Deep TechTomorrow's greatest founders. Featured today.
Anto Biosciences · The drug that worked everywhere else
There is a drug sitting somewhere in the graveyard of failed US clinical trials that was fully approved and performing in China. A billion dollars lost. A therapy that worked - just not here.
For decades, no one could explain it with any precision. The compound was the same. The disease was the same. The patients were human. The result was not the same.
The standard explanation - population differences, genetic variation, regulatory divergence - was always unsatisfying. Technically true, but too vague to act on. You couldn't optimize around an explanation that diffuse.
Arvid Gollwitzer can now tell you exactly why. Not approximately. Not probably. The specific microbial strains responsible for metabolizing that drug differently in American gut microbiomes versus Chinese ones. The mechanism identified. The failure explained.
He's 27. His company, Anto, is four people. And he just demonstrated something the pharmaceutical industry has been trying to do for decades.
Arvid grew up outside Munich - quiet, precise, the kind of city that produces engineers who build things that last. He spent part of high school in Washington DC at Thomas Jefferson, one of the most selective science and technology high schools in the country. The exposure early to serious technical ambition shaped something.
Undergrad took him to ETH Zurich - the MIT of Europe - where he began doing what would become a career-long obsession: building computational tools to process biological data that existing hardware wasn't designed to handle. Not studying biology. Not studying computer science. Building the infrastructure at the intersection of both, because the intersection is where the unsolved problems live.
Graduate school brought him to Cambridge in the UK, then to MIT. That's where he met David de Gruijl.
David was a wet lab experimentalist. Arvid was a computational researcher working on foundation models. In most labs, those two don't have much to say to each other. In their lab, they found the overlap immediately - because they were both running into the same wall.
"We don't have enough data," Arvid describes the frustration. "The experiments are too slow and they're incredibly expensive. And once you spend all that money and time, you still don't have enough data to get to really fundamentally novel insight."
It's a structural problem with how biological research works. You design an experiment, you run it, you get results, you publish. But the results are often narrow — one compound, one bacterial strain, one cohort — and the data generated rarely scales to the kind of insight that changes how the field thinks. You move slowly because the process forces you to move slowly.
Arvid and David didn't want to move slowly.
The conventional answer in biotech is to generate more proprietary data. Build bigger experiments. Lock up exclusive datasets. This is the worldview that has driven a generation of biotech companies - the belief that data exclusivity is the moat.
Arvid thought that was wrong.
"The data set you want is probably sitting somewhere in the public domain," he says. "You can't really build a moat around proprietary or exclusive access to data sets anymore. Although that is the controversial, old-fashioned view that pharma companies sometimes still have."
His bet was different: the real advantage isn't in owning data. It's in building algorithms sophisticated enough to make sense of data that already exists - messy, noisy, distributed across repositories in the US, China, Europe, and around the world - and turning that into something a foundation model can actually learn from.
This is the insight that became Anto's technical foundation. Not "we have data others don't." But "we built the computational solution to use data others can't."
The distinction matters enormously to anyone allocating capital. Proprietary data can be replicated, bought, or regulated away. A better algorithm compounds.
MetaOmics-10T: The Foundational Dataset to Unlock Causal Modeling of Microbial Ecosystems (research paper by Arvid)
Anto is a full-stack frontier research lab - they pre-train models, fine-tune them, and generate their own data for targeted downstream applications. But the centerpiece is what Arvid calls a multimodal foundation model for the gut microbiome: the first of its kind.
The microbiome is the ecosystem of bacteria, microbes, and other organisms living in your gut. It's not passive. It's metabolically active - constantly processing everything you ingest, including drugs.
When you swallow a small molecule pharmaceutical, it doesn't go directly to work. It lands in your gut microbiome first, where it's digested and transformed by whatever microbial community happens to live there. That transformation determines the effective concentration of the drug, whether it remains active, whether it becomes toxic. In short: it determines whether the drug works.
The problem is that nobody's gut microbiome is the same. And population-level microbiome differences - between countries, ethnicities, diets, geographies - are significant. Significant enough to explain why a drug approved in one population fails in another.
This is the mechanism behind that billion-dollar clinical trial failure. The drug wasn't defective. The trial design wasn't flawed. The microbiome of the US patient population metabolized the compound differently than the Chinese population it had been developed and tested in. And until Anto, no one had the tools to identify the specific strains responsible or predict how that metabolism would vary across populations.
