How developability results compare to the bench

You have a developability result. Here is what it supports.

• Most of what this tool returns is a hand-written sequence rule, not a published method. tap_score is not the Therapeutic Antibody Profiler — the served code says so in a comment and in a tap_score_method.caveat field on every result. aggregation, immunogenicity, estimated_viscosity_class, charge_patches, clearance_risk and overall_developability_score are threshold rules over Kyte–Doolittle hydropathy and residue-count charge. None of them has a published concordance figure, because none of them is a published method.
• A sequence-motif liability flag is a flag, not a measured degradation rate. Motif-only prediction of Asn/Asp degradation has a published false-positive rate of 0.31 for Asp and 0.41 for Asn on a 37-antibody forced-degradation set. In a 131-antibody set, only 47.0% (8/17) of CDR-H2 NG motifs actually deamidated, only 19.0% (4/21) of CDR-H2 DG motifs isomerized, and 0 of 32 DT and 0 of 6 DH motifs did anything at all.
• ⊘ Every cysteine is flagged critical, including the four conserved framework cysteines. A normal paired VH+VL therefore always carries four critical liabilities and a fixed −0.60 penalty on the composite. On the trastuzumab sequence the repository itself commits as its reference case, the shipped rules return a composite of about 0.29 and developability_class: "problematic". This is a known defect, stated below with the fix it needs.
• Do not compare overall_developability_score between runs that requested different metrics. Every penalty term is conditional on its metric being requested, so asking for less returns a higher score for the same antibody. And no published composite developability score predicts the measured panel: the best-characterised one correlates with the 12 clinical-stage assays at |Spearman ρ| ≤ 0.29.
• Confirm with the developability assay panel: AC-SINS and CIC for self- and cross-interaction, DSF/DSC for Tm, SEC plus accelerated stress for aggregation, HIC for hydrophobicity, a cone-and-plate or capillary viscometer at formulation concentration, and forced degradation with peptide-map mass spectrometry for the chemical liabilities.

Jump to agreement with wet lab or against other in silico methods. The general framing for every tool is in reading an in silico result against the bench, and the warning there about property predictors being graded one step removed from the bench applies to this page directly.

How in silico is used currently

Computed developability is used to order a candidate list and to site-direct a re-engineering question, never to clear a molecule. In a live antibody program that means three things:

• Liability triage before lead selection. You have more candidates than you can express and stress-test, and you want the ones carrying obvious chemical liabilities — an NG in CDR-H2, a DG in CDR-H3, an N-glycosylation sequon in a CDR — pushed down the list or fixed by mutation first. This is the use with the most published support, and it is support for the motif list, not for a score.
• Site-direction for a forced-degradation study. The flag tells the analytical chemist which peptide to look at in the map. Whether the site degrades is what the mass spectrometer says.
• Re-profiling after a sequence change. Humanization and affinity maturation both move charge and hydrophobicity, so the field re-runs the profile after BioPhi or a CDR change. The value here is the delta, which is more defensible than the absolute number, because the same systematic bias sits on both sides of the subtraction.

What the field does not do — and what the published record will not support — is treat a computed developability score as a gate. The most recent blind benchmark says this in as many words: "the true predictive power of models submitted for the competition remains insufficient for reliable, decision-making fidelity."

Against other in silico methods

TAP (Therapeutic Antibody Profiler). The method this tool's tap_score field is named after, and does not implement. Real TAP computes five metrics — total CDR length, patches of surface hydrophobicity (PSH), patches of positive charge (PPC), patches of negative charge (PNC), and the structural Fv charge symmetry parameter (SFvCSP) — four of which require a 3D model, which TAP builds internally with ABodyBuilder before evaluating anything. It also produces no scalar score. It returns green, amber or red flags per metric, with amber set at the extreme 5% of the clinical-stage distribution and red at previously unobserved values; the shipped thresholds are total CDR length > 60, PSH < 83.84 or > 173.85 on the Kyte–Doolittle scale, PPC > 3.16, PNC > 3.50 and SFvCSP < −20.40.

TAP's own validation does not include a single correlation or AUC against a measured assay. It was validated three ways: against the sequence and structure distributions of 242 post-phase-I therapeutics versus a naive human repertoire; against a held-out set of 105 further clinical-stage therapeutics, of which 8 (7.69%) drew a red flag, an average of 0.08 red flags each; and against six antibodies from two programs as case studies, where the flags agreed with observed aggregation and expression phenotypes. The authors are explicit about the limits: "Our simple TAP guidelines will not capture the whole spectrum of developability issues. For example, they will not detect sources of immunogenicity or more subtle mechanisms that lead to poor stability." And: "As with the Lipinski rule of five, the thresholds themselves should not be interpreted as hard-and-fast rules."

