How GenMol results compare to the bench

You have a GenMol result. Here is what it supports.

• Every molecule in your output is a design proposal. Nothing about any of them has been measured, and no published study has synthesized or assayed a GenMol-designed compound. The searches behind that statement are listed in agreement with wet lab.
• The quality metrics describe a set of strings, not a set of compounds. Validity, uniqueness and diversity are properties of the SMILES. The published "quality" figure of 84.6% is defined as valid, unique, QED ≥ 0.6 and SA score ≤ 4 — a filter pass rate, which NVIDIA's own model card lists among performance metrics alongside the warning that "the model may produce molecules that are difficult or impossible in synthesis".
• The SA score is a heuristic, not a route. sa_score comes from RDKit's SA_Score, which scores fragment frequency and structural complexity. It ranks molecules within a chemical series far better than across series, so a low number is a hint to look, not evidence of makeability. See interpreting outputs.
• The rates and the returned set answer different questions. Validity and uniqueness come back as rates over everything generated, while enforce_validity and enforce_uniqueness — both on by default — decide what is in the set you receive. A validity rate below 100% does not mean an invalid molecule reached you.
• Confirm makeability with a retrosynthesis plan (AiZynthFinder on this platform) and then a chemist. Confirm activity by making the compound and assaying it. Nothing on this platform converts a generated structure into a measured property.

Jump to agreement with wet lab or against other in silico methods. The general framing for every tool is in reading an in silico result against the bench.

How in silico is used currently

Fragment-level generative design is used to enumerate candidate structures inside a human-run program, not to decide what to make. In practice that is three things:

• Local enumeration around something you already have. Scaffold decoration, linker design, motif extension and superstructure generation take a fragment or core you have reason to believe in and vary the rest. This is the mode with the least to go wrong, because the surrounding chemistry is familiar and the hypothesis being tested is the series, not the molecule.
• Chemical-space expansion. De novo sampling to produce a diverse set for a downstream computational filter cascade — property prediction, docking, retrosynthesis — where the per-molecule bar is low and the set is what matters.
• Goal-directed optimization against a computational oracle. The PMO mode maximizes a scoring function. This is the setting GenMol's published figures describe, and the thing being maximized is always a calculator or a classifier, never an assay.

What generative design is not used for, by anyone reporting their validation honestly, is concluding that a molecule is active or makeable. Every published program in this class that reached a real compound put a chemist between the generator and the flask.

Against other in silico methods

This is the only section where GenMol has numbers, and every one of them is graded by a computational oracle. They say nothing about bench translation.

De novo generation, author-reported, 1,000 molecules, means and standard deviations over 3 runs. At the paper's quality-favouring setting (one token unmasked per step, softmax temperature 0.5, randomness 0.5):

• GenMol: validity 100.0 ± 0.0%, uniqueness 99.7 ± 0.1%, quality 84.6 ± 0.8%, diversity 0.818 ± 0.001, sampling time 21.1 ± 0.4 s.
• SAFE-GPT, the autoregressive model over the same representation: validity 94.0 ± 0.4%, uniqueness 100.0 ± 0.0%, quality 54.7 ± 0.3%, diversity 0.879 ± 0.001, 27.7 ± 0.1 s.

⚠︎ Quality and diversity trade off against each other, and the trade-off is a parameter you set. At a diversity-favouring setting (temperature 1.5, randomness 10) the same model gives quality 39.7% and diversity 0.911. The single most useful calibration in that paper is its own control: 100,000 real molecules sampled at random from the held-out test set score 38.2% quality and 0.897 diversity. Real drugs mostly fail the quality filter. An 84.6% quality figure therefore measures how tightly the sampler sits inside the QED ≥ 0.6 / SA ≤ 4 box, not how good the molecules are.

Goal-directed hit generation, on the Practical Molecular Optimization benchmark — 23 objectives, a budget of 10,000 oracle calls, scored as area under the curve of the mean top-10 property score, summed across tasks:

GenMol is best on 19 of the 23 tasks. ⊘ The REINVENT row is not REINVENT 4. That figure is taken from the PMO benchmark, whose implementation is the original 2017 RNN-plus-reinforcement-learning REINVENT, and where it ranked first of 25 methods. No published benchmark compares GenMol against REINVENT 4.

