An ADC knowledge base that checks every new design against every construct built before it
Every ADC built here, searchable — each new design checked against every one before it.
See one case, screen by screen ↓A structured, searchable record of every antibody–drug conjugate a team has built: target, antibody format, linker, payload class, conjugation chemistry, drug-to-antibody ratio and its distribution, aggregate, free payload, plasma stability, efficacy, organ findings and why each construct stopped. ADC Design Knowledge Base builds it from your own reports, with the document, page and sentence behind every field.
How to choose AI solutions for life sciences →
Updated
An ADC group has usually answered the design question in front of it already, two or three programmes ago. The answer sits in a repeat-dose study's ophthalmic section nobody re-reads, a conjugation and characterisation report whose size-exclusion results never came inside the aggregate limit, and a decision record that explains why the ratio came down. Inherited programmes add another team's templates and another team's words — “conjugate load”, “monomer loss”, “ocular surface finding”.
So when a new construct comes to the design board, “has this been tried here?” is answered from memory. A hydrophobic linker at a ratio near 8 can come back as a fresh proposal, approved and made, and eighteen months later its four-week monkey study shows the finding a stopped construct showed in 2021, for the same physical reason. Published reviews say this is the normal failure mode: toxicity is the leading named cause of ADC clinical failure, and the dose-limiting finding usually belongs to the payload and the chemistry rather than to the target.
Estimated split for checking one new ADC design before the design board, by hand and with the solution. Hours.
Six specialist agents read, extract, map and check every report, then answer design questions and score each new design; the design board chair decides.
Northwind Bio has built 14 antibody–drug conjugates across five payload classes, with 186 reports behind them — 34 inherited with an acquired programme. Dr. Maya Chen wants a second run at NRX-7 with the payload that gave the strongest growth inhibition in the portfolio. Here is what happens next, screen by screen, in the working solution.
The home screen is the design space: 14 constructs by drug-to-antibody ratio across the page and payload class down it, height in a lane for aggregate at release, colour for how each one ended. The band at a ratio of 6.5 and above is shaded. Beside it: what is open on the record, and the design board's agenda for Thursday — NB-ADC-119 is on it, with no prior-art check attached.
“14 constructs · drug-to-antibody ratio across the page, payload class down it. Height inside a lane is aggregate at release; dashes join one programme.”
The report reader read NB-ADC-118's two-week pilot study and the construct extractor wrote 34 facts; 31 matched confirmed values and 3 went to Dana. Aggregate at release 1.3% by size-exclusion chromatography, no ocular findings at any dose including 12 mg/kg, and 91% of the material at DAR 4 by HIC — each with its document, page and section, and a Confirm or Correct beside it.
“Confirm it and it becomes searchable, correct it and the correction is on the record.”
NB-ADC-104 came across with the acquired Halden programme, 34 documents in that team's templates, with six inherited terms mapped and the original wording kept. Its inherited report gives aggregate at release as 2.6%; the transfer summary gives 3.1% for the same batch. The conflict checker raises it and keeps both values until a steward chooses one.
The conflict checker never averages two values and never resolves a conflict itself.
Maya asks inside the plot. Everything dims except NB-ADC-103, NB-ADC-106 and NB-ADC-116, and a cited lesson appears beneath: both hydrophobic-linker constructs above a ratio of 6 released outside the 3.0% aggregate limit — 6.4% and 5.8% — while NB-ADC-116 at the same ratio of 7.8 on the PEG-extended linker released at 2.3%.
“Above a ratio of 6, the linker decides — not the ratio.”
The source chip opens CMC-0916, NB-ADC-116's conjugation and characterisation report, at page 19, §4.2: at an average DAR of 7.8 the PEG-extended linker held aggregate at 2.3% and unconjugated payload at 0.07%, and clearance matched the unconjugated antibody within 15%.
Maya opens the full answer to “Which of our topoisomerase-1 payload constructs had eye findings, and what did they have in common?”. Two of the five: NB-ADC-103 (grade 2–3 corneal findings in 4 of 6 animals at 10 mg/kg) and NB-ADC-106 (grade 2 in 3 of 6). Both share the hydrophobic tetrapeptide linker, stochastic cysteine and a ratio near 8; the three clean constructs use the PEG-extended linker at a ratio of 4. Nine facts, five sources, confidence 0.93.
