Skip to content
Methods · version 1.1 · not peer reviewed

Designing molecules against an aging state, and reporting what the controls killed.

A complete description of the engine as it runs today: how aging is represented, how candidates are assembled so a synthesis route exists by construction, which search operators survived their own adversarial controls, and which did not.

Abstract

Molecule generation has outpaced its own credibility, and geroscience adds a second problem on top: there is no endpoint to optimise that anyone can measure in under three years. GeroQubit takes a different object as the thing being searched. Aging is carried as a 22-component state, 10 tissue components superposed with 12 hallmark components, built from 15,705 genes with a measured aging direction across 4 independent datasets. The registry lists nine operators: eight are integrated and QNS remains a prototype that the integrated pass does not call. The master flag is off by default. Candidates are assembled as reaction recipes rather than emitted as structures, so each delivered molecule carries its proposed route. We report one operator with a measured search-efficiency result, a +22.9% improvement in products per reaction evaluation with a 95% confidence interval of [+16.2%, +33.1%] over 24 paired runs, and we report in the same detail the three operators and one entire subsystem removed after their pre-declared controls failed to separate them from chance. Five pre-registered retrospective tests of our efficacy signal showed no effect or were inconclusive, and all five are described here rather than omitted.

Changes from version 1.0

Most of the previous note describes a system we no longer run.

This is a rewrite rather than a revision. The parts that changed are the parts the old document was built on, so there was nothing structural left to amend.

Componentv1.0v1.1
Aging representationAn 8-gene signature in a 128-D embeddingA 22-component state over 15,705 genes
Search layerNone. A genetic algorithm on fitness aloneNine registered operators: eight integrated; QNS remains a prototype
Reaction choiceFixed template orderRanked by measured template applicability. +22.9% more products per evaluation
Targets11 lifespan targets21 runnable across two programmes, with quality tiers published
Efficacy claimA k-NN signal reported as the efficacy signalFive pre-registered retrospectives, all no-effect or inconclusive, published
ReproducibilityNot testedByte-identical across processes after two string-hash defects were fixed
1

The aging state

Aging is not one number and we stopped producing one. The engine carries a state vector whose two axes are held in superposition rather than averaged, so a candidate that helps liver and does nothing for brain stays legible as exactly that, all the way to the end of the run. Collapsing it to a scalar is a measurement we defer until a human asks for one.

|TISSUE⟩ · 10cnshypothalimmunemuscleadiposeheartlivergutthymusvascular⊗|HALLMARK⟩ · 12→|A⟩22 componentsThe two axes stay separate through the whole search. They are not three filters applied in sequence.
Tissue axis · genes with a measured aging direction
cns
8,595
hypothalamus
7,650
immune
7,257
muscle
6,243
adipose
6,090
heart
4,599
liver
4,102
gut
3,907
thymus
2,605
vascular
1,707

Every tissue is covered in the thousands, thymus included. That last one is worth a sentence: the curated aging database we started from returns zero thymus genes, so this platform ran with no thymus signal at all until a second, cell-type-resolved source was merged in.

That fix reached the aging direction reference and not the tissue clock. The clock that actually scores a molecule in thymus is synthesised from the immune clock by four hand-typed offsets and contains no thymus measurement. So thymus has 2,605 genes of aging direction behind it and zero behind its clock, and a thymus score is labelled derived-from-immune wherever we show one. The honest fix is thymus data, not better constants.

Hallmark axis · gene set sizes
chronic inflammation
2,972
loss of proteostasis
1,904
intercellular communication
1,615
genomic instability
1,211
mitochondrial dysfunction
1,066
epigenetic alterations
1,062
disabled autophagy
677
dysbiosis
478
deregulated nutrient sensing
344
stem cell exhaustion
272
telomere attrition
195
cellular senescence
99

The twelve are not equally well observed, and we do not draw them as though they were. Chronic inflammation has 2,972 genes behind it and cellular senescence has 99.

Sources are merged by precedence rather than by vote, and we tested whether voting would be better. Of 6,175 genes with a direction from two or more sources, the signs agree on 3,070 and clash on 3,105, which is a coin flip. Keeping only the concordant genes improved the downstream metric, and so did keeping only the discordant ones, so the effect was coverage rather than agreement and the agreement claim was withdrawn.

2

Nine registered operators; eight integrated

The registry contains nine entries, but they do not all execute. Eight are integrated behind the default-off QX master flag; QNS is a prototype helper that qx_step does not call. Integrated diagnostics describe the selected batch and do not change a hard gate, score or final ranking. QRC decoder fallback reordering is a separate opt-in path.

QASE
target-state distance

Compares a 22-component representation with a target-state prior.

QCP
pathway prior summary

Summarises target-to-hallmark priors and their source counts.

QF
failure-field margins

Reports candidate gate margins and population decode no-fire shares.

QNS
credible-unknown prototype

A helper exists, but the integrated pass emits no QNS result.

QTFS
tissue contrasts

Separates shared tissue signal from residuals and pairwise contrasts.

QHA
hallmark covariance

Reports static hallmark covariance with an estimability check.

QRF
batch-relative field

Summarises coherence minus conflict relative to the batch.

QRC
template applicability

Which reaction template to try first on the fragment in hand. Applicability is whether the template APPLIES to the structure, not whether the reaction runs in a flask.

QDE
pair hypotheses

Ranks molecule-pair hypotheses; it does not estimate synergy.

QRC decoder reordering is opt-in under a separate flag and changes fallback order, not the reachable product set. The QX master flag itself is off by default. QNS is registered for transparency but is not called by the integrated pass.

3

Reaction-first assembly

The engine designs syntheses, not structures. What it searches over is a route — a scaffold and building blocks brought together by named reactions — and running that route is what produces the molecule, so the two arrive together. A generator that emits a structure first has to guess a route afterwards, and sometimes there is not one.