Goal-directed lead optimization. Across 30 tasks (five targets — parp1, fa7, 5ht1b, braf, jak2 — with three seed molecules each), GenMol improved the seed in 26 of 30 while satisfying QED ≥ 0.6, SA ≤ 4 and a Tanimoto similarity floor of 0.4 or 0.6. ⚠︎ "Improved" here means a lower docking score in kcal/mol, not a measured affinity. A docking score is a scoring function's opinion; the molecules were never made.

One independent evaluation exists, and it is less flattering. A 2026 docking-oracle benchmark of 25 tasks across hit generation, lead optimization and a new selectivity task evaluated GenMol against Saturn, Genetic-guided GFlowNet and ChemLlama under a strict oracle budget with held-out test tasks. Its findings on GenMol, in the authors' words:

• "no method is superior in all tasks", and "there is no single generalist method that performs consistently better across all the tasks".
• On hit generation, "Saturn and ChemLlama are separated from GenMol and Genetic GFN" by their confidence intervals — GenMol is in the lower pair.
• Hyperparameters do not transfer across targets: "Saturn shows clear cross-protein alignment, GenMol shows essentially none."
• Hit-task and lead-task performance are unrelated for it: "Saturn shows the strongest alignment, GenMol is near zero."
• The practical consequence, stated directly: GenMol "cannot justify reuse across task types and require tuning on a sufficiently similar validation task before test deployment."

Read against the platform's defaults, that last point is the operational one: a GenMol configuration that worked on one target is not evidence it will work on the next.

Gao W, Fu T, Sun J, Coley CW. Sample Efficiency Matters: A Benchmark for Practical Molecular Optimization. Advances in Neural Information Processing Systems. 2022. doi: 10.48550/arXiv.2206.12411

Simonyan G, Abrahamyan T, Abelyan N, Fahradyan T, Khachatrian H. PMO-Dock: Benchmarking Docking, Specificity, and Generalization in Molecular Optimization. 2026. doi: 10.48550/arXiv.2610.19011

In silico vs. wet lab

The standard wet-lab approach for the question GenMol answers is a turn of the Design-Make-Test-Analyze cycle: a medicinal chemist designs a compound, a chemist synthesizes it, and an assay measures it. That is a process, not a measurement technique, and it changes how the table reads.

There is no RMSD or AUC here. A generated molecule is not a prediction about a specific experiment, so there is nothing to grade it against the way a docked pose is graded against a co-crystal. What the literature measures instead is one of three different things:

1. Makeability — does retrosynthesis software find a route, and would a chemist call it feasible?
2. Recovery — does the model regenerate compounds a real project actually made and advanced?
3. Prospective outcome — of the molecules designed, how many were made, and how many were active?

Read each cell against its own question. GenMol has no published figure for any of the three.

On that reading, the honest position for this chain is: seconds instead of months to produce design ideas, with no measured accuracy of any kind attached to them. The chain's value is throughput of hypotheses. The bar it has to clear is a chemist's judgement, not a benchmark.

Two asymmetries matter. Synthesis can fail outright for a given structure and no amount of effort fixes it, which is the largest single reason a generated molecule never becomes data. And the wet lab reports what a specific compound does, while GenMol reports what its training distribution finds plausible; those are not interchangeable at any temperature setting.

Agreement with wet lab

There is none. No wet-lab result for GenMol exists in the literature, and this is the weakest evidence base of any tool on the platform.

That is a stated absence, not a gap in sourcing. The searches run, on 2026-10-22:

No peer-reviewed record of this model exists at all. The primary source is a preprint; the checkpoint is distributed on Hugging Face under NVIDIA's open model licence; the only independent evaluation is another preprint, and it is a docking benchmark.

NVIDIA states the limitation itself. The model card for the checkpoint this platform runs lists, under known risks, that "the model may produce molecules that are difficult or impossible in synthesis", and under technical limitations that the "model may not perform well on sequences that are highly divergent from the ZINC-15 dataset". Its performance metrics are given as validity, uniqueness, diversity, central distance and qualified ratio — all computational. It also places the obligation on you: "Users are responsible for ensuring the physical properties of model-generated molecules are appropriately evaluated."

What the class's evidence looks like, for calibration

These figures are for other generators and are reported here so the absence above has a scale, not as GenMol figures. Do not read any of them as this tool's performance.

