Developability and T-cell (MHC class II) immunogenicity from a VH/VL sequence, in one report — before you commit bench time. Many in-silico tools stop at humanness; we predict which CDR peptides bind MHC class II — the presentation step upstream of a T-cell-dependent ADA response. Free scanner; immunogenicity module in early access, no sales call.
A large fraction of clinical-stage antibodies flag on at least one developability assay — often only after months at the bench (Jain et al. 2017). PresciaBio flags them from sequence alone, before any bench work.
Paste a VH or VL domain and the scanner flags chemical-liability and PTM-hotspot motifs as you type — a sequence-level motif scan (no structure, no CDR weighting) that runs entirely in your browser; nothing is uploaded. The production API adds IMGT numbering, CDR weighting and structure-aware scoring.
Sequence-level motif screen — solvent exposure and CDR context are not assessed here. The production engine adds CDR-weighting and structure-aware scoring.
When connected, the API returns IMGT numbering, per-chain CDRs, region-weighted liabilities and a developability score. It's live behind an access key — the panel below shows the real response shape from POST /v1/developability.
The production engine numbers every sequence in IMGT, Kabat, Chothia, Martin & Aho and weights liabilities by region — live today. A framework flag is down-weighted; the same motif in a CDR is weighted 8× — it can cost binding and stability. Exposure weighting comes from the structure route, not this one.
numbering schemes, auto-aligned
Scores update as you type in the browser demo; high-throughput batch scoring is live in the production engine — up to 200 candidates per call, ranked and threshold-filtered.
Humanness alone is a weak predictor of clinical ADA — we measured it ourselves at AUROC 0.63, with a confidence interval that spans chance. So we predict which CDR peptides bind MHC class II instead: the presentation step upstream of a T-cell response. Framework binders are excluded — they are germline, and T cells are centrally tolerant to them.
The five TAP biophysical metrics (total CDR length, PSH, PPC, PNC, SFvCSP), banded against the published Raybould 2019 guidelines — live; they are structure-based, so they run on a modelled Fv. Every row states which model mode it came off and how much its band is worth. Chemical-liability and epitope flags use separate, motif- and predictor-specific criteria.
One call returns numbering and scores as JSON for your screening scripts (early access).
Asn deamidation (e.g. NG/NS), Asp isomerization (e.g. DG/DS/DT), DP-bond fragmentation, Met oxidation, Trp oxidation, N-glycosylation sequons (N-X-S/T, X≠Pro), and unpaired cysteines — surfaced and ranked, with the residue positions called out.
The in-browser demo runs sequence-level liability checks today. CDR-weighting, structure-based metrics and the immunogenicity module all run in the production engine today, behind an access key while we scale. For example, it flags trastuzumab's well-documented CDR-H2 Asn-Gly deamidation and CDR-H3 Asp-Gly isomerization motifs.
Run it on every clone, every round.
Maps the MHC class II epitope content that drives T-cell-dependent ADA — an epitope map and CDR-epitope count, not an ADA score.
One report per candidate: a developability read and a T-cell (MHC class II) epitope map, the flagged positions, and where each metric falls within the clinical-stage antibody distribution (Raybould 2019). An illustrative example — sample values, not a live computation:
How it works: paste paired VH+VL, or POST a JSON panel of up to 200 candidates; the engine returns a per-candidate report — flagged positions, ranked shortlist, suggested edits — as JSON. FASTA/CSV upload and a branded PDF export are (Future Functionality). No humans in the loop, no scheduling. Your sequences are never used to train models, never shared, and never sold — here is exactly what happens to them. The free scanner is live now; the production engine is live behind an access key.
Example figures. Every number in this table is made up for illustration, not computed from a molecule. A live report on your own sequence runs in the sections below.
| Position | Region | Finding | Risk |
|---|---|---|---|
| N55–G56 | CDR-H2 | Asn deamidation (NG) | High |
| D102–G103 | CDR-H3 | Asp isomerization (DG) | High |
| 9-mer @ 52 | CDR-H2 | Predicted MHC-II binder · DRB1*04:01 · %rank 0.8 · 3 of 8 panel alleles | no band assigned |
| M107 | CDR-H3 | Met oxidation (exposed) | Moderate |
Recommendation: engineer out the CDR-H2 NG and CDR-H3 DG motifs; the strongest predicted MHC-II binder sits in CDR-H2 (%rank 0.8, 3 of 8 panel alleles) — a candidate for deimmunization (deimmunizing suggestions are Future Functionality; developability engineering suggestions are live). The values above are illustrative; the interactive scanner earlier computes the developability (chemical-liability) portion live in your browser.
