GeneGenie: enhancing biomedical question-answering with agentic graphs

Area of Science:

• Biomedical Informatics
• Artificial Intelligence in Medicine
• Computational Biology

Background:

• Large language models (LLMs) show promise in biomedical research but suffer from hallucinations and lack precision for multi-hop reasoning.
• Existing AI models struggle to balance generative capabilities with the factual rigor required for complex biomedical analysis.

Purpose of the Study:

• Introduce GeneGenie, a model-agnostic, multi-agent framework using a directed acyclic graph architecture.
• Address the gap between AI generative capacity and factual accuracy in biomedical research.
• Orchestrate query planning, retrieval-augmented generation, and bioinformatics tool execution.

Main Methods:

• Implemented a deterministic five-node pipeline within a directed acyclic graph architecture.
• Integrated curated databases (GenCC, HGNC, UniProt) and bioinformatics tools (NCBI E-Utilities, BLAST+).
• Evaluated performance on the 1600-pair GeneTuring benchmark, comparing standalone LLMs with the agentic Graph Mode.

Main Results:

• The graph-based architecture consistently outperformed single-model baselines across all metrics.
• Gemini 2.5 Pro in Graph Mode achieved 72.375% accuracy, significantly higher than the best baseline (15.8%).
• LLM-as-Judge assessment confirmed enhanced lexical accuracy, completeness, and factual grounding with the agentic approach.

Conclusions:

• GeneGenie establishes a robust, reproducible paradigm for biomedical AI systems.
• Tool-augmented orchestration via a graph architecture is superior to relying solely on LLM scale.
• Future biomedical AI systems can benefit from this agentic, tool-integrated approach, despite limitations in protein-coding gene recognition.