StartingmaterialsStep 1named reactionIntermediateisolableStep 2named reactionCandidatewith its routeTHE REACTIONS ARE CHOSEN, NOT RECONSTRUCTED AFTERWARDS

Because the reactions are chosen as part of the design, the route is not a post-hoc reconstruction. It is the thing that was optimised. Both assembly steps must take a real reaction template; a candidate where both fell back to a direct bond is dropped, because its route would be fictitious.

The library
Namedevery active template is named after the bond its pattern actually forms
Reagentedeach carries the reagents and conditions that bond needs
Refuseda template that fires but cannot be classified is not admitted

That refusal is the feature. A template whose name and reagent profile disagree with the bond it forms prescribes a route that cannot run, and we shipped one for months that had never once fired on anything.

4

QRC, projecting chemistry onto the reactions that can run

The template-applicability operator, and the only one of the nine granted a steering role. It projects the functional-group state of a fragment onto the reactions that state can support, and tries those first. Nothing is skipped and the reachable product set is identical; only the order changes. That constraint is deliberate, because an operator that changed what is reachable would be changing the chemistry rather than the search.

050%100%27 templates never fired

77 reaction templates ranked by measured success rate over 127,946 recorded decode attempts. 29,047 products, 22.7% overall. 16.1% of all attempts went to templates that have never once produced a product.

Predicting whether a reaction will fire

Held out on 9,173 reactant pairs never seen in training. The scale starts at chance, because a bar drawn from zero makes a coin flip look half competent.

Hand-written rule table0.6372
Per-reaction marginals0.8754
Identity-keyed memo0.8754
QRC0.9460

We first scored these predictors on log loss and the hand-written rule table came out below the per-reaction averages, which cannot be right for a table that encodes real chemistry. The rule emits two probabilities and log loss punishes that calibration rather than the ranking, so the comparison was measuring the wrong thing. Repeated on a rank metric, which is what the engine actually consumes.

Paired A/B
+22.9%

more products per reaction evaluation

95% CI [+16.2%, +33.1%]

Pairs
24 (8 targets × 3 seeds)
Record
24 won, 0 tied, 0 lost
Reaction calls
−4,675 [−5,479, −3,887]
Molecules delivered
unchanged (23 ties)

Separate process per arm, hash seed pinned, and the persisted caches snapshotted and restored around every arm. This is an efficiency result about a search. It says nothing about whether any molecule works in an animal.

What we do not claim from this run: a potency improvement. The peak interval excludes zero, but there are two regressions in the 24 pairs and the mean is carried by one of them. Scaffold diversity moved in a direction whose interval includes zero, so it is not a finding either way.

5

More compute produced fewer and worse molecules

A pathology we measured in our own engine, and the reason this section exists: scaled up, the same search returned fewer molecules and the ones it did return scored lower. The cause was that the search optimises a softer feasibility surface than the verifier enforces, so a population can sit inside a region where everything is rejected and feel no gradient out of it.

The mechanism that closed it runs while candidates are being proposed, and that placement is a constraint rather than a detail. It cannot alter a score, a hard gate or the final ranking, so it has no route by which to make a molecule look better than it is. It changes which candidates get tried; the verifier it is trying to satisfy is unchanged by it.

One target is excluded from it by default, because it regressed there. A mechanism that adds exploitation helps a search that is stuck and hurts one that has already converged, and after five instances of that pattern we treat a lever that helps only a few targets as failing rather than as promising.

6

What the controls removed

Three operators and one entire subsystem were built, measured against controls declared before the measurement, and deleted. The slot each occupied is empty rather than filled with a third variant, because trying variants until one passes is how a no-effect result becomes a result.

QIST
state multiplicity

Beat its permutation control only at grid resolutions that collapsed 25 molecules into 2–16 bins. Exactly 1.000 wherever they stayed distinct.

QST
grid-free replacement for QIST

Bandwidth read off the data. Still below the control: effective n 35.62 against 38.06, −23.6 sd.

QRM
reaction memory on reactant identity

Held-out gain +0.0448 on pairs already seen and +0.0000 on unseen pairs. All of it was recall. Demoted to a cache.

QEA layer
625 lines, 8 mechanisms

Grover diffusion measured classical at shipped strength; the anti-collapse operator fired 4 times in a whole production run; the interference register gave byte-identical output on 8 of 9 paired runs. 0 wins.

The subsystem in the last row was 625 lines of production code with a scoreboard of zero wins, three measured negatives and seven mechanisms never tested at all. It is preserved at a version tag so every experiment stays reproducible by checkout, but negative results belong in a ledger and not in executable code.

7

What is quantum here, stated precisely

A molecule is amplitude-encoded, and the similarity kernel is built from the one-body reduced density matrix marginals of that state. The derivation is genuinely quantum. The object it produces is not, and we are the ones who measured that.

The projected kernel is a radial basis function

The projected quantum kernel is a radial basis function on L1-normalised count fingerprints. One scikit-learn call reproduces it to machine precision. Measured against the standard library implementation on our shipped reference set, the largest absolute difference is 1.44e-15.

Its +39% improvement over plain Tanimoto is real and stays. The cause is count weighting and a nonlinearity, both of them classical. The tell sat in our own documentation for months: the genuinely quantum forms gave no gain, and only the projected one worked.

Fidelity kernel|⟨ψᵢ|ψⱼ⟩|²no gain
Ochiai kernelnormalised overlapno gain
Projected kernelone-body RDM marginals+39%
The one term that is not

Feature co-occurrence is invisible to any function of the one-body marginals alone. If anything here earns the name, it is this term. It moves a rank metric independently by +0.0112, 95% CI [+0.0083, +0.0145], and costs two matrix products through an identity that never forms the full density matrix.

Two-body structure has now been tested three times in this project. It failed twice: in one regime the sample size makes a no-effect result mean the coupling cannot be estimated, and in the other it does not. This is the single win, and we quote the tally rather than the win.

