Computational Antibody Developability vs the Bench Assays That Grade It

A computational developability screen orders a candidate list and directs a bench assay. It does not replace one, and the published correlations are why. The reference panel for this field expressed 137 clinical-stage antibody variable regions on a common IgG1 constant region and measured every one of them on a dozen biophysical platforms (Jain et al., PNAS 2017). When sequence-trained models were fitted against that panel, hydrophobicity came out at R² 0.391 and accelerated-stress aggregation at R² 0.086 with p = 0.282 — not significant (Hebditch & Warwicker, PeerJ 2019). When a composite developability score was correlated post hoc against all twelve assays, every value was small, with a maximum |ρ| of 0.29 (Negron et al., MAbs 2022). So the honest division of labour is this: compute to rank candidates and to decide which peptide the mass spectrometer looks at, then measure self-association, aggregation, Tm, viscosity and hydrophobicity on material.

The reason to compute first is not accuracy. It is that the panel needs expressed, purified antibody at formulation concentration, and at the point in a program where a developability question is most useful, you do not have it.

At a Glance#

Read that table by row, never down a column. "R² 0.391 against a measured assay", "no published scheme exists" and "this is the measurement" are three different kinds of statement, and conflating them is how a screen gets promoted into a gate.

What the Bench Assays Actually Grade#

The developability panel is not one technique. It is a dozen, each answering a different question, and a computed descriptor is only ever graded against one of them at a time.

AC-SINS — self-association. Affinity-capture self-interaction nanoparticle spectroscopy reads a plasmon wavelength shift when antibody-coated gold nanoparticles approach each other, so it reports self-interaction at low concentration on very little material. It is also the assay computation has the least purchase on. The one published figure correlating a composite developability score against measured AC-SINS is ρ = −0.11, and the authors of a large structure-descriptor study report that AC-SINS 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" (Park & Izadi, MAbs 2024).

SEC after accelerated stress — aggregation. Size-exclusion chromatography on stressed material measures monomer loss and high-molecular-weight species. This is the property sequence-level prediction failed on: the non-significant row in the sequence-trained panel models, R² 0.086 at p = 0.282. There is no sequence-level substitute for it.

DSF / DSC — thermostability. Differential scanning fluorimetry or calorimetry gives a measured melting temperature. Tm is the number programs ask for most, and sequence-trained models reached R² 0.13 against measured DSF on the 137-antibody panel — the second-weakest row in that study. The platform's developability job does not report a Tm at all, which is the more honest position than reporting a weak one.

Viscometry — injectability. Cone-and-plate or capillary viscometry at formulation concentration decides whether a subcutaneous product is possible. Every published classification in this literature is binary, at 15, 20 or 30 cP depending on the paper, and the best-known sequence-level rule emits a continuous cP value fitted at 180 mg/mL in one arginine buffer, whose authors state the coefficients "are specific to this buffer system and the respective protein concentration" (Sharma et al., PNAS 2014). Charge descriptors do earn rank-correlation keep on uniform panels: against measured viscosity on 83 scaffold-consistent antibodies, Spearman ρ reached −0.51 for sequence pI and −0.49 for Fv net charge, though those authors note they "do not have a true hold-out set" (Mock et al., MAbs 2023).

HIC — hydrophobicity. Hydrophobic interaction chromatography retention time is the one developability property computation predicts at a genuinely useful level: R² 0.391 from sequence on the 137-antibody panel, the strongest row in that study. If you are going to trust one computed developability axis to sort a long list, this is the one — with a caveat about the hydrophobicity scale, below.

Forced degradation plus peptide-map LC-MS/MS — chemical liabilities. Stress at defined pH and temperature, then tryptic digest and mass spectrometry, tells you which asparagine deamidated and which aspartate isomerized, and by how much. This is what a liability flag is graded against, and it is the one place the computed half has real numbers behind it.

The One Computed Output With Published Grading#

Of everything a sequence-level developability screen returns, the sequence-liability motif scan is the part with published figures against a measurement. Take it seriously, and take its false-positive rate equally seriously.

In 37 therapeutic monoclonals stressed and read by tryptic LC-MS/MS, motif-only prediction over the same motif vocabulary these tools use reached a

true-positive rate of 1.0 with a false-positive rate of 0.31 for aspartate, and 0.91 with 0.41 for asparagine. A structure-based classifier built on homology-model ensembles with conformational and secondary-structure parameters reached Asp TPR 0.94 / FPR 0.02 and Asn TPR 0.95 / FPR 0.04 on the same data — an order-of-magnitude cut in false positives, bought entirely with structural information a sequence rule does not have. Read the sample size before the improvement: of all 559 Asn and Asp residues in those 37 Fv regions, 60 (11%) showed quantifiable modification at all (Sydow et al., PLoS One 2014).

