AI for Bio · Model

Saturn

Generative molecular design model aimed at sample-efficient search under a limited oracle budget

Small moleculesAcademic labSwitzerlandActive中文

Overview

NameSaturn · aliases: Saturn
Full nameThe paper uses the model name only; no expanded name
Ambiguous names (not auto-linked)None
FamilyNot a family entry. The predecessor algorithm and downstream synthesizability work have no entity records of their own
OrganizationEPFL (Laboratory of Artificial Chemical Intelligence)
Organization typeAcademic lab
CountrySwitzerland
Model classMolecular generation with reinforcement-learning optimisation
DomainSmall molecules (also: Molecule generation)
First release · paper revision
Basis: arXiv 2405.17066 v1 submission date source
Other milestones
  • First public evidence: · arXiv v1 source
  • Preprint v1: · arXiv v1 source
  • Journal online: · Journal version online (registry-recorded created/issued date) source
  • Official announcement: † · News item from the university AI centre source
Latest update · journal
Basis: Journal publication
Journal version published (DOI 10.1038/s42256-026-01200-4); afterwards the repository merged reaction-condition constraints (2026-06-17) and documentation commits (2026-10-26), counted as secondary code activity source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · Journal publication and repository commits within 12 months
OpennessCode: Open
Weights: Unclear
Code is public on GitHub; licence text and whether weights are released separately were not retrieved this round
Benchmarks reported0 counted · none of the 44
Wet-lab validationNone recorded
Unverified fields† open_status.weights, open_status.code_license, release_milestones.official_announcement
As of2026-11-16

What it does

Combines a state-space (Mamba) sequence model with a replay-plus-augmentation reinforcement-learning algorithm to optimise molecules when calls to an expensive oracle (docking, quantum calculations) are budget-limited, aiming to satisfy multi-objective constraints with fewer calls. Like REINVENT 4 it is a reinforcement-learning optimiser and uses it as the comparison, but it emphasises sample efficiency.

Benchmarks reported

None of the 44 benchmarks. This model is on the zero-pairing list.

How it evaluates itself: Docking MPO and Hit Ratio tasks; PMO

Other evaluations (outside the 44)

  • Own docking multi-objective and hit-ratio tasks (Own task set) Self-reported: Docking multi-objective tasks on three receptors and hit-ratio tasks on five targets, 1,000-call oracle budget with multiple seeds; the abstract claims it beats 22 models, with baselines taken from another paper's table
  • Sample-efficiency context citation (Comparison against baselines only) Self-reported: PMO is cited only as context for the sample-efficiency problem; no scores on that suite are reported

Developers report: a state-space model plus memory-manipulation algorithm beats prior methods on drug-relevant multi-objective optimisation under limited oracle budgets.

Independent evidence: No independent evidence found. No third-party evaluation of this model on a site benchmark was found

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesPretrained on a public compound library (ChEMBL 33) source
b. Private or licensed dataNo source
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationNoNo wet-lab work reported source

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: A laboratory for artificial chemical intelligence inside the chemical sciences and engineering institute of a Swiss technical university, working on AI-accelerated molecular and materials discovery and synthesis; the group was founded in 2022.
  • Funding: Funded by two Swiss national centres of competence in research (catalysis and separations), a national science foundation starting grant and collaborations with chemical and semiconductor companies; in 2025 the group received a corporate early-career award and a university academic award.
  • Products and partnerships: Builds on the group's earlier memory-augmented reinforcement-learning algorithm; this model is an early use of a state-space architecture in generative molecular design, released as open code, with follow-up work on synthesizability constraints.
  • Related models: REINVENT 4

Reading notes

  1. The budget framing matters more than absolute results: the paper uses a 1,000-call oracle budget while PMO's standard budget is 10,000, so the two are not comparable.
  2. Most comparison numbers are taken from another paper's table and only some models were re-run by the authors, so versions and tuning differ.
  3. No wet lab: synthesizability and real activity are unverified.
  4. Nearly two years separate the preprint (2024-05) from the journal version (2026-03), so the comparison set may be dated.

Caveats

  1. Zero pairings: no scores reported on any site benchmark.
  2. The PMO pairing does not hold; the paper cites it only as background.
  3. "Beats 22 models" refers to the authors' own task set, not a public leaderboard position.
  4. The model name is an ordinary word, so site auto-linking must use the common-word guard.

Sources

  1. Preprint arXiv 2405.17066 (preprint)
  2. Journal version (journal)
  3. Code repository (repository)
  4. University AI centre news (official site)

Corrections

1 correction
FieldWasNowBasisFound
benchmarks_reportedCandidate pairing PMO (#19)Not a pairing: the original paper reports no score on this suite and cites it only as backgroundsource2026-11-16

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