BetaDescribe: Providing rich descriptions from protein sequences

Area of Science:

• Bioinformatics
• Computational Biology
• Protein Science

Background:

• Deciphering protein function is crucial for medicine and biotechnology but experimentally challenging.
• Existing large language models (LLMs) struggle with accurate protein functional and structural prediction.
• Current bioinformatics tools often rely on sequence or structure similarity for protein annotation.

Purpose of the Study:

• To introduce BetaDescribe, a novel model for generating comprehensive textual descriptions of protein properties.
• To address the limitations of general LLMs in bioinformatics by incorporating biological knowledge.
• To provide an alternative method for protein function exploration independent of sequence homology.

Main Methods:

• BetaDescribe is built upon the LLAMA2 model and trained on combined biological and English text datasets.
• The model accepts protein sequences as input and outputs detailed functional and structural descriptions.
• Validator models and a judge component refine and rank the generated descriptions for accuracy.

Main Results:

• BetaDescribe successfully generated descriptions for proteins with low sequence similarity to known ones.
• In silico mutagenesis experiments confirmed BetaDescribe utilizes functionally important protein regions.
• The model demonstrates the ability to identify critical regions for protein functionality without homologous sequences.

Conclusions:

• BetaDescribe offers a powerful new tool for exploring protein functionality.
• The model augments existing annotation transfer methods by providing sequence-independent functional insights.
• BetaDescribe has significant potential to advance bioinformatics and related fields.