BetaDescribe: Providing rich descriptions from protein sequences
Significance:
Determining protein function remains a major bottleneck in biology, particularly for proteins lacking close homologs with known annotations. This work introduces BetaDescribe, a protein-to-text modeling framework that generates rich, human-readable functional descriptions directly from protein sequences. The proposed architecture relies on a specialized multicomponent framework featuring protein-to-text generation, multiple validators, and a judge model. We show that this integrated system significantly boosts performance and ensures description reliability. By integrating biological and natural language knowledge, BetaDescribe captures functional signals beyond sequence similarity and highlights regions critical for protein activity. This approach provides a complementary strategy to traditional annotation methods, enabling deeper exploration of protein function across uncharacterized regions of the proteome universe.
Abstract:
Deciphering protein function is fundamental to advancements in medicine and biotechnology. However, conventional experimental characterization remains resource-intensive. Public large language models (LLMs), though proficient in natural language processing, often fail to accurately interpret and predict the functional and structural properties of proteins, limiting their utility in bioinformatics. To address this gap, we introduce BetaDescribe, designed to generate detailed and rich textual descriptions of proteins, including their function, catalytic activity, involvement in specific metabolic pathways, subcellular localizations, and the presence of specific domains. The trained BetaDescribe model receives protein sequences as input and outputs a textual description of these properties. BetaDescribe starting point was the LLAMA2 model, which was trained on trillions of tokens. Our model was next trained on datasets containing both biological and English text, which allowed the incorporation of biological knowledge. In addition to the description generator, BetaDescribe comprises multiple validator models and a judge, which together enable accurate ranking of alternative generated descriptions. We demonstrate the utility of BetaDescribe by providing descriptions for proteins that share little to no sequence similarity to proteins with functional descriptions in public datasets. Using in silico mutagenesis, we further show that BetaDescribe relies on functionally important regions, as part of its prediction, suggesting that the model identifies regions of importance for the protein functionality without needing homologous sequence. BetaDescribe offers a powerful tool to explore protein functionality, augmenting existing approaches such as annotation transfer based on sequence or structure similarity.
image/png
BetaDescribe workflow. The generator processes the protein sequences and creates multiple candidate descriptions. Independently, the validators provide simple textual properties of the protein. The judge receives the candidate descriptions (from the generator) and the predicted properties (from the validators) and rejects or accepts each description. Finally, BetaDescribe provides up to three alternative descriptions for each protein.
Dotan, E., Lyubman, I., Ehrlich, M., Bacharach, E., Pupko, T., and Belinkov, Y. 2026. BetaDescribe: providing rich descriptions from protein sequences. Proceedings of the National Academy of Sciences (PNAS)
Description:
FUNCTION$ The leader peptide is a component of released, infectious virions and is required for particle budding, & The transmembrane protein (TM) acts as a class I viral fusion protein. Under the current model, the protein has at least 3 conformational states: pre-fusion native state, pre-hairpin intermediate state, and post-fusion hairpin state. During viral and target cell membrane fusion, the coiled coil regions (heptad repeats) assume a trimer-of-hairpins structure, positioning the fusion peptide in close proximity to the C-terminal region of the ectodomain. The formation of this structure appears to drive apposition and subsequent fusion of viral and target cell membranes. Membranes fusion leads to delivery of the nucleocapsid into the cytoplasm, SUBCELLULAR LOCATION$ Endoplasmic reticulum membrane.
FUNCTION$ Envelope glycoprotein that forms spikes at the surface of virion envelope. Essential for the initial attachment to heparan sulfate moities of the host cell surface proteoglycans. Involved in fusion of viral and cellular membranes leading to virus entry into the host cell. Following initial binding to its host receptors, membrane fusion is mediated by the fusion machinery composed at least of gB and the heterodimer gH/gL. May be involved in the fusion between the virion envelope and the outer nuclear membrane during virion egress, SUBCELLULAR LOCATION$ Virion membrane, SUBUNIT$ Homotrimer; disulfide-linked. Binds to heparan sulfate proteoglycans. Interacts with gH/gL heterodimer, SIMILARITY$ Belongs to the herpesviridae glycoprotein B family.
FUNCTION$ The surface protein (SU) attaches the virus to the host cell by binding to its receptor. This interaction triggers the refolding of the transmembrane protein (TM) and is thought to activate its fusogenic potential by unmasking its fusion peptide. Fusion occurs at the host cell plasma membrane, & The transmembrane protein (TM) acts as a class I viral fusion protein. Under the current model, the protein has at least 3 conformational states: pre-fusion native state, pre-hairpin intermediate state, and post-fusion hairpin state. During viral and target cell membrane fusion, the coiled coil regions (heptad repeats) assume a trimer-of-hairpins structure, positioning the fusion peptide in close proximity to the C-terminal region of the ectodomain. The formation of this structure appears to drive apposition and subsequent fusion of viral and target cell membranes. Membranes fusion leads to delivery of the nucleocapsid into the cytoplasm, SUBCELLULAR LOCATION$ Cell membrane. SUBUNIT$ The mature envelope protein (Env) consists of a trimer of SU-TM heterodimers attached by noncovalent interactions or by a labile interchain disulfide bond
Description:
FUNCTION$ Probably involved in the RNA silencing pathway and required for the generation of small interfering RNAs (siRNAs), CATALYTIC ACTIVITY$ a ribonucleoside 5'-triphosphate + RNA(n) = diphosphate + RNA(n+1), SIMILARITY$ Belongs to the RdRP family.
Description:
FUNCTION$ DNA-dependent RNA polymerase catalyzes the transcription of DNA into RNA using the four ribonucleoside triphosphates as substrates, CATALYTIC ACTIVITY$ a ribonucleoside 5'-triphosphate + RNA(n) = diphosphate + RNA(n+1), SIMILARITY$ Belongs to the RNA polymerase beta' chain family.
FUNCTION$ DNA-dependent RNA polymerase catalyzes the transcription of DNA into RNA using the four ribonucleoside triphosphates as substrates. Specific core component of RNA polymerase III which synthesizes small RNAs, such as 5S rRNA and tRNAs, SUBCELLULAR LOCATION$ Nucleus, SUBUNIT$ Component of the RNA polymerase III (Pol III) complex consisting of 17 subunits, SIMILARITY$ Belongs to the eukaryotic RPC3/POLR3C RNA polymerase subunit family.
FUNCTION$ Decapping enzyme for NAD-capped RNAs: specifically hydrolyzes the nicotinamide adenine dinucleotide (NAD) cap from a subset of RNAs by removing the entire NAD moiety from the 5'-end of an NAD-capped RNA, SUBCELLULAR LOCATION$ Nucleus, COFACTOR$ a divalent metal cation, SIMILARITY$ Belongs to the DXO/Dom3Z family.
Citation
bibtex
1@article{Dotan_BetaDescribe_Providing_Rich_2026,
2 author = {Dotan, Edo and Lyubman, Iris and Ehrlich, Marcelo and Bacharach, Eran and Pupko, Tal and Belinkov, Yonatan},
3 doi = {10.1073/pnas.2537345123},
4 journal = {Proceedings of the National Academy of Sciences},
5 month = jun,
6 title = {{BetaDescribe: Providing Rich Descriptions for Protein Sequences}},
7 volume = {123},
8 year = {2026}
9}
