Including the tests we failed.
We match the strongest open generators on the standard optimisation benchmark. Where our accuracy breaks down, the number is on this page at the same size as the ones that flatter us — five pre-registered tests came back with no effect, and they are in the same table.
ROC-AUC bars measure skill above the 0.50 chance floor. Rows without current validation show a status instead of a bar. The two binder-recovery rows are a retrospective recovery score, not a hit rate, they cover 11 of our 21 targets, and no molecule behind any number on this page has been synthesised or assayed. Detail and methodology for every row is below.
The exact task, and the metric, with its chance floor where one exists.
Held-out splits, dropped compounds, and analogue-exclusion rules, stated before the run.
An interval, or a plain statement that we do not have one.
The regime the number does not cover. Every benchmark here has one.
Where we lead, and where we don’t.
Compared against the two categories that matter commercially: large pharma-AI platforms, and structure-first generators. Open any row for the reasoning.
We compare against categories rather than named companies on purpose: we will not publish a competitor’s figures we cannot independently verify.
Route-space (recipe) novelty
Novelty measured on the synthesis recipe itself. The signal that maps onto freedom-to-operate rather than onto structural resemblance.
Without a ground-truth recipe there is nothing to measure against, so recipe-level novelty cannot be computed at all.
Design-time genotoxic (ICH M7) screen
DNA-reactive reagents are flagged before synthesis, including a genotoxin that is consumed during the route and absent from the final product.
Product-only mutagenicity screening cannot see reagent or intermediate risk, which then surfaces in process chemistry.
Calibrated interval + published failure modes
Every efficacy call carries a measured confidence interval, and the regimes where the model breaks are published alongside the ones where it works.
Accuracy is typically reported for the regime that flatters it. The novel-chemotype collapse has been reported independently on six ADMET tasks, and it is rarely printed.
Geroscience-native (aging is the objective)
Designs are specified on target × tissue × hallmark, so aging biology is the objective function rather than a filter applied afterwards.
General-purpose platforms retarget oncology and kinase chemistry, where the data and the chemical matter both already exist.
Candidates in minutes, not overnight
A full run finishes in minutes and reaches its top molecules in roughly 1,000 oracle calls, so iterating on a target costs an afternoon, not a quarter.
Overnight training runs put cost and turnaround per campaign in a different bracket.
Not listed as a lead: every candidate we deliver carries a synthesis route. That is real, and it is not a differentiator — reaction-space generation with guaranteed routes is publicly available, so it separates us from structure-first tools and from nobody else. We would rather say so than count it twice.
Competitive with the field's best, while every molecule stays makeable.
QED is a near-saturated oracle where the strongest unconstrained methods bunch at ~0.94–0.95, and we land in that band while every molecule we score carries a route. ⚠ Designing through reactions is not ours alone — it is an active research area with published methods — so read this as a like-for-like score, not as an edge. We do not publish per-method competitor decimals we cannot independently verify.
Our accuracy drops on exactly the molecules we are built to make.Here are both numbers.
This is a known hard problem and it is usually reported only in the regime where it looks good. The same drop has been reported independently on six public safety-model tasks, and the robustness methods tested there did not fix it. We report both regimes and tell you which one your molecule is in.
Similarity to compounds confirmed to bind. Lower = further from anything known.
Our lifespan output sits roughly ≈ 2.4–3.1× further from binder space than known binders sit from each other; our rejuvenation output roughly ≈ 1.7–2.2×. Published as a single band until 2026-10-23, it averaged two programmes that differ by about 40%. About 7.8% of stored molecules reach 0.35 or above, which is inside the applicability domain rather than outside it.
Aging models are built on published old-vs-young expression references. We hold four, so we can ask what most groups cannot: when two of them cover the same gene, do they agree on which direction it moves with age? Across 13,784 genes, 55% are carried by a single source, and of the 6,171 where agreement can be assessed at all, they agree 49.7% of the time.
That includes the two targets our own lifespan programme is built around: CD38 (3 sources) and MTOR (4 sources) are both contradicted. We publish it because it constrains our numbers as much as anyone else’s.
Not a claim we found this first — the field is already asking whether aging clocks are ready, and peer-reviewed reviews conclude they are not validated as clinical surrogate endpoints. We add a measurement. And agreement is not correctness: two mouse-derived sources can agree because they share a bias.
Generators report up to 98% novel scaffolds, measured on the Bemis-Murcko framework. For any engine that assembles molecules from catalogued parts that number cannot fail: couple a known scaffold to a known building block and the combined framework is new by construction. Ours is 82% by the same definition.
One level down the question can fail. Against a complete seed catalogue of 29,124 structures and 710 distinct ring systems, 94% of our molecules contain no ring system the catalogue did not already have. Only 6.3% carry an uncatalogued one.
We deliver a new arrangement of known rings, which is what a reaction does. This corrects our own earlier figure: we previously reported 27.5%, flagged as an upper bound because the seed enumeration was incomplete. Completing it moved the number to 6.3% — the old one was 4.4× too high, and the error flattered us. “Uncatalogued” means absent from our own seed lists, never unknown to chemistry.
A property of the regime, not our implementation. So we detect and disclose rather than claim it away.