Makeability. Against ASKCOS retrosynthesis (route depth ≤ 9, starting materials buyable at ≤ $100/g from a 2018 eMolecules/Sigma-Aldrich list, 60 s per molecule, first route found wins), 30.2% of top-100 molecules had a route on hard GuacaMol objectives for ChEMBL-trained models and 32.7% for MOSES-trained models, pooled across SMILES LSTM, SMILES GA, Graph GA and two baselines. On trivial objectives the figures were 60.1% and 63.5%. Biasing generation by SA_Score raised the hard-objective figure to 80.2% at a cost of more than 0.2 on the normalized objective for some tasks. For reference, the same pipeline judged 68.3% of ChEMBL, 89.8% of MOSES, 60.8% of ZINC and 3.5% of GDB17 synthesizable. A different study using a different retrosynthesis oracle found 98% of GuacaMol top-100 molecules routable and concluded the benchmark was too easy to be informative — which is why neither number is "the" synthesizability rate.

Chemist judgement, and its reproducibility. Seven chemists blindly labelled 30 generated molecules feasible or not; 12 of 30 were called feasible. On a Pi3K/mTOR multi-parameter task, the top 10 molecules from unconstrained generation scored 0 on the panel — not one was called feasible by any of the seven. Adding any synthetic-accessibility constraint moved that to 0.58–0.63. In the same work, SA_Score reached AUCs of 1, 0.33 and 0.96 against chemist labels within each of three separate generations but only 0.63 across all 30 together: it ranks within a series and fails across series.

The referee is itself only fair-to-moderately reproducible. The largest study of chemist judgement collected over 5,000 pairwise preferences from 35 medicinal chemists at one company; agreement, measured in two pilot rounds of 220 pairs with 9 and then 14 chemists, was Fleiss κ 0.40 then 0.32 between chemists and Cohen κ 0.60 and 0.59 within a chemist, with a per-chemist range of 0.16 to 1.00 and 4 of 23 chemist-rounds not significant even at p < 0.1. The task was deliberately reduced to a binary "which of these two do you prefer" precisely because earlier work found agreement lower on anything harder, so those are the optimistic values. The prompt spanned "drug-likeness, synthetic accessibility, or other criteria", so it measures agreement on which molecule to follow up, not on whether one can be made.

Prospective outcome. One campaign in this class is fully reported with denominators, and it used DrugEx v3 with an AiZynthFinder-derived in-house synthesizability score. Its funnel: 100,000 molecules sampled, 32,907 passing both an activity and a synthesizability classifier at probability 0.8, 20,055 with an in-house retrosynthesis route, 4,675 with a route of five steps or fewer, 950 passing Lipinski, 609 after a novelty cut, then three chosen by chemists for diversity, "chemical eye" and one-to-two-step routes. All three were made and assayed against monoglyceride lipase. One had an IC50 of 1 µM; the other two were around 100 µM. At the ≤ 10 µM bar conventionally used for hit finding, one of three was active. The authors attribute scaffold choice to their QSAR model rather than the generator, noting that it rediscovered 145 known actives and that "key scaffolds for the target are re-used, and the sidechains are algorithmically altered".

Lee S, Kreis K, Veccham SP, Liu M, Reidenbach D, Peng Y, Paliwal S, Nie W, Vahdat A. GenMol: A Drug Discovery Generalist with Discrete Diffusion. 2025. doi: 10.48550/arXiv.2501.06158

Noutahi E, Gabellini C, Craig M, Lim JSC, Tossou P. Gotta be SAFE: a new framework for molecular design. Digital Discovery. 2024;3(4):796-804. doi: 10.1039/D4DD00019F

Gao W, Coley CW. The Synthesizability of Molecules Proposed by Generative Models. Journal of Chemical Information and Modeling. 2020;60(12):5714-5723. doi: 10.1021/acs.jcim.0c00174

Parrot M, Tajmouati H, da Silva VBR, Atwood BR, Fourcade R, Gaston-Mathé Y, Do Huu N, Perron Q. Integrating synthetic accessibility with AI-based generative drug design. Journal of Cheminformatics. 2023;15(1):83. doi: 10.1186/s13321-023-00742-8

Choung OH, Vianello R, Segler M, Stiefl N, Jiménez-Luna J. Extracting medicinal chemistry intuition via preference machine learning. Nature Communications. 2023;14(1):6651. doi: 10.1038/s41467-023-42242-1

Hassen AK, Šícho M, van Aalst YJ, Huizenga MCW, Reynolds DNR, Luukkonen S, Bernatavicius A, Clevert DA, Janssen APA, van Westen GJP, Preuss M. Generate what you can make: achieving in-house synthesizability with readily available resources in de novo drug design. Journal of Cheminformatics. 2025;17(1):41. doi: 10.1186/s13321-024-00910-4