Anto's model can do both.
The clearest demonstration of what Anto's models can do isn't a simulation. It's a retroactive analysis of a real drug failure.
Arvid and his team took a compound that had been approved and performing in China, failed clinical trials in the US, and represented a material loss for the pharma company that ran the trial. They fed it through their model. And for the first time, they identified the specific microbial strains in representative US patient cohorts responsible for metabolizing the drug differently - the precise mechanism behind the failure.
That analysis is now a paper under review. It's the kind of result that, in the words of the pharma buyers Anto talks to, makes scientists stop and pay attention.
"What we're doing is based on our publications," Arvid explains of the go-to-market logic. "We've combined over a decade of research experience in this space. And those works are really what's exciting to pharma."
There are currently eight Anto papers under review at major AI conferences and journals, including Nature.
Pharma is not a market you close top-down. It doesn't work that way. The buyers aren't executives who saw a LinkedIn post - they're scientists inside large organizations who have spent careers on these problems and know immediately when something is real versus noise.
You reach them through peer-reviewed science. You build credibility in the lab before you build it in the boardroom. Then - when the scientists are excited, when the papers are circulating internally, when someone finally says "we need to talk to these people" - you have a conversation.
That's Anto's go-to-market. Bottom-up. Publication-driven. Slow by design and fast in result.
The structure of the commercial relationship once that conversation happens: pilots, priced at $300–500K each, designed to demonstrate what the model can find on a specific drug asset. Pilots convert to joint development partnerships — longer-term, larger-scope engagements where Anto's models are integrated into the drug development process.
They currently have $750K in Letters of Intent. For a company this early, that number matters less than the signal it carries: pharma scientists are saying yes.
Link to all of Arvid’s research papers
Arvid was a part of the W26 YC batch. His partner is Nicolas Dessain.
When he talks about why he applied, he's precise about what YC is actually useful for - and it isn't what most people think.
"Often as researchers, you publish this work, you build new foundation models, but then you don't do that last translational step where you turn it into a product." The research-to-product gap is where most frontier science dies. The papers get published. The results get cited. The technology sits in a lab.
YC forces you through that gap. It's not about the check. It's about the accountability structure that makes researchers become builders - that pushes you to stop optimizing the model and start figuring out who is going to pay for what it can do.
For Arvid and David, that meant building a commercial organization around the foundation models they'd spent years developing - not at the cost of the science, but in service of it.
Anto isn't planning to build new models. The foundation is in place. The next phase is application.
The focus: using existing models to discover new biomarkers that predict drug response variability, identify the specific mechanisms driving that variability, and optimize existing drug assets for broader efficacy across different population microbiomes. That optimization problem, once you have the mechanism identified, is - in Arvid's words - "quite easy."
The pharma partnerships are structured around exactly this workflow. A company brings an asset. Anto runs it through the model. They identify the population-level variability in drug response, explain the microbial mechanism, and propose optimization approaches to improve efficacy across a broader patient base.
It's a new category of service in drug development. And the market, measured by the $100 billion-plus spent annually on acquiring drug assets from abroad, is large enough that even early commercial traction is meaningful.
If you want to understand what Anto has built at a technical level - written for a broader audience, not just researchers - the paper to start with is Making the Microbiome Computable. It covers the foundation model architecture, the data pre-processing approaches Arvid's team developed to make public microbiome data useful for model training, and the computational solutions behind training on microbiome data at scale.
It's the clearest articulation of the scientific thesis in one place. Worth an hour.
Additional research from the Anto team:
Anto paper on their foundational model: Darwin-7B: Making the Microbiome Computable via Sparsified Quality-Aware Tokenization
There's a version of this story that is a biotech story. There's a version that is an AI story. Arvid would tell you it's neither - that the most interesting work is always at the intersection of fields that don't usually talk to each other.
A computational researcher and a wet lab experimentalist, frustrated by the same wall, building a different door.
The drug that failed in America already told them they were right.
Arvid Gollwitzer is the co-founder of Anto, a full-stack microbiology foundation model lab and YC W26 company. If you're building in biotech, drug development, or longevity and want an introduction, reach out.
FoundersBrief - Tomorrow's greatest founders. Featured today.