We describe this as quantum-derived and classically computed, and never as quantum computation or quantum advantage. A company named GeroQubit calling a classical kernel quantum would be exactly the inflation the rest of this document exists to prevent. The whole system runs on ordinary processors with no accelerator.

8

A density-matrix applicability domain

Our own retrospective binder recovery reaches 0.941 on scaffold-disjoint held-out actives and falls to 0.681 on genuinely novel chemotypes. That collapse is the regime we operate in by design, so the useful lever is detecting it rather than claiming to have solved it. Every applicability measure in common use is a distance to a nearest neighbour and is therefore blind to direction; treating the reference set as a density matrix is not.

How far our molecules sit from known binders
Real binders, compared to each other0.41 to 0.53
Our output — lifespan programme0.171 median
Our output — rejuvenation programme0.240 median

Our molecules sit roughly three times further from binder space than real binders sit from each other. That distance is what a novel chemotype is, and it is also why any similarity-based score reads low on our output. We publish the low number and explain it rather than loosening the threshold that produces it.

A description nothing else computes

Spectral leakage is the fraction of a query state lying outside the subspace the reference spans. About 60% of one of our molecules points where the reference has no mass, against about 20% for a real binder.

Held-out real actives0.18\u20130.30
Our generated molecules0.56\u20130.64
0.0 · entirely inside the reference1.0 · entirely outside

The same construction says a 200-compound reference set spans an effective rank of 8.45. Two hundred compounds are not two hundred independent directions, and nothing else we ship could say so.

A tie at n = 6 with no interval, so it is quoted as a tie. It earns its place as a description, not as a detector, and it never gates anything.

This collapse is not unique to us. It has been independently reported on six public ADMET tasks, and the robustness methods tested in that work did not resolve it. That report is a preprint on one domain, so it cannot carry a stronger word than this one, and neither can we.

9

Two programmes, and why they are not one list

Extending lifespan and pushing cells younger are different questions on different evidence. We measured the separation rather than assuming it, and it is close to total.

LifespanRejuvenationNote
In GenAge (known longevity genes)11 / 110 / 6circular. The 11 were selected from this literature
Has ageing-perturbation data1 / 116 / 6
Statusshippedall ten rejuvenation-only targets run

The top row is circular and we say so: the lifespan targets were selected from the longevity gene literature, so of course they are all in a longevity gene database. The rows that carry the finding are the other two. Nothing about how the rejuvenation targets were chosen referenced that database, and the perturbation data row was measured rather than selected on. Exactly one gene, SIRT6, appears in both programmes, which is why 11 and 11 give 21 runnable targets and not 22.

Rejuvenation targets, by quality tier

A flat count would hide that three sit under the anchor floor and two carry chemistry caveats, so the tier travels with the number.

generate · 6
METTL3 · SIRT3 · EGLN1 · EP300 · MAPK14 · CXCR2

Anchors buildable, and the available chemistry does not contradict the direction.

thin · 3
ALKBH1 · SOD1 · SIRT6

Real aging evidence, under the 50-binder floor a reliable anchor set needs.

off-target chemistry · 1
KAT7

Its best available anchors are KAT6A compounds, so they are not really its chemistry.

direction conflict · 1
IDH1

Thousands of binders and every one an inhibitor, while the biology asks for activation. Carried as a pathway, with no direction claimed.

CBX4 · CLOCK · FOXO3 · opportunity · not runnable

Essentially no usable small-molecule chemistry: zero actives in ChEMBL above the usual potency threshold, and no ligand-bound structure of any of them in the PDB. FOXO3 does have a published direct binder of its DNA-binding domain (eLife 2019), but it is not in ChEMBL and the whole reported series is ~26 compounds — below our own 50-binder admission rule. They stay visible because the gap is the finding, and no job can be launched against them.

HIF1A · retarget · not runnable

You raise HIF-1α by inhibiting the prolyl hydroxylase, so EGLN1 is the target that actually runs.

Case study · CXCR2 in osteoarthritis

A target with four human safety datasets, and an indication nobody pointed it at.

1 · What the clinic already knows

CXCR2 signalling reinforces cellular senescence from outside the cell: it is the receptor arm of the senescence-associated secretory phenotype (Acosta et al., Cell 133:1006, 2008). Four antagonists have been taken into humans against it. Every one was developed for lung disease or cancer, and every one was dosed systemically.

Navarixin
Merck
COPD, asthma, psoriasis, solid tumours
Phase 2 closed at interim for lack of efficacy. Dose-limiting grade 4 neutropenia.
Danirixin
GSK
COPD
Phase 2 unsuccessful. Severe adverse events at high dose ended a trial.
AZD5069
AstraZeneca
Asthma, bronchiectasis
Cut airway neutrophils as designed. No clinical benefit.
Reparixin
Dompé
CXCR1/2, transplant and oncology
Developed elsewhere, never in osteoarthritis.

The recurring problem across the class is neutropenia. Blocking CXCR2 impairs neutrophil trafficking, so it is a consequence of hitting the target rather than of any particular molecule. A new structure does not fix it, and we do not claim it does.

2 · The gap
Systemic dosing is the constraint

Neutropenia limited the dose in every trial. The exposure that causes it is body-wide.

A joint is not systemic

Osteoarthritis allows dosing directly into the knee. UBX0101 established that route in this exact indication, whatever else that trial did not show.

Nobody has run the combination — and the direction is contested

CXCR2 antagonism plus local joint delivery has not been tested. But the cartilage literature argues the other way: CXCR2−/− mice develop MORE severe osteoarthritis with increased chondrocyte apoptosis (Sherwood, Ann Rheum Dis 74:2207, 2014), and the one intra-articular OA result on this pathway used an engineered AGONIST, not an antagonist (Caxaria, EMBO Mol Med 15:e16218, 2022). An open question, and today the published evidence points against us.

