Interviews • Sponsored by SYDRA
Target-Agnostic Longevity Drug Discovery: How SYDRA Uses AI and Whole-Animal Testing to Hunt for Geroprotectors
Aging is the biggest single risk factor for most chronic diseases, yet drug development still goes after them one at a time, and regulators do not treat aging itself as an indication. Animal studies with rapamycin, acarbose and, most recently, semaglutide show that lifespan can be extended with drugs, even when treatment starts late in life. But finding such molecules systematically, and then working out what they actually do, is still slow and expensive.
Zurich-based SYDRA AG, founded in 2024 by biogerontologist Alexander Dakhovnik and machine-learning researcher Pavlo Mozharovskyi, is one of the companies working on that problem. Instead of starting from a target, SYDRA uses its own AI models to select and generate molecules predicted to act as geroprotectors. It tests them for lifespan extension in C. elegans and in naturally aged mice, and only then works backwards to the targets and mechanisms, an approach it calls "hit-to-target" discovery. Its two lead compounds, SY23F and SY23H, have shown encouraging gains in remaining lifespan in mice that began treatment at 22 months.
We spoke with Alexander Dakhovnik, SYDRA’s co-founder, CEO and CSO, about how the company started, what is actually new about its platform, what the mouse data do and do not show, and where pharma and regulators stand on aging as a drug target.
Andrii: How did SYDRA start? What were you working on when it became clear this needed to be a company and not just a research paper?
Alexander: SYDRA began during my postdoc at ETH Zurich, in Collin Y. Ewald’s laboratory, where I worked on the molecular mechanisms of aging. Two projects helped shape the idea. One used transcriptomic signatures of extracellular matrix remodeling to identify potential geroprotectors. The other explored an AI-predicted mTOR inhibitor that extended lifespan in C. elegans.
While running lifespan assays, I started asking myself whether we could train a neural network to recognize potential geroprotectors simply from their molecular structures. Could we give the initial screening of millions, eventually billions, of compounds to a computational system and concentrate our experiments on its best predictions? An experienced medicinal chemist learns to recognize promising structures. I wanted to explore how much of that useful information a model could learn.
To turn that idea into a working AI, I joined forces with my co-founder, Pavlo Mozharovskyi, a professor of machine learning at Télécom Paris. We combined my knowledge of aging biology with his expertise in machine learning to develop the system together.
Honestly, the first ambition was a paper in a leading journal. I did not begin with much confidence that the idea would work. The biological dataset was tiny compared with what we would normally want for training a large neural network. We also had to respect differences between organisms: a molecule that works in mice may not work in flies or worms. And even a promising prediction would leave us with the practical question of which concentration to test.
Transfer learning gave us a way forward. We could adapt broader chemical knowledge to our specific biological problem instead of expecting a small geroprotector dataset to teach the model everything from scratch. We chose C. elegans as our starting point because it offered useful experimental data, a short lifespan and a practical way to test predictions in a whole organism.
When the first computational tests began to look promising, the ambition changed. We started to see a system that could produce successive generations of candidates. That was the point at which building a company made sense.
Andrii: Explain the approach the way you would explain it to a skeptical drug hunter. You look for molecules that make animals live longer and only afterwards go hunting for what they actually do in terms of mechanism, am I getting the idea right? Why is that a better bet than starting from a target?
Alexander: Yes, you have the idea right. We start by looking for compounds that produce a useful phenotype in a living organism. For us, the initial phenotype is lifespan extension. We then work backwards to identify the targets and mechanisms responsible and determine which disease applications deserve development.
Part of the inspiration came from drugs such as rapamycin and acarbose. They were developed for particular medical uses and subsequently showed lifespan effects in animal studies. More recently, a study starting semaglutide in 20-month-old female mice reported approximately 12% longer median total lifespan, alongside improvements in several functional measures. These observations make the reverse question interesting: can we start with the longevity phenotype and use it to discover useful therapeutic biology?
If a compound reproducibly changes survival in a whole organism, there is a biological effect worth investigating. It may involve a particular target, several targets or a broader cellular response. The task is to establish what causes the effect and whether that biology can be used against an age-related disease. Lifespan activity gives us a reason to pursue the compound; the disease application still has to be demonstrated.
That is what we mean by hit-to-target discovery. The hit is the molecule. Target deconvolution and mechanism studies help us understand what it does and where it might become a medicine. Ultimately, we need a defined disease indication and clinically meaningful endpoints.
I would not claim that this is universally better than starting from a target. It changes the order in which we tackle uncertainty. We obtain evidence of activity in a whole organism early, while accepting the substantial challenge of finding the mechanism afterwards. It does not remove the later risks of translation, safety or development.
We also have to be realistic about the model. Worms are useful for lifespan screening, but they cannot represent every human disease. Target-based discovery, cell models, organoids and mammalian studies all have important roles. Our approach adds another route into that process.
