AI for Bio · Model

MassiveFold

Large-scale sampling and parallelisation around AF2, not a new model; hard interfaces gain most

Structure & co-foldingAcademic labFranceActive中文

Overview

NameMassiveFold · aliases: MassiveFold, AFmassive
Full nameOfficial materials use MassiveFold only; no expanded name
Ambiguous names (not auto-linked)None
FamilyNot a family entry. A single tool; its two inference back ends are not separate models
OrganizationLille University UGSF, CNRS
Organization typeAcademic lab
CountryFrance
Model classSampling and parallelisation framework
DomainStructure & co-folding
First release · preprint
Basis: Preprint of the tool paper (journal version 2024-11-11)
Note: The site had treated 2025-05 as the first release; that is the CASP16-CAPRI data paper source
Other milestones
  • First public evidence: · Preprint of the tool paper source
  • Preprint v1: · Tool paper preprint v1 source
  • Journal online: · Tool paper journal publication source
  • Code release: No sourced date† · No sourced date for the first tagged release source
Latest update · code release
Basis: GitHub release v1.8.5
v1.8.5 tracks the new inference back end and refreshes install files and docs; the v1.8.x series began 2026-05, and the CASP16-CAPRI data paper appeared 2025-08-28 source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · Still releasing as of 2026-11
OpennessCode: Open
Weights: Open
Code is open source, inference uses the official AF2 weights, and the companion dataset ships under an open licence
Benchmarks reported2 counted (board 2)
Wet-lab validationNone recorded
Unverified fields† benchmarks_reported[1].source_url, release_milestones.code_release.date
As of2026-11-16

What it does

Runs AF2-Multimer repeatedly with many parameter sets, generating up to 8,040 structures per target and then selecting a few by confidence. It is a sampling and parallelisation tool rather than a new model, so it cannot be ranked against co-folding models.

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. 2 of 2 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
01 CASPOfficial leaderboard or resultsCASP16, five models per target selected by confidenceBoardGroup took part in CASP16 †unverified · no source link
22 CAPRIOfficial leaderboard or resultsCASP16-CAPRI Round 57 (Round 58 data also released)BoardTook part in Round 57 (CASP16-CAPRI); self-reports 31 targets and 65 interfaces, with no rank. †partly verified · evidence

Other evaluations (outside the 44)

  • Sampling-volume ablation (self report) Self-reported: The paper self-reports that cutting the per-target prediction count sharply costs little accuracy, and that difficulty can be predicted from ordinary confidence, so the tool can be reserved for hard cases (no figures)
  • Failure of the selection step (self report) Self-reported: The paper concludes good models are present in the sampled set but confidence cannot rank them to the top, which is the output bottleneck
  • Compute footprint (resource use) Self-reported: The project began at a 2023 hackathon and its large-scale sampling used a 300,000 GPU-hour national supercomputer allocation (per the paper)

Developers report: Developers report that large-scale sampling helps hardest interfaces most and that difficulty is predictable from ordinary confidence, so it can be used selectively (self-reported).

Independent evidence: No independent evidence found. No official placing is reported and no third-party rerun was found

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesUses the official AF2 weights (PDB, UniRef and others); no new training source
b. Private or licensed dataNo
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationNo

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: An academic team with no company: the UGSF laboratory of Lille University and CNRS (GBLille team), not Strasbourg.
  • Funding: No funding amount is public; its CASP16 sampling used a 300,000 GPU-hour national supercomputer allocation (per the paper).
  • Products and partnerships: The tool is built on AF2-Multimer and publishes its CASP16-CAPRI dataset; it trades compute for accuracy rather than changing the model.
  • Related models: AF2-Multimer, AlphaFold 3

Reading notes

  1. It shows sampling volume raises the ceiling on the hardest interfaces; the bottleneck is selection.
  2. A sampling tool is not a new model; do not rank it alongside co-folding models.
  3. Thousands of structures per target cost far more compute than a single prediction.
  4. No official placing is reported; participation itself is the strongest evidence available.

Caveats

  1. The CASP line rests on the site's own page and is shown as unverified.
  2. The first release had been recorded as the data paper's date and is corrected to the tool preprint.
  3. Inference uses the official AF2 weights, so accuracy is capped by the underlying model.

Sources

  1. MassiveFold tool paper (journal) ·
  2. CASP16-CAPRI data paper preprint (preprint)
  3. CASP16-CAPRI data paper, journal version (journal)
  4. MassiveFold GitHub releases (repository)
  5. CASP16-CAPRI dataset (archive)

Corrections

1 correction
FieldWasNowBasisFound
first_release2025-05 (the data paper preprint)2024-04-30, the tool paper preprint; 2025-05 is the CASP16-CAPRI data papersource2026-11-16

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