The questions we get asked,
including the awkward ones.

The unusual part is the order of operations. Most systems draw a molecule first and work out afterwards whether anyone can build it. Ours searches over recipes instead, so a molecule only exists in the output if a route to it already exists.

A design is specified on three axes rather than one: the target, the tissue you care about, and which kind of age-related damage you are trying to address. That combination is what makes it a geroscience tool rather than a general-purpose molecule generator.

It is deliberately front-of-funnel. The job it does is turning "we believe this target matters in aging" into "here are candidate structures, each with a route, an ADMET profile and an explicit statement of how much to trust the numbers".

It is not a replacement for a medicinal chemistry team, a CRO, or an assay. It is the step before all three.

So the useful question is not just "does this hit the target" but "does it hit the target, in the tissue that matters, for the damage type we care about". That is why every design here is aimed at a target, a tissue and a hallmark together.

The second difference is dosing. A geroprotector is a chronic, largely preventative medicine given to people who are not yet ill, so the safety bar is closer to a statin than to an oncology drug. That changes which trade-offs are acceptable — potency cannot buy back a liver signal — and it is why our scores multiply rather than average.

Structure prediction has been genuinely transformative for knowing the shape of a target. It does not, on its own, tell you which molecule to make, whether that molecule can be made, or whether hitting the target helps.

We do not currently consume predicted structures at all, because our scoring is 2D and ligand-based. That is a deliberate limitation with a measured reason behind it — see the docking question below.

The reason to accept that trade-off is coverage. Purchasable libraries are enormous but heavily biased toward chemistry that has already been made, which for most aging targets means the chemistry of a different, better-funded therapeutic area.

The reason to be sceptical of it is that generated molecules carry more uncertainty, and we quantify exactly how much: our own ability to tell a real binder from a decoy degrades sharply the further a molecule sits from known chemistry, and we publish that curve rather than hiding it.

Designing molecules as synthesis routes is not unique to us — it is an active research area with several published methods, and we would be wrong to claim we invented it.

What we have not found anyone else doing is the combination: designing on a target-tissue-hallmark basis for aging specifically, screening for genotoxic risk across the reagents in the route rather than only the final compound, and publishing the regime where our own predictions stop working.

A reactive fragment on a starting material is a real handling and impurity risk even when it never appears in the product. Only a system that knows the route can see it.

We built a 3D shape-and-electrostatics comparison — including desolvation — and put it up against plain 2D structural similarity on exactly the kind of novel molecules we produce. The 2D method scored higher.

We are not claiming docking is useless. It clearly earns its place in many programmes, and running one docking pass over our final selections is still on our roadmap. We are reporting what happened in our own hands, on our own output, and being guided by it. Our honest expectation is that it fails again, and we intend to publish that too.

When we compare our output to the compounds already known to bind a given target, our molecules sit about three times further away than those known compounds sit from each other. In chemistry terms they are not close analogues of anything in the reference set.

There is a real cost to that distance and we publish it: the further a molecule sits from known chemistry, the worse our ability to predict whether it will work. That trade-off is the honest centre of this whole approach, and it is why the platform reports a "frontier distance" next to every candidate.

One thing this rules out, incidentally, is the most common failure of aging-focused AI tools — quietly rediscovering rapamycin, metformin and quercetin and presenting them as discoveries.

Structural novelty is one minus the highest similarity to any compound in our reference sets of known actives and known lifespan-extending compounds. It is the number that matters for "is this a new chemotype".

Route novelty matters for a different reason: two molecules can look different and still be made the same way, which means they will fail the same way at the bench and are weak as a portfolio. Route novelty is the signal that catches that, and it maps directly onto freedom-to-operate thinking.

"By construction" is a strong claim about the route existing and a weak claim about it working first time. Yield, selectivity and purification are not things our system predicts, and any real synthesis may need a chemist to change the conditions.

We have also found and fixed the failure mode where a template produces chemistry that cannot run: a reaction named for one bond that actually forms another, an aromatic substitution proposed on a ring that would not undergo it. Those are now gated and asserted rather than assumed, because they survived being read and only failed when run.

We are actively asking contract research organisations to judge sample routes as though they were incoming orders. A negative answer there would be decisive, and cheap.

The two practical points we can speak to are factual rather than legal. First, novelty: our output sits far from known chemistry by measurement, and we give you that number rather than an assurance. Second, prior art: we check each candidate against PubChem, and we report when a structure appears to be already known.

A caution on that second check, from our own history. It once reported "known compound (CID 0)" for genuinely novel molecules because of a truthiness bug in how the identifier was tested, which made novel matter look unoriginal. It is fixed, and our published credibility figure was re-baselined downward as a result rather than quietly left high.

The route-novelty signal is the one most directly relevant to freedom-to-operate conversations, because it describes how a compound is made rather than only what it is.

Nothing proprietary and nothing scraped. That is partly principle and partly practicality: a claim you cannot point someone at is not a claim they can check.

The building-block and scaffold pools are assembled from approved-drug chemistry and fragmented public compounds, with structural-alert filtering applied before assembly rather than after.

To be clear about the framing: this is not a criticism of DrugAge, which does exactly what it says and is a genuine service to the field. It is a statement about what happens when we use it as a training label for a mammalian question.

Our own analysis found that the maximum-lifespan-extension label we train on is roughly 90% invertebrate. Invertebrate effect sizes run about three times larger, so when a compound has been tested in several species the maximum is almost always the worm or fly number and the mouse data is discarded by the aggregation.

We measured whether the invertebrate signal transfers. Across compounds tested in both, the rank correlation between worm and mouse effects was about zero, with an interval wide enough to include a strong effect in either direction. So we cannot show that it transfers, and we cannot show that it does not.