Where TAP has been independently benchmarked against measured assays, its hydrophobicity metric came last. On a 37-antibody viscosity set, "TAP's CDR.PSH performed the worst"; on HIC retention for 137 antibodies, "CDR_PSH remains the lowest performing descriptor in this dataset as well"; on a 15-antibody SEC monomer-loss set, CDR_PSH "exhibited poor correlation." Its charge metric fared similarly on polyspecificity: two other charge descriptors "significantly outperformed TAP's CDR.PPC" on heparin retention and PSR. The authors of that evaluation note TAP ran at a protocol disadvantage — an older ABodyBuilder flavour and no conformational sampling — so part of the deficit is parity, not method.

DeepSP. Often cited as a sequence-only aggregation and viscosity predictor, which it is not. It is a CNN surrogate that predicts 30 spatial descriptors (SAP_pos, SCM_neg, SCM_pos across ten domains) from sequence, trained on 20,530 sequences, and its headline average linear correlation of 0.87 (range 0.76–0.96) is against molecular-dynamics-computed descriptors, not against any measurement — and the MD was itself run on ABodyBuilder-ML homology models. Its authors say so: "the goal of this study is not to directly predict stability properties … The aim of this study is to expedite the MD simulation process." Its one measured-assay result is aggregation rate on n = 21 antibodies at 150 mg/mL, where a downstream support-vector model over three of the 30 descriptors reached r = 0.97 in-sample and r = 0.75 under leave-one-out — and where MD-derived features in the same group's prior work scored marginally better out of sample (0.77). It reports nothing against viscosity, Tm, AC-SINS, HIC or SEC.

SAP, SCM and the structure-based surface descriptors. These are the methods a serious aggregation or viscosity descriptor belongs to, and this platform implements none of them. The relevant independent finding is about their stability rather than their accuracy, and it is covered under What the figure doesn't cover.

What is on this platform instead. Humatch and BioPhi answer humanness, not manufacturability, and neither carries a translatability page yet. Paragraph maps the paratope so a liability fix can avoid a binding residue. Nothing on this platform computes TAP, SAP, SCM, DeepSP or a structure-based charge descriptor, so there is no in-house alternative to compare against.

Raybould MIJ, Marks C, Krawczyk K, Taddese B, Nowak J, Lewis AP, et al. Five computational developability guidelines for therapeutic antibody profiling. Proc Natl Acad Sci USA. 2019;116(10):4025-4030. doi: 10.1073/pnas.1810576116

Kalejaye L, Wu IE, Terry T, Lai PK. DeepSP: Deep learning-based spatial properties to predict monoclonal antibody stability. Comput Struct Biotechnol J. 2024;23:2220-2229. doi: 10.1016/j.csbj.2024.05.029

In silico vs. wet lab

The standard wet-lab approach is the developability assay panel, and the first thing to understand is that it is not one technique but a dozen, each answering a different question. The reference panel for this field is Jain et al.'s, which expressed 137 clinical-stage antibody variable regions on a common IgG1 constant region and measured all of them on twelve platforms: AC-SINS and CSI-BLI for self-interaction, PSR, BVP-ELISA, ELISA and CIC for cross-interaction and polyspecificity, HIC and SMAC for hydrophobicity, SGAC-SINS for salt-gradient solubility, SEC under accelerated stress for aggregation, DSF for thermostability, and HEK titer for expression. Chemical liabilities are measured separately, by forced degradation followed by tryptic peptide-map LC-MS/MS.

That panel is the comparator for every figure below. There is no single agreement threshold for this class of output — unlike pose prediction, where RMSD ≤ 2 Å settles it — so agreement is reported as a correlation or an AUC, one descriptor against one assay, and the values differ enormously between them.

Read the accuracy column by cell, never down the column. A measured assay value, a Spearman correlation against that assay, and "no figure exists" are three different kinds of statement. What the table supports is this: seconds and a cent instead of weeks and purified material, at an accuracy that is unmeasured for this implementation and, for the published methods it resembles, tops out around ρ = 0.7 on the single easiest assay and around ρ = 0.3 on the rest.

The asymmetry that matters most here is different from the one on a docking page. Crystallography can fail for a given complex; the developability panel essentially always returns a number. What it needs is expressed, purified antibody at concentration — which you do not have at the point in the program where this tool is useful. That, and not accuracy, is the honest reason to compute first.

Agreement with wet lab

This section has to be read in two halves, because the published literature and this platform grade different objects.

What this platform actually computes

Verified against the currently deployed service, 2026-10-20:

So the honest mapping is short. Of the descriptors this tool returns, exactly one — the sequence-liability motif scan — has published figures grading it against a measurement. Everything else is either a measurement of your own input (structure_surface, cdr_lengths) or a rule with no published comparison.

The sequence-liability flags, graded against forced degradation

This is the number the index page's rule demands and the one a reader holding a liability list needs.

