect pose. The selected interface mask and model version affect the summary.

Use with ipTM, model/seed convergence, and structure inspection as pose-confidence diagnostics; none establishes binding.

Lower can prioritize poses for review under the same protocol. There is no validated iPAE threshold for real binding or specificity.

[0, 1]; raw RF2 classifier output, not a calibrated binding probability

An interaction-classifier score emitted by RF2. It is an exploratory model feature, not an answer to whether a binder will bind an antigen.

Classifier head on RF2 features, returning one scalar per modeled pair.

Dataset shift and calibration error can produce high scores for non-binders and low scores for binders; the value is not K_D or specificity.

Use only as a same-protocol review feature and retain model/version provenance.

Do not use a universal threshold or advance candidates from p_bind alone. Require orthogonal binding and specificity experiments.

Ų; the current Interface report uses partner-averaged buried area

Surface area occluded at the interface in the predicted pose. It describes modeled geometry, not whether binding occurs or how tight it is.

Shrake-Rupley SASA: reported BSA = [SASA(Ab alone) + SASA(Ag alone) - SASA(complex)] / 2. Per-partner buried areas may also be reported.

A wrong or experimentally non-binding pose can still have a large BSA. BSA alone does not establish affinity, specificity, or promiscuity.

Use as a pose-geometry diagnostic alongside PAE/ipTM, model or seed agreement, and contact inspection; experimental calibration is required for biological claims.

Do not rank affinity from BSA alone. Compare values only under the same prediction and SASA protocol, then inspect the modeled interface.

Lower = more antibody-like; typical 2-6 for human repertoire sequences

How surprised an antibody language model is by your sequence. Low perplexity = looks like a natural human antibody.

AbLang model forward pass; perplexity is exp(mean negative log-likelihood) over the sequence.

Designed sequences may be 'too natural' (memorizing repertoire) or unusually low for de novo binders — interpret in context.

Strong prior for expressibility. Sequences with perplexity above the repertoire distribution are higher risk of expression failure.

Use as a developability prior. Heavy chain (VH) perplexity carries more weight than light (VL) for VHH and scFv.

Per-metric — TAP returns continuous values plus PASS/AMBER/RED flags

Five canonical developability filters: total CDR length, patches of surface hydrophobicity (PSH), patches of positive charge (PPC), patches of negative charge (PNC), structural Fv charge symmetry parameter (SFvCSP). Hardcoded gates derived from clinical-stage antibody distributions.

TAP 2.0; structural analysis of Fv with patch detection. Compares against the clinical-stage antibody (CST) distribution.

Calibrated on traditional mAbs; VHH and engineered scFv distributions differ. Some 'red' flags are false positives for unconventional formats.

Industry-standard for IgG developability assessment. Take 'red' flags seriously; investigate per-flag rationale.

Use as hard gates: any TAP RED flag → drop unless rescued. For mAb format, all five metrics in CST range = developable.

kcal/mol; negative = stabilizing; -2 to +2 typical range for designs

Predicted stability change from mutations. Negative = the mutation stabilizes the fold.

ThermoMPNN forward pass on each mutation site, comparing wt vs mutant energy.

Trained on monomer thermal denaturation data; less calibrated for interface stability or chronic-expression contexts.

Strong directional signal but absolute ΔΔG values have ~1 kcal/mol error.

For each candidate, compute mean predicted ΔΔG; prefer negative averages. Use as a tie-breaker after interface metrics.

Continuous; lower = less aggregation propensity

Structure-aware aggregation prediction. Identifies aggregation-prone patches on the folded structure.

AggreScan3D rolls a hydrophobicity-weighted score along the molecular surface; reports aggregation hotspots.

Static structural — doesn't capture dynamic exposure of hydrophobic patches at higher temperature.

Standard tool; correlates moderately with measured aggregation propensity.

Lower is better. Pair with TAP PSH for a more complete aggregation picture.

