The vast protein universe remains largely uncharted. While gene sequencing projects have unveiled billions of protein sequences, fewer than one percent have ever been functionally characterized in a laboratory. The experimental painstaking—crystallizing structures, assaying enzymatic activities, mapping interaction networks—cannot scale to the trillions of proteins that nature has sculpted over billions of years. Even cutting-edge large language models, which have revolutionized textual understanding, have stumbled when asked to read the language of amino acids and spit out a coherent biological function. Now, a team from the Technion – Israel Institute of Technology and Tel Aviv University has broken that barrier. In a study published in the Proceedings of the National Academy of Sciences, they unveil BetaDescribe, an artificial intelligence system that directly translates a raw protein sequence into a detailed, natural-language description of what that protein does and how it behaves.
BetaDescribe departs sharply from conventional homology-based annotation pipelines. Standard computational tools infer function by searching for sequence similarity against databases of known proteins. If a protein looks like a kinase, it is labeled a kinase. But such approaches fail spectacularly for the vast “dark proteome” of sequences that have diverged too far from characterized relatives. BetaDescribe instead marries a generative model with an internal verification architecture. At its core sits a transformer stack that jointly embeds protein sequences and textual descriptions into a shared representational space. The encoder digests the amino acid chain, capturing local physicochemical motifs and long-range dependencies; a dedicated decoder then autoregressively generates a multi-sentence caption that details catalytic residues, substrate preferences, metabolic pathway membership, and allosteric binding pockets. Crucially, a second module cross-checks this output against structure-prediction surrogates and evolutionary coupling constraints, flagging inconsistencies and iteratively refining the description until it converges on an evidence-grounded narrative.
The technical sophistication lies in how the system integrates multiple layers of biological knowledge without ever seeing a crystal structure. It leverages pre-training on millions of unlabeled sequences, akin to the protein language models ESM-2 and ProtT5, but adds a supervised fine-tuning stage on carefully curated sets of experimentally annotated proteins. During inference, the verification loop consults predicted contact maps and residue-conservation scores from multiple sequence alignments, ensuring that a claimed active site is physically plausible and evolutionarily preserved. The final output is not a simple gene ontology term but a rich paragraph, for example, explaining that a previously mysterious bacterial protein is a metal-dependent hydrolase involved in cell-wall remodeling and likely binds zinc ions via a conserved HExxH motif, with potential implications for antibiotic development.
The researchers demonstrated the power of BetaDescribe on six genuinely uncharacterized proteins selected from metagenomic datasets. For each, the system proposed concrete functional hypotheses, ranging from a novel thermophilic lipase to a transcription factor responsive to oxidative stress. Laboratory validation of the first two targets confirmed the predicted activities with high accuracy, underscoring that the AI was not merely hallucinating plausible-sounding jargon but reasoning over genuine sequence-encoded features. This success points toward a future where functional genomics can be done computationally at scale, generating testable hypotheses within seconds rather than months.
The broader significance for medicine and biotechnology is profound. Consider the blockbuster drug Ozempic, a glucagon-like peptide-1 receptor agonist whose development was inspired by a peptide discovered in the saliva of a Gila monster. Nature is brimming with such molecular treasures, but our ability to identify them has been bottlenecked by slow experimental characterization. BetaDescribe could systematically scan millions of environmental DNA sequences, flag those with predicted therapeutic properties—such as glucose regulation, antimicrobial activity, or plastic degradation—and prioritize them for synthesis and testing, radically compressing the timeline from discovery to clinical candidate.
Agriculture, too, stands to benefit. The system can probe soil and plant microbiomes for enzymes that detoxify pollutants, fix nitrogen more efficiently, or confer drought tolerance. By describing a protein’s optimal pH, temperature range, and cofactor requirements, BetaDescribe provides engineers with the specifications needed to incorporate those proteins into biotechnological workflows immediately, bypassing lengthy wet-lab trial-and-error.
The interdisciplinary collaboration was led by doctoral student Edo Dotan, jointly supervised by Prof. Yonatan Belinkov from the Technion’s Computer Science faculty and Prof. Tal Pupko from Tel Aviv University’s School of Life Sciences, with contributions from Prof. Eran Bacharach, Prof. Marcelo Ehrlich, and doctoral student Iris Lyubman. Belinkov, an expert in interpretable natural language processing, and Pupko, a leading evolutionary biologist, combined insights from both worlds to ensure the model was not just linguistically fluent but biologically faithful. The research was supported by the Israel Science Foundation.
With BetaDescribe, the protein universe becomes legible in a new, profoundly useful way. Instead of gazing at a string of letters and seeing only encrypted chemistry, scientists can now have a conversational partner that proposes, in plain English, what a protein is up to—and backs its claims with evolutionary logic. As the cost of DNA sequencing continues to plummet and the tide of unannotated sequences rises, such AI interpreters will transition from marvels to essential tools, turning the dark proteome from an impenetrable mystery into a vast, organized library of molecular function.
Subject of Research: Protein sequence analysis and functional annotation using artificial intelligence
Article Title: BetaDescribe: Providing rich descriptions from protein sequences
News Publication Date: 29-Jun-2026
Web References: https://www.pnas.org/doi/abs/10.1073/pnas.2537345123
References: Dotan, E., Lyubman, I., Bacharach, E., Ehrlich, M., Pupko, T., & Belinkov, Y. (2026). BetaDescribe: Providing rich descriptions from protein sequences. Proceedings of the National Academy of Sciences, DOI: 10.1073/pnas.2537345123.
Image Credits: Prof. Yonatan Belinkov photo by Hadas Parush; Prof. Tal Pupko photo by Chen Galili; doctoral student Edo Dotan photo credit not specified; system diagram and example protein illustrations generated using ChatGPT.
Keywords: artificial intelligence, protein function prediction, bioinformatics, deep learning, generative models, drug discovery, computational biology, protein language models, dark proteome, metagenomics
