AI for Bio · Model

Chemformer

Pretrained molecule-string sequence model for reaction prediction and retrosynthesis, long used as a baseline

Small moleculesPharmaSweden, United KingdomArchived中文

Overview

NameChemformer · aliases: Chemformer
Full nameThe paper gives no expanded name
Ambiguous names (not auto-linked)None
FamilyNot a family entry. The organisation's multi-step retrosynthesis search tool is a separate project and is not merged into this record
OrganizationAstraZeneca (Molecular AI, Discovery Sciences)
Organization typePharma
CountrySweden, United Kingdom
Model classRetrosynthesis / reaction prediction (string sequence model)
DomainSmall molecules (also: Molecule generation)
First release · paper revision
Basis: Preprint posted (per the registration agency record)
Note: The site used the 2021-12 / 2022 journal dates as the release; first release follows the preprint source
Other milestones
  • First public evidence: · Preprint posted source
  • Preprint v1: · Preprint v1 source
  • Journal online: † · Journal record issued date (the registration record's created date is 2021-12-07) source
Latest update† partly verified · archived
Basis: Code repository switched to read-only archive
The original code repository was archived read-only; no new version of this model has appeared since the 2022 journal paper and later development is said to have moved to another model library of the same group. This is the latest sourced event
Note: Previously recorded as a code release whose description cited another project's version release from the same organisation; that event has been removed source
The most recent sourced event, not a claim that nothing changed since.
StatusArchived · The original repository is archived read-only and there is no new version
OpennessCode: Unclear
Weights: Unclear
The original repository is archived and returned unavailable through the interface this round, with no licence stated on the page; the organisation's search tool is MIT but belongs to another project
Benchmarks reported1 counted (self 1)
Wet-lab validationNone recorded
Unverified fields† release_milestones.journal_online, latest_update, open_status.code, open_status.weights, training_data.b_private_licensed
As of2026-11-16

What it does

An encoder-decoder Transformer pretrained on large molecule-string corpora and fine-tuned for forward reaction prediction, retrosynthesis and molecular optimisation; its single-step predictions can be called by multi-step search tools to assemble routes. Relative to RetroDFM-R and Uni-Mol3 it serves mainly as the comparison target rather than a new method.

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
07 USPTO family (+PaRoutes)Developer report, official protocol—Self-reportedUSPTO-50K; the Uni-Mol3 re-quote is comparison only · evidence

Other evaluations (outside the 44)

  • Third-party single-step retrosynthesis benchmark (Comparison against baselines only) Third-party: A 2023 third-party comparison ranks it the best template-free model and also evaluates a company-internal reaction set that outsiders cannot obtain (some figures are second-hand compilations)
  • Multi-step route success study (Own task set) Self-reported: A study from the same organisation reports higher multi-step route success than template baselines on a proprietary compound set, with a same-direction result on a public compound set; it trains on tens of millions of proprietary reactions and cannot be reproduced externally

Developers report: after large-scale string pretraining one encoder-decoder model covers forward, reverse and optimisation sequence tasks and beats the template-free methods of the time on the official retrosynthesis split.

Independent evidence: A 2023 third-party single-step retrosynthesis comparison rates it the best template-free model, and later model papers list it as the stronger comparator in their own tables (no figures transcribed).

Training data

Group in Figure 7: Private or licensed data.

ClassValueDetail
a. Public databasesYesPublic patent reaction sets and public molecule-string libraries for pretraining source
b. Private or licensed dataYesA company-internal reaction dataset is used for comparisons and downstream work; its scale and terms are asserted but not detailed publicly source
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationNoThe papers report no wet-lab validation source

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: The developer is a large pharma company headquartered in the UK with three strategic R&D centres in the UK, the US and Sweden; this model comes from the molecular-AI group inside its discovery-sciences organisation in the Gothenburg area. Its 2025 annual report states revenue of about $58.7 billion and about 96,100 employees.
  • Funding: Funded from the company's internal R&D budget with no separate financing; the papers list no external grants.
  • Products and partnerships: The group's open-source line also includes a multi-step retrosynthesis search tool (MIT, able to call single-step models) and the molecule-generation tool REINVENT 4; these are separate projects whose releases are not events of this model. On-site, Uni-Mol3 and RetroDFM-R use it as a baseline.
  • Related models: REINVENT 4, Uni-Mol3, RetroDFM-R

Reading notes

  1. It is the steadiest sequence baseline in retrosynthesis comparisons: later models concede in their own tables that its single-step result is higher.
  2. Differences on the small USPTO split are tiny and the split leaks templates, so rankings mean little.
  3. Second-hand reports note that models trained only on patent reactions do poorly on simple medicinal-chemistry reactions, so baseline scores do not extrapolate to real chemistry.
  4. Multi-step route success rates come from proprietary company data and cannot be reproduced externally, and no new version has appeared since the repository was archived.

Caveats

  1. The original repository is archived and licence and weights status are unclear.
  2. The site previously recorded another project's release from the same organisation as an update to this model.
  3. The internal reaction dataset's scale and terms are asserted only.
  4. The journal online date differs between records and is marked partly verified.

Sources

  1. Model journal paper (journal)
  2. Registration agency record (date basis) (archive)
  3. Archived code repository (repository)
  4. Third-party single-step retrosynthesis comparison (third-party paper)
  5. Multi-step route study from the same organisation (journal)

Corrections

3 corrections
FieldWasNowBasisFound
latest_update2025-12-09, a version release of another project (a multi-step search tool)2026-02-11, this model's code repository archived read-only; the other project's event was removedsource2026-11-16
latest_update.event_kindcode-releaserepository-archivedsource2026-11-16
first_release2022 (journal year)2021-07-15 (preprint)source2026-11-16

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