On stage in Kreuzberg, Paul Kohlhaas put a familiar number to the room: it costs about $2.5 billion to bring a drug to market. Then he reframed it. That money doesn’t buy discovery. It buys the bureaucracy built around it. He had the figures to back it: a 90 to 96% trial failure rate, ten to fifteen years from hypot...
On stage in Kreuzberg, Paul Kohlhaas put a familiar number to the room: it costs about $2.5 billion to bring a drug to market. Then he reframed it. That money doesn’t buy discovery. It buys the bureaucracy built around it. He had the figures to back it: a 90 to 96% trial failure rate, ten to fifteen years from hypothesis to patient, roughly $28 billion a year lost to research that can’t be reproduced, and about half of all trials never published. (watch)
The fifth edition of DeSci Berlin wasn’t really about any single one of those numbers. Across two days and twenty-odd talks, one shift kept surfacing: science is starting to behave like software. Fundable, ownable, composable and now agent-operable. The conference was where you could watch that happen in real time, demos and caveats included.
Here is what we saw.
The loop is starting to close
For most of biology’s history, the path from idea to experiment ran through a human, slowly. The clearest theme in Berlin was the attempt to make that loop run on its own design as a molecule, test it, learn from the result, and feed that learning back into the next design without waiting on anyone.
The most concrete version came from BIOS, Bio’s AI scientist. Geeve George showed binders generated on a GPU in real time and validated in Robiosis, an “autonomous lab in a box” roughly a thousand candidates in a couple of hours, about $10 for a verified design, on robots built for $10 to 15k against an incumbent’s $120k. (watch) Johannes Weniger ran the pipeline live pick a target, predict the binding sites, generate and score candidates for about $35 of compute, hand the best to the wet lab. ()
Newsletter
Get the latest updates from the Bio ecosystem
New launches, research progress, and ways to get involved across the Biosphere.
On day two, Rafael Rolli pointed the same idea at peptides. PeptAI, he said, is “an agentic fleet of swarms that designs, validates, pays for and learns from peptide experiments”, running against real targets like GLP1R and the orexin receptor, minting each candidate on-chain so its provenance is owned from the first step. And inside Bio’s own projects, Elliott Brunet introduced Gaia, an agent that holds a live view of the whole ecosystem, tracks each project against its milestones, and acts on what it finds. (watch)
What made it credible was that the people building this were the first to say where it stops. Julian Englert of Adaptyv Bio the cloud lab much of this work is sent to was blunt about the limits.
Biology is hard. Wet labs are hard.
In about 95% of cases, he noted, a human still decides which candidates actually get made researchers want “an emotional bond” with the proteins they test. The loop is closing. It is not closed.
Funding is catching up to science
If science is becoming programmable, the money around it has been slow to follow. Several talks were really about fixing that. Kevin Noessler of Molecule made the case for treating scientific IP as a liquid asset class rather than something trapped in tech-transfer offices and was honest that the first attempts overshot: “crypto built beautiful tech castles in the sky.”
Benjamin Snipes laid out a compliant path for token holders to become equity holders, while being candid that it “doesn’t solve how token holders get value” on its own. And the news beat of the weekend was OpenLabs: a funnel that turns a raw hypothesis into a funded project, with quality gates to filter the noise and vaults that route staking yield to the science. Ryan Noble framed it conservatively “no DeFi-flavored gambling yet.”
The proof that this isn’t theoretical came from Michael Torres, whose DeSci-funded company is advancing a genuine drug program and whose previous venture, Crossbridge Bio, was acquired by Eli Lilly for $300 million.
The patient is becoming the principal investigator
A quieter thread ran underneath the infrastructure: the people science is for are moving from the end of the pipeline to the start of it. Dongsinne Sohn made the sharpest version of the point with DermaDAO and Biofy real, formulated products that turn use and lived experience into research.
The product is not the end point. It's the entry point.
Dongsinne Sohn, DermaDAO
Alex Miloski showed an encrypted longevity app built on the same instinct your biology, in one place, that the company itself can’t see. His warning about everyone else was: “if you connect your wearables somewhere else, you’re probably the product.” And Dr. Mohsen Soofian gave the room the clinical reality of the peptide boom what works, what’s noise, and the honest gap underneath it: “there’s no test to tell you which peptide you need.”
Whoever controls the model controls the science
If science runs on AI, then access to AI becomes the question. Bradley Clark Royes of the Foresight Institute told a story that landed hard: a frontier model his group relied on had its access restricted with a day’s notice.
If we don't host the model, someone else controls the off switch.
Bradley Clark Royes, Foresight Institute
He put numbers to the asymmetry roughly 90% of notable frontier models come from industry, and about 90% of academics can’t access them, and argued the answer is physical: on-prem compute and real institutions, not just protocols. Migle Rakitaite came at the same problem from the data side, drawing on a two-year ordeal getting her own cross-border medical records. Her proposal inverts the usual model: don’t hand over the dataset, “bring the code to the data.”
The most credible people on stage undercut their own hype
This is the part worth sitting with. In a field that rewards loud claims, the strongest talks did the opposite. Aubrey de Grey, joining remotely, reported that combining four anti-aging interventions in middle-aged mice extended life by only about four months, then explained why that modest result is the real news: it’s proof of principle that the effects stack. Nico Alavi pointed out that the data needed to make AI models generalize in biology is still sparse, and might be “a couple more years” away. The legal panel agreed an AI still can’t hold a patent on its own, and that liability always lands on a human somewhere.
Even Rolli, mid-demo, drew the line plainly: “closing the loop is a demo, owning the data at each turn is the moat.” That candour is the signal, not the caveat. As Kohlhaas put it in the opening: hype mints attention; outcomes generate trust.
So where does that leave things? The loop isn’t closed. The wet lab is still the only real gate, the data is still thin, and the law still trails years behind. But the direction across two days was unmistakable: the distance between an idea and a tested result is collapsing, and the rails to fund, own, and run that work are being built in the open, much of it on Bio.