3 · Three candidates, with the route that makes each one

All three are 1-aryl-pyrazole-4-carboxamides built in two steps from commercial starting materials. The engine evolves the recipe, so the route exists before the molecule does rather than being proposed afterwards.

GQ-SM-06
O=C(c1cnn(-c2cccc(F)c2)c1)N1CCNCC1
Route
1pyrazole-4-carboxylic acid+1-chloro-3-fluorobenzene→ Cu N-arylation
2+piperazine→ amide coupling
MW
274.3
logP
1.06
TPSA
50.2
QED
0.891
SA
2.20
Ro5
0
maxTan
0.185
overall
0.155
hERG
0.30
AMES
0.22
DILI
0.75
ClinTox
0.13
HIA
1.00
bioavail
0.94
GQ-SM-07
O=C(NC1CCNCC1)c1cnn(-c2cccc(O)c2S(=O)(=O)C2CCC2)c1
Route
1pyrazole-4-carboxylic acid+cyclobutylsulfonyl chlorophenol→ Cu N-arylation
2+4-aminopiperidine→ amide coupling
MW
404.5
logP
1.39
TPSA
113.3
QED
0.694
SA
2.80
Ro5
0
maxTan
0.250
overall
0.286
hERG
0.30
AMES
0.14
DILI
0.65
ClinTox
0.23
HIA
0.97
bioavail
0.76
GQ-SM-08
CC1=NCCN1C(=O)c1cnn(-c2cccc(F)c2)c1
Route
1pyrazole-4-carboxylic acid+1-chloro-3-fluorobenzene→ Cu N-arylation
2+2-methylimidazoline→ amide coupling
MW
272.3
logP
1.89
TPSA
50.5
QED
0.838
SA
2.60
Ro5
0
maxTan
0.207
overall
0.237
hERG
0.14
AMES
0.08
DILI
0.89
ClinTox
0.06
HIA
1.00
bioavail
0.97

All three clear the hard gates: hERG and AMES both below 0.5, zero Rule-of-Five violations. DILI is read against a 0.6 floor rather than 0.5, because the model scores many safe approved drugs high on it.

4 · One of the three can be ordered today

GQ-SM-06 already exists in a make-on-demand catalogue. It is listed as 1-[1-(3-fluorophenyl)-1H-pyrazole-4-carbonyl]piperazine, supplied by Enamine through Chemspace as CSMS01090540378, at 90% purity with a 30-day lead time to India.

So the usual objection to a computational candidate does not apply to this one. Testing it against CXCR2 starts at $190 for 10 mg, with no synthesis step at all. The other two are not in any catalogue we can find and would need to be made.

This does not soften the novelty figure. The 0.185 is maximum Tanimoto to known CXCR2 binders: it says nobody has tested this chemotype against this target, not that nobody has ever made the molecule. Both are true at once.

GQ-SM-06 · listed price
1 mg$163
2 mg$166
5 mg$175
10 mg$190
20 mg$220
Enamine Ltd. · 90% purity · 30-day lead time to India
Chemspace CSMS01090540378
List price only. Excludes shipping and duties, and a make-on-demand listing is a synthesis promise rather than a shelf item.
5 · How different are they, really

The obvious claim would be that these are structurally novel, so they escape what stopped the others. The measurement does not support it.

0.00.20.40.6failed drugs,to each other0.11–0.17ours, to thenearest failed drug0.18–0.25TANIMOTO SIMILARITY

The four failed drugs are barely more similar to each other than ours are to them. CXCR2 antagonists are already a structurally diverse class, and all of them failed, so a new scaffold is not the argument. The argument is the indication and the route of administration, which is a different thing and is testable.

6 · What this does not show
Nothing has been made or tested

These are designs. There is no synthesis, no assay, no animal data, and no wet lab behind them.

No claim about neutropenia

That liability follows from blocking the target. Nothing here addresses it and nothing here should be read as addressing it.

No joint-residence model

The local-dosing argument is about the indication, not about these molecules. We do not model intra-articular pharmacokinetics, and two of the three carry basic amines that would clear a joint quickly.

No joint tissue in the engine

Ten tissues are modelled and none of them is cartilage or synovium. These were designed in whole-body and immune context.

Direction is designed for, not measured

The activity score is similarity to known binders, and a target’s binder set contains both agonists and antagonists.

The activity axis is weak this far out

Similarity to known binders collapses on novel chemotypes by construction, which is why overall reads low. That is the measurement working, not a hidden problem.

10

Reproducibility

The engine did not reproduce itself across fresh processes, and we did not know until we checked. Three runs of an identical configuration returned three different answers.

Before
0.0714
spread across 3 processes
After
0.0000
identical molecule sets

The obvious suspect, floating point addition across threads, was not the cause. It was string hashing, which the language randomises per process by design.

The two defects
  1. 1

    BRICS.BRICSDecompose returns a set of SMILES strings, so the building-block pool came back in a different order in every process, and two identical runs did not agree.

  2. 2

    A bond chosen by hash((smiles_a, smiles_b)), under a comment claiming two runs on the same pair produce the same bond. They did not.

Historical A/B deltas are unaffected: every rig in the repository already pinned PYTHONHASHSEED=0, so both arms of every comparison saw a deterministic engine. Absolute peak and yield values from before the fix do not reproduce and are not quoted as current.

11

Results, including the five that failed

Five retrospective tests of the efficacy signal were pre-registered, with the specification recorded before the result in every case. Four showed no effect and one is inconclusive. None of them is omitted here, and none of them is described as trending.

1Can the lifespan model rank the NIA-ITP mouse winners above the ITP no-effect compounds?AUC 0.645 [0.447, 0.826]Inconclusive

n = 35. A Hanley-McNeil calculation done afterwards shows the design could only ever have reached significance at AUC 0.695, so it was underpowered by construction. That does not rescue the model; it means this particular test can no longer settle it.