Motif-only prediction over-predicts by design. In a set of 37 therapeutic mAbs stressed at pH 6.0 and 40 °C for two weeks and read by tryptic LC-MS/MS, sequence-motif prediction using exactly the motif vocabulary this tool uses — NG, NS, NT and DG, DS, DT, DD, DH — reached:

• Asp: true-positive rate 1.0, false-positive rate 0.31.
• Asn: true-positive rate 0.91, false-positive rate 0.41.

The same study's structure-based classifier, built on homology-model ensembles with 20 conformational and secondary-structure parameters, reached Asp TPR 0.94 / FPR 0.02 and Asn TPR 0.95 / FPR 0.04 — an order-of-magnitude reduction in false positives, bought entirely with structural information this tool does not use. Read the small sample before you read the improvement: the hotspot classes were 8 Asp and 11 Asn hotspots against 285 and 188 non-hotspots, and the authors say so. Across the whole Fv set, only 60 of 559 Asn and Asp residues (11%) showed quantifiable modification at all, of which 21 were hotspots at the ≥3% level, and 15 of 37 antibodies carried at least one.

Per-motif hit rates, from a 131-antibody set. The larger study stressed 131 isotype-matched clinical-stage variable regions at pH 8.5/40 °C for one week (deamidation) and pH 5.5/40 °C for two weeks (isomerization). Its per-motif, per-CDR modification fractions are the clearest statement available of what a flag is worth:

Two readings follow, and both are load-bearing. First, the same motif in a different CDR has a different hit rate — NS runs 80% in CDR-L1 and 14.6% in CDR-L3; DG runs 53.8% in CDR-H3 and 19.0% in CDR-H2. This tool weights neither CDR nor position, and scans the framework identically to the CDRs. Second, two of the canonical isomerization motifs never degraded at all in either study, DT in 32 occurrences and DH in 6; the DS motif at position H61 was unmodified in all 37 occurrences, which the authors attribute to that position sitting in a structurally constrained turn in all but 34 of 942 antibody structures examined.

Sequence is not even sufficient to separate two near-identical antibodies. Rituximab deamidated 9.8% at H54 where ocrelizumab, with an identical CDR-H2, reached 3.5%. Gemtuzumab accumulated 17.3% at H54 where lintuzumab, differing at two positions outside the motif, was unmodified. Pembrolizumab modified 16.9% at H54 where ozanezumab, with the same NG motif at H54 and three CDR-H2 differences, did not. Tabalumab deamidated 4.3% at an NH motif where patritumab and zanolimumab, with identical CDR-H2 sequences, showed none; the authors ran the three through a published structure-based deamidation predictor and report that it "leads to very similar values for the key descriptors … and thus no suggestion that they may differ in their deamidation propensity." A sequence rule cannot see this, and the structure-based methods tested could not either.

Two further corrections to how these flags read. The apparent N-glycosylation sequon NHS in tabalumab, patritumab and zanolimumab "does not lead to actual glycan occupancy in any of the three cases" — a flagged sequon is not an occupied glycosylation site. And across the whole 131-antibody set, "correlation between Tm and chemical stability (as assessed in this work) does not reach statistical significance," so a liability list says nothing about thermostability, which this tool does not report at all.

Taken together: 48% of the 131 antibodies (63 of 131, by the paper's Discussion) were modified under one stress or the other, and 33 deamidated while 28 isomerized, with only 7 doing both.

Methionine and tryptophan oxidation. This tool flags every M and every W as minor oxidation risk, with no exposure test. The relevant published study forced-oxidized 121 clinical-stage mAbs with 0.1% H₂O₂ and found that the experimental oxidation "correlates well with theoretical predictions based on the solvent-accessible surface area of the methionine side-chains." That is a structure-based predictor, not a residue-presence flag, and no numeric correlation is available.

Lu X, Nobrega RP, Lynaugh H, Jain T, Barlow K, Boland T, Sivasubramanian A, Vásquez M, Xu Y. Deamidation and isomerization liability analysis of 131 clinical-stage antibodies. MAbs. 2019;11(1):45-57. doi: 10.1080/19420862.2018.1548233

Sydow JF, Lipsmeier F, Larraillet V, Hilger M, Mautz B, Mølhøj M, et al. Structure-based prediction of asparagine and aspartate degradation sites in antibody variable regions. PLoS One. 2014;9(6):e100736. doi: 10.1371/journal.pone.0100736

Yang R, Jain T, Lynaugh H, Nobrega RP, Lu X, Boland T, et al. Rapid assessment of oxidation via middle-down LCMS correlates with methionine side-chain solvent-accessible surface area for 121 clinical stage monoclonal antibodies. MAbs. 2017;9(4):646-653. doi: 10.1080/19420862.2017.1290753

Jain T, Sun T, Durand S, Hall A, Houston NR, Nett JH, et al. Biophysical properties of the clinical-stage antibody landscape. Proc Natl Acad Sci USA. 2017;114(5):944-949. doi: 10.1073/pnas.1616408114

Per-descriptor correlations, for descriptors graded in the literature

These grade published descriptors, not this tool's. They are the right order of magnitude to expect from any sequence-level developability number, and three of them bear directly on descriptors this tool does compute.

