Protein2Text: Resampling Mechanism to Translate Protein Sequences into Human-Interpretable Text

Area of Science:

• Bioinformatics
• Computational Biology
• Artificial Intelligence in Life Sciences

Background:

• Proteins are essential biological molecules, but most known sequences are uncharacterized due to experimental limitations.
• Accelerating protein characterization is crucial for advancing biological understanding and drug discovery.

Purpose of the Study:

• To introduce Protein2Text, a multimodal large language model designed for interpreting protein sequences.
• To generate informative text addressing open-ended questions about protein functions and attributes, assisting experimentalists.

Main Methods:

• Utilized an adapted LLaVA framework integrated with a resampling mechanism to map protein sequences into a language-compatible space.
• Trained the model on a newly curated dataset derived from PubMed articles.
• Developed and employed four comprehensive benchmarks for rigorous evaluation, including in-domain and cross-domain assessments.

Main Results:

• Protein2Text demonstrated superior performance in open-ended question-answering tasks compared to existing models.
• The model effectively interprets protein sequences and generates relevant textual information.
• Highlighted limitations in current evaluation metrics for template-based approaches, advocating for unbiased assessment.

Conclusions:

• Protein2Text offers a powerful new tool for accelerating protein characterization and hypothesis generation in biological research.
• The model's ability to handle complex queries and generate informative text represents a significant advancement in bioinformatics.
• Public availability of model weights and datasets facilitates further research and development in protein informatics.