2Can it rank the Million Molecule Challenge worm winners above its no-effect compounds?AUC 0.532 [0.391, 0.671]No effect

n = 133. The damaging figure is the secondary one, not the AUC: Spearman against the continuous outcome is -0.0226 at p = 0.796, and that arm was not underpowered. Ranking the same compounds by another laboratory’s measured lifespan scores no better, so the test had no headroom left in it.

3Does a molecule’s target perturbation point cells toward a younger state?ρ = -0.086, p = 0.159No effect

The first run of this read p < 0.0001 and replicated beautifully. It was an artifact: the two contrasts being correlated share their Old group, and all 593 caloric-restriction records came from one study. Dropping that study collapsed the result. When two contrasts share a group, correlating them is not a test.

4Does our aging direction predict which way a gene moves with age in GTEx?0.547 [0.487, 0.606]No effect

A real bug was found and fixed on the way: the harness took the median of six age brackets and correlated six points, discarding about 98% of the samples. Fixing it moved the number the predicted way and it still showed no effect.

5Can the aging reference order individual people by age?Two tissues above chance, five belowNo effect

Not underpowered: every interval could have resolved a 0.60 effect and five of nine resolve a departure from chance. The diagnosis is exact, and it is in section 11.

The one diagnosis that came out of them

Test five has an exact answer. Our aging reference returns the median change between an old group and a young group, which is a population-level contrast, and ordering individuals needs an individual-level weight. Fitting weights to individuals on the same samples, the same genes and the same protocol moves the score from 0.483 to 0.676, above chance in 45 of 45 donor-disjoint splits.

Every published aging clock fits weights to individuals. None reads them off a group difference. That is a category error rather than a tuning problem, and it explains all five tissues that sat below chance.

What still holds
  • Aggregate marker-level use of the reference, which reproduces textbook aging direction on a pre-specified gene list at p = 2e-5.
  • Synthesizability by construction, which depends on no efficacy model.
  • Route novelty, genotoxic screening at design time, and developability, for the same reason.

The efficacy signal itself reports a leave-one-out correlation of 0.158 over 775 compounds. That is weak in absolute terms and we say so wherever it appears.

12

Limitations

Stated in full, because a limitation a reader finds for themselves costs more than one they were handed.

Nothing has been made

No candidate this system has designed has been synthesised or assayed. We have no wet laboratory. Every number on this page is computational.

The efficacy signal is weak and unvalidated externally

Two pre-registered retrospectives on two organisms found no demonstrated external ranking ability. The binding constraint is cross-dataset label disagreement rather than sample size, so a third public database would not settle it.

We have no time axis

Every model here is static. The aging state is a difference between an old group and a young group, with no rate and no relaxation time, because we have no longitudinal data to attach one to. Claiming a dynamical result from static contrasts would be inflation.

Two hand-typed priors are live

A tissue coupling matrix and a hallmark conservation table are hand-assigned and uncited. Their directions reflect real consensus; their decimals are not measured and are never published.

Seven of the nine operators have no A/B behind them

One has a measured steering effect and one survives its statistical controls. The rest inform the search without a paired experiment demonstrating that they help.

One licence is unresolved

One of the four aging sources is licensed for academic use only and we are commercial. It covers 5% of entries, no tissue depends on it for coverage, and it is the only one of the four that passes its own validation test. That is a business conversation, not a code change.

13

Data, code and availability

Validation draws on public datasets. The engine internals do not ship.

Public data used

DrugAge for measured lifespan, ChEMBL for bioactivity, GTEx for tissue expression, Aging Atlas and Tabula Muris Senis for aging direction, Open Targets and GenAge for target evidence, and the López-Otín hallmark framework.

Not disclosed

Scoring weights, the search operators, the generator encoding and the reaction template set are proprietary. Aggregate validation results are reported in full.

Available on request

Representative candidate structures with their routes, under a research agreement. The code is available for audit on request.

No human or animal subjects were involved. This platform is a research tool for hypothesis generation and is not intended for clinical, diagnostic or therapeutic use. This document has not been peer reviewed.

References
  1. 1.López-Otín C, Blasco MA, Partridge L, Serrano M, Kroemer G. Hallmarks of aging: an expanding universe. Cell. 2023;186(2):243–278.
  2. 2.López-Otín C, Blasco MA, Partridge L, Serrano M, Kroemer G. The hallmarks of aging. Cell. 2013;153(6):1194–1217.
  3. 3.Gems D, de Magalhães JP. The hoverfly and the wasp: a critique of the hallmarks of aging as a paradigm. Ageing Res Rev. 2021;70:101407.
  4. 4.Flórez-Ablan M, Roth M, Schnabel J. On the similarity of bandwidth-tuned quantum kernels and classical kernels. arXiv:2503.05602. 2025.
  5. 5.Barardo D, et al. The DrugAge database of aging-related drugs. Aging Cell. 2017;16(3):594–597.
  6. 6.Harrison DE, et al. Rapamycin fed late in life extends lifespan in genetically heterogeneous mice. Nature. 2009;460:392–395.
  7. 7.Swanson K, et al. ADMET-AI: a machine learning ADMET platform for evaluation of large-scale chemical libraries. Bioinformatics. 2024;40:btae416.
  8. 8.Gao W, Fu T, Sun J, Coley CW. Sample efficiency matters: a benchmark for practical molecular optimization. NeurIPS. 2022.
  9. 9.Bickerton GR, Paolini GV, Besnard J, Muresan S, Hopkins AL. Quantifying the chemical beauty of drugs. Nat Chem. 2012;4:90–98.
  10. 10.Ertl P, Schuffenhauer A. Estimation of synthetic accessibility score of drug-like molecules. J Cheminform. 2009;1:8.
  11. 11.Rogers D, Hahn M. Extended-connectivity fingerprints. J Chem Inf Model. 2010;50:742–754.
  12. 12.Deb K, Pratap A, Agarwal S, Meyarivan T. A fast and elitist multiobjective genetic algorithm: NSGA-II. IEEE Trans Evol Comput. 2002;6:182–197.
  13. 13.Oh HS-H, et al. Organ aging signatures in the plasma proteome track health and disease. Nature. 2023;624:164–172.
  14. 14.Pridham G, Rutenberg AD. Network dynamical stability and biological aging. J Gerontol A Biol Sci Med Sci. 2024;79(10):glae021.
  15. 15.Kedlian VR, Dönertaş HM, Thornton JM. The widespread increase in inter-individual variability of gene expression in the human brain with age. Aging. 2019;11(8):2253–2280.
  16. 16.Pyrkov TV, et al. Longitudinal analysis of blood markers reveals progressive loss of resilience and predicts human lifespan limit. Nat Commun. 2021;12:2765.
  17. 17.Sen P, et al. Histone acetyltransferase p300 induces de novo super-enhancers to drive cellular senescence. Mol Cell. 2019;73(4):684–698. PMID 30773298.
  18. 18.Freund A, Patil CK, Campisi J. p38MAPK is a novel DNA damage response-independent regulator of the senescence-associated secretory phenotype. EMBO J. 2011;30(8):1536–1548. PMID 21399611.
  19. 19.Acosta JC, et al. Chemokine signaling via the CXCR2 receptor reinforces senescence. Cell. 2008;133(6):1006–1018. PMID 18555777.
  20. 20.Wang Q, et al. Aging Atlas: a multi-omics database for aging biology. Nucleic Acids Res. 2021;49(D1):D825–D830.
  21. 21.Tabula Muris Consortium. A single-cell transcriptomic atlas characterizes ageing tissues in the mouse. Nature. 2020;583:590–595.
Earlier version
v1.0 · superseded

