VFB-MCP: Natural-Language Access to Drosophila Neuroscience Grounded by an Expert-Curated Ontology-Led Knowledgebase

Technical Summary: VFB-MCP

Problem Statement
Biological databases, such as the Virtual Fly Brain (VFB), contain vast amounts of expert-curated knowledge essential for neuroscience research. However, accessing this data effectively traditionally requires navigating complex web interfaces or utilizing APIs, which demands significant domain-specific expertise to formulate precise queries. This creates a substantial barrier to entry for new researchers, particularly in fields undergoing rapid paradigm shifts like connectomics. While Large Language Models (LLMs) offer a potential solution for natural-language interaction, standard LLMs lack the specific grounding required to retrieve accurate, verifiable, and quantified data from specialized scientific repositories, often leading to hallucinations or imprecise responses.

Methodology
To address these limitations, the authors implement the Model Context Protocol (MCP) to bridge LLMs with the VFB knowledgebase. VFB is an ontology-backed system specifically designed for Drosophila neuroscience, integrating expert curation with structured data to support complex queries, including those involving recently integrated connectomes.

The proposed VFB-MCP system allows an LLM to interact directly with the VFB ontology and database through a standardized protocol. This setup enables the model to:

1. Access structured, expert-curated data rather than relying on unstructured training data or general web search.
2. Execute precise queries grounded in the specific ontology of the field.
3. Generate responses that are verifiable against the source data and appropriately quantified.

Key Contributions

• Implementation of MCP for Neuroscience: The paper demonstrates the integration of MCP with an ontology-led knowledgebase (VFB) to facilitate natural-language access to complex biological data.
• Benchmarking Framework: The authors establish a comparative evaluation framework testing three distinct approaches on 30 specific neuroscience tasks:
  1. A bare LLM (no external tools).
  2. A web-search-assisted LLM.
  3. The VFB-MCP-equipped LLM.
• Ontology-Grounded Precision: The work highlights how grounding LLMs in expert-curated ontologies via MCP improves the retrieval of precise, verifiable, and quantified information compared to general retrieval methods.

Results
The evaluation on 30 neuroscience tasks yielded the following performance metrics:

• VFB-MCP LLM: Produced precise, verifiable, and appropriately quantified answers on 25 out of 30 tasks.
• Web-Search-Assisted LLM: Achieved success on 14 out of 30 tasks.
• Bare LLM: Achieved success on only 2 out of 30 tasks.

Statistical analysis using Wilcoxon signed-rank tests with Holm correction confirmed that the performance differences were statistically significant (p < 0.01) across all pairwise comparisons. The advantage of the MCP approach was particularly pronounced in tasks requiring data quantification, where VFB-MCP achieved an 89% success rate compared to 11% for the web-search approach.

Significance and Claims
The paper posits that utilizing MCP over ontology-backed knowledge graphs represents an effective method for enhancing the quality of LLM responses in specialized domains like neuroscience and connectomics. By moving beyond casual browsing and general web search, this approach provides the necessary precision to make complex, recently integrated datasets (such as connectomes) accessible to researchers who may lack the deep technical expertise required to query traditional database interfaces. The work establishes a pathway for improving data accessibility in scientific fields facing paradigm shifts by leveraging structured, expert-curated ontologies as the grounding mechanism for AI interaction.