Design a new binder to a chosen target

Comparisons

Each table is one study or protocol with its own truth set and scoring. Compare tools within a table, not across tables.

Held for review

Recorded but not shown as evidence until the reason below is resolved.

Not covered yet

• No conventional comparator exists in the retrieved evidence. Not one of the sources runs directed evolution, library display selection or grafting from a known complex on the same targets as a design method, with the same expression and assay pipeline and the same hit definition. This is the primary gap and it cannot be closed by extracting more of these papers; it needs a pass over the display and directed-evolution literature, or a measurement.
• No source anywhere runs two design methods and a conventional comparator on the same targets with one assay and one hit definition, so the use case's own question cannot currently be answered.
• BenchBB (benchbb.bio) was not retrieved. Its seven-target list, protocol, assay and whether any method has been run on it are unknown, and no results on BenchBB were found for any method.
• ProtDBench (arXiv:2605.04118 v2) was not retrieved; only its abstract was read in the search pass. It is the only candidate standing benchmark for binder design found, and nothing is claimed about it here.
• The single-target binder design literature was not retrieved. The search pass surfaced a large population of one-off papers, each reporting a hit rate for one target with its own assay (Francisella virulence factor, deltacoronavirus, Nipah, Langya and measles receptor-binding domains, GDF15-GFRAL, RH5, canine and human TNF-alpha, HSP90, CCL25, BRD3). Taken together they are the largest untapped pool of measured outcomes and the place where publication bias will be worst.
• Adaptyv per-method and per-category hit rates were not extracted. The round 2 repository publishes one row per design with a binding flag and an affinity, but neither the repository nor the paper states an aggregation over those rows, the method labels are entrants' free text, and the three binding thresholds behind the Figure 8 counts are never defined.
• The Adaptyv round 1 repository, raw kinetic curve package, neutralisation package and structure predictions were not retrieved, and the 60 measured affinities in Figure 10 were not read.
• BindCraft's per-target numerators for CrSAS-6 and CbAgo were not obtained. They are printed only inside the Figure 1b graphic, which was not read, and Supplementary Data 1 does not let them be recomputed: the row filter that reconciles with the prose for nine targets fails for PD-L1, CrSAS-6 and CbAgo.
• BoltzGen (bioRxiv 10.1101/2025.11.20.689494) Tables 2 and 3 were not extracted. It reports a cold-start against warm-start split (6 of 9 novel targets against 4 of 5 benchmark targets with known binders) that is exactly the protocol limit the use case needs, and some of its data are stated to be temporarily confidential.
• Filius et al. 2025 (ACS Chem Biol 10.1021/acschembio.5c00774) Supporting Information Table S3 was not extracted. It is a third-party wet-lab test of BindCraft on four targets that reports the two targets which failed outright, which is the kind of independent evidence this use case is shortest of.
• The Nature Communications minibinder paper (doi 10.1038/s41467-026-76760-5) was not extracted. It is the closest thing found to an independent replication: one group, one assay, three targets of differing difficulty with the same method.
• PXDesign's own measured values (bioRxiv 10.1101/2025.08.15.670450, Tables 1 and 2) were not extracted. Its baseline column is deliberately excluded: those numbers are copied from publications and partly adjusted by the PXDesign authors using their own HTRF data, which is a third party's reanalysis.
• The values in AlphaProteo's 'Other design methods' rows and its 'RFdiffusion (published)' row are recorded only as a transcription claim, not as results, for the same reason: they are copied from publications this pass did not retrieve, and Section S4.2 states that the asterisked values were re-measured by the AlphaProteo authors and that for PD-L1 they substituted a different design for the published best one. The primary sources behind those cells (Procko et al., Gainza et al., Cao et al., Berger et al., Yang et al., Watson et al.) were not retrieved.
• Junker and Schoeder 2026 (PLOS ONE 10.1371/journal.pone.0355549) is the most careful cross-method comparison found and has no wet-lab arm, so it supplies no outcome. Nanda 2026 (doi 10.21203/rs.3.rs-11106209/v1) on amyloid fibrils is likewise in silico only. Amyloid fibrils and GPCR peptide targets are target classes for which no measured binder design outcome was found at all.
• No source reports uncertainty for any hit rate or affinity. Every rate in this use case is a point value over a denominator between 6 and 9,000, with no interval.
• No source connects a designed binder to a downstream outcome such as efficacy, developability or manufacturability.
• Automated source review only. Independent human scientific review remains outstanding, and nothing here has been experimentally replicated.

Contribute evidence or propose a correction

Question, sources and review