Improvements in viral gene annotation using large language models and soft alignments.

Area of Science:

• Molecular Biology
• Bioinformatics
• Artificial Intelligence

Background:

• Protein sequence annotation is a significant challenge in molecular biology, especially for viral proteins.
• Traditional homology search methods (alignment, k-mer, profile-based) struggle with viral sequences due to limited homology.
• Large Language Models (LLMs) present a novel approach for protein sequence annotation using embeddings.

Purpose of the Study:

• To introduce a novel methodology for protein sequence annotation using LLM embeddings.
• To address the limitations of traditional methods in annotating viral proteins.
• To enhance the efficiency and interpretability of protein annotation.

Main Methods:

• Development of a soft alignment algorithm leveraging amino acid embedding similarity.
• Bypassing traditional scoring matrices by using embedding similarity.
• Creating transparent, BLAST-like alignment visualizations for interpretability.

Main Results:

• The soft alignment algorithm surpasses pooled embedding-based models in efficiency and interpretability.
• The method allows users to trace homologous amino acids and provides clear alignment visualizations.
• The novel approach successfully annotated sequences that blastp and pooling-based methods failed to detect, as validated on Virus Orthologous Groups and ViralZone databases.

Conclusions:

• LLM-based embedding approaches hold significant potential for improving protein sequence annotation, particularly in viral genomics.
• This method offers a more efficient and accurate pathway for protein function inference in molecular biology.
• The findings highlight a promising advancement in combining AI with traditional biological research for enhanced annotation.