After AI Finds a Candidate, What Gets Paid For? ↗
Adds application-specific acceptance criteria and comparison methods to the business question of who performs validation.
R&D Allocation · TOPIC v1
Does an antibody that binds well make a good development candidate?
How reliably can research AI lead to antibody candidates that meet the requirements of a particular application?
Read article → Does an antibody that binds well make a good development candidate?
2026-10-13 13:52:46.719 UTC · Last recorded update
CURRENT READING
A common-condition study of 511 sequences from 29 organizations found different performance in improvement and ranking tasks. Cradle’s high-affinity candidate passed the aggregate threshold but retained concerns in individual developability assays.
Recorded editorial state, not a maturity score
- Blinded comparison under common conditionsRecorded
- Compare application-specific requirementsUnverified
- Reproducibility across targetsUnverified
New question opened · v1
Distinguished antibody design, selection and development decisions, and recorded aggregate scoring alongside individual concerns.
Permanent link to this revision ↗01 / WHAT WE TRACK
How reliably can research AI lead to antibody candidates that meet the requirements of a particular application?
02 / EVIDENCE & TECHNOLOGY
Antibody-design AI / affinity maturation / candidate ranking / developability assessment
03 / EDITORIAL READING
Assess usefulness through reliability, time and total cost to obtain candidates meeting non-negotiable requirements, not only peak affinity or an aggregate pass.
04 / WHAT REMAINS UNKNOWN
The August paper assessed methods from the first quarter of 2025 on a single target with extensive prior data. It does not rank all current AI or all targets. An unfavorable assay does not establish clinical failure, and the study did not measure cost-effectiveness.
05 / NEXT CHECK
Reproducibility across targets, yield of candidates meeting required conditions, synthesis and assay counts, and cost and time to a development decision.
A hypothesis to keep honest
HYPOTHESIS · NOT YET TESTEDPredefining application-specific acceptance conditions may reveal development value more clearly than model rankings alone.
What would change our view: Reconsider the evaluation approach if explicit criteria do not improve reproducibility or decisions, or discard too many promising candidates.
A prospective test proposal, not a prediction score. A graph connection does not validate this hypothesis.
Linked people & organizations
Open an entity to find its cardsCOMPANYCradle ↗
Mentioned in this issue
INSTITUTION / INFRASTRUCTUREAIntibody benchmark ↗
Mentioned in this issue
Related questions, optional follows
Map ↗Building experimental data together for cell AI ↗
Antibodies and cells differ, but both raise the question of evaluation outside training conditions.
Latest card updates
Latest 5 · editorial record dates- v1 · 2026-10-13 13:52:46.719 UTCNew question opened
Distinguished antibody design, selection and development decisions, and recorded aggregate scoring alongside individual concerns.
Reading at the time: A common-condition study of 511 sequences from 29 organizations found different performance in improvement and ranking tasks. Cradle’s high-affinity candidate passed the aggregate threshold but retained concerns in individual developability assays.
Earlier updates (0)
Earlier entries will appear here once more than five reader-facing revisions have accumulated.
Background readings
Source dates and verification scope
Biophysical properties of the clinical-stage antibody landscape ↗
PNAS / PubMed Central · · peer-reviewed-study
Cradle explanation of its antibody benchmark results ↗
Cradle · 2026-10-12 · company-research-commentary
A blinded, prospective benchmark of in silico antibody discovery anchored to experimental affinity and developability ↗
Nature Biotechnology · 2026-08-19 · peer-reviewed-study
This issue’s insight
Research AI cannot be priced by the number of candidates it produces →
More candidates and experiments do not automatically produce faster decisions. Buyer-defined criteria and knowledge retained across providers offer another way to assess research AI.