AI for Bio · Model

MolE

Molecular-graph foundation model adapting disentangled attention to graphs for property prediction

Small moleculesAI-native biotechUnited StatesMaintained only中文

Overview

NameMolE · aliases: MolE
Full nameThe paper gives no expanded name
Ambiguous names (not auto-linked)MolE model, Mole
FamilyNot a family entry. The sibling successor MiniMol has its own record; no family entry
OrganizationRecursion
Organization typeAI-native biotech
CountryUnited States
Model classProperty prediction (molecular-graph foundation model)
DomainSmall molecules
First release · paper revision
Basis: arXiv 2211.02657 v1 submitted
Note: The site treated the 2024-11 journal publication as the release; first release follows the preprint v1 source
Other milestones
  • First public evidence: · arXiv v1 source
  • Preprint v1: · arXiv v1 source
  • Journal online: · Journal publication (DOI 10.1038/s41467-024-53751-y) source
  • Code release: † · First commit in the public code repository source
Latest update · journal
Basis: Journal publication
Journal publication; the public code repository's first commit was 2024-10-04 and the last around 2025-06. This is the latest sourced substantive event source
The most recent sourced event, not a claim that nothing changed since.
StatusMaintained only · Only maintenance commits since publication, and a successor model exists
OpennessCode: Open · CC-BY-NC
Weights: Closed
Code is open for non-commercial use and the full pretrained weights are not released; a third-party container offers a smaller pretrained checkpoint
Benchmarks reported1 counted (board 1)
Wet-lab validationNone recorded
Unverified fields† benchmarks_reported[0], release_milestones.code_release, background.products_partnerships
As of2026-11-16

What it does

Works directly on molecular graphs with atom identities as tokens and graph topological distance as relative position, pretraining first self-supervised (masking atoms and predicting their neighbourhood) and then with graph-level supervision, for downstream property and activity prediction. The sibling successor MiniMol self-reports better results on most comparable tasks with roughly a tenth of the parameters, which makes this model a reference point alongside ADMET-AI and KERMT in ADMET work.

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
04 TDC (ADMET + DTI)Official leaderboard or results—BoardPaper compares with a 2023-09 board snapshot; no row on the live boards †partly verified · evidence

Other evaluations (outside the 44)

  • Pretraining scale ablation (Own task set) Self-reported: Controls on smaller molecular-graph subsets assessing how self-supervised scale affects downstream tasks; author-reported
  • Comparison against classical baselines (Comparison against baselines only) Self-reported: On some tasks fingerprint plus gradient-boosting baselines remain competitive, a point made both by the authors and by later work

Developers report: moving disentangled attention to molecular graphs with two-stage pretraining beat the then-published best results on most tasks of a public ADMET task group (the comparison is a 2023-09 board snapshot).

Independent evidence: No independent evidence found. No independent third-party evaluation was found; the successor's comparison comes from the same organisation and is not independent

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesSelf-supervised on about 842 million molecular graphs from a public compound library (with smaller subsets as controls), then supervised pretraining on about 456,000 molecules with labels from a public bioactivity database source
b. Private or licensed dataNo source
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationNoThe paper reports no wet-lab validation source

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: The developer is an AI-native biotech founded in Salt Lake City in 2013-11 and listed on Nasdaq in 2021-04, grown out of machine learning on high-content cell microscopy; it completed the acquisition of a UK AI drug-design company on 2024-11-20, after which headcount figures differ by source (roughly 600 to 800). Its machine-learning research centre launched 2023-07.
  • Funding: A listed company. In 2023-07 a GPU vendor invested $50 million through a private placement alongside a foundation-model collaboration.
  • Products and partnerships: Drug-discovery collaborations with two large pharma companies (from 2021-12 and 2023-09; announcements through 2026-08 report over $200 million in upfront and milestone payments received). A trade outlet says this model was released jointly with a Danish pharma company while the paper's affiliations read differently; the two accounts do not agree. A new chief executive took office 2026-01-01.
  • Related models: MiniMol, ADMET-AI, KERMT

Reading notes

  1. The "first on most tasks" claim belongs to its 2024 paper and a 2023-09 board snapshot only: this round's sweep of all 22 official sub-task leaderboards found no row for the model, so it cannot be presented as a current board position.
  2. Preprint and journal versions count differently, showing how sensitive such rankings are to the snapshot date.
  3. The full weights are not released, so outsiders cannot reproduce the whole model and must use a third-party smaller checkpoint.
  4. On some TDC tasks classical fingerprint methods remain competitive, so a task-group ranking is not a general advantage.

Caveats

  1. No row for the model was found on the official sub-task leaderboards, so the board relation is shown as unverified.
  2. The comparison is a 2023-09 board snapshot and has since been surpassed.
  3. The full pretrained weights are not released.
  4. Media and paper accounts disagree on which organisations co-released the model.

Sources

  1. Journal paper 2024 (journal)
  2. Journal full text (public archive) (journal)
  3. Preprint v1 (preprint)
  4. Code repository (repository)
  5. Official task-group leaderboard overview (basis of this round's check) (leaderboard)

Corrections

2 corrections
FieldWasNowBasisFound
benchmarks_reportedFirst on 10 of 22 tasks (stated as a leaderboard position)The claim belongs only to its 2024 paper and a 2023-09 snapshot; a full scan of the 22 official per-task boards found no row for it, so it is not a current leaderboard positionsource2026-11-16
first_release2024-11 (journal publication)2022-11-03(arXiv v1)source2026-11-16

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