Sequence-only models against all 12 Jain assays (n = 137, author-reported). These are whole-model R² over 35 sequence features, not single-descriptor correlations — and they are not held-out numbers. A 50-times-repeated 10-fold cross-validation was used to choose each assay's algorithm by mean absolute error, but the chosen model was then "used to predict the entire experimental dataset to obtain predicted values for each of the proteins", so the R² below is a re-prediction over the same 137 antibodies the model was trained on. Read every value as an optimistic ceiling:

The single most useful row is the last one: the aggregation assay is the one the sequence could not predict at all. The authors' own reading of the weak rows is that "these deficits, for example in DSF (Tm) and HEK (expression titer) may highlight where sequence fails to capture salient structural features." Their HIC model, applied to an independent 97-variant set, reached r = 0.35 (p = 0.00044) against that set's own authors' 0.46.

The composite-score question, answered directly. A logistic-regression developability classifier over five structure-and-sequence descriptors separated 339 clinical antibodies from 4,929 repertoire antibodies at AUC 0.80 on a 20 + 20 held-out set — but that is a class-label task, not an assay task. When its authors took the same score and correlated it post hoc against all twelve Jain assays, every value was small:

Maximum |ρ| = 0.29, and the authors note the model was never fitted on assay data. No n, p-value or confidence interval is given for any cell. That is the published ceiling for a composite developability score against the measured panel, and it is the figure to hold against overall_developability_score.

Hydrophobicity scale choice, which bears on hydrophobicity_cdr_h3 directly. Two independent evaluations converge on the same conclusion about Kyte–Doolittle, the scale this tool uses everywhere. Across four scales on 674 clinical-stage therapeutics, "the widely used KD scale shows weak correlations with all other scales," and "the commonly used KD scale demonstrated weaker correlations with hydrophobicity-related biophysical assays." An independent re-implementation study testing eleven scales against measured HIC retention on four datasets found that "the relatively old scales by Eisenberg as well as Kyte and Doolittle do not perform well in our analysis," and later: "the Kyte-Doolittle and Eisenberg scales produce very poor results, even though they are still widely used." The scales that worked were Jain's own HIC-optimized scale and Wimley–White, both of which score tryptophan and the aromatics as strongly hydrophobic where Kyte–Doolittle scores tryptophan at −0.17 and tyrosine at −0.31.

The choice is not cosmetic: on one Fab, "based on BM and WW scales, the largest hydrophobic patch … is in the variable domain, whereas using KD and EI scales, it is located at the elbow." It can flip a conclusion's sign — whether clinical-stage antibodies are more or less hydrophobic than natural human antibodies depends on which of four scales you pick.

Viscosity: the net-charge-plus-hydrophobicity rule is the closest published analogue to what this tool does, and it does not replicate. The original rule combined Fv net charge at pH 5.5, a charge-symmetry parameter (FvCSP) and a CDR hydrophobicity index, and its authors reported Pearson r = −0.8 for Fv charge alone, −0.8 for FvCSP and 0.6 for hydrophobicity index against cone-and-plate viscosity at 180 mg/mL — on n = 14 antibodies, in one arginine buffer, with a three-descriptor model reaching r = 0.9 in-sample and 0.8 under leave-one-out. Two of its three regression coefficients have standard errors larger than the coefficient itself, and the authors state the coefficients "are specific to this buffer system and the respective protein concentration."

Independent re-testing is unkind to it. On the largest independently measured set — 229 antibodies at 150 mg/mL, binary at 20 cP — that method scored 52.8% accuracy (108 misclassifications) against 62.9% for TAP, 68.1% for DeepSCM and 80.7% for the evaluating authors' own model. The majority class in that set is 70.7%, so 52.8% is below the do-nothing baseline. A separate independent group found that "the predictions based on the Sharma model with default parameters, as reported in the original publication, were significantly off from the actual measurements and showed negative correlation with experimental data from both the PDGF38 and Ab21 test sets"; re-fitting the same functional form to their own data recovered Spearman 0.55–0.67 and ROC-AUC 0.73–0.78 out of sample.

Where charge descriptors do earn their keep is as rank correlates on a uniform set: against measured viscosity at ~150 mg/mL on 83 antibodies of one format, Spearman ρ reached −0.51 for sequence pI, −0.51 for CDR net charge and −0.49 for Fv net charge — though those authors note "we do not have a true hold-out set." On a balanced 96-antibody set, in-silico charge and hydrophobicity descriptors alone "provided only 66% and 71% accuracy" for the low- and high-viscosity groups.