The answer writer makes no claim without a source, and says when the evidence is thin and what is missing.
NB-ADC-119 — NRX-7, topoisomerase-1 payload, hydrophobic tetrapeptide linker, stochastic cysteine, ratio 8.0 — is scored against all 14 constructs. The nearest precedents are NB-ADC-103 (91% close, stopped on safety) and NB-ADC-106 (88%, stopped on manufacturability). Four risks, each with its source: ocular findings, aggregate above the release limit, NRX-7's need for a high antigen density, and free payload above 0.35%.
“Two constructs with this exact chemistry were stopped — one on safety, one on manufacturability. The portfolio also shows a design that works.”
The check proposes three changes the portfolio shows work. Maya applies the first — the PEG-extended cleavable linker, on which none of the four constructs had ocular findings and aggregate stayed at 1.1–2.3% even at a ratio of 7.8. The score re-runs to 51, the precedents re-order with NB-ADC-109 now closest, and the verdict reads “Still close to the designs we stopped”.
Bringing the ratio to 4 (−18) and conjugating site-specifically (−12) take the score to 21. The design now sits with NB-ADC-109, NB-ADC-112 and NB-ADC-118 — the three topoisomerase-1 constructs with no eye findings and aggregate near 1%. The one risk left is NRX-7's threshold of about 100,000 copies per cell, a patient-selection question. Maya sends it with the check attached; Priya accepts it for the 8 October board.
“Priya Raman decides. The check proposes; the board chair accepts it or sends it back, and either way it is on the record.”
The same portfolio, seen by the five people who build on it — what their week looked like, and what it looks like now.
Six specialist agents read, extract, map and check every report, then answer design questions and score each new design; the design board chair decides.
Reads study reports, conjugation records and papers, and works out which construct and which study each one is, keeping the page and section for every sentence it passes on.
Pulls the construct-level facts out of each report — target, format, linker, payload class, conjugation, ratio and distribution, aggregate, free payload, efficacy and findings — each with its page, section and sentence.
Maps inherited and published naming onto the house vocabulary — “conjugate load” to drug-to-antibody ratio, “monomer loss” to aggregate at release — keeping the original wording on the record.
Compares every new fact with the confirmed value for that construct and raises a conflict, with both sources attached, when they disagree beyond the field's tolerance.
Scores a new design against every construct built before it on target, payload class, linker, conjugation and ratio, names the nearest precedents and the risks they carry, and proposes the changes the portfolio shows work.
Answers questions in plain words from confirmed facts, direct answer first, with a citation on every claim and a comparison table where it helps.
Accepts or sends back each check. The agents propose; a named person decides.
14 constructs, 186 reports, every answer cited — from confirmed facts only. Each citation opens the report at its page.
Where does aggregate at release go above the limit, and what drives it?
Two constructs sit above the 3.0% release limit: NB-ADC-103 at 6.4% and NB-ADC-106 at 5.8%. Both carry a hydrophobic payload on a hydrophobic linker, conjugated stochastically to a ratio near 8. The ratio alone does not explain it — NB-ADC-116 runs at 7.8 with the PEG-extended linker and sits at 2.3%. What drives aggregate here is hydrophobicity carried at a high ratio.
Which conjugation chemistry gives the tightest ratio distribution?
Enzymatic conjugation is tightest: NB-ADC-112 holds 95% of the material at DAR 4. Site-specific cysteine is close behind at 91%, and both keep aggregate near 1%. Stochastic cysteine puts material on the even ratios only, and at a ratio of 2 leaves a third of the antibody carrying nothing. Stochastic lysine spreads across every ratio from 0 to 7.
Has anyone here made an ADC against NRX-7 work?
Twice tried, never landed. NB-ADC-114 reached only 48% tumour growth inhibition in a model carrying 40,000 copies per cell; NB-ADC-106 reached 79%, only in the antigen-high model, and was stopped at the manufacturability gate. NRX-7 needs roughly 100,000 copies per cell before this chemistry works. Nothing here has tested a PEG-extended, site-specific design against it.