The larger study stressed 131 isotype-matched clinical-stage variable regions and reported per-motif, per-CDR modification fractions. They are the clearest statement available of what a flag is worth (Lu et al., MAbs 2019):

Two readings follow, and both change how you use a flag list. First, the same motif in a different CDR has a different hit rate — DG runs 53.8% in CDR-H3 and 19.0% in CDR-H2 — while a sequence scan weights neither CDR nor position and reads the framework identically to the CDRs. Second, two canonical isomerization motifs never degraded at all in that work: DT in 32 occurrences and DH in 6, consistent with the smaller study.

Sequence is not even sufficient to separate near-identical molecules. In that same set, rituximab deamidated 9.8% at H54 where ocrelizumab, with an identical CDR-H2, reached 3.5%; pembrolizumab modified 16.9% at H54 where ozanezumab, with the same NG motif at that position, did not. The authors ran a published structure-based deamidation predictor over one such trio and report it "leads to very similar values for the key descriptors … and thus no suggestion that they may differ in their deamidation propensity." Neither the sequence rule nor the structure-based predictor saw it.

One further correction worth carrying: across that whole 131-antibody set, "correlation between Tm and chemical stability (as assessed in this work) does not reach statistical significance." A liability list says nothing about thermostability.

What a Composite Score Cannot Tell You#

Most of what a sequence-level developability screen returns — an aggregation propensity, an immunogenicity risk, a viscosity class, a clearance risk, an overall developability score — is a threshold rule over hydropathy and residue counts. That is not a criticism of any one implementation; it is what the class of method is. What matters is that a rule with no published concordance figure has no published concordance figure, and the composites that have been graded set a low ceiling.

The composite ceiling is |ρ| ≤ 0.29. A logistic-regression developability classifier over five descriptors separated 339 clinical antibodies from 4,929 repertoire antibodies at AUC 0.80 on a held-out set — a class-label task. When its authors took the same score and correlated it against all twelve measured assays, the values ran: ELISA −0.29, BVP-ELISA −0.28, polyspecificity reagent −0.26, HEK titer 0.21, Fab Tm 0.15, CIC −0.15, accelerated stability slope −0.14, AC-SINS −0.11, SGAC-SINS 0.05, SMAC 0.05, HIC retention −0.04. No n, p or confidence interval is given per cell, and the model was never fitted on assay data. That is the published ceiling for a composite against the panel.

The clinical-versus-repertoire premise behind those classifiers is itself contested. Repertoire antibodies are never labelled undevelopable, so the separation is not a true-positive / true-negative task. The authors of the composite 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", and the authors of the five-guideline method note that "not every human antibody would make a good therapeutic" (Raybould et al., PNAS 2019).

TAP is a different object from a developability score, and it is worth being precise about. The Therapeutic Antibody Profiler computes five metrics — total CDR length, patches of surface hydrophobicity, patches of positive charge, patches of negative charge, and the structural Fv charge symmetry parameter — four of which require a 3D model that it builds internally before evaluating anything. It emits amber and red flags, not a scalar: amber at the extremes of the clinical-stage distribution, red at previously unobserved values. Its own validation contains no correlation or AUC against any measured assay; it was calibrated on the sequence and structure distributions of 242 clinical-stage therapeutics against a naive human repertoire, and its authors write that "as with the Lipinski rule of five, the thresholds themselves should not be interpreted as hard-and-fast rules" and that the guidelines "will not capture the whole spectrum of developability issues … they will not detect sources of immunogenicity." TAP is not offered on .bloom*, and no field this platform returns is a TAP metric.

Hydrophobicity scale choice is not cosmetic, and it bears on the one axis that works. Two independent evaluations converge on the same conclusion about Kyte–Doolittle, the scale most sequence-level tools use. Testing eleven scales against measured HIC retention on four datasets, one group writes that "the Kyte-Doolittle and Eisenberg scales produce very poor results, even though they are still widely used" (Waibl et al., Front Mol Biosci 2022). The scales that worked score tryptophan and the aromatics as strongly hydrophobic, where Kyte–Doolittle does not. So the one developability axis computation predicts well is commonly computed on the wrong scale for it.

