‘Show me the error bars before you sell me longevity’

Deerfield’s Jim Flynn on shrinking error bars, cutting through longevity hype, and why AI needs better loss functions.

The Deerfield Managing Partner calls longevity “a misnomer”, is building a not-for-profit longevity center in New York to cut through the noise, and thinks AI models that can’t run their own experiments won’t learn much. He spoke to Longevity.Technology on the sidelines of ARDD 2026.

Jim Flynn has spent 38 years in what he calls “the observation business”. As Managing Partner and CEO of Deerfield Management, the New York healthcare investor, he backs companies from university license to commercial launch, and has just watched one of them, Nuvalent, go to GSK for around $10.6 billion.

When the longevity sector asks whether big healthcare money is coming, Flynn is a useful person to ask. His answer is nuanced: he is skeptical of “longevity” as a frame, enthusiastic about healthspan, and building a not-for-profit center designed to tell people what actually works.

Shrinking the error bars

Deerfield’s edge, Flynn says, starts with a simple idea: every investment reduces to an analysis of probabilistic cash flows.

“No number that you collect, no survey that you do, no scientific information that you get is an exact number. It all has variance around it,” he says. “And if you have no certainty at all around what your return would be, your job is not to get better numbers, but to reduce the error bars. Our entire information apparatus is designed to reduce our uncertainty.”

Genetics is one concrete example. In 2015, GSK’s head of genetics, Matt Nelson, published work in Nature Genetics showing that genetic support for a drug target roughly doubles the probability of clinical success. Deerfield hired Nelson and built a genetics team around him to sharpen its read on whether a trial will work.

The rest of the Deerfield model covers the end market: how many people have the disease, how many are diagnosed, what payers will pay, how long and how costly recruitment will be. That analysis is the same whether a drug is in discovery or on the market, Flynn says. What changes is the work needed to push a company to its next value stage.

De-risk first, build companies later, embrace China

At the early end, Deerfield has tried to rework the classic biotech start-up. The usual route is to license a target, lease a lab and hire a team. “As soon as you do that, it’s really hard to fail for less than $50 or $100 million,” Flynn says, “yet these are all 95% likely to fail when you license them in.”

Deerfield’s answer is to own the labs and the discovery team, spread their cost across every program, and only form a company once a target is de-risked. That work sits in a New York building it calls Cure, which houses preclinical and clinical development teams, its own labs and around 30 university partnerships. Ideas come from those partners, from Deerfield itself and, increasingly, from China.

Flynn sees no reason to treat Chinese assets as a threat. Chinese groups have advanced molecules against a very broad range of targets, he says, but most cannot take them to the US themselves. “I see it as hugely synergistic. It doesn’t have to be a competitive US-versus-China thing. It’s drugs against targets that matter, and the job of the rest of us is to advance that and get it to people.”

Jim Flynn at ARDD 2026 (credit Michael DeStefano) Healthspan first, longevity later

Asked whether the longevity movement merits Deerfield’s time, Flynn is blunt. “I just see it as a misnomer.”

He does see investment opportunities in healthspan, but he has noticed a pattern in longevity biotech presentations: companies hunting for drugs that might extend healthspan, “but which they can get approved for something else first”.

“It’s an interesting admission,” he says. “They’re talking about something they can’t really measure or define well enough to have a clinical endpoint relevant for approval.” An investor, therefore, has to believe the acute indication is worth enough on its own, with any longer-term upside coming from later observation.

Longevity.Technology’s own modeling of the small- and mid-cap longevity biotech pipeline values it at $230 billion a year in risk-adjusted peak sales on approved indications, with a further $173 billion a year of upside once a gerotherapeutic regulatory pathway opens.

Should investors price that option in? Flynn wants something concrete. Multi-factorial biology means a “muddy drug”, the antithesis of a gene therapy that replaces one missing protein. Gene therapy is “kind of a straight line”, he says; this is not. “From an investment standpoint, what are my error bars? How do I assess that?”

On exits, Deerfield’s rule is equally clear. “If someone will pay above our probabilistic value, it’s gone. Private, public, it doesn’t matter. If it’s below, we’ll keep building it.” Nuvalent shows the logic: Deerfield founded it, owned 100% for five years, floated it at around $800 million and held it for four more before GSK’s multi-billion acquisition.

Flynn finds BioAge’s exercise-derived research, partnered with Novartis, “really interesting”, but adds: “It will probably take a really long time to get a drug that you know what it’s for. I just don’t know how I would practically invest in something like that.”

