AI for Bio · Model

REINVENT 4

Open-source small-molecule generative design framework driven by reinforcement, transfer and curriculum learning

Small moleculesPharmaSweden, United KingdomActive中文

Overview

NameREINVENT 4 · aliases: REINVENT4, REINVENT 4
Full nameThe name is a software name; the paper gives no expanded form
Ambiguous names (not auto-linked)REINVENT
FamilyPart of the REINVENT family
OrganizationAstraZeneca Molecular AI (Discovery Sciences, R&D)
Organization typePharma
CountrySweden, United Kingdom
Model classMolecule generation and optimisation framework
DomainSmall molecules (also: Molecule generation)
First release · code release
Basis: First public repository release (v4.0.11)
Note: The site recorded 2024-02-21, which is the journal paper and v4.1.8 date, not the first public version source
Other milestones
  • First public evidence: · First public release source
  • Journal online: · Journal paper online source
  • Code release: · v4.0.11 source
  • Weights release: · Priors moved to a public data repository (v4.6.22) source
Latest update · new version
Basis: v4.9 release
v4.9 adds modular diversity filters, a clustering component, intrinsic rewards and a similarity scoring component, and removes deprecated settings source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · A new version was released within 12 months
OpennessCode: Open · Apache-2.0
Weights: Open · Apache-2.0
Code and priors are Apache-2.0, with priors hosted in a public data repository from v4.6.22
Benchmarks reported0 counted (cited only 1)
Wet-lab validationNone recorded
Unverified fields† training_data.a_public.detail
As of2026-11-16

What it does

Built on recurrent and transformer priors with reinforcement learning, driven from the command line with file-based task configuration, it covers de novo design, R-group replacement, linker design, scaffold hopping and optimisation, merging four previously separate tools into one repository. Unlike prompt-style general generators such as GenMol it needs a scoring function defined per task, and it sits opposite Saturn on the same reinforcement-learning line.

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. 0 of 1 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
19 PMO (Practical Molecular Optimization)Cited as comparison only (not counted)—Third-partyThe PMO paper used the prior-generation REINVENT as a baseline; REINVENT 4 itself reports no PMO score. †partly verified · evidence

Other evaluations (outside the 44)

  • Feature comparison with two similar tools (Comparison against baselines only) Self-reported: The paper gives a table of design modes and learning styles, not performance scores

Developers report: Developers report one open framework covering several molecule design modes, already used in production on internal projects.

Independent evidence: No independent evidence found. Public evaluations feature the previous generation's implementation; no independent evaluation of this version was found

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesPriors were trained unsupervised on SMILES from public molecule libraries; the paper lists no library versions and the versions given by secondary compilations could not be confirmed in the original source
b. Private or licensed dataNoThe paper discloses no internal data used to train the public priors; internal project use is application, not training data source
c. Own wet-lab data used for trainingNo source
d. Wet lab used only for validationNo source

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: The developer is a molecular-AI group inside a large pharmaceutical company's R&D organisation, based in Sweden with members in the UK and funded from internal R&D budget.
  • Funding: Not disclosed
  • Products and partnerships: The same group also maintains a retrosynthesis planning tool; this framework merges four previously separate tools and its code is in a public repository.
  • Related models: Saturn, GenMol

Reading notes

  1. Its value lies in engineering maturity and production use, not in benchmark standing.
  2. REINVENT figures in public evaluations usually come from a third party's implementation of the previous generation and must not be read as this version's results.
  3. The paper compares with only two similar tools and only by feature table, a narrow scope.
  4. No public wet-lab data corresponds directly to this model.

Caveats

  1. The site's recorded first-release date was the journal paper date and is corrected to the first public version.
  2. Under this round's relation rules it enters the zero-pairing list: it has no countable pairing with any benchmark on this site.
  3. The versions of the public libraries behind the priors have no primary source, so second-hand version numbers should not be repeated.

Sources

  1. Journal paper (journal) · 2024-02-21
  2. Release list (repository)
  3. Code repository (repository)
  4. Third-party benchmark paper (preprint)

Corrections

2 corrections
FieldWasNowBasisFound
first_release2024-02-212023-11-07 (first public release v4.0.11)source2026-11-16
benchmarks_reportedHas results on this optimization benchmark (third-party)Only the previous generation is used by third parties as a baseline; this version has no self-reported result, so the relation is changed to cited only and not countedsource2026-11-16

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