Flagged liabilities and epitopes sit on the antibody's structure. This is a real intact IgG (PDB 1IGT, a murine IgG2a used here as a reference scaffold), colored by residue charge — acidic (Asp/Glu) red, basic (Lys/Arg) blue, neutral cyan (His is ~neutral at pH 7.4) — the same charge coloring as the hero above. Drag to rotate.
Reference structure shown (PDB 1IGT). Score your own candidate below — its ABodyBuilder2 model is rendered in place of this one.
Paste a paired VH+VL. The engine predicts the structure and returns the validated SAP + TAP metrics with the model — no login, no call. In early access; an access key is required while we scale.
Exposed hydrophobic and charged patches on the modelled Fv — surface hydrophobicity, positive and negative charge patches, the VH/VL charge dipole, and total CDR length — placed against the 137 clinical-stage antibodies of Jain 2017. We band on that distribution rather than on the published TAP thresholds because we measured what those thresholds do to our values: 52% of clinical-stage antibodies fall below the published hydrophobicity floor, and not one reaches either charge-patch threshold. A threshold that flags half of all marketed drugs, or can never fire at all, is not a verdict. The published guideline is shown beside every row as context.
Reads the surface block off a finished structure job — no second prediction. Total CDR length is the exception that keeps its published threshold: counting residues needs no calibration. Charge is measured on the unrefined model and hydrophobicity on the refined one, and each percentile is computed against the reference column for that same model. Reference: Jain et al., PNAS 2017, doi:10.1073/pnas.1616408114. An access key is required while we scale.
Paste a mouse or chimeric VH+VL. The engine grafts your CDRs onto the nearest human germline, generates the back-mutation ladder (graft → Vernier → framework tiers → binding-rescue), and scores each variant's humanness — so you can pick the human/binding trade-off you want. Built on open OGRDB germlines, not IMGT — commercial-clean by construction.
Validated against BioPhi-humanized reference antibodies (matched to ~0.98 sequence identity). In early access; an access key is required while we scale.
Paste a VH and/or VL, choose a direction and an optional target isoelectric point, and the engine proposes a ranked path of conservative, one-charge-unit surface mutations to get there. CDR positions are protected so binding is preserved, and every step shows the cumulative pI. Better formulation stability and viscosity, without touching the paratope.
Framework-only, conservative substitutions (K/R→Q, N→D, Q→E and the reverse). In early access; an access key is required while we scale.
Enter candidate mutations by IMGT position and the engine re-runs the entire developability scan — so you see precisely which liabilities a fix removes, whether it introduces new ones, and what it does to pI and the overall score. Recomputed, never estimated.
Positions are IMGT, so a mutation means the same thing across numbering schemes. In early access; an access key is required while we scale.
Log an experimental finding against an IMGT position — a fix that worked, a mutation that hurt binding, a liability that turned out benign — and it surfaces automatically the next time any molecule touches that same position. Append-only and position-keyed, so institutional knowledge compounds across programs instead of living in one scientist's head.
Position-keyed, append-only. A finding at IMGT 62 on the heavy chain surfaces for any future molecule that touches the same position — nothing here is molecule-specific.
Stored by chain + IMGT position, not by molecule name. In early access; an access key is required while we scale.
Paste every candidate from the round and get one ranked table — developability score, CDR-located liabilities, pI per chain, paratope hydrophobicity. Set thresholds and the panel filters itself. Each row is the same computation as a single-candidate report, so the shortlist and the report can never disagree — and a candidate that fails to number is reported, never silently dropped.
Score is 100 minus weighted liability penalties — CDR-located liabilities weighted 8× framework — plus smaller penalties for a pI outside the 4.8–9.9 window, high paratope hydropathy and high CDR net charge. An access key is required while we scale.
Search your candidate against ~177,000 paired antibody sequences from patents and papers (PLAbDab), ranked by IMGT-aligned identity, each hit carrying its source reference and any target annotations. A near-identical hit is a freedom-to-operate flag worth knowing before you spend a program on it.
Reference set: PLAbDab (Abanades et al., Nucleic Acids Research 2024, doi:10.1093/nar/gkad1056). A search heuristic and a starting point for review — not an exhaustive search, and not a legal opinion. In early access.
Paste a panel of candidates: rank them by nearest-neighbour identity, align them on shared IMGT columns and clade them into a dendrogram to check diversity, or diff them against a reference in a mutation grid. Equivalent positions line up even when CDR lengths differ.
Identity is IMGT-aligned, so indels and CDR-length differences count properly. In early access; an access key is required while we scale.
Score every position in a VH or VL against the human germline repertoire it descends from. A rare residue in a framework position is an unusual, often engineerable feature worth a second look; a rare residue in a CDR is exactly what you'd expect, so it is never flagged. Built entirely on our shipped OGRDB germline sets.
Licence-clean — OGRDB germlines, no IMGT. In early access; an access key is required while we scale.