And structure-based descriptors are not stable enough to quote without their pipeline. The same descriptor, over the same 592 clinical antibodies, differing only in which predictor built the Fv, agrees across predictor pairs at R² as low as 0.34 for a positive-electrostatic CDR descriptor and 0.60–0.78 for a hydrophobic-patch descriptor, with systematic distribution shifts that "result in significant inconsistency in categorically flagging antibodies based on predefined cutoff threshold." Averaging over a short molecular-dynamics window lifts cross-model agreement but does not uniformly improve agreement with the assays — "conformational sampling consistently reduces the performance of all structure models in predicting the HIC assay" — and "none of the structure models outperformed the others across all biophysical datasets simultaneously." If you are comparing a structure-based developability number to a published one, the structure pipeline is part of the number. The same caution applies to anything built on a predicted fold, which is the subject of ESMFold vs AlphaFold.

When to Use Each#

Compute when:

• You have more candidates than you can express, and you want the ones carrying obvious chemical liabilities pushed down the list or fixed by mutation first. This is the use the published record supports.
• You need a site list to direct a forced-degradation study. The flag tells the analytical chemist which peptide to look at; the mass spectrometer decides whether it degrades.
• You are re-profiling after a sequence change — humanization or affinity maturation both move charge and hydrophobicity — and you read the delta rather than the absolute number. The same systematic bias sits on both sides of the subtraction, and a liability that appeared or disappeared is a real sequence fact.
• You want CDR delineation, CDR lengths, net charge and isoelectric points as inputs to your own analysis. Those are computed from your sequence, not predicted.
• You are sorting a long list by hydrophobicity and have nothing better.

Measure when:

• The question is self-association or polyreactivity. Run AC-SINS. Nothing computed is a substitute, and the measured assay carries information the descriptors do not.
• The question is aggregation you will act on. Run SEC after accelerated stress.
• The question is Tm. Run DSF or DSC.
• The question is injectability. Run viscometry in your buffer at your intended concentration. A viscosity answer is buffer- and concentration-specific, so a class with no concentration attached is not an answer to it.
• You are gating a lead, full stop. No composite developability score in the published record correlates with the measured panel above |ρ| = 0.29, and none of them was fitted on assay data in the first place. A score is a sort order, not a decision.

Limitations to Keep in View#

• Most composite fields are rules, and the result says so. The platform's developability result marks the heuristic fields is_heuristic and carries a per-field caveat — the tap_score, aggregation, immunogenicity, viscosity and overall-score methods each disclaim themselves in the payload. Those fields are the authority on a result you are holding.
• Don't compare composites across different metrics selections. Each penalty term is conditional on its metric having been requested, so asking for fewer metrics returns a higher score for the same antibody. Fix one metric set and keep it fixed across a candidate panel.
• Every cysteine is flagged. The liability scan matches a bare C with no disulfide-pairing or exposure test, so the conserved framework cysteines of a normal paired Fv are flagged as critical alongside any genuinely free one. In practice a well-characterised marketed antibody comes back with a long liability list and a low composite. Subtract the framework cysteines yourself, and read the liability list rather than the class.
• Positions are linear offsets, not numbering-scheme positions. Liability positions index the submitted chain from 1, and the aggregation, epitope and charge-patch coordinates index the VH‖VL concatenation, so a patch reported past the end of the heavy chain is in the light chain.
• net_charge_ph7 is a residue-count sum, and pi_fv is the isoelectric point of the concatenated VH+VL string rather than of a modelled Fv. Useful as a relative sort key across one panel; not a formulation parameter.
• No Tm, and no self-association descriptor. Two of the five assays this article is set against have no computed counterpart here at all. That is a statement about what the tool reports, and it is better than a weak number with a confident label.
• Immunogenicity is not an immunogenicity prediction. The epitope count is a hydrophobic-window count with no allele and no MHC-II binding model, and no figure in this literature grades such a count against measured anti-drug antibody incidence. For humanness — a different question — see BioPhi and Humatch.
• Antibody numbering and CDR delineation move the answer. CDR boundaries differ between schemes, so a motif that sits in CDR-H2 under one convention can be framework under another — and the per-CDR hit rates above are reported under the source study's convention, not necessarily yours.

How to Run It on .bloom*#

The platform runs developability as an asynchronous CPU job, not as a notebook you maintain. You supply a VH sequence — variable region only, 90 to 130 residues — optionally a VL, and the metrics you want. Alternatively, name a known therapeutic and the platform resolves the VH/VL for you, so you can profile a marketed antibody without pasting sequences.

What it computes. CDR sequences and lengths under the numbering scheme you request (imgt or chothia), isoelectric point per chain and for the Fv, net charge, CDR-H3 hydropathy, the sequence-liability list, and the rule-based aggregation, immunogenicity, viscosity-class, charge-patch, clearance-risk and composite fields that the result marks as heuristic. Pass the metrics enum values exactly — tap, charge, pI, hydrophobicity, cdr_length, liabilities, aggregation, immunogenicity, viscosity, or all.