What the figure doesn't cover

• No wet-lab result exists for this tool, at any level. Not a synthesis, not an assay, not a chemist panel, not a retrosynthesis census. Every number on this page is either a string statistic or a computational oracle score.
• No peer review. The model's paper is a preprint, and so is its only independent evaluation. The checkpoint's documentation is a vendor model card.
• Nothing on this platform has been validated for GenMol. This page is literature only. The platform's internal validation harness is not evidence of anything about this tool, and no figure here was measured on a platform run.
• "Quality" is a filter pass rate with a stated threshold. It means valid, unique, QED ≥ 0.6 and SA ≤ 4. Real test-set molecules score 38.2% on it. Treat a high quality figure as evidence that the sampler is well-centred, never as evidence that the molecules are good.
• Both halves of that definition are computable on a run, and neither makes the molecules good. The platform returns QED and an SA score for every molecule, so the published quality metric can be reconstructed — which measures how tightly the sampler sits inside that box, nothing more.
• Training-distribution overlap sets the applicability domain. The checkpoint was trained on the SAFE corpus — 1.1 billion SAFE strings drawn from ZINC and UniChem, spanning drug-like compounds, peptides, multi-fragment molecules, polymers, reagents and non-small molecules. NVIDIA's own caution is that performance may degrade on sequences highly divergent from that space. Novelty against training data is not reported by the platform either: the novelty_rate field is never populated.
• Benchmark oracles are not assays, and docking scores are not affinities. The lead-optimization result is 26 of 30 docking scores improved. If you want an affinity, an assay is the only route; MM-GBSA on this platform returns no binding free energy.
• Hyperparameter fragility is documented. Temperature and randomness move quality from 84.6% to 39.7%, and the independent benchmark found GenMol's best settings essentially non-transferable between targets and between task types. A configuration is not a result.
• Reward exploitation is the characteristic failure of the PMO mode. When a computed oracle is the only judge, the optimizer finds that oracle's blind spots. A 0.995 on the DRD2 oracle is a statement about the DRD2 oracle.
• Pipeline attribution is the error to guard against hardest. No generated molecule has ever become a confirmed active on its own. In the one fully reported prospective campaign in this class — a different generator — the molecule that worked passed through a QSAR classifier, a synthesizability classifier, a retrosynthesis route check, a route-length cut, a Lipinski filter, a novelty cut and a chemist's selection before anyone touched a flask. That compound belongs to the campaign.

Go / no-go

Go — act on the result when:

• You want structures to put in front of a chemist, and the chemist is the gate.
• You are enumerating locally — scaffold decoration, linker design, motif extension, superstructure generation, or lead optimization from a seed you trust — where the surrounding chemistry is familiar and makeability is the smaller risk. See inputs and parameters.
• You are feeding a filter cascade and judging the set rather than the molecule: generate broadly, then screen with ADMET-AI, check routes with AiZynthFinder, and treat the survivors as the output.
• You filter on is_valid yourself, and you inspect the 2D structures. Spot-checking is the cheapest control available and it catches the failures that no metric on the card reports.
• You treat the set as a hypothesis pool whose cost is seconds, and you budget the real cost — synthesis and assay — against it honestly.

No-go — do not act on the result when:

• You are reading sa_score as a makeability verdict rather than a within-series ranking heuristic, or comparing it across chemically unrelated molecules.
• You are relying on min_qed or apply_druglikeness_filter without setting them. Both are off unless you ask for them, outside lead_optimization, which applies its own QED floor and SA ceiling.
• You are treating a high quality, validity or diversity figure as evidence that the molecules are good, makeable, or likely active. Real drugs score 38.2% on the quality metric.
• You are ordering or registering molecules straight from the output without a retrosynthesis plan and a chemist reading the structures.
• You are quoting a PMO or docking-score result as a potency claim, or carrying a GenMol configuration from one target to another as though it were validated there.
• You need a hit rate to plan against. There is no published one for this tool, and the honest answer is that nobody knows.

Caveats on reading this page: every figure above is quoted from the cited source with its original metric and sample, and each is labelled as author-reported or independent. Figures from different studies are not comparable, because generators, CASP oracles, objectives and judgement criteria differ between them. Where a number for GenMol does not exist, this page says so and lists the searches rather than substituting a figure from another tool.