An antibody language model scores every position against the rest of the domain — not against a per-position germline average. A residue can be common in germline and still be wrong in this context, which is precisely the case a frequency table cannot see. Click any residue to get the substitutions the model would tolerate there, with each alternative's probability relative to what you actually have.
Built on AbLang2 (Olsen et al., Bioinformatics 2024; BSD-3). Only framework residues are flagged — unusual CDR residues are what somatic hypermutation produces. We tested whether these likelihoods predict measured biophysics across the 137 clinical-stage antibodies of Jain 2017, and they do not — strongest association ρ=0.24 against a corrected significance floor of 0.27 (n=137, 30 comparisons). So we present this as a naturalness and engineerability signal, not a developability predictor. Everyone else in this market ships language-model scores with no validation attached; we would rather tell you what ours is worth. In early access; an access key is required while we scale.
Place a candidate against the 137 clinical-stage antibody therapeutics of Jain et al. 2017 — the same set we validate the engine on. You get a percentile for every property, so you can see at a glance whether your molecule is ordinary or unusual for something that reached the clinic. And where a property genuinely tracks a wet-lab assay in that set, you also get what the most similar clinical antibodies actually measured.
Most sequence-level properties do not clear the significance bar against these assays, and we show an empty cell rather than a correlation we can't defend — the strong predictor on this same set is structural (SAP vs HIC, ρ=0.60), not sequence-level. Every ρ is recomputed from the reference data at build time, corrected for searching all properties against all assays, and never quoted from a paper. Reference: Jain et al., PNAS 2017, doi:10.1073/pnas.1616408114; assay values via the FLAb benchmark. In early access; an access key is required while we scale.
Build a full-length human IgG1 or an scFv from your variable domains — on UniProt-verified secreted constants, not reconstructed ones. Or go the other way and dissect the VH/VL out of full-length chains, exactly, using the numbering rather than motif guessing.
Knobs-into-holes places its CH3 mutations by EU numbering, which the engine self-verifies against twelve landmarks before it will mutate anything. KiH heterodimerises the heavy chains only — correct light-chain pairing still needs a CrossMab or common-light-chain design, which is not yet offered. In early access; an access key is required while we scale.
Most developability tools return a number and ask you to trust it. We ran our metrics on the 137 clinical-stage antibodies of Jain et al. 2017 and measured how well they rank the actual laboratory assays — using the published TAP and SAP metrics computed on ABodyBuilder2 structures.
Retrospective, rank-based correlation (Spearman ρ) against measured biophysical assays, via the FLAb benchmark. The aggregation result (0.60) is strong for this field; self-interaction (|ρ|=0.33, PNC vs AC-SINS) is solid but field-level — even large protein-AI models do not yet consistently predict developability. Each metric is reported from the model mode its validation supports: SAP from the refined model, charge from the unrefined one. The structure-based engine is live behind an access key; validation methodology and code are available on request. Metrics: TAP (Raybould 2019, PNAS 10.1073/pnas.1810576116) and SAP (Chennamsetty 2009, PNAS 10.1073/pnas.0904191106); benchmark: Jain 2017, PNAS 10.1073/pnas.1616408114.
Immunogenicity scoring earns skepticism for good reason — most tools return a score with no experimental anchor. Every epitope we report is shown with its position, in the form MAPPs and DC–T-cell assays can be checked against, and the model is scored against published clinical ADA rates as a guardrail — a prediction you can check against the bench, not a number to take on faith.
MHC class II binding across an 8-allele HLA-DR reference panel (DRB1*01:01, *03:01, *04:01, *07:01, *08:01, *11:01, *13:01, *15:01; DQ and DP not yet covered), using our own pan-DR two-head model — trained on IEDB eluted-ligand and affinity data, evaluated cluster-held-out. It reaches AUROC 0.789 against real measured non-binders of the same allele, and 0.915 against proteome decoys. We lead with the harder number because almost nobody reports it.
Our output is a peptide list, so it maps onto what MAPPs (MHC-associated peptide proteomics) recovers and what DC–T-cell and T-cell proliferation assays respond to — that is the readout to check us against. We have not run these assays ourselves yet; paired in-silico / wet-lab results will be posted here as we generate them (Future Functionality).
We scored this model against the published clinical ADA rates of 64 approved antibodies as a guardrail, not a calibration: it reaches AUROC 0.686 — real signal by the permutation test, but it does not beat a plain hydrophobicity count of the binding loops at 0.691. So we deliberately report no ADA score, no risk band and no predicted ADA rate. A low/moderate/high band derived from this epitope map scores 0.501 — chance — so we removed it rather than recalibrate it.
