AI for Bio · Model

GenMol

NVIDIA's masked discrete-diffusion molecule generator covering several generation tasks in one model

Molecule generationBig-tech labUnited StatesActive中文

Overview

NameGenMol · aliases: GenMol, NV-GenMol, NV-GenMol-89M-v2, GenMol NIM
Full nameThe paper gives no expanded name
Ambiguous names (not auto-linked)None
FamilyNot a family entry. v1 and v2 are version axes within one record
OrganizationNVIDIA
Organization typeBig-tech lab
CountryUnited States
Model classMolecule generation (discrete diffusion)
DomainMolecule generation (also: Small molecules)
First release · paper revision
Basis: arXiv preprint v1 submitted source
Other milestones
  • First public evidence: · arXiv preprint v1 source
  • Preprint v1: · arXiv preprint v1 (v3 on 2025-07-22) source
  • Weights release: † · v1 weights on NGC (month-level evidence only) source
Latest update† partly verified · new version
Basis: GenMol NIM 2.0.0 release notes
NIM 2.0.0 moves to the NV-GenMol-89M-v2 checkpoint; GitHub has had only maintenance commits since
Note: Release notes say 2026-04-24 and the NGC container listing 2026-05-07; v2 weights had appeared in the repository earlier source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · New version released within 12 months
OpennessCode: Open · Apache-2.0
Weights: Open · NVIDIA Open Model License
Code Apache-2.0 and weights under the NVIDIA Open Model License; the NIM microservice uses the NVIDIA software licence
Benchmarks reported1 counted (self 1)
Wet-lab validationNone recorded
Unverified fields† latest_update, release_milestones.weights_release
As of2026-11-17

What it does

Writes molecules as SAFE fragment sequences and generates them in parallel by masked discrete diffusion, with fragment remasking to explore chemical space; one pretrained model switches by prompting between de novo, fragment-constrained generation and lead optimisation. Unlike reinforcement-learning frameworks such as REINVENT 4 it needs no retraining per task; it belongs to the NVIDIA BioNeMo / Clara 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. 1 of 1 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
19 PMO (Practical Molecular Optimization)Developer report, official protocol—Self-reportedPMO 23 tasks, original protocol · evidence

Other evaluations (outside the 44)

  • De novo and fragment-constrained generation (Own task set) Self-reported: Validity, uniqueness and quality metrics, author-reported; most models are near saturation on these
  • Lead optimisation tasks (Own task set) Self-reported: Success on 30 lead-optimisation tasks, author-reported

Developers report: one discrete-diffusion model unifies several molecule-generation tasks, leads prior methods on PMO and ships open weights.

Independent evidence: No independent evidence found. No third-party rerun in a shared environment or wet-lab validation was found

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesThe SAFE dataset from ZINC and UniChem, about 946 million molecules source
b. Private or licensed dataNo source
c. Own wet-lab data used for trainingNo source
d. Wet lab used only for validationNo

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: NVIDIA, founded in 1993 and based in Santa Clara, California, is a GPU and accelerated-computing company; Clara and BioNeMo are its open model and tool platforms for biology and drug discovery.
  • Funding: A listed company whose drug-discovery work sits in its internal healthcare unit; its venture arm NVentures invests in drug-discovery companies such as Terray and Iambic.
  • Products and partnerships: GenMol is offered as a NIM microservice on build.nvidia.com and NGC; KERMT, built with Merck, sits in the same BioNeMo line.
  • Related models: REINVENT 4, Saturn, KERMT

Reading notes

  1. The PMO result is author-reported with baselines copied from other papers, so the margin should be discounted.
  2. v2 scores below v1 on PMO; cite the version.
  3. There is no wet lab, so synthesisability and activity of generated molecules are unknown.
  4. The strength is multitask unification and open weights rather than best-in-class on any one task.

Caveats

  1. Baselines were not rerun in the same environment.
  2. Two sources disagree on the latest version date (2026-04-24 vs 2026-05-07).
  3. No GuacaMol or MOSES results are reported.

Sources

  1. arXiv preprint (preprint)
  2. Code repository (repository)
  3. NIM 2.0.0 release notes (official site)
  4. NGC v1 resource page (package registry)
  5. NGC v2 resource page (package registry)

Corrections

2 corrections
FieldWasNowBasisFound
first_release2025-012025-01-10 (arXiv v1 submission date)source2026-11-16
latest_updateNGC v2 weights updated 2025-102026-04-24 NIM 2.0.0 (the NGC container page says 2026-05-07)source2026-11-16

Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-17.