Which linker held up best in plasma?
In cynomolgus plasma at seven days: non-cleavable thioether 98% intact, PEG-extended cleavable 94%, valine–citrulline 88%, hydrophobic tetrapeptide 76% and disulfide 71%. The non-cleavable linker is the most stable but gives no bystander killing; the PEG-extended cleavable linker is the only chemistry here that has held both stability and bystander killing at once.
12 screens from the working solution, on its sample data. Pick one to see it large.
How answers are made, who confirms what and what every prior-art check flags are settings — and the team can add a rule in plain words.
| Setting | Default | Choose from |
|---|---|---|
| Answer only from confirmed facts | On | On · Off |
| Confidence floor | 0.85 | 0.75 · 0.85 · 0.90 |
| Include external literature | On — always labelled | On · Off |
| Who confirms what | Steward: extracted facts · toxicologist: safety facts · CMC scientist: characterisation facts · board chair: accepts or sends back a check | Assigned person by person, by name |
| Attach a check to every proposal | On — no proposal reaches the board without one | On · Off |
| Flag a payload class with a known organ finding | On — the studies are attached to the risk | On · Off |
| Flag a target with a failed earlier attempt | On — with the reason it was stopped | On · Off |
| Flag a target that already has a live construct | On — the proposal must say what is different | On · Off |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
“demo” = seen in the working solution, on its sample portfolio of 14 constructs and 186 reports · “target” = the design goal, measured in the live solution · “estimated” = our estimate · “typical” = published figures (Nguyen et al., Cancers 2023, and “Current and next-generation ADCs: successes and failures” — 97 clinically failed ADC candidates, toxicity the leading named cause; Hamblett et al., 2004 — cAC10-MMAE at 8 drugs per antibody cleared about three times faster than at 4). Research use; quoted facts link back to the original study record. People, companies, programmes and constructs named on this page are fictional — characters and sample data in the working solution.
A structured, searchable record of every antibody–drug conjugate a team has built: target, antibody format, linker, payload class, conjugation chemistry, drug-to-antibody ratio and its distribution, aggregate, free payload, plasma stability, efficacy, organ findings and why each construct stopped. ADC Design Knowledge Base builds it from your own reports, with the document, page and sentence behind every field.
The prior-art reasoner scores the proposal against every construct on target, payload class, linker, conjugation and ratio, pulls the studies behind the closest ones and applies your rules. It returns a repeat-risk score, the nearest precedents and how they ended, the risks with their sources and the changes the portfolio shows work. Applying a change re-scores the design.
Yes. The vocabulary mapper maps the other team's terms onto yours — “conjugate load” to drug-to-antibody ratio, “monomer loss” to aggregate at release — and keeps the original wording on every field. A term with no mapping is proposed to the steward, not invented.
The conflict checker raises it with both sources attached, for example aggregate at release recorded as 2.6% in an inherited report and 3.1% in a transfer summary. Both values stay on the record until a steward keeps one; no agent averages them.
Not by default. Answers come from facts a named steward has confirmed; facts still waiting are shown separately and marked. Below the confidence floor the answer says the evidence is thin and names what is missing.
Yes. Stewards confirm or correct every extracted fact, and the prior-art check only proposes — the design board chair accepts it or sends it back, and either way it is on the record.
Yes, as a setting. Published papers sit beside your own reports and are always labelled as external.
The Agentic Solution Engine builds and deploys it from your requirements and documents — your study, CMC and tox reports, decision records and design rules — and it goes live once every quality gate has passed. We will walk you through it on your own constructs first.
Sourced answers from papers, patents and reports in seconds, our own disagreements first.
1-min demoCompetitive Pipeline TrackingEvery rival trial, readout and approval in a sourced brief each Monday
1-min demoPreclinical Study Report AutomationA draft GLP report in hours, every number traced to raw data.
1-min demoWe'll run ADC Design Knowledge Base on a sample of your own study, CMC and tox reports.