Supply a structure and you get a measurement. Pass a PDB and the job computes solvent-accessible surface over the Fv chains that match your submitted VH/VL. That block is a measurement of the structure you provided rather than a prediction, and it is deliberately excluded from the composite so it cannot quietly move a score. A structure with no chain matching your VH/VL fails the job rather than guessing.

How to use it in a program. Profile a panel with one fixed metric set, read the liability list and the hydrophobicity ranking, then take the top candidates to AC-SINS, DSF, SEC and HIC. Re-profile after humanization with BioPhi or Humatch and read the delta. Before you fix a flagged liability by mutation, map the paratope with Paragraph so the fix does not land on a binding residue.

Start with the developability quickstart, the inputs reference for the full parameter list, and the outputs reference for every field the job returns. For where developability sits in the wider antibody workflow, see antibody engineering. The same compute-then-measure logic in small molecules is in the ADMET prediction guide.

Access. The developability job runs without an account.

Key Takeaways#

• Hydrophobicity is the one developability axis computation predicts well (R² 0.391 from sequence against measured HIC) — and it is commonly computed on a hydrophobicity scale the benchmarks call poor.
• Aggregation is the one it failed outright (R² 0.086, p = 0.282). Run SEC after accelerated stress.
• Sequence-motif liability flags are real triage with a known over-prediction rate (FPR 0.31 for Asp, 0.41 for Asn), and per-motif hit rates swing from 0% to 54% depending on the motif and the CDR it sits in.
• A composite developability score tops out at |ρ| = 0.29 against the measured panel. Use it to sort, never to gate.
• AC-SINS and Tm have no computed stand-in. Self-association is orthogonal to the calculated descriptors, and this platform reports no Tm at all.
• TAP is five structure-based flags, not a score, and it is not offered on .bloom*. Don't read any scalar as a TAP score.

Profile the panel in minutes, rank it, flag the sites — then spend your purified material on the assays that decide. Run an antibody developability assessment on the .bloom* platform.

References#

1. Jain, T., Sun, T., Durand, S. et al. "Biophysical properties of the clinical-stage antibody landscape." Proceedings of the National Academy of Sciences 114, 944–949 (2017). doi:10.1073/pnas.1616408114
2. 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 7, e8199 (2019). doi:10.7717/peerj.8199 · code: maxhebditch/abpred
3. Raybould, M.I.J., Marks, C., Krawczyk, K. et al. "Five computational developability guidelines for therapeutic antibody profiling." Proceedings of the National Academy of Sciences 116, 4025–4030 (2019). doi:10.1073/pnas.1810576116
4. Lu, X., Nobrega, R.P., Lynaugh, H. et al. "Deamidation and isomerization liability analysis of 131 clinical-stage antibodies." MAbs 11, 45–57 (2019). doi:10.1080/19420862.2018.1548233
5. Sydow, J.F., Lipsmeier, F., Larraillet, V. et al. "Structure-based prediction of asparagine and aspartate degradation sites in antibody variable regions." PLoS One 9, e100736 (2014). doi:10.1371/journal.pone.0100736
6. Negron, C., Fang, J., McPherson, M.J. et al. "Separating clinical antibodies from repertoire antibodies, a path to in silico developability assessment." MAbs 14, 2080628 (2022). doi:10.1080/19420862.2022.2080628
7. Sharma, V.K., Patapoff, T.W., Kabakoff, B. et al. "In silico selection of therapeutic antibodies for development: viscosity, clearance, and chemical stability." Proceedings of the National Academy of Sciences 111, 18601–18606 (2014). doi:10.1073/pnas.1421779112
8. Mock, M., Jacobitz, A.W., Langmead, C.J. et al. "Development of in silico models to predict viscosity and mouse clearance using a comprehensive analytical data set collected on 83 scaffold-consistent monoclonal antibodies." MAbs 15, 2256745 (2023). doi:10.1080/19420862.2023.2256745
9. Waibl, F., Fernández-Quintero, M.L., Wedl, F.S. et al. "Comparison of hydrophobicity scales for predicting biophysical properties of antibodies." Frontiers in Molecular Biosciences 9, 960194 (2022). doi:10.3389/fmolb.2022.960194
10. Park, E. & Izadi, S. "Molecular surface descriptors to predict antibody developability: sensitivity to parameters, structure models, and conformational sampling." MAbs 16, 2362788 (2024). doi:10.1080/19420862.2024.2362788
11. Dunbar, J. & Deane, C.M. "ANARCI: antigen receptor numbering and receptor classification." Bioinformatics 32, 298–300 (2016). doi:10.1093/bioinformatics/btv552 · code: oxpig/ANARCI