⊘ And no published scheme predicts a three-level viscosity class at all. Every published classification in this literature is binary, and the cutoff varies by paper: 15 cP, 20 cP or 30 cP. The original rule's output is a continuous cP value, not a class. So this tool's three-level low/moderate/high has no published scheme to be graded against — the honest form would be a binary call at a named cP threshold.

Clearance: the pI heuristic is refuted, three times, independently. This tool's clearance_risk adds +0.35 for pi_fv > 9.0 and +0.15 for > 8.5. The published record does not support that:

• The paper that introduced the sequence-only clearance rule reports its own null: "no clear correlation was observed between the calculated mAb pI or HI and clearance." Its actual rule is an Fv-charge window (0–6.2) plus a CDR hydrophobicity sum, not a pI threshold — and the authors call their criteria "arbitrary." On cynomolgus clearance it reached 77% of fast and 70% of normal clearers on a 47-antibody held-out set.
• Applied to human clinical clearance on 64 antibodies, those same criteria "captured only 24 (50%) of our normal clearing antibodies … not indicating any significant advantage in prioritizing antibodies within the box", and the hydrophobicity term "does not help discriminate". Across 40 in vitro and in silico properties, "we found no strong correlation of any single property with linear clearance"; the best in-silico descriptor was pI_3D at |r| = 0.31, and the best of any kind a wet assay at r = 0.51.
• On 59 antibodies with package-insert pharmacokinetics, there was "no significant correlation between human PK parameters with either calculated mAb charge at pH 6.0 or pI (|r| < 0.2)" — measured pI against clearance r = −0.113 (n = 29) — and the authors conclude there is "no PK penalty incurred when selecting mAbs with high pI or charge at pH 6.0."

Two further shape problems. The published relationship is not monotonic: fast-clearing antibodies "were found to be clustered primarily in the lower and higher end ranges" of the charge descriptor, which is why that benchmark reports a within-range accuracy (82.81%, with boundaries fitted on the same 64 molecules and "intended to maximize accuracy") and no correlation coefficient. This tool's one-sided additive thresholds are the wrong shape for a two-sided relationship. And its hydrophobicity term never fires at all, because the value passed in is the whole-Fv mean Kyte–Doolittle, which is negative for every real antibody.

⚠︎ The thresholds in this literature are not even stable between the originating group's own papers: the CDR hydrophobicity-sum cutoff is given as > 4.0 in one and > 5.2 in another, the fast-clearance definition as ≥ 10 and > 8 mL/kg/day, and an independent group had to stretch the hydrophobicity sum to < 10 to make it work on human data. There is no canonical threshold to implement.

Measured AC-SINS has no computed predictor worth the name. Only one published figure correlates a computed descriptor against measured AC-SINS, and it is the composite score's ρ = −0.11 in the table above. One group reports outright that "AC-SINS related attributes showed weak correlation to the calculated structural features, indicating that AC-SINS is orthogonal to the structural features and provides information not captured by them." The best blind-benchmark result for self-association was ρ = 0.356. Nothing this tool computes addresses self-interaction at all.

Hebditch M, Warwicker J. Charge and hydrophobicity are key features in sequence-trained machine learning models for predicting the biophysical properties of clinical-stage antibodies. PeerJ. 2019;7:e8199. doi: 10.7717/peerj.8199

Negron C, Fang J, McPherson MJ, Stine WB, McCluskey AJ. Separating clinical antibodies from repertoire antibodies, a path to in silico developability assessment. MAbs. 2022;14(1):2080628. doi: 10.1080/19420862.2022.2080628

Waibl F, Fernández-Quintero ML, Wedl FS, Kettenberger H, Georges G, Liedl KR. Comparison of hydrophobicity scales for predicting biophysical properties of antibodies. Front Mol Biosci. 2022;9:960194. doi: 10.3389/fmolb.2022.960194

Park E, Izadi S. Molecular surface descriptors to predict antibody developability: sensitivity to parameters, structure models, and conformational sampling. MAbs. 2024;16(1):2362788. doi: 10.1080/19420862.2024.2362788

Sharma VK, Patapoff TW, Kabakoff B, Pai S, Hilario E, Zhang B, et al. In silico selection of therapeutic antibodies for development: viscosity, clearance, and chemical stability. Proc Natl Acad Sci USA. 2014;111(52):18601-18606. doi: 10.1073/pnas.1421779112

Kalejaye LA, Chu JM, Wu IE, Amofah B, et al. Accelerating high-concentration monoclonal antibody development with large-scale viscosity data and ensemble deep learning. MAbs. 2025;17(1):2483944. doi: 10.1080/19420862.2025.2483944

