ViroBioTree: A Tree-Structured Biological Evidence Retrieval Framework for Viral Protein Function Annotation.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Accurate viral protein function annotation is crucial for genomic surveillance.
• Conventional retrieval-augmented generation (RAG) pipelines fragment biological evidence, disrupting critical relationships.
• Existing methods struggle to integrate diverse evidence types like ORF, annotations, structure, and attention data.

Purpose of the Study:

• To introduce ViroBioTree, a novel tree-structured framework for organizing and retrieving biological evidence for viral protein function review.
• To enhance the traceability and interpretability of evidence used in viral protein annotation.
• To provide a robust system for evidence-aware reranking and report generation.

Main Methods:

• ViroBioTree converts diverse biological evidence (sequence, annotation, structure, attention) into typed nodes and traceable edges.
• The framework employs deterministic multi-channel recall, evidence-aware reranking, and rule-based verification.
• It builds a tree structure on ORF-level predictions and annotations, organizing evidence for downstream review.

Main Results:

• ViroBioTree demonstrated strong performance in structure-explanation tasks (Precision@K=1.0, Recall@K=1.0).
• It showed advantages over LlamaIndex vector retrieval for conflict detection and evidence retrieval in SARS-CoV-2 and Influenza A Virus (IAV) data.
• The framework successfully represented IAV evidence namespaces and exposed data gaps.

Conclusions:

• ViroBioTree offers a significant advancement in organizing and retrieving complex biological evidence for viral protein analysis.
• The framework supports traceable evidence review and report generation, aiding genomic surveillance efforts.
• While excelling in evidence organization and retrieval, ViroBioTree complements, rather than replaces, primary annotation and wet-lab validation.