[0, 1]; >0.85 native-human-like; <0.6 likely to elicit ADA

How similar your sequence is to the natural human antibody repertoire.

OASis: percentile-rank of 9-mer subsequences vs the Observed Antibody Space database.

Excellent for traditional mAbs; less validated for VHH (camelid origin). High humanness doesn't guarantee no ADA — clinical ADA is multifactorial.

Industry standard. Strongly correlates with clinical immunogenicity for traditional antibodies.

Higher is better. For VHH, supplement with cross-reactivity counter-screen against human Ig.

[0, 1]; lower = fewer MHC-I-binding epitopes

Predicted T-cell epitope load. High = more MHC-I-presentable peptides → higher immunogenicity risk.

DeepImmuno predictor over all 9-mer fragments of the binder.

Only covers MHC-I (CD8 response). Doesn't capture B-cell ADA or MHC-II / CD4 responses.

Useful supplement to humanness; not a substitute for full immunogenicity panel.

Lower is better. Combine with humanness to triage.

Boolean (pass / fail)

ANARCI-based check that CDR lengths fall within canonical IMGT ranges.

ANARCI numbering → extract CDR-H1/H2/H3 + L1/L2/L3 lengths → compare against IMGT canonical ranges.

Length-only check; doesn't validate loop conformation, just that the length is plausible.

Hard gate — failures usually indicate the design has produced something that won't fold like an antibody.

Boolean filter. Drop candidates with cdr_geometry_ok = false unless intentionally non-canonical.

Std-dev in residue positions; typical 1.5-4 for good binders

How spread out the antigen-contacting residues are along HCDR3. Concentrated contacts can indicate a non-canonical paratope.

ANARCI annotate → identify HCDR3 residues → find contacts (within 5 Å of antigen heavy atoms) → std-dev of contact positions.

5 Å threshold is arbitrary; doesn't distinguish productive contacts from incidental ones. Heavy-atom-only — ignores hydrogen-bond geometry.

Useful sanity check; not a primary ranker.

Prefer 1.5-4. Very low (<1) means contacts cluster on one or two residues (possibly non-canonical). Very high (>4) means HCDR3 is so long the spread is meaningless.

Count of motif matches; 0 = clean; >0 in CDR = red flag

N-glyco motif N-X-S/T (X != P). In CDR = developability red flag for mAb format (affects PK, aggregation).

Regex motif scan or Tamarind n-linked-glycosylation tool. CDR overlap computed via ANARCI numbering.

Motif scan flags potential, not confirmed glycosylation. Actual glycosylation depends on cell line and local context.

Conservative — flags more than actually glycosylate. Treat CDR hits as hard fails for mAb format; framework hits as soft warnings.

0 in CDR = preferred. Any CDR hit for mAb format = drop unless intentional.

MW: mAb ~150 kDa, scFv ~25, VHH ~13. pI: 6-9 typical; <6 or >9 = formulation challenges

Molecular weight (kDa) and isoelectric point (pI). Basic format-fit + PK proxy.

BioPython ProtParam — exact MW from residue composition; pI from titration curve.

Doesn't capture glycosylation effects on apparent MW. pI doesn't tell you about local charge distribution (use TAP PPC/PNC for that).

Exact computation; reliable. Limited interpretive power on its own.

Use as a format-fit hard gate (e.g. for in-vivo CAR mRNA, total CAR construct ≤ 1.5 kb). pI in 6-9 range is preferred for manufacturability.

Planned Fc–FcRn complex ipTM; no validated affinity or half-life threshold

Model confidence for a generated Fc–FcRn arrangement. It does not measure pH-dependent affinity, recycling, pharmacokinetics, or serum half-life.

Planned Boltz-2 prediction of an Fc + FcRn complex; backend is not wired. A sequence-only structure prediction does not itself encode or validate endosomal pH behavior.

Even a confident pose may be non-binding or have the wrong pH dependence. Fc sequence, glycosylation, kinetics, exposure, and clearance all affect PK.

