RAG-Based Multi-Ontology Annotation Framework

This repository contains the code, experiments, and analysis for a study proposing a new Retrieval-Augmented Generation (RAG) approach tailored for the automated annotation of biological sample metadata.

Overview

Traditional metadata annotation pipelines often rely on single-ontology retrieval or focus on isolated biological sample categories, which limits the richness and robustness of the resulting annotations. In this study, we introduce a two-stage RAG-based method designed to overcome these limitations:

1. Interpretation Stage The GPT-5-mini model infers and extracts contextual information from each biological sample label, producing an enriched and standardized interpretation.

2. Annotation Stage This refined interpretation is used as input to generate more accurate, ontology-aligned annotations with GPT-4o-mini.

To support the retrieval phase, we use a reduced and semantically focused subset of four major biological ontologies. This compact knowledge base minimizes noise, improves similarity search precision, and avoids the computational overhead of large-scale ontology retrieval.

Unlike most previous studies:

• We annotate across up to four ontologies, enabling richer multi-dimensional metadata standardization.
• We cover three biological entity categories — cell lines, cell types, and anatomical structures — expanding the scope beyond single-type annotation approaches.
• Our pipeline is designed to automate the majority of the annotation workflow, while keeping human validation as a final quality-control layer.

---

Repository Structure

analysis/
│── error_analysis/
│   └── Scripts for detailed error inspection.
│── evaluation_scripts/
│   └── Scripts for metric computation and performance evaluation.
│── method_comparison/
│   └── Tools for comparing the proposed method with other approaches.

k_candidates/
│── 5K/
│── 10K/
│── scripts/
│── candidate_positions.tsv
    Experiments and resources used to determine the optimal number of retrieved candidates
    for the RAG component.

RAG_process/
│── RAG_reduced_info_results/
│   └── Output annotations obtained using the reduced-information ontology subset.
│── scripts/
    Scripts implementing the RAG pipeline, including prompt templates and RAG logic.

---

Folder Descriptions

RAG_process/

Contains all code and results related to the main RAG system:

• scripts/ Scripts implementing the two-stage RAG pipeline (interpretation → annotation), prompt templates, retrieval configuration, and ontology preprocessing.
• RAG_reduced_info_results/ Generated annotations using the reduced semantic ontology subset.

Here is the README for this directory.

k_candidates/

Includes experiments conducted to determine the optimal number of retrieved candidates for the similarity search stage:

• Tests for 5K and 10K candidate configurations.
• candidate_positions.tsv logs candidate distributions and behavior.
• scripts/ contains all associated experimental code.

Here is the README for this directory.

analysis/

Contains the full evaluation suite for the study:

• error_analysis/ — inspection of annotation errors and qualitative assessment.
• evaluation_scripts/ — computation of precision, recall, and other metrics.
• method_comparison/ — comparisons with other annotation methods mentioned in the paper.

Here is the README for this directory.

Summary of Contributions

• A two-stage RAG methodology that improves annotation precision without requiring fine-tuning.
• Use of a semantically filtered multi-ontology knowledge base for efficient, noise-reduced retrieval.
• Annotation across four biological ontologies, capturing complementary biological dimensions.
• Application to three distinct biological sample categories.
• A complete experimental and evaluation pipeline for reproducibility.