Andrii: What was the hardest part of the whole setup to get working? I ask because a lot of what gets called a "discovery platform" is really just a workflow, and so I am curious what is the innovation behind your platform? Also, two mouse lifespan studies is a lot for a relatively small team, how did you manage to arrange that?
Alexander: There was no single hardest part. It changed with the stage. Initially, the main challenge was creating an AI model tailored to our biological question. Assembling the dataset took longer and involved plenty of tedious work, but developing the architecture was the harder conceptual task. Training then brought its own problems: access to suitable GPUs, time, electricity costs and hardware reliability.
The innovation starts with the question we built the AI to answer: can we recognize a potential geroprotector from its molecular structure without choosing a target beforehand? Pavlo and I developed our own neural-network architecture and used transfer learning to make that possible with limited biological data.
Together, we then developed and trained our own chemical foundation model to generate new molecules. We keep generation and prediction of geroprotective activity as separate tasks. A chemically plausible structure does not automatically make a useful geroprotector. Generated molecules go through chemical filtering, biological prioritization and medicinal chemistry refinement before synthesis and experimental testing.
That gives us two complementary routes to candidates: searching existing chemical space and creating new structures to investigate. We have already taken AI-selected compounds into nematode and aged-mouse experiments. Whatever a model predicts, the molecule eventually has to do something useful in a living organism.
When I call SYDRA a platform, I mean a system built to produce successive generations of candidates. Our proprietary AI determines what chemistry we explore, whole-animal experiments determine what deserves further investment, and the observed biology guides the search for targets and indications.
The experimental infrastructure was another substantial challenge. Setting up the Worm Lifespan Machine and building our own CNC-style lifespan robot took real engineering work. But obtaining naturally aged mice was particularly difficult, both financially and in terms of time.
We were fortunate to meet the team at Bienta, part of Enamine. They had mice that had aged naturally while waiting for another project that did not start, and they agreed to test some of our molecules. That opportunity made the mouse work possible.
Dose selection was difficult. In worms, we could explore several concentrations relatively quickly. With these mice, we had a much more limited opportunity. We used molecular comparisons and our experimental experience to select doses. Of three compounds tested, two produced encouraging trends in remaining lifespan. The groups were approximately 25 to 27 animals, and the results remain exploratory.
At that stage, it was just Pavlo and me, bootstrapping and trying to secure external funding. Today, the hardest task is raising the resources for target deconvolution and the next experimental stage. Access to appropriately aged animals remains a challenge.
Andrii: Your two lead compounds, SY23F and SY23H, are the ones that came through the mouse work. You started dosing at 22 months, which is already old for a mouse, and on your published numbers the treated animals had roughly 40% and 30% more life left than the controls. But you still don’t know what either molecule is actually hitting. Where does that stand now: what do you know, and what is still open?
Alexander: Yes, and the distinction matters: these were differences in remaining lifespan after treatment began at 22 months. The same gain expressed relative to the animals’ entire lives would be smaller. We should be precise about that.
Our public figures are 41.7% and 29.2% greater remaining median lifespan for SY23F and SY23H, respectively. The reported p values are 0.2945 and 0.0733. Neither reaches the conventional 0.05 threshold, so these are exploratory observations that require confirmation.
What interests me is that we started when substantial aging had already occurred. From a translational perspective, that is a relevant question: can we still make a useful difference when treatment begins in middle or later life? We cannot build a practical development strategy around the assumption that people will start taking a medicine as teenagers and continue for decades.
There is also an interesting connection to Peter Fedichev and colleagues’ work on the loss of physiological resilience with age. Think of a ball sitting in a valley: after a disturbance, it rolls back toward a stable position. In their framework, that restoring capacity weakens with age, making recovery slower and the organism more vulnerable to losing stability. Helping preserve that stability is an appealing therapeutic idea. Whether our compounds actually do that remains to be established.
Rapamycin provides a concrete precedent for late intervention. The landmark mouse study began treatment at around 20 months and still extended lifespan. A subsequent comparison with treatment starting at nine months did not detect a significant survival advantage from the earlier start at the dose tested. Those were separate cohorts, so I would not turn that into a universal rule about the best age to intervene. But it shows why a treatment started late in life deserves serious attention.
As for SY23F and SY23H, we do not yet know their molecular targets or mechanisms of action. We observed modest positive trends in grip strength and rotarod performance, but those functional findings also need confirmation. I would not claim that we are developing a treatment for sarcopenia or metabolic syndrome on that basis. At this stage, that would be speculation.
Our next milestone is to connect those observations to mechanism. We plan to use omics to identify affected pathways, alongside experimental target-fishing assays and computational work to prioritize possible interactions. Candidate targets will then need experimental confirmation. That should help us identify which disease applications are worth pursuing and design the next, more focused studies.
