Overview
| Name | ReaSyn v2 · aliases: ReaSyn v2, NV-ReaSyn-AR-166M-v2, NV-ReaSyn-EB-174M-v2 |
|---|---|
| Full name | A code name plus version, with no expanded form in the paper |
| Ambiguous names (not auto-linked) | ReaSyn |
| Family | This record centres on the second version; the two papers carry different titles and v2 no longer describes the reinforcement-learning step of v1
|
| Organization | NVIDIA, KAIST |
| Organization type | Mixed |
| Country | United States, South Korea |
| Model class | Molecule generation (synthesis route prediction) |
| Domain | Small molecules (also: Molecule generation) |
| First release | · paper revision Basis: Preprint v1 Note: The v2 checkpoints went up in January 2026 source |
| Other milestones | |
| Latest update | · docs update Basis: README edit on the default branch The last substantive change on the default branch is a README update; decoding and index speed-ups exist only on three unmerged pull-request branches, and an October 2026 commit only updates citation details. Note: The site previously described unmerged branch work as a delivered update source The most recent sourced event, not a claim that nothing changed since. |
| Status | Maintained only · Only documentation and citation updates on the main branch; no new weights |
| Openness | Code: Open · Apache-2.0 Weights: Open · NVIDIA Open Model License Code under Apache-2.0 with two second-version checkpoints in a public model hub under the developer's own open model licence |
| Benchmarks reported | 1 counted (modified protocol 1) |
| Wet-lab validation | None recorded |
| Unverified fields | † release_milestones.weights_release, release_milestones.official_announcement, training_data.b_private_licensed |
| As of | 2026-11-16 |
What it does
Generates synthesis routes as reasoning chains with an encoder-decoder, finding, for an unsynthesisable molecule, a close analogue assembled from catalogue building blocks and reaction templates; the second version adds bottom-up and top-down tree search plus a discrete-flow model that edits a route by insertion, deletion and substitution. It sits on the same molecule-generation line as GenMol from the same organisation but targets synthesisability rather than general generation.
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.
| Benchmark | Relation | Eval set / protocol | Source type | Note and evidence |
|---|---|---|---|---|
| 19 PMO (Practical Molecular Optimization) | Developer report, modified protocol | — | Self-reported | Uses 15 TDC oracle tasks rather than the full 23 †partly verified · evidence |
Other evaluations (outside the 44)
- Synthesis route reconstruction (Own task set) Self-reported: Route reconstruction on 1,000 molecules from each of three sources, against two synthesis-constrained methods
- Hit expansion and surrogate optimisation (Own task set) Self-reported: Analogue expansion and optimisation scored by two computational surrogates, with route accessibility checked by a retrosynthesis tool
- Absent from a third-party comparison (Public benchmark not among the 44) Third-party: A conference paper comparing several models in this area does not list it
Developers report: keeping optimisation on a par with unconstrained methods while enforcing synthesisability, and beating existing synthesis-constrained methods at route reconstruction and hit expansion.
Independent evidence: No independent evidence found. No third-party evaluation was found, and a comparison in the same area does not list it
Training data
Group in Figure 7: Public data only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | 211,220 purchasable building blocks from one supplier catalogue and 115 reaction templates, following an existing published setup source |
| b. Private or licensed data | Not disclosed | |
| c. Own wet-lab data used for training | No | The 10.5 million in-house training routes are all computationally generated source |
| d. Wet lab used only for validation | No |
Wet-lab validation
None. The developers report no wet-lab validation.
Background
- Organization: Produced by a US compute company's research team with a South Korean university, with the lead work done during an internship there; the company's life-science business runs through an open development platform and an open model family.
- Funding: Internal corporate research with no separate funding.
- Products and partnerships: Released alongside an RNA model and several chemistry tool libraries; in January 2026 the company and a US pharmaceutical company announced a joint AI lab with up to $1B over five years, with the split undisclosed.
- Related models: GenMol, Saturn, REINVENT 4
Reading notes
- Matching an unconstrained optimiser means the constraint costs nothing, not that it beats other optimisers.
- Reconstruction depends heavily on where the molecules come from and looks far better on the set that shares a source with the training blocks.
- Synthesisable here means definable by templates and a catalogue, not demonstrably makeable; the paper contains no experimental synthesis.
- The latest substantive event is a documentation update, with the speed-up code still sitting on unmerged branches.
Caveats
- Its pairing here covers only part of that protocol's task set.
- Synthesisability is judged by computational tools with no experimental check.
- The training routes are entirely computationally generated.
- Model card and repository give checkpoint release dates one day apart.
- The platform announcement date was not reverified this round.
Sources
- Preprint v1 2025 (preprint)
- Preprint v2 full text (preprint)
- Code repository (repository)
- Second-version checkpoint model card (repository)
Corrections
1 correction
| Field | Was | Now | Basis | Found |
|---|---|---|---|---|
| latest_update | 2026-07-18 (speed-up commit on an unmerged branch) | 2026-01-11 (README update on the default branch); event kind changed to documentation update | source | 2026-11-16 |
Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.