Rai BK, Apgar JR, Bennett EM. Low-data interpretable deep learning prediction of antibody viscosity using a biophysically meaningful representation. Sci Rep. 2023;13(1):2917. doi: 10.1038/s41598-023-28841-4

Mock M, Jacobitz AW, Langmead CJ, Sudom A, et al. Development of in silico models to predict viscosity and mouse clearance using a comprehensive analytical data set collected on 83 chemically and biophysically diverse monoclonal antibodies. MAbs. 2023;15(1):2256745. doi: 10.1080/19420862.2023.2256745

Paidi SK, Ibrahim J, Stepurska K, Zarzar J, et al. Self-interaction nanoparticle spectroscopy predicts high-concentration viscosity of therapeutic antibodies. MAbs. 2026;18(1):2709948. doi: 10.1080/19420862.2026.2709948

Grinshpun B, Thorsteinson N, Pereira JN, Rippmann F, et al. Identifying biophysical assays and in silico properties that enrich for slow clearance in clinical-stage therapeutic antibodies. MAbs. 2021;13(1):1932230. doi: 10.1080/19420862.2021.1932230

Kingsbury JS, Saini A, Auclair SM, Fu L, et al. A single molecular descriptor to predict solution behavior of therapeutic antibodies. Sci Adv. 2020;6(32):eabb0372. doi: 10.1126/sciadv.abb0372

The blind benchmark, which is the strongest evidence in this field

In 2025 a blinded competition benchmarked developability predictors on a single industrial assay platform: a public training set of 246 clinical antibodies, a held-out test set of 80 antibodies selected from paired repertoire data for sequence and germline diversity, and 113 teams from 25 countries. Top Spearman ρ on the held-out set, per assay:

Three things to take from this table. It is a ceiling, not a typical result — each value is the best of 113 submissions. The generalization gap is the finding: self-association fell from 0.653 in cross-validation to 0.356 held out, titer from 0.640 to 0.310, polyreactivity from 0.559 to 0.337, which the organizers read as "substantial overfitting to training data." And hydrophobicity is the only assay anyone predicts well, which the organizers attribute to it reflecting "physicochemical features … [that] can be reasonably approximated from the structure of a single folded molecule."

TAP and DeepSP were both included as baselines in that benchmark, but their individual per-assay values appear only inside Figure 3, so no baseline figure can be quoted from it.

van Niekerk L, Moller J, Ritter S, Quintero-Cadena P, Cohen R, Channing G, et al. Ginkgo Datapoints Antibody Developability Competition outcomes: limited model performance and a call for data standardization. MAbs. 2026;18(1):2634216. doi: 10.1080/19420862.2026.2634216

What the figure doesn't cover

• ⊘ Known defect: every cysteine is flagged critical. The liability scan matches a bare C with no disulfide-pairing or exposure test, so the two conserved intra-domain cysteines of every VH and every VL are each reported as a critical "free cysteine — disulfide scrambling risk". A normal paired Fv therefore always returns four critical liabilities and a fixed −0.60 penalty on overall_developability_score. Re-running the served rules by hand on the trastuzumab VH/VL that the repository commits as its own reference case gives 21 liabilities — 4 critical, 2 moderate, 15 minor — an aggregation score of 0.4628 ("moderate"), and a composite of about 0.294 under the default metric set or 0.204 under metrics: ["all"]. Both are below the 0.4 cut, so the shipped code returns developability_class: "problematic" for a marketed, well-characterised antibody — while the committed validator at the same path expects "developable" and no_critical_liabilities: true. The fix needed is a pairing or exposure test before the critical severity is assigned, so that only genuinely unpaired cysteines flag. Until then, subtract the four framework cysteines yourself and read the composite as uninterpretable.

• ⊘ Known issue, now fixed: CDR lengths were not derived from your sequence. Before the 2026-10-20 fix, this tool did not apply a numbering scheme when measuring CDR lengths, so the values it returned were effectively the same for every antibody submitted. The row above describes the corrected behaviour, which uses real ANARCI numbering under the scheme you request. ⚠︎ If you are holding a developability result from before that date, its cdr_lengths are not measurements of your antibody — re-run it. Earlier results have not been reprocessed or withdrawn automatically, so nothing will flag them for you.

• ⚠︎ Known issue: the metrics selector does not narrow what is computed. You receive the full set of liability measures on every run regardless of which metrics you ask for, so including or omitting four of the options changes nothing in the result. Nothing is missing or wrong because of it — read the fields you care about and ignore the rest. Either gating the selector or removing those options is pending.

• ⊘ Known defect: the composite is not comparable across metrics selections. Each penalty term is conditional on its metric having been requested, so requesting fewer metrics yields a higher overall_developability_score for an identical antibody. The fix is to normalise by the terms actually evaluated, or to refuse to emit a composite unless a fixed metric set was computed. Today, metric_provenance is the only record of which terms were in play.

