It turns a vast, undifferentiated landscape into candidates worth investigating.
For as long as biology has been practiced, the gap between a protein's existence and humanity's understanding of it has been vast — billions of sequences in nature, only hundreds of thousands ever characterized by hand. A team from Technion and Tel Aviv University has now built an AI system called BetaDescribe that reads raw protein sequences and writes back plain-language descriptions of what those proteins do, how they behave chemically, and where drugs might bind to them. The work, published in the Proceedings of the National Academy of Sciences, does not replace the laboratory, but it reorients the researcher's first question — from 'where do I even begin?' to 'is this hypothesis worth testing?'
- Billions of proteins exist in nature while science has formally characterized only a fraction, leaving drug discovery and biotechnology perpetually constrained by what remains unknown.
- Previous AI attempts to bridge this gap stumbled because protein function demands inference, not just pattern-matching — a distinction BetaDescribe was specifically designed to address.
- The system pairs a generative language model with verification mechanisms, allowing it to describe proteins that share no close relatives with anything previously studied.
- In a direct test, BetaDescribe successfully characterized six proteins that had never been described before, producing actionable hypotheses about function, catalytic activity, and binding sites.
- The technology is a proof of concept, not a finished tool — its real measure will come as researchers deploy it against the full, uncontrolled diversity of proteins found in the wild.
In any laboratory working with proteins, the first obstacle is often the simplest and most stubborn: a researcher holds a sequence of amino acids and has no reliable way to know what it does. Experimental characterization is slow, expensive, and selective — science has formally described only hundreds of thousands of proteins while nature contains billions. That gap has long acted as a ceiling on drug discovery and biotechnology. A team from Technion and Tel Aviv University, led by doctoral student Edo Dotan under supervisors Yonatan Belinkov and Tal Pupko, has built an AI system called BetaDescribe to push that ceiling higher.
Published this week in the Proceedings of the National Academy of Sciences, BetaDescribe converts raw protein sequences into natural-language descriptions covering function, chemical activity, metabolic roles, and potential binding sites for drug molecules. The system's design sets it apart from earlier large-language-model approaches, which tended to rely on similarity to already-known proteins. BetaDescribe instead combines a generative model with verification and evaluation layers, enabling it to reason about proteins that have no close relatives in any existing database — the very proteins most likely to be overlooked.
The researchers grounded their motivation in a vivid example: Ozempic, now a cultural phenomenon, traces its origins to a peptide found in Gila monster saliva. That discovery required someone to ask the right question about an unusual animal. Most proteins never receive that kind of attention. BetaDescribe is an attempt to extend that attention systematically, at scale.
To test the system, the team applied it to six previously uncharacterized proteins. It described all six successfully, generating hypotheses specific enough to guide laboratory follow-up. The result does not prove what those proteins do — experimental validation remains essential — but it transforms open-ended exploration into hypothesis-driven investigation. A biologist no longer faces a blank sequence; they face a set of informed guesses worth testing.
The broader implications touch drug development timelines, biotechnology pipelines, and agricultural research. What the team has demonstrated is that artificial intelligence can translate the language of proteins into something researchers can read and act upon. How far that translation carries — and how reliably — will become clearer as BetaDescribe moves from controlled validation into the full complexity of proteins found in the world.
Somewhere in a laboratory, a researcher holds a protein sequence—a long string of letters representing amino acids—and faces a familiar problem: What does this thing actually do? How might it bind to a drug molecule? Could it treat disease? The answers exist, but finding them requires months of expensive experimental work, if the answers can be found at all. Now, a team from Technion and Tel Aviv University has built an artificial intelligence system that skips past some of that bottleneck by reading a protein's genetic code and writing back a description of what it does.
The system, called BetaDescribe, was published this week in the Proceedings of the National Academy of Sciences. It works by converting raw protein sequences into natural-language text that describes the protein's function, its chemical activity, how it participates in cellular metabolism, and where other molecules might attach to it. The breakthrough matters because science has characterized only hundreds of thousands of proteins in the laboratory, while billions or even trillions exist in nature. Most of what is out there remains unknown. The gap between what we know and what exists has always been a constraint on drug discovery and biotechnology. BetaDescribe narrows it.