The original methods note. It is kept available because withdrawing a document people may have read is worse than labelling it, but it describes an earlier engine: an eight-gene signature instead of a state, no search layer, and an efficacy figure that has since been re-baselined downward after we found a data bug that had inflated it. Read this page instead.

↓ v1.0 PDF
v1.1 · manuscript

This page as a typeset preprint: abstract, numbered sections, tables and references, formatted for submission rather than for the browser. Six pages.

↓ v1.1 preprint PDF
Current numbers

Benchmarks, every pre-registered test in full, and every interval are on the benchmark page.

The QX search layer

Not another molecule generator.A different way to search.

Most design systems score a molecule against a list of objectives and pick the winner. GeroQubit treats discovery as one search through biological and chemical state at the same time, and moves compute toward the directions that hold up.

What the search actually evolves · a route, not a drawing
SCAFFOLD
+
BLOCK 1
AMIDE COUPLING
+
BLOCK 2
SNAr
CANDIDATE

The reactions are chosen as part of the design, so the route exists before the molecule does. A generator that draws a structure first has to guess the route afterwards, and sometimes there is not one.

01
Explore
thousands of candidate interventions at once, not one at a time
02
Converge
compute moves toward the directions that keep surviving checks
03
Commit
one molecule, with the reactions and reagents that build it

Nine registered operators · eight integrated

|A⟩AGING STATEtarget · tissuehallmarks · objectiveQASEaging-state evolutionQCPpathway convergenceQFfailure-field guidanceQNScredible noveltyQTFStissue separationQHAhallmark structureQRFrejuvenation fieldQRCtemplate applicabilityQDEmulti-intervention

This diagram indexes the registered names; it is not an execution trace. Eight operators are integrated behind the default-off QX master flag. QNS is a prototype helper that the integrated pass does not call. The displayed state and priors are descriptive, not causal evidence or measured molecule effects.

Behind the tissue axis

Genes with a measured aging direction, per tissue. Every tissue in the state is covered in the thousands, including thymus, which one of our four sources does not carry at all. Direction only: thymus has 2,605 genes here and a clock synthesised from immune, so a thymus score is labelled derived-from-immune rather than presented as measured.

cns
8,595
hypothalamus
7,650
immune
7,257
muscle
6,243
adipose
6,090
heart
4,599
liver
4,102
gut
3,907
thymus
2,605
vascular
1,707

Behind the hallmark axis

The twelve hallmark gene sets are not the same size, and we do not draw them as though they were. Chronic inflammation has 2,972 genes and cellular senescence has 99, so the two are not observed equally well.

chronic inflammation
2972
proteostasis loss
1904
intercellular comms
1615
genomic instability
1211
mitochondrial dysfunction
1066
epigenetic alterations
1062
disabled autophagy
677
dysbiosis
478
nutrient sensing
344
stem cell exhaustion
272
telomere attrition
195
cellular senescence
99
15,705
genes with an aging direction
4
independent aging datasets
21
targets you can run today
10 × 12
tissue by hallmark state
The name

Aging is a state, not a score.

GERO is geroscience. QUBIT is how we hold it: aging carried as a quantum-derived state vector, ten tissue components superposed with twelve hallmark components, and never collapsed to a scalar until a human asks for one. Below is a real candidate from a real run, read out both ways.

ONE CANDIDATE, AS BUILT

The molecule assembled in the recipe above, carried through the search as a state rather than a score.

COLLAPSED TO ONE NUMBER
0.354

Its own hallmark mean. Everything to the right disappears into it.

TISSUE COMPONENTS · 10
adipose
0.58
gut
0.57
liver
0.57
vascular
0.57
hypothalamus
0.57
immune
-0.00
cns
-0.01
heart
-0.02
muscle
-0.03
thymus
-0.13

Five tissues respond. Five do not. Brain is one of the five that do not.

HALLMARK COMPONENTS · 12
nutrient sensing
0.95
disabled autophagy
0.90
proteostasis loss
0.70
cellular senescence
0.45
intercellular comms
0.45
mitochondrial dysfunction
0.40
stem cell exhaustion
0.40
genomic instability
0.00
telomere attrition
0.00
epigenetic alterations
0.00
chronic inflammation
0.00
dysbiosis
0.00

Runs 0.95 to 0.00. Four hallmarks are untouched, and a mean cannot say which four.