• ⊘ Known defect: the "high" viscosity class and two tap_score penalty arms are effectively unreachable. estimated_viscosity_class returns "high" only when mean Kyte–Doolittle over the whole Fv exceeds 0, which no soluble variable domain does (trastuzumab's Fv is −0.334); in practice the field is "moderate" below |net charge| 5 and "low" above it, which assigns low viscosity to the highest-charge antibodies with no upper bound and no charge-symmetry term. tap_score's hydrophobicity arm fires on CDR-H3 mean Kyte–Doolittle > 0.5 (trastuzumab's CDR-H3 is −0.808) and its charge arms on |net charge| > 5, so tap_score is 0.9 or 1.0 for most real antibodies — near-constant, and therefore uninformative. The same near-constancy affects aggregation_class, whose "high" band begins at 0.55 on a quantity that sits around 0.44–0.48 for any soluble Fv. The fix is to recalibrate each threshold against a real distribution of antibody values, or to stop reporting the class.

• Positions are not numbering-scheme positions. Liability position values are 1-based linear offsets into the submitted chain, not IMGT or Chothia positions, despite the output schema's example reading "H52" and despite the tool accepting a numbering_scheme. aggregation.hot_spots, immunogenicity.epitope_positions and charge_patches coordinates all index the VH‖VL concatenation with no chain label, so a patch reported at 156–168 is actually in the light chain. aggregation.hot_spots is also silently truncated to 10 entries with no count field, and on trastuzumab it returns buried framework windows such as LRLSCAA and GTLVTVS.

• net_charge_ph7 ignores pH. It is a residue-count sum with no terminal contribution and no pH term, despite the name. pi_fv is the pI of the concatenated VH+VL string, not of a modelled Fv.

• Structure-based descriptors are not stable, so the published figures are not either. The same descriptor, on the same 592 clinical antibodies, differing only in which structure predictor built the Fv, correlates across predictor pairs at R² as low as 0.34 for a positive-electrostatic CDR descriptor (0.85 at best) and 0.60 to 0.78 for a hydrophobic-patch descriptor. The cause is systematic distribution shifts that "result in significant inconsistency in categorically flagging antibodies based on predefined cutoff threshold." Averaging over just 5 ns of accelerated molecular dynamics lifts cross-model agreement to 0.86 and 0.91 — but it does not uniformly improve agreement with the assays, and "conformational sampling consistently reduces the performance of all structure models in predicting the HIC assay." The same study found that "none of the structure models outperformed the others across all biophysical datasets simultaneously," and that a static model's apparent superiority "may be due to inadvertent error cancellation rather than being mechanistically meaningful." Any structure-based developability figure quoted without its structure pipeline is unusable.

• The clinical-versus-repertoire premise is contested. Both TAP and the composite classifier are calibrated on the assumption that clinical-stage antibodies are enriched for developability and repertoire antibodies are not. Repertoire antibodies are never labelled undevelopable, so the separation is not a true-positive/true-negative task — TAP's own authors note "not every human antibody would make a good therapeutic", and the classifier's authors flag that published work "challenges this assumption, as B-cell derived antibodies display a distribution of biophysical properties that overlaps with the approved antibody set."

• A three-level viscosity class has nothing to be graded against. Every published viscosity classification in this literature is binary, at 15, 20 or 30 cP depending on the paper, and the originating sequence-only rule emits a continuous cP value rather than a class. estimated_viscosity_class reports low/moderate/high with no concentration, no buffer and no cP threshold attached, so there is no published scheme it corresponds to and no figure that could validate it. The fix is to emit a binary call at a named cP threshold and concentration, or to stop emitting a class. Viscosity is also buffer- and concentration-specific: the rule this most resembles was fitted at 180 mg/mL in one arginine buffer and its authors state the coefficients "are specific to this buffer system and the respective protein concentration."

• immunogenicity is not an immunogenicity prediction, and nothing here is graded against clinical immunogenicity. t_cell_epitope_count counts 9-mers whose summed Kyte–Doolittle exceeds 10.0; there is no allele, no MHC-II binding model, and no allele is used (it was previously mislabelled DRB1*0101). No published figure on this page grades any computed epitope count against measured anti-drug-antibody incidence, and TAP's authors state outright that their guidelines "will not detect sources of immunogenicity." For a humanness read, which is a different question, see Humatch and BioPhi.

• No thermostability output at all. Tm is the panel's most-requested number and this tool does not compute it. The published figures for predicting it are weak in any case: R² 0.13 from sequence on 137 antibodies, and a best-of-113 blind ρ of 0.392.

• Aggregation is the hardest assay and this tool's aggregation score is the least grounded. Sequence-only prediction of accelerated-stress aggregation reached R² 0.086 at p = 0.282 — the one non-significant row in that study. Against a 15-antibody SEC set, even purpose-built structural descriptors led its authors to conclude that "relying on a single descriptor alone may not be sufficient to achieve a strong correlation."