Unvalidated against SPR/BLI or PK data. Treat only as a future structure-review feature.

Do not rank half-life or choose Fc engineering from ipTM. Require pH-dependent FcRn binding assays and PK evidence.

Per-ortholog ipTM; ratio = cyno_iptm / human_iptm (or analogous); no validated cutoff

Compares model confidence for generated poses against antigen orthologs. It does not determine cross-species binding or pharmacology-species suitability.

Re-predict the complex with cyno / mouse antigen ortholog (from UniProt). Report per-species ipTM + retention ratio. Status: data file + runner pending.

Ortholog construct boundaries, model bias, and different generated poses can dominate the ratio. Similar scores can occur without shared binding, and dissimilar scores do not prove loss of binding.

An uncalibrated model-review signal only. It cannot establish toxicology relevance or predict whether cyno/mouse PK mirrors human PK.

Do not use the ratio as an IND gate. Confirm ortholog binding, kinetics, epitope conservation, function, and PK with appropriate experiments.

Δ ipTM = target_ipTM - max(homolog_ipTM); no validated specificity threshold

Difference between target and homolog pose-confidence outputs under one modeling protocol. It is not measured selectivity or proof of target preference.

Run the same model, constructs, seeds, and settings on (binder, target) and (binder, each homolog); report the ipTM gap. Example BCMA panel: TACI, BAFF-R.

Model/construct bias, incomplete off-target panels, and different pose modes can create or erase the gap. A confident off-target pose is not proof of cross-binding.

Use as a manual-review flag only. Experimental selectivity and functional counter-screen panels are required.

Do not apply a universal cutoff or automatically redesign/deprioritize from ΔipTM. Compare only under the same protocol and seek orthogonal evidence.

[0, 1]; lower = safer; <0.3 acceptable, >0.6 high risk

Composite proxy for binder-driven CAR clustering on T-cell surface without antigen present. Drives tonic signaling (T-cell exhaustion).

Composite: surface hydrophobicity + net charge at pH 7.4 + Fv-Fv self-association. Heuristic, NOT validated against experimental data.

First-principles heuristic with no clinical or experimental calibration. Marked as 'v0' for a reason.

Use as a soft warning only. Don't kill candidates on this score alone until calibrated against IASO internal data.

Lower is better, but treat as a flag, not a hard filter.

[0, 1]; higher = more CAR-T-compatible (hand-tuned)

Aggregate of size + stability + expression prediction + aggregation + immunogenicity for CAR-T format-fit.

Hand-tuned weighted composite. Not learned, not calibrated against clinical CAR-T performance.

Public literature has no validated CAR-fit predictor. Our composite weights are best-guesses.

Treat as 'heuristic v0'. Useful for quick triage; not a substitute for wet-lab CAR-T validation.

Higher is better, but always check the underlying components separately.

Positive integer; 1 = best (non-dominated), higher = dominated by some candidate

Multi-objective rank: a candidate is Pareto rank 1 if no other candidate is better on every dimension simultaneously.

Non-dominated sort across included model dimensions: pose confidence (ipTM), stability (ddG), humanness, developability, immunogenicity, and aggregation.

The frontier inherits every input limitation: ipTM is not binding or affinity evidence, and highly correlated dimensions can distort the frontier.

Mathematically clean but UX-dangerous if interpreted as 'rank 1 = best candidate'.

Use only for exploratory multi-model triage. Re-rank biological decisions with calibrated affinity and orthogonal experimental evidence—not ipTM alone.

[0, 1]; higher = better overall

Weighted linear combination of normalized score dimensions per scoring preset (permissive / standard / strict).

Sum of weight_i × normalized(score_i). Default weights in scoring_stage.py DEFAULT_WEIGHTS.

Linear combination hides which dimension drove the score. Two candidates with same composite can be very different.

Convenient single-number ranker but always drill into components before promoting a candidate.

Higher is better. Use as the default sort on /candidates; always cross-reference Pareto rank + individual dimensions.