ProBEiT combines protein sequence and structure for efficient engineering predictions
Published in Artificial Intelligence, Bioinformatics, Journals, Tools.
Published October 28, 2026.
An October 26 paper in Communications Biology introduces ProBEiT, a protein language model that combines amino-acid sequences with structural information. Li and colleagues aim to improve protein-property prediction without relying solely on increasing model size.
ProBEiT jointly represents sequences and tokenized structures through an architecture that specializes parts of the model for different information types. Trained on 1.7 million AlphaFoldDB proteins, the 400-million-parameter system outperformed larger comparison models on six evaluated fitness-prediction and functional-annotation tasks.
The team also tested predictions experimentally while engineering the thermostability of glycerol-3-phosphate oxidase. The model identified improved variants involving positive epistasis, where combinations of changes have favorable interactions.
The work offers a tool for prioritizing protein-engineering experiments and studying sequence–structure–function relationships. Benchmark performance and an enzyme demonstration do not establish equivalent accuracy for every protein family or therapeutic application. Predicted structures, training-data coverage and task selection remain relevant constraints; experimental validation is still necessary.
Source: Li et al., Communications Biology, October 26, 2026. DOI: 10.1038/s42003-026-10816-6.