Andrii: Say everything works. Do you hand the molecule to a pharma company, or do you want to take it into the clinic yourselves?
Alexander: Honestly, yes, we would like to take our molecules into Phase I and II ourselves. Whether we can do that will depend heavily on access to capital. If we have the resources to fund and manage those trials properly, that is the route we want to take.
If we don’t, we could out-license a candidate with its preclinical package. And the partner does not have to be Big Pharma. A midsize biotech with sufficient funding and the right development capabilities could be a good fit.
But keep in mind that SYDRA is a platform. We have other potential candidates coming through the pipeline, so out-licensing our first couple of molecules would not define our business model forever. We could partner on an early programme and take a later candidate further ourselves as our resources and capabilities grow.
We are willing to take on the risks of clinical development. Whether we do that with our first candidates or with later ones will depend on the data, the funding and what we can realistically execute. The ambition is there; we need to build the means to support it.
Andrii: On a general industry note, "mainstream" pharma has stayed away from aging research/longevity for a long time. Is that changing? Do you see any interest from big players in age-related projects, and in your platform in particular?
Alexander: You touched a sensitive topic. As a biogerontologist, I find the pace frustrating. Yes, there is movement, but I do not yet see the level of commitment to targeting aging itself that I would like.
Part of the difficulty is that humans live "too long". That is a wonderful fact, but it makes longevity research exceptionally challenging. Measuring whether an intervention improves human survival takes substantial time and resources. Naturally, everyone wants a faster readout.
Dietary interventions can lower some circulating AGEs and alter inflammatory markers associated with cellular senescence. These measurements can respond over relatively short periods. But we still need to establish whether those changes reliably predict a treatment’s effects on disease, function or survival. An improved laboratory readout does not automatically tell us how many additional healthy years a person will gain.
Drug interventions provide another example: older animal studies reported reduced neuronal lipofuscin after meclofenoxate treatment. Those findings remain preclinical, and later cell-culture work questioned whether the drug clears existing pigment or mainly slows further accumulation. They do not establish the same effect in humans or a human longevity benefit.
Still, we have practical ways forward. The proposed TAME trial is an important example: its design focuses on whether metformin delays major age-related diseases or death in older adults, with additional measures of function. Its proposed clinical outcomes were developed with FDA input. That is a constructive precedent for studying several consequences of aging together.
We do not need to wait until every participant dies. When the treatment and control Kaplan-Meier survival curves clearly separate, and the prespecified analysis confirms a meaningful survival benefit, the message is clear: the treatment improves survival. These trials take years, yes. But so does the standard drug development process, doesn’t it?
There are examples of substantial pharmaceutical investment in exploring broader applications of existing drugs. Novo Nordisk tested semaglutide in two large Phase III Alzheimer’s trials. Unfortunately, they did not demonstrate slower clinical progression, despite improvements in some disease-related biomarkers. That is both evidence of willingness to investigate an ambitious hypothesis and a concrete illustration of why biomarker improvements need clinical confirmation.
I also welcome efforts to rethink access, such as Montana’s changes to its Right to Try framework. But access to an experimental treatment and evidence that it works are separate achievements. Those changes do not establish an FDA-approved route for treating aging.
What I would like to see is more coordinated ambition from pharma and regulators around preventing several age-related diseases through their shared biology. Medicine already practises prevention; geroscience asks how much further we can take it. With a car, we understand the value of changing the oil, tyres and battery before a breakdown. I want us to investigate that preventive logic more seriously in aging.
For SYDRA, we have to do our homework within the development pathways available today. Yes, we see interest and curiosity from pharmaceutical companies. Our next task is to connect the observed biology to targets, mechanisms of action and a credible disease application, supported by reproducible evidence. That is the package we need to turn those conversations into serious development opportunities.
Andrii: What’s the next important result or milestone you are waiting on, and what will it change for SYDRA?
Alexander: The next major scientific milestone is to understand the targets and mechanisms behind our lead compounds and use that knowledge to nominate a disease indication.
We plan to run omics studies to map the biological response, establish DARTS followed by mass spectrometry to identify candidate targets, and use in silico work to investigate and prioritize possible interactions. We would then use SPR to test direct binding independently. Together, these results should help us understand what drives the activity we have observed and which disease applications deserve focused testing.
From there, we want to advance into disease-specific validation, PK/PD and ADMET studies, including appropriate animal experiments, and assemble the evidence needed for a solid preclinical package. For me, the important change is that we would have a defined therapeutic programme: a candidate, a proposed mechanism, a disease indication and the data supporting its further development.
Of course, carrying out that programme requires capital. Our immediate business milestone is to secure our pre-seed round with a biotech VC or family office that understands the development path. That investment would allow us to execute the next experimental stage and build a much stronger basis for development and partnering discussions. If you are interested in exploring it, reach out!
Topic: AI in Bio