• Thera-SAbDab resolution is a live third-party fetch. therapeutic_name is resolved on the Next side by downloading the Thera-SAbDab CSV from Oxford OPIG with an in-memory cache; there is no local data file, so an OPIG outage surfaces as a service error rather than "not found". The resolved sequence must still clear the 90–130 aa bound.

• numbering_scheme accepts only imgt and chothia. A kabat request fails validation and is non-recoverable, even though the shared boundary table supports kabat and aho. The restriction is this tool's, not the underlying data's. Note that CDR definitions move the answer: the cold-spot at H61 "would be considered part of the framework" under IMGT.

• Version and basis. Everything above about this implementation was read from the served Modal app at commit 9d67198 (v108, deployed 2026-10-20), verified as unmodified against the local repository tree. A later deploy can change any of it, and the honesty fields on the result — is_heuristic, tap_score_method.caveat, viscosity_method.caveat, overall_score_method.caveat, metric_provenance — are the authority on a result you are holding, not this page.

Go / no-go

Go — act on the result when:

• You want a list of candidate chemical-liability sites to confirm by forced degradation and peptide mapping, and you treat the list as the thing that directs the mass spectrometer. This is the one use the published record supports.
• You are re-profiling after a sequence change and reading the delta rather than the absolute number — the same systematic bias sits on both sides, and a liability that appeared or disappeared is a real sequence fact.
• You need CDR delineation, CDR lengths and sequence pI/net charge as inputs to your own analysis. Those are computed, not predicted, and the numbering is real ANARCI output that fails loudly rather than guessing.
• You supplied a PDB and want structure_surface — Shrake–Rupley SASA over the Fv chains matching your sequences. That is a measurement of the structure you provided, and it is deliberately excluded from the composite so that it cannot silently change a score.
• You are sorting a long candidate list by hydrophobicity and you have nothing better. Hydrophobicity is the only developability property anyone predicts well (blind ρ = 0.708), though Kyte–Doolittle is the wrong scale for it.

No-go — do not act on the result when:

• You are gating a lead on developability_class or overall_developability_score. The composite always carries four spurious critical cysteine liabilities, is not comparable across metrics selections, and the best-characterised published composite correlates with the measured panel at |ρ| ≤ 0.29. Trastuzumab reads problematic.
• You are reading tap_score as a TAP score. It is not TAP, it needs no structure, and it is 0.9 or 1.0 for most antibodies. Real TAP returns five flags and no scalar, and requires a 3D model.
• You are treating a liability flag as a degradation rate. Motif-only prediction has a published false-positive rate of 0.31–0.41; DT and DH never degraded in two studies; and two antibodies with identical CDR-H2 sequences differ in measured deamidation. Flagged N-glycosylation sequons are not necessarily occupied.
• You are deciding a formulation question from estimated_viscosity_class. Its "high" class is unreachable, it assigns "low" to the highest-charge antibodies, and no published scheme predicts a three-level class at all. The closest published rule scored 52.8% on 229 measured antibodies — below that set's 70.7% majority-class baseline — and anti-correlated with measurement when run with its own published parameters. Measure viscosity at your intended concentration and buffer.
• You are acting on clearance_risk, or on pi_fv as a clearance signal. The pI heuristic is independently refuted three times: a stated null in the rule's own source paper, |r| < 0.2 and pI-vs-clearance r = −0.113 on human pharmacokinetics with "no PK penalty incurred when selecting mAbs with high pI", and a 50% capture rate when the original criteria were re-tested on human clearance. The best in-silico descriptor against human clearance reaches |r| = 0.31, the relationship is two-sided rather than monotonic, and this tool's hydrophobicity term never fires.
• You need an immunogenicity read. t_cell_epitope_count is a hydrophobic-9-mer count with no allele and no binding model, and no figure on this page grades any such count against clinical anti-drug-antibody data.
• You need Tm, or an aggregation answer you will act on. Tm is not computed. Aggregation is the one assay sequence-only methods failed to predict at all (R² 0.086, p = 0.282).
• You are comparing a number here to a published TAP, SAP, SCM or DeepSP figure. Different objects. Even between published structure-based methods, the same descriptor shifts enough across structure predictors (R² 0.34–0.85) to flip categorical risk flags.

Caveats on reading this page: every published figure above is quoted from the cited source with its original metric and sample size, read from the body text, tables or figure captions rather than from an abstract, except where a NEEDS REVIEW marker says otherwise. Figures from different studies are not directly comparable, because assay platforms, antibody sets, structure pipelines and metrics differ between them. Where a descriptor this platform computes has no published comparison, this page says so rather than borrowing a figure from a different method — and for most of this tool's outputs, that is the answer.