The motivation is concrete. Ozempic, the obesity and diabetes drug that has become a cultural phenomenon, was developed by studying a peptide found in the saliva of a Gila monster—a rare desert lizard. That discovery required someone to look at an unusual animal and ask the right question. But most proteins never get that kind of attention. They sit in nature, uncharacterized, their potential unrealized. Large language models have tried to bridge this gap before, but with limited success. They can recognize patterns in text, but protein function is not purely a matter of pattern recognition. It requires inference—the ability to reason about what a protein might do even when it looks nothing like anything previously studied.
BetaDescribe takes a different approach. Rather than relying solely on similarity to known proteins, the system combines a generative model—the part that actually writes the descriptions—with verification mechanisms and evaluation processes that check whether the predictions make sense. This allows it to infer function even for proteins that have no close relatives in the existing database of characterized sequences. The system also goes beyond simple function labels. It identifies specific properties: catalytic activity, metabolic involvement, potential binding sites. These details matter for drug design and for understanding how a protein might fit into a larger biological system.
The researchers tested BetaDescribe on six proteins that had never been characterized before. The system successfully described all of them, generating hypotheses about their function that could guide future laboratory work. This is not the same as proving what a protein does—that still requires experimental validation—but it is a significant step. It transforms the research process from open-ended exploration into hypothesis-driven investigation. Instead of a biologist staring at a sequence and wondering where to start, they now have a detailed, evidence-based guess about what to look for.
The implications ripple outward. Drug discovery could accelerate. Biotechnology companies could move faster from identifying a protein target to developing a therapeutic. Agricultural research could benefit. The path from basic discovery to clinical application, which can take years and hundreds of millions of dollars, could become shorter and more efficient. The system does not replace laboratory work—it cannot. But it makes that work more focused, more purposeful. It turns a vast, undifferentiated landscape of unknown proteins into a landscape of candidates worth investigating.
The work was led by doctoral student Edo Dotan under the supervision of Yonatan Belinkov from Technion's computer science faculty and Tal Pupko from Tel Aviv University's School of Life Sciences. The team included Eran Bacharach, Marcelo Ehrlich, and Iris Lyubman, also from Tel Aviv. What they have built is not a finished product but a proof of concept—a demonstration that artificial intelligence can read the language of proteins and translate it into something humans can understand and act upon. The real test will come in the months and years ahead, as researchers begin using BetaDescribe to guide their own investigations and as the system encounters proteins in the wild, far from the controlled conditions of its initial validation.
Citações Notáveis
The ability to rapidly generate evidence-based hypotheses regarding unknown protein functions could significantly shorten the path from basic discovery to medical and industrial applications.— Research team (paraphrased from paper)
A Conversa do Hearth Outra perspectiva sobre a história
Why does it matter that we can describe proteins we've never seen before? Don't we already know how to study proteins in the lab?
We do, but it's slow and expensive. There are trillions of proteins in nature and we've only characterized a tiny fraction. BetaDescribe lets us make educated guesses about unknown proteins before we spend months and money testing them. It's like having a map instead of wandering blind.
But how does the system know what a protein does if it's never seen anything like it before?
It doesn't rely on similarity alone. It uses a generative model trained on known proteins, then verifies its own predictions. It's reasoning about structure and chemistry, not just pattern-matching. It can identify binding sites, catalytic activity, metabolic roles—the details that matter for drug design.
The source mentions Ozempic came from a lizard. Is BetaDescribe trying to replace that kind of serendipitous discovery?
Not replace it. Accelerate it. Right now, most proteins never get discovered because no one thinks to look. BetaDescribe could flag proteins worth investigating before anyone stumbles upon them. It turns luck into strategy.
What happens after the system describes a protein? Does that description count as proof?
No. It's a hypothesis. A very good one, backed by the system's reasoning about chemistry and structure, but it still needs laboratory validation. The real value is that it tells researchers where to focus their experiments, not that it eliminates the need for them.
Who benefits most from this?
Drug companies, biotechnology firms, agricultural researchers—anyone trying to find new proteins with useful properties. But also basic science. It democratizes protein discovery. You don't need a rare desert lizard or a lucky hunch anymore.