A 2026 benchmark found robust training and much larger models move frontier error by ≈ zero. Bars show skill above the 0.50 floor.
Decoration is not the cause, building the molecule out moves it toward binder space, not away.
A pooled number can hide the target where it fails.So here is every target.
A retrospective recovery score: given a target’s confirmed binders hidden among decoys at a realistic ~4.9% active fraction, how well does our similarity model rank the real ones to the top. It is not a hit rate, and it is not a prediction that any molecule works. No molecule from any run on this page has been synthesised or tested in an assay.
Two splits, two regimes, four numbers — never one figure standing for both. Every ROC-AUC is drawn as skill above the 0.50 chance floor, because a bar drawn from zero renders a coin flip half full.
The split changes the answer, so both are here. Enrichment is real and survives a scaffold-disjoint split, and it drops hard once the held-out actives are ones no reference compound resembles: 0.941 to 0.681 on the split we publish, 0.959 to 0.720 on the easier one. The novel slice is 36% of the usable range above chance against 88% in domain — far from a coin flip and far from usable on its own. A 2D similarity score cannot generalise to scaffolds unlike anything it has seen, which is the regime our own output occupies by design.
The scaffold-disjoint novel slice is 101 actives and the random split’s is 83 — both thin, both printed, because an AUC over a hundred-odd compounds is a different object from an AUC over a thousand and the reader cannot tell the two apart from the bar alone.
These figures cover the lifespan programme only. The harness opened one reference file and never the second, so the 10 rejuvenation-only targets have never had their binder recovery measured at all — not measured and weak, simply not measured:
SIRT6 is measured because it sits in both programmes, which is why the gap is 10 and not 11. The harness now reads both reference files, so the run that closes this is a run, not a rebuild. Until it exists, no recovery number on this site covers a rejuvenation target, and a pooled figure that implied otherwise would hide exactly the target where it might fail.
9 of 11 targets recover at 0.948 or better. The 2 that do not are the 2 with the fewest known binders — SIRT6 on 22 and p21 on 16 held-out actives, against 60 for every other target. That is a data-scarcity limit, disclosed, not a modelling claim, and it is why every row carries its n. Bars measure skill above the 0.50 chance floor.
⚠ These are the random split. The harness does not export a per-target breakdown of the scaffold-disjoint split or of the novel slice, so the two figures we lead with are pooled across the 11 and we cannot yet say whether 0.681 holds target by target. Stated because it is the question a reader should ask next.
The discipline: a method ships only if it beats the strongest honest baseline measured on the same split. Pairing our scaffold-disjoint score with a baseline from the easier random split would flatter both numbers, and this table did exactly that until 2026-10-28. The gain over the centroid baseline is real and it is also modest — it is a better ranking of known chemistry, which is not the regime our own molecules live in.
We tested the lifespan model against confirmed mouse negatives.
Above chance, but the interval spans it. Nothing is demonstrated either way.
Then we ran it again, on four times the compounds.
The ITP test was too small to settle anything, so we took the largest open lifespan screen that publishes its negatives 160 rows from Ora Biomedical · Million Molecule Challenge (open data), and asked the same question in C. elegans. The method was fixed in writing before the data was touched.
Barely above the coin flip, and the interval swallows it whole.
This is the number that matters, and it is worse than the AUC. The AUC test was underpowered. We calculated that in advance and it could only have passed above 0.632. The correlation was not underpowered: across all 133 compounds our published in-house figure of ρ 0.158 should have shown up. It did not. What the model predicts and what the worms did are unrelated here.
MMC is a single-lab, fixed-dose, 20 °C screen, and its own no-effect results include several established geroprotectors. That widens what a no-effect result here can mean. It does not turn it into a pass. Raw extracted table and the scoring code are committed alongside the result, so the extraction can be audited independently of us.
Two pre-registered external tests. Neither showed the ability. The lifespan model has no demonstrated ranking ability outside its own training set, and we publish that rather than wait to be asked.
It does not touch the rest of the platform: every candidate still carries a verified synthesis route, a novelty measurement and a design-time genotoxicity screen, and none of those depend on the lifespan model. But if you are here to ask whether we can predict what will extend life, today, honestly, the answer is that we have tried twice to show it and failed twice, and the next test that could settle it is a wet-lab one.
A front-of-funnel engine that de-risks a program, not a full pipeline.
We lead on makeability, novelty, safety-at-design, and calibrated honesty. We're not a clinical organisation, and our efficacy signals are computational and retrospective today, prospective wet-lab validation is the roadmap. On chemistry outside geroscience a specialised library scores low by design PMO DRD2 is 0.000, a dopamine-receptor task our building-block library was never built to reach. All disclosed, none hidden.
We have no longevity-specific benchmark. Every benchmark below is a borrowed one.
PMO, MOSES and the TDC oracles measure molecular generation in general. None of them measures whether a molecule affects ageing.
One ageing-specific benchmark now exists. LongevityBench, 17 tasks across 15 models, and it is built for large language models, which we are not. So it is not a leaderboard we can honestly claim to top, and we will not pretend otherwise.
If we publish one, it follows the rule the two published results already followed: the specification is committed before the result exists, so the test cannot be shaped to its own outcome.