GERO
The biology

Twelve hallmarks, ten tissues, 15,705 genes with a measured aging direction, and 21 targets across two programmes.

QUBIT
The representation

A state vector, not a number. Twenty-two components held in superposition, read by quantum-derived operators, and measured at the end rather than averaged at the start.

GEROQUBIT
Both at once

A candidate that helps liver and does nothing for brain stays legible as exactly that, all the way through the search.

The structure runs classically on ordinary CPUs. What is quantum here is the formalism the operators are derived from, not the hardware, and we say so on every surface.

Measured · QRC, one of the nine

A quantum operator that reads chemistry before it tries it.

QRC is the template-applicability operator in the QX layer, and the only one of the nine we let steer the engine. It projects a fragment onto the reactions its functional groups can support, so the search stops spending itself on chemistry that was never going to work. Below is why that matters, measured on our own corpus.

050%100%27 templates never fired77 REACTION TEMPLATES, RANKED BY MEASURED SUCCESS RATE127,946 recorded decode attempts · 29,047 products · 22.7% overall

Head of the curve: mitsunobu 98% · amide coupling 66% · buchwald–hartwig 63%. Tail: templates that have never yet produced a product on our chemistry, which is what the reordering below is for.

What QRC does

It reads the functional groups on the fragment in hand and puts the reactions most likely to work on it first. Nothing is skipped and nothing new becomes reachable. Only the order changes, so a workable route turns up earlier and fewer attempts land in that flat tail.

16.1% of all reaction attempts used to go to templates that never once produced a product.

How well it predicts

Hand-written rules0.637
Per-reaction averages0.875
QRC0.946

Predicting whether a reaction will fire, on 9,173 reactant pairs never seen in training. Scale starts at 0.50, which is chance.

+22.9%
more candidates for the same search effort
95% CI [+16.2%, +33.1%]

24 paired runs, 8 targets, 3 seeds. 24 won, 0 lost.

A search-efficiency result. It says nothing about whether a molecule works in an animal.

How a candidate is built

Assembled from real chemistry, so the route exists before the molecule does.

The engine evolves recipes, not pictures of molecules. Nine quantum operators read the aging state you chose and steer that search; the chemistry underneath is ordinary named reactions with the reagents each bond needs.

01 · You choose
TargetmTORC1
TissueWhole-body
Hallmark12 scored
02 · Engine assembles
Scaffold
+
Block
Amide coupling
One of 73 named reactions. Each with the reagents that bond needs
03 · You get a route
12
Fragments
real, catalogued
Reactions
named + verified
Candidate
makeable
2 steps✓ every candidate carries a route

We build from the reaction, so the route is not a prediction. That is a real property of the output and not a differentiator — reaction-space generation with routes is publicly available, so it separates us from structure-first tools and from nobody else.

Published · benchmarked · disclosed

The proof, in three numbers.

Backed by a methods preprint, including the number where we fail.

Binder recovery
0.941
ROC-AUC

Recovers known binders from decoys on a scaffold-disjoint held-out split.

On novel chemotypes (chance 0.50)0.681 · disclosed
Sample efficiency
0.938
PMO AUC-top10 · QED

Level with the strongest unconstrained generators (0.94–0.95), and ours stay makeable.

vs SOTA cluster0.94–0.95
Calibrated honesty
90%
Interval coverage

The lifespan prior is weak but real (ρ 0.158), always shipped with its interval.

775 measured-lifespan compounds42 NIA-ITP
See all the benchmarks →PMO · recovery · MOSES · out-of-domain, reported in full, reproducible.
Pre-registered validation

We write the test down before we run it.

Tested twice against compounds already known to fail 19 in mice, then 133 in worms. Neither test showed we can rank them. Published in full, both times.

  1. Protocol fixed
  2. Analysis run
  3. Result published
0.645
95% CI [0.447, 0.826]
ROC-AUC · n = 35
Inconclusive
Measured against the field

Nine quantum-derived operators, rebuilt for chemistry.

Quantum methods are written for qubits, not for molecules. Every operator we use had to be rebuilt before it worked on a search over real synthesis routes. We do not publish the rebuilds. We publish what they score.

Standard oracle benchmark (PMO, QED)0.938
Known-binder recovery, scaffold-disjoint0.941 AUC
Candidates delivered with a working synthesis route100%

The first bar is the industry’s own generation benchmark, and we reach it while every molecule is constrained to a real route, a restriction the methods in that range do not carry. The third is not a score but a property: a route exists by construction, so it is 100%, and a structure-first generator cannot report it at all.

1.44e–15

We audit our own quantum claims

A similarity model at the centre of our stack turned out to be reproducible by classical means to within 1.44e–15. We found it, published it, and withdrew the claim rather than sell it.

+0.0112

What survived that audit

One operator captures structure that no conventional formulation can express. It moves a ranking metric by +0.0112, 95% interval [+0.0083, +0.0145], excluding zero. Small, real, and the interval travels with it.

12 / 12

Genotoxicity screened at design time

ICH M7 alerts are run over the route reagents, not only the final molecule, catching a DNA-reactive handle that is consumed during synthesis and invisible to product-only screening. 12/12 known positives flagged, 6/6 benign clean.

Quantum-inspired, classically computed. Benchmark range from Gao et al., NeurIPS 2022. Where a result is a tie or a loss we publish it as one. The full ledger, including what did not work, is on the benchmark page.

How a run works

Three steps, and you watch each one.

01

Pick a target

Eleven, ranked by human genetic evidence.

02

Set the scope

A single organ, or pan-aging across all ten.

03

Get candidates

A route and a stated confidence on each.

Who it's for

Built for teams hunting real longevity chemistry.

Longevity biotechs

Novel, makeable scaffolds for hit-to-lead, before you commit synthesis budget.

