AI System BetaDescribe Generates Natural Language Descriptions of Protein Functions

Original post

Check out our work on generating rich textual descriptions from protein sequences, now appearing in @PNASNews https://www.pnas.org/doi/abs/10.1073/pnas.2537345123

I’ll share some background and more thoughts in the thread.

Technion Israel@TechnionLive

AI is learning the language of proteins. 🧬

Researchers led by Edo Dotan, under the supervision of Prof. Yonatan Belinkov (Technion's Henry and Marilyn Taub Faculty of Computer Science) and Prof. Tal Pupko (Tel Aviv University), developed BetaDescribe—an AI system that explains protein functions in natural language.

Read more: https://www.technion.ac.il/en/blog/article/the-language-of-proteins/

#AI #ProteinScience #Biotech #Technion

2:41 PM · Jul 3, 2026 · 48 Views

Yonatan Belinkov@boknilev

Another observation: validation of rich descriptions is hard. We had three experienced biology professors go through dozens of these manually.

12m91

LIKES1REPLIES1

Yonatan Belinkov@boknilev

We initially considered two routes to address this: (1) use a protein encoder and train a text decoder, possibly as an adapter on an existing LLM. (2) continue training an existing LLM.

Yonatan Belinkov@boknilev

Check out our work on generating rich textual descriptions from protein sequences, now appearing in @PNASNews https://www.pnas.org/doi/abs/10.1073/pnas.2537345123

I’ll share some background and more thoughts in the thread.

12m610

Yonatan Belinkov@boknilev

We opted for (2) based on insights from other multi modal models. But we didn’t have the resources for full A/B testing so would love to see alternatives. (Indeed there are alternatives that came out while the paper was in the works).

12m61

Yonatan Belinkov@boknilev

One result from our work shows that in remote proteins, our generator produces better descriptions. When the sequence is similar to one seen in training, in contrast, a retrieval approach is better or same.

Yonatan Belinkov@boknilev

12m410

Yonatan Belinkov@boknilev

An interesting lesson is that a pipeline system is important: generate multiple candidates, run several validators, score candidates with an LLM judge, and produce diverse alternatives for an expert user. This system architecture works well with different underlying generators.

12m21

Yonatan Belinkov@boknilev

Finally, ending with mixed feelings about the review process: The paper was under a long review in a top journal, eventually being rejected with hardly any evidence or reasoning. Very disappointing. But we also got valuable feedback from other reviewers that improved the work.

12m81