Technion and Tel Aviv University AI writes plain-language profiles of unknown proteins

Israeli researchers have built an artificial intelligence system that turns a protein's raw sequence into a written description of its likely function. The tool, called BetaDescribe, was presented in the Proceedings of the National Academy of Sciences at the end of June 2026 and reported by Ynetnews on 4 July. The work was led by doctoral student Edo Dotan under the joint supervision of Professor Yonatan Belinkov of the Technion's Taub Faculty of Computer Science and Professor Tal Pupko of Tel Aviv University's School of Life Sciences, with Tel Aviv colleagues Eran Bacharach, Marcelo Ehrlich and Iris Lyubman.

Proteins carry out nearly every task in living cells, but working out what a particular protein does in the laboratory is slow and expensive. As a result, only a few hundred thousand proteins have been carefully characterised, while sequencing has revealed billions more, and possibly trillions exist in nature. Standard computer methods guess a protein's role mainly by comparing it with similar, already-known proteins, and they struggle when there is no close relative to compare against.

BetaDescribe pairs a generative language model with checking and scoring steps that test its own suggestions. Given a sequence, it can describe a protein's function, the chemical reactions it may catalyse, its role in metabolism and likely binding sites, the kind of information drug developers and bioengineers need. The team demonstrated it by describing six proteins that had never been characterised.

The project combines natural-language processing, Belinkov's speciality, with evolutionary biology, Pupko's field, and was supported by the Israel Science Foundation. The researchers say tools like this could help prioritise which unknown proteins are worth studying experimentally.