AI for Bio · Model

NeuralPLexer3

Flow-matching co-folding model emphasising stereochemistry and physical validity

Structure & co-foldingAI-native biotechUnited StatesActiveSite entity page (Chinese site)中文

Overview

NameNeuralPLexer3 · aliases: NeuralPLexer3, NP3
Full nameThe name is a product name plus version; the paper gives no expanded form
Ambiguous names (not auto-linked)NeuralPLexer
FamilyNot a family entry. Earlier NeuralPLexer 1 and 2 fall outside this record, which covers the third generation only
OrganizationIambic Therapeutics
Organization typeAI-native biotech
CountryUnited States
Model classCo-folding (flow matching)
DomainStructure & co-folding
First release† partly verified · product update
Basis: Developer's NP3 Beta preview post
Note: The site recorded the arXiv v1 date, 2024-12-14, as first release source
Other milestones
  • First public evidence: † · Developer Beta preview post source
  • Preprint v1: · Preprint v1 source
  • Journal online: † · Conference proceedings entry (December 2025 meeting) source
  • Official announcement: † · Developer preview post source
Latest update† partly verified · journal
Basis: Main-track entry in the conference proceedings
Entered a machine-learning conference's main-track proceedings in December 2025; no weights or main code release was found.
Note: The meeting date is month-level; the event is recorded as a proceedings publication source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · A proceedings event within 12 months, but no code or weights release
OpennessCode: Closed · BSD-3-Clause
Weights: Closed
Weights and main code are not released; only the NPBench evaluation set is open under BSD-3
Benchmarks reported3 counted (self 1, modified protocol 1, retrospective 1)
Wet-lab validationNone recorded
Unverified fields† first_release, release_milestones.journal_online, latest_update, training_data.c_own_wetlab_for_training, benchmarks_reported
As of2026-11-16

What it does

Generates protein-ligand, protein-protein and peptide complex structures with flow matching, with the paper stressing stereochemistry and physical validity. It is one of two core models on the developer's platform, the other being the property-prediction model Enchant; it differs from AlphaFold3 in its flow architecture and its emphasis on physical validity, and outside users reach it only through partnerships.

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. 3 of 3 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
05 PoseBustersDeveloper report, official protocol—Self-reportedV2 set, with validity checks included · evidence
25 RNA-PuzzlesDeveloper report, modified protocol—Self-reportedCASP15 RNA 8-target subset; AF3 baseline copied from its paper, not rerun · evidence
01 CASPRetrospective dataset use—Third-partyMatrix marks a self-report, but page evidence is only CASP15 RNA, duplicated from the RNA-Puzzles column †partly verified · evidence

Other evaluations (outside the 44)

  • Developer's own evaluation set (Own task set) Self-reported: A self-built set of 1,143 targets covering protein-protein interfaces, peptides and conformational change
  • Comparison with docking software (Comparison against baselines only) Self-reported: The paper places its results alongside two classical docking programs and one diffusion model

Developers report: a higher share of poses meeting both RMSD under 2 Å and physical validity than AlphaFold3 on the V2 pose set, near-perfect stereochemistry and second-scale inference.

Independent evidence: No independent evidence found. Weights and main code are not public and no third-party rerun of the model was found

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesA protein structure archive (before 2020-09-01), about 270,000 distilled entries from a public protein set, and about 12.7 million quantum-chemistry conformers source
b. Private or licensed dataNoThe paper does not mention the developer's private data source
c. Own wet-lab data used for trainingNot disclosed
d. Wet lab used only for validationNo

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: The developer is an AI-native oncology drug company founded in 2019 in California by two chemistry professors from a US institute of technology and a UK university, growing out of a quantum-chemistry software company; its platform spans structure prediction, property prediction and synthesis planning. It listed on a US exchange in September 2026.
  • Funding: A $100M Series B plus a $50M extension in June 2024; a September 2026 listing of 9.4 million shares at $16 raised about $150M, with the prospectus allocating roughly $20M to training new versions of its two core models.
  • Products and partnerships: Its pipeline includes a Phase 1 oral HER2 inhibitor and two preclinical programmes; partners include a Japanese pharmaceutical company (reported milestones up to $1.7B), a US oncology company ($25M for custom training) and another large pharmaceutical company.
  • Related models: AlphaFold 3, Boltz-2, Enchant v1/v2

Reading notes

  1. The comparison with AlphaFold3 rests on numbers the paper recomputed itself; other papers report widely differing figures, so any citation needs its source.
  2. Without the physical-validity requirement the two are close, so the advantage sits in validity alone.
  3. With no public weights nothing can be rechecked externally; only the open evaluation set is reproducible.
  4. The RNA comparison uses eight targets with baselines quoted from another paper, so it carries little weight.

Caveats

  1. Weights and main code are unreleased; outside use runs through partnerships.
  2. First release is taken from the developer preview post, with the paper v1 later; both are recorded.
  3. The proceedings event is pinned only to the month of the meeting.
  4. The CASP pairing reuses the RNA evidence and remains to be rechecked.
  5. Own wet-lab training data is not disclosed, which is not the same as none.

Sources

  1. Preprint 2024 (preprint)
  2. Preprint full text v2 (preprint)
  3. Conference proceedings entry 2025 (paper)
  4. Developer Beta preview post (company blog)

Corrections

1 correction
FieldWasNowBasisFound
first_release2024-12-14(arXiv v1)2024-11-01 (developer beta preview post); arXiv v1 kept as paper_v1source2026-11-16

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