Scope: ADA is multifactorial. This addresses the T-cell-dependent (MHC-II / CD4⁺) sequence component only — not B-cell / conformational epitopes, pre-existing anti-drug antibodies, or aggregation-, impurity-, dose- and route-driven immunogenicity. Coverage: the three cores at each chain terminus are not scored — there is no centred 15-mer register for them, and a %rank read off a background that does not apply is worse than no number. Those are framework positions. Results are panel-relative and do not account for an individual patient's HLA background. Provenance: the model is ours and so is the validation. It is trained on peer-reviewed public data (IEDB), evaluated cluster-held-out, and scored against published clinical ADA rates as a guardrail — but every headline number here is our own internal work, not yet peer-reviewed. The full benchmark write-up is available on request and is being prepared for publication.
No. The scanner runs entirely in your browser — sequences you paste are analyzed locally and are not uploaded to or stored on our servers. If you later use the API, anything you send is treated as confidential — never used to train models, and deleted on request.
Sequence-level liability motifs — deamidation (e.g. NG/NS), isomerization (e.g. DG/DS/DT), Met/Trp oxidation, N-glycosylation sequons (N-X-S/T, X≠Pro), unpaired cysteines, DP fragmentation — plus estimated pI and net charge. CDR-weighting, structure-based metrics and the immunogenicity module run in the production engine; access is key-gated while we scale.
MHC class II binding and CD4⁺ T-cell epitope screening across a panel of common HLA-DR alleles, mapped to what wet-lab assays measure (MAPPs, DC–T-cell). It addresses the sequence- and T-cell-driven component of anti-drug-antibody risk; aggregation, impurities, and dose are separate contributors.
No. It is for triage and prioritization — it supports, but does not replace, experimental characterization or regulatory assessment.
No. Reports are computed by the engine and returned to you — no consulting engagement, no scheduling, no readout call. Submit sequences, get the report back.
The free scanner is live now. The paid tiers run on the production engine — CDR-weighted, structure-aware scoring plus the immunogenicity module — which is in early access. Request early access and we'll onboard you when it ships. Pricing is published so there are no surprises.
The scanner is free to use. Paid access is $5,000 a year for your whole organization (unlimited antibodies, unlimited users), or you can buy a single report per antibody; prices are published up front and in US dollars, and paid access is in early access. See Pricing.
Free academic tools each cover one liability. Enterprise suites cover everything, behind a sales process. We give you developability and ADA risk in a single report, priced in the open.
| Capability | Free point tools | Enterprise suites | PresciaBio |
|---|---|---|---|
| Developability flags | ✓ | ✓ | ✓ |
| Immunogenicity / ADA risk | — | ✓ | Early access |
| Both in one report | — | Sometimes | Early access |
| Instant self-serve scanner | ✓ | — | ✓ |
| Published pricing | Free | — | ✓ |
| REST API | — | ✓ | Early access |
| Runs in your browser, no upload | — | — | ✓ |
A generalized comparison of tool categories, not specific vendors. “Early access” = built and running today; access is granted by key while we scale. Anything not yet built is marked (Future Functionality).
A liability caught after a lead reaches the bench is expensive to fix; catching it from sequence is not. Screen as many candidates as you want for free. Then pay $5,000 a year for unlimited antibodies and unlimited users across your organization, or buy a single report to try it first. No per-seat licensing.
All prices are in US dollars. Pricing is published up front. The free scanner is live now; the paid tiers run on the production engine, which is in early access. Everything is self-serve — no sales call, no consulting engagement. Request early access and we'll onboard you when it ships.
Paste a VH/VL, or POST a JSON panel of up to 200 candidates to the API. No integration, no setup, no call. FASTA/CSV upload (Future Functionality).
IMGT/Kabat/Chothia/Martin/Aho numbering, CDR-weighted chemical-liability scoring, and MHC class II epitope prediction — computed, not interpreted. TAP biophysical metrics and salt bridges run on the structure job.
A per-candidate report — flagged positions, ranked shortlist, suggested edits — as JSON, straight back to you. Branded PDF export (Future Functionality).
The free scanner runs this way today. The full engine (CDR-weighting, structure-based metrics, immunogenicity) is in early access.
Numbering & CDR-weighted developability as a clean REST endpoint — the production engine, not the in-browser heuristic. Structure-aware metrics run as a separate async job (POST /v1/structure → POST /v1/surface-properties). Request an access key; self-serve signup and usage billing are (Future Functionality).
Request API access# live · key-gated · real request / response shape curl https://api.presciabio.com/v1/developability \ -H "Authorization: Bearer $KEY" \ -H "Content-Type: application/json" \ -d '{"vh":"EVQLVESGGG…","schemes":["imgt"]}' # → { "developability":{"score":82,"verdict":"low-risk"}, "chains":[{"liabilities":[{"pos":"55-56", "region":"CDR2","type":"deamidation (high)", "severity":6}]}] }
Three decades of experience in antibody engineering.