Academic geroscience labs

Tool compounds to probe a target you have already validated biologically.

Medicinal chemistry teams

A ranked shortlist with routes, ADMET and an explicit confidence on each.

Every candidate is a longevity hypothesis
Target × tissue × hallmark

We fix the biology first, not the molecule.

12
hallmarks of aging
Scored on every candidate
  • Genomic instability
  • Telomere attrition
  • Epigenetic alterations
  • Loss of proteostasis
  • Disabled autophagy
  • Deregulated nutrient sensing
  • Mitochondrial dysfunction
  • Cellular senescence
  • Stem-cell exhaustion
  • Altered intercellular comm.
  • Chronic inflammation
  • Dysbiosis
14
tissues
Pick one, or run whole-body
  • CNS / brain
  • Heart
  • Liver
  • Skeletal muscle
  • Immune (PBMC)
  • Adipose
  • Vascular
  • Hypothalamus
  • Gut
  • Thymus
  • Kidney
  • Lung
  • Skin
  • Pancreas

Hallmarks follow López-Otín et al., Cell 2023. The readout is rule-based and reported as a prior, not a measurement.

A worked case

A target the clinic already knows, in a disease nobody pointed it at.

CXCR2 is the receptor that reinforces cellular senescence. Four antagonists have been through human trials against it, all for lung disease or cancer, all dosed through the whole body. None was taken into osteoarthritis, where the drug can be put into the joint instead. We designed three candidates for that gap.

GQ-SM-06
Smallest of the three, cleanest predicted profile.
CXCR2 · 274 Da
N-arylation → amide coupling
GQ-SM-07
Sulfone-substituted, the most elaborated of the set.
CXCR2 · 404 Da
N-arylation → amide coupling
GQ-SM-08
Lowest predicted cardiac and mutagenicity flags.
CXCR2 · 272 Da
N-arylation → amide coupling

Nothing here has been tested. One of the three can already be ordered from a catalogue, which puts a real experiment three weeks away. Full study, scores and routes on the methods page.

What makes it different

Built for aging biology, not retrofitted from oncology.

The whole stack is built around one idea: hold the aging state whole. Quantum formalism is what makes that practical, because a state vector is the natural object for something with twenty-two components that interact.

Built for aging

Twelve hallmarks and ten tissues, carried as one state. Not an oncology retrofit.

Honest uncertainty

Efficacy calls ship a calibrated interval, and we publish the tests we fail.

Minutes, not months

A design run finishes in minutes. The comparison is a med-chem cycle, not a page load.

Two programmes

11 lifespan targets, 10 rejuvenation targets, 14 tissues, nine quantum operators.

Programme one · lifespan · eleven targets, genetics-ranked

The pathways longevity actually runs on.

Design against one mechanism, or switch to whole-body mode and optimise across all fourteen tissues at once.Each tag is the direction that program is designed toward, its reference compounds all act that way. Confirming it takes an assay.

CD38Inhibitor

NAD⁺ → SIRT1 / PGC-1╬▒ metabolic axis

SIRT6Activator

Chromatin guardian; NF-╬║B co-repression

NAMPTNAD⁺ salvage

Rate-limiting enzyme of the NAD⁺ salvage pathway

Bcl-2PPI inhibitor

Senolytic clearance of senescent cells

mTORC1Inhibitor

Autophagy gate; canonical longevity axis

AMPKActivator

Cellular energy sensor; ADaM-site activation

NF-╬║B / IKK╬▓Inhibitor

Suppress the SASP / inflammaging

p21 (CDKN1A)Senostatic

Silence the senescence secretory phenotype

MEK1Inhibitor

RAS–MEK–ERK inflammaging axis

NRF2Activator

Master antioxidant-response regulator

IGF1RInhibitor

IGF-1 / insulin (daf-2) longevity axis

Your target not here?Tell us the mechanism. We scope new targets per program.Talk to us →
Programme two · cellular rejuvenation · ten targets

A different question: can it push cells younger?

Lifespan and rejuvenation are not the same programme, and we measured it rather than assumed it. Every one of our eleven lifespan targets is a known longevity gene; almost none carries direct aging-perturbation data. The targets that do carry it are epigenetic enzymes, and none of them was on the lifespan list.This programme is research. It designs candidates today; nothing in it has been synthesised or assayed.

METTL3m6A writer

Epitranscriptomic ageing; highest-scoring admitted target

KAT7H3K14ac writer

The Zhang Sci Transl Med 2021 rejuvenation target

EP300 / p300Histone acetyltransferase

Drives the senescent state (Sen, Mol Cell 2019)

MAPK14 / p38╬▒Stress kinase

Master regulator of the SASP (Freund, EMBO J 2011)

CXCR2Chemokine receptor

The SASP receptor (Acosta, Cell 2008)

EGLN1 / PHD2Prolyl hydroxylase

Raises HIF-1╬▒; hypoxic adaptation

SIRT3Mitochondrial deacetylase

Mitochondrial protein acetylation with age

IDH1Metabolic enzyme

2-HG and the metabolic arm of ageing

ALKBH1DNA/RNA demethylase

Nucleic-acid methylation marks

SOD1Superoxide dismutase

Redox balance

We design against targets a screen has almost nothing to start from.

A library screen needs known ligands to begin from. SOD1 has 18 measured binders in ChEMBL and ALKBH1 has 45, enough to recognise the chemistry, nowhere near enough to screen against. Building a molecule needs no starting library, which is exactly why these stay on our list rather than falling off it.

Honesty-first
We don't ship a score we can't measure. Where the signal is strong, we say so. Where it's weak, we show the interval, and let your chemists decide.

A field burned by overpromising. We measure, or we don't ship.

Built by
Dinesh K BSc Chemistry·H. Swetha MSc Chemistry
Numbers current as of
Who we are →

Pick a programme. We'll bring the chemistry.

Eleven aging programmes, fourteen tissues, and a synthesis route on every candidate.