AI for Bio · Model

Mosaic

Open multi-objective protein design library optimising binders through combined scorers

ProteinsAI-native biotechUnited StatesActive中文

Overview

NameMosaic
Full nameThe repository uses the tool name only; no expanded name
Ambiguous names (not auto-linked)Mosaic
FamilyNot a family entry. A single open-source library with no family members
OrganizationEscalante Bio
Organization typeAI-native biotech
CountryUnited States
Model classDesign (multi-objective differentiable optimisation library)
DomainProteins (also: Molecule generation)
First release† unverified · code release
Basis: Repository creation and first commit (no earlier public announcement or paper found this round)
Note: Policy forbids repository creation as a first release, so this date must display as unverified; the only tagged version is v0.1.0 (2026-10-06) source
Other milestones
  • First public evidence: † · First repository commit (not admissible as first-release evidence) source
  • Code release: · First tagged version v0.1.0 (for an archive deposit) source
Latest update · code release
Basis: Repository commits: structure-model contract and sampler options
Adds a structure-model contract and exponential/geometric sampler options; recent commits added two structure backends and an enzyme-design backend and removed an older one; v0.1.0 tagged 2026-10-06 source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · Tagged release and ongoing commits within 12 months
OpennessCode: Open · MIT
Weights: Unclear
The library has no weights of its own and depends on third-party scorer weights under their own terms
Benchmarks reported1 counted (board 1)
Wet-lab validation2 recorded campaigns
Unverified fields† first_release, release_milestones.first_public_evidence, wetlab_validation[1].date, background.products_partnerships, open_status.weights
As of2026-11-16

What it does

Not a single model but a library that combines co-folding predictors and other scoring terms into one differentiable multi-objective loss, then optimises sequences by gradient to design binders, with swappable scorers that track newer models. Like BindCraft it backpropagates through scorers, but its backends are interchangeable; it differs from the trained generative route of BoltzGen, and typical backends come from models such as Boltz-2.

Benchmarks reported

Every row is one pairing from the survey's relation file; a score belongs to the benchmark page's ranking, where it is shown with metric, eval set, source type and date. 1 of 1 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
28 Adaptyv competitions / BenchBBOfficial leaderboard or results—BoardEntrant in the 2025 de novo track with results on the organiser site; hit rates are self-reported · evidence

Other evaluations (outside the 44)

  • Own two-target minibinder test (Wet-lab validation) Self-reported: Three design arms across two targets, reporting the best arm's per-target hit rate
  • Scorer interchangeability demonstration (Case study) Self-reported: The repository demonstrates plugging different structure and enzyme-design backends into the same optimisation loop

Developers report: assembling off-the-shelf scorers into a differentiable objective already yields minibinders that bind at a high rate, with top-ranked designs in a public competition's de novo track.

Independent evidence: The organiser's results page records design-level outcomes for that competition and can be checked against the entry, but it gives no per-tool hit rate and so is not an independent evaluation of the library.

Training data

Group in Figure 7: Public data + wet lab for validation only.

ClassValueDetail
a. Public databasesNoNo training of its own; relies on the weights of the models it calls source
b. Private or licensed dataNo
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationYesDesigns assayed by a protein measurement service provider, for validation only source

Wet-lab validation

Every rate carries its numerator and denominator. Claims that could not be verified are shown as unverified, without numbers; different paper versions are separate entries.

Competition entry (de novo track) (developer blog)
Targets 1 · Designs tested 10 · Hits 9 · Success 9/10 · Best affinity —

Denominator basis: Ten designs, one target, self-reported
The "1st-4th" framing means four of its designs ranked top-four de novo by affinity, not participant placings; the organiser's participant ranking differs. Keep the n=10 single-target caveat
Executor: Executed by the organiser and its measurement provider; hit counts self-reported

Self-reported · † · verified · source
Own minibinder tests (two targets) (developer blog)
Targets 2 · Designs tested 60 · Hits — · Success 8/10 and 7/10 (best of three design arms) · Best affinity —

Denominator basis: Ten designs per method per target, three methods across two targets, 60 sent in total
Both rates are the best method arm per target and must be stated as selected across three arms
Executor: Executed by a protein measurement service provider

Self-reported · † · verified · source

Background

  • Organization: A small protein-design company focused on computational minibinder design and open-source tooling; its jobs page calls it a well-funded startup, while founding year, funding and headcount are not publicly disclosed and so are not stated.
  • Funding: Not disclosed
  • Products and partnerships: The product is the open library itself, alongside open work porting an open binder-generation model to another framework (per a developer social post, second-hand); no pharma partnerships are disclosed.
  • Related models: BindCraft, Boltz-2, BoltzGen

Reading notes

  1. The 9/10 figure rests on ten designs against one target, so the confidence interval is wide.
  2. Distinguish the ranking: four of its designs placed top-four, which is not a participant placing.
  3. The 8/10 and 7/10 figures are best-of-three arms and must not be read as one method's hit rate.
  4. It is a library, not a model, so performance tracks the chosen scorers; its first-release date is inadmissible under policy and displays as unverified.

Caveats

  1. The first-release basis is repository creation or first commit, which policy disallows; it must show as unverified.
  2. Hit rates and placings come from the developer blog, and the competition ranks by affinity.
  3. The library has no training or weights of its own, so performance cannot be attributed to it alone.
  4. The self-test blog page carries no sourced date.

Sources

  1. Code repository (repository)
  2. Releases API (repository)
  3. Developer competition blog post (company blog)
  4. Developer minibinder blog post (company blog)
  5. Competition results page (leaderboard)

Corrections

3 corrections
FieldWasNowBasisFound
wetlab_validationTook first to fourth place in de novo (participant placings)Its four designs rank top four by affinity; the organizer's participant ranking is differentsource2026-11-16
wetlab_validationHit rates of 8/10 and 7/10 on two targetsSame numbers, but it must be noted that they come from the best of three design armssource2026-11-16
first_release2025 (no basis)2025-01-12, based on repository creation / first commit, which the policy does not accept; shown as unverifiedsource2026-11-16

Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.