PlasGO: enhancing GO-based function prediction for plasmid-encoded proteins based on genetic structure.

Area of Science:

• Microbiology
• Bioinformatics
• Genomics

Background:

• Plasmids are mobile genetic elements crucial for bacterial trait transfer, like antimicrobial resistance.
• Accurate annotation of plasmid-encoded proteins using Gene Ontology (GO) is vital for functional studies and classification.
• Existing GO prediction methods face challenges due to functional diversity and limited high-quality annotations for plasmid proteins.

Purpose of the Study:

• To develop and evaluate PlasGO, a novel computational tool for predicting Gene Ontology (GO) terms for plasmid-encoded proteins.
• To address the limitations of current annotation methods by leveraging advanced language models.

Main Methods:

• PlasGO employs a hierarchical architecture combining a protein language model for local protein context and a BERT model for global plasmid context.
• A self-attention confidence weighting mechanism is integrated to allow users to control prediction precision.
• The tool was rigorously evaluated and benchmarked against seven state-of-the-art methods.

Main Results:

• PlasGO achieved commendable performance in GO term prediction for plasmid proteins.
• The tool significantly expanded annotations, assigning high-confidence GO terms to over 95% of previously unannotated proteins.
• High precision values (0.8229, 0.7941, 0.8870) were demonstrated across the three GO categories on a novel protein test set.

Conclusions:

• PlasGO effectively expands plasmid protein annotations with high-confidence GO terms, utilizing a hierarchical approach with protein language models and BERT.
• The generated annotations are compiled into a database, offering a valuable resource for downstream plasmid analysis and research.