The honest summary is that the fitted noise is roughly twice the signal. We say so in the interface rather than only in the methods, because a number with that much noise around it is dangerous if it is presented cleanly.

The interval is not a guess. It is calibrated so that the true value lands inside it about nine times out of ten, and we measured that coverage on held-out compounds rather than assuming it.

This prediction is deliberately kept out of the design loop. Letting it steer generation would pull every run toward things that already look like known geroprotectors, which is exactly the novelty trap the platform exists to avoid. It is computed afterwards, for display and for final ranking only.

No molecule we have designed has ever been made or tested, so no score on this platform has ever been checked against reality. That sentence is the most important one on this page.

The axes are combined by multiplying rather than averaging, on purpose. Averaging lets a molecule be excellent on one axis and poor on another and still look good. Multiplying means one weak axis drags the whole score down, so potency cannot buy back bad safety.

Because that multiplication includes three "does it look like a known binder" axes, a genuinely novel molecule scores low almost by definition. That is why the interface also shows a developability grade computed over only the similarity-independent axes — it answers "is this makeable, drug-like and safe?" separately from "does it resemble something already known?". A candidate can honestly be poor on the first number and strong on the second, and we show both rather than picking the flattering one.

We put that on the homepage rather than in the footnotes, because it is the first thing a scientist should know and the last thing anyone should have to discover for themselves.

What we do have is a synthesis route for every candidate, which means the step from a design to a real compound is a purchase order rather than a research project.

Closing this gap is the roadmap, and the cost driver is synthesis rather than assay. Worm healthspan screening runs a few thousand dollars per compound; first-time synthesis of a novel molecule is the expensive part, and notably the price is close to flat between ten milligrams and five hundred, because labour dominates.

The US National Institute on Aging runs a programme that tests compounds for lifespan extension in mice at three independent sites and — unusually — publishes its failures as well as its successes. That gives a set of real, replicated negatives, which is the part most predictors never face.

We asked whether our model ranks the compounds that worked above the compounds that did not, scoring each compound with its own record and its close analogues removed. It scored above chance on the central estimate, but thirty-five compounds is a small set and the uncertainty around that estimate comfortably includes chance. So the test settles nothing in either direction: we cannot show the model has this ability, and neither did we show it lacks it. Being underpowered is not the same as failing, and we are careful not to describe it as either.

Two details matter more than the headline. Removing close analogues made the score go up rather than down, so the result is not propped up by the model recognising compounds it had already seen. And the ranking is unconvincing in its specifics: the top-ranked compound of all thirty-five is one that failed in mice, while rapamycin — the most firmly established geroprotector there is — lands seventh.

The likeliest explanation is our own label rather than our method, for the reason given in the DrugAge question above.

That last one deserves unpacking, because it is the least obvious. When a programme is labelled as aiming to activate or inhibit a target, that describes the design intent: the reference compounds behind it all act in that direction and they drive the scoring. But the activity score itself measures resemblance to known binders without distinguishing activators from inhibitors. Nothing in the pipeline predicts functional direction, and only an assay can confirm it.

We take that seriously enough to have renamed a programme over it. One target was labelled as an activator while its reference set was led by two well-known inhibitors, which pulled candidates toward the opposite chemotype. Rather than relabel and move on, we found that the validated activator class for that target is essentially a single compound, so there is no honest five-compound activator set to build — and the programme is now named for the pathway instead of claiming a direction its references cannot support.

On chemistry outside geroscience, a specialised library scores low by design. On one standard benchmark task involving an unrelated receptor we score zero, and we publish that rather than omitting the row.

Everything this system produces is a hypothesis. The scores are priors computed from public data, not measurements, and the failure mode of this whole category of tool is that a clean-looking number gets treated as evidence.

What it genuinely saves is the expensive early loop: enumerating chemistry that turns out to be unmakeable, or unsafe in a way that only surfaces once someone tries to order the reagents. Catching a DNA-reactive starting material at design time is cheap; catching it in process chemistry is not.

The judgement calls — is this series worth pursuing, is this route sensible, is this target right — are chemistry and biology decisions, and we build the interface on the assumption that a person makes them.

The reason for that order is cost asymmetry. A chemist rejecting a route costs an afternoon. A binding assay costs less than a synthesis. A worm healthspan run costs a few thousand dollars. A mouse lifespan study costs years. Each step should kill as many candidates as it can before the next one.

The most informative early experiment is not the most flattering one. Target engagement on a small set — including the candidates our own scores rank poorly — tells you far more about whether the platform works than testing only the top-ranked molecule.

If you want a decisive negative cheaply: send three routes to a contract research organisation as if they were incoming orders and ask whether they would accept them.

Short of that: a chemist telling us the routes do not survive contact with a real bench. We are asking contract research organisations to judge sample routes as though they were incoming orders, precisely because a negative answer there would be decisive and cheap.

We have a track record of publishing the answers that did not go our way — a pre-registered mouse-data test that came back inconclusive, a physics-based scoring method that lost to a simpler one, and an accuracy figure we revised downward after finding a bug in our own data. That is the standard we intend to be held to.

The practical consequence is cost. There is no cluster and no training bill, so the money in this project goes to chemistry rather than to compute. A full design run finishes in seconds on a laptop-class processor.

The honest counterpoint is that this is a constraint as well as a feature: it rules out approaches that genuinely need scale. We think that trade is correct for this problem, given that more model capacity has repeatedly failed to fix the frontier problem that actually limits us.

We would rather talk to you about the target than sell you a login. If your target is not among the eleven, tell us the mechanism — new programmes are scoped individually and the target list is not a hard limit.