For Flynn, the most underrated part of healthspan is mobility. “When you look at exercise physiology and activities of daily living, your ability to actually engage and be mobile is super, super important. But where is that in a cellular clock sort of thing? It doesn’t really come up.” He reads mobility partly as “a code name” for very large markets such as osteoarthritis and cartilage repair.

A “true north” longevity center

Flynn is not content to observe. He is building a longevity center in New York, built around what is actually known.

“Almost everybody who’s in this space is selling something,” he says. “Should I take this peptide? Should I be micro-dosing GLP-1? Cold plunges, cryo, infrared – where do you go for a true north review of what has been shown biologically in humans, in terms of healthspan? What has been done to even show that this is safe?”

The center will start with a digital platform giving people neutral information. Users will be able to upload their own diagnostic data and work with an agentic system that tailors options to their profile and constraints – “I have four hours to exercise a week, I hate this, I love that” – with “no bias behind it”. The physical center will then let people access those interventions.

Deerfield is also building a compounding lab within Cure, both to guarantee what is delivered and to run trials on it. “We intend to actually test this stuff,” Flynn says. “None of this is for profit. This is all for true north information – and it’s mostly because I’m just really tired of all the misinformation.”

The digital layer should be low cost and widely accessible; physical services will be tiered and paid out of pocket. Full build-out is 18 months to two years away, with capacity aimed at thousands of people. In his view, clinical findings such as hypertension, an occluded artery or a tumor need a path into primary care, while the healthspan side is different but complementary. “Your bone density is not going to be improved by going to the physician. It will be improved if you jump a lot.”

Prevention isn’t new

The longevity sector often holds up GLP-1s as proof that preventive therapeutics can become blockbusters. Flynn isn’t persuaded the story is new.

Byetta, the GLP-1 drug Amylin brought to market with Lilly around 20 years ago, was approved for type 2 diabetes, needed twice-daily injections and delivered far less weight loss, he notes; today’s drugs refined an idea that had been around for well over a decade. Big pharma largely shunned the obesity space for years after fen phen, whose fenfluramine component was sold by American Home Products, was linked to heart valve damage and triggered multi-billion-dollar litigation. “Everyone ran away from the obesity space. GLP-1 is not a new idea at all. I’ve been in this observation business 38 years, and it doesn’t seem new to me.”

That doesn’t make him dismissive of prevention. “If someone comes up with something that prevents osteoarthritis, that is going to be a colossally huge drug.” But he argues it need not be a drug at all.

A foundation model of the joint

Deerfield has been working with the Hospital for Special Surgery (HSS) on foundation models of the joint. As with Parkinson’s disease, Flynn says, joints let you link imaging with clinical histories, biomarkers and genetics to build generative models that are highly predictive.

He offers a striking statistic: a 16-year-old girl is significantly more likely than a 16-year-old boy to tear her ACL in the same sport. The reason is anatomical variation, and how that variation combines with the way someone runs, cycles or moves through daily life. “That doesn’t matter whether you’re taking a drug or not. It’s actually the combination of your daily activities and how you do them.”

Link a foundation model to a wearable and it could deliver haptic feedback that corrects movement. “If you want to prevent osteoarthritis in ways that might require knee replacement, an ACL tear or hip replacement, it doesn’t have to be a drug. It could be a wearable.”

That is different from today’s consumer devices, he stresses. “The problem with Oura Ring and Whoop is they’re giving you information relative to averages. It has no idea how much sleep you need.”

Mobility suits this approach because much of it is physics. “You can create functional predictive models that course-correct where you end up, in a way that may be much more profound than a drug.”

AI needs a loss function

Flynn’s most pointed view is on AI drug discovery, including the wave of foundation-model companies setting up their own labs. His critique: they can’t create a useful loss function.

“What made ChatGPT work was that they used words rather than letters. When you look at chemistry, it’s very hard to use atoms for the same reason.” Mining historical data won’t get you far, he argues. A model has to make a prediction, have that molecule made and tested, learn whether it was right, and repeat. “No matter how they build their lab, it’s going to take them three-month cycles to make anything. They’re not going to learn anything for ages. So our approach is to focus on high throughput tokenized synthesis of molecules so that you can actually have a loss function that a foundational model can actually learn from.”

The bigger prize may be in the clinic. Flynn expects foundation models of disease to create a new generation of trial endpoints, accepted by the FDA, that compare patients against a neutral digital model. Measuring a qualitative sign such as tremor in Parkinson’s requires long trials with many patients. Asking instead whether a patient is ahead of or behind where the model says they should be could mean “smaller trials, faster trials, more accurate trials”.

For a sector still fighting to define its endpoints, that may be the most useful message Flynn has to offer.

Read more of our on-the-spot coverage of ARDD 2026 – CLICK HERE.