AI for Bio · Model

TamGen

Target-aware molecule generator that conditions on a protein pocket and refines compounds

Small moleculesMixedChina, United StatesMaintained only中文

Overview

NameTamGen · aliases: TamGen
Full nameThe paper uses the model name only; no expanded name
Ambiguous names (not auto-linked)None
FamilyNot a family entry. A single model with no family members or successor
OrganizationMicrosoft Research AI for Science, Global Health Drug Discovery Institute (GHDDI)
Organization typeMixed
CountryChina, United States
Model classTarget-conditioned molecule generation (chemical language model + pocket encoder)
DomainSmall molecules (also: Molecule generation)
First release · paper revision
Basis: Preprint posted (registry record)
Note: Previously recorded as 2024-01 (month only); upgraded to day precision from the registry record source
Other milestones
  • First public evidence: · Preprint posted source
  • Preprint v1: · Preprint v1 source
  • Journal online: · Journal publication (DOI 10.1038/s41467-024-53632-4) source
Latest update† partly verified · journal
Basis: Journal publication
Journal publication is the latest sourced substantive event; the repository has no releases and its last commit (2025-09-24) is maintenance whose content was not inspected
Note: Maintenance commits post-date the journal event, but the policy on counting maintenance commits is unsettled, so this is flagged unverified source
The most recent sourced event, not a claim that nothing changed since.
StatusMaintained only · Maintenance commits only, the last on 2025-09-24
OpennessCode: Open
Weights: Unclear
Code is open under the developer's GitHub organisation; licence text and weights release were not retrieved this round
Benchmarks reported1 counted (self 1)
Wet-lab validation1 recorded campaign
Unverified fields† latest_update, open_status.weights, open_status.code_license, wetlab_validation[0].affinity_best
As of2026-11-16

What it does

Adds a protein-pocket context encoder to a chemical language model for target-conditioned generation and refinement of existing molecules; the paper runs a full design-refine-test loop, with the big-lab side generating compounds and a nonprofit drug-research institute doing binding analysis, structure-activity work and enzyme assays. Unlike GenMol and Saturn it closes a real wet-lab loop but does not compete on public optimisation suites.

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
32 CrossDocked2020Developer report, official protocol—Self-reportedSelf-reported on the 100 test pockets, with values in the supplement †partly verified · evidence

Other evaluations (outside the 44)

  • Own metric combination for target-conditioned generation (Own task set) Self-reported: The paper aggregates drug-likeness and docking metrics with a mean reciprocal rank; that aggregation is not an official protocol
  • Viral target case (Case study) Self-reported: The developer page also mentions a coronavirus target; a case description without evaluation data

Developers report: target-conditioned generation plus refinement yields compounds that can go straight to assay, and ranks near the top on several drug-likeness and docking metrics on a public pocket set.

Independent evidence: No independent evidence found. No third-party evaluation or replication found; assays were run by the partner institute and compiled by the authors

Training data

Group in Figure 7: Public data + wet lab for validation only.

ClassValueDetail
a. Public databasesYesDecoder pretrained on 10 million SMILES sampled from a public compound library; target conditioning trained on CrossDocked2020 and similar complex sets source
b. Private or licensed dataNo source
c. Own wet-lab data used for trainingNoNo claim of feeding assay results back into training source
d. Wet lab used only for validationYesIn-vitro enzyme inhibition against one tuberculosis protease target, used for validation only source

Wet-lab validation

Every rate carries its numerator and denominator. Claims that could not be verified are shown as unverified, without numbers; different paper versions are separate entries.

In-vitro protease inhibition, tuberculosis target (journal paper)
Targets 1† · Designs tested 296† · Hits 14† · Success — · Best affinity about 1.9 μM, but that compound is a commercial analogue, not a model-generated molecule†

Denominator basis: 2,612 / 8,635 designs generated with 296 selected onward; the 14 active compounds comprise 5 commercial analogues, 3 structure-activity syntheses, and 6 of 8 directly synthesised compounds (in-vitro inhibition below the 40 μM level)
No overall success rate: the hits span three different sources (commercial analogues, SAR syntheses, direct syntheses) and cannot be merged into one ratio; the published "16 tested" claim does not exist in the paper and is removed
Executor: Assays run by the partner nonprofit drug-research institute; not an independent evaluation

Self-reported · · †partly verified · source

Background

  • Organization: A science-AI research division of a US software company (founded 2022) together with a nonprofit drug-discovery institute founded in Beijing in 2016 by a global-health foundation, the Beijing government and a Chinese university; co-authors also come from two Chinese universities.
  • Funding: The nonprofit institute received about $100M of five-year support initially, and in 2023-06 the foundation announced a further $50M over five years with municipal matching; the paper states partial support from that foundation.
  • Products and partnerships: The two organisations announced a collaboration in 2024-01: the big-tech lab generates candidates and the institute runs the experiments, with stated plans for lead optimisation, ADMET improvement and preclinical work; terms undisclosed.
  • Related models: GenMol, Saturn

Reading notes

  1. The 14 actives are not 14 model-generated molecules: five are commercial analogues and three are structure-activity syntheses, and the most potent one is a commercial analogue.
  2. The real denominators are 2,612 / 8,635 designs with 296 selected; the site's earlier "16 tested" is not in the paper.
  3. In-vitro enzyme inhibition is early hit-level evidence; no cell or animal data, so it is far from a drug.
  4. On CrossDocked the per-metric values live in a figure and the supplement with none in the main text, so cross-paper comparison needs care.

Caveats

  1. Wet-lab hits span three source types and cannot be merged into a single success rate.
  2. The latest-update framing depends on the unsettled maintenance-commit policy and is flagged unverified.
  3. The site said there was no site-benchmark pairing; one self-reported pairing on the official protocol does exist.
  4. The aggregate ranking method is the authors' own, not the benchmark's official protocol.

Sources

  1. Journal paper 2024 (journal)
  2. Journal full text (public archive) (journal)
  3. Preprint registry record (preprint)
  4. Code repository (repository)
  5. Developer research blog (company blog)

Corrections

3 corrections
FieldWasNowBasisFound
wetlab_validation16 compounds tested, 14 strong inhibitors, best IC50 1.88 μMRemoved "16 tested" (the paper does not say so); the 14 actives are 5 commercial analogues + 3 SAR syntheses + 6 of 8 direct syntheses, and the most potent is a commercial analogue; 2,612/8,635 designed, 296 selectedsource2026-11-16
first_release2024-012024-01-08 (preprint posted)source2026-11-16
benchmarks_reportedNo pairing with a benchmark on this siteCrossDocked: self-reported on the official protocolsource2026-11-16

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