VFB-MCP: Natural-Language Access to Drosophila Neuroscience Grounded by an Expert-Curated Ontology-Led Knowledgebase
Alex D. McLachlaniD, Rob CourtiD, Clare PilgrimiD, Kit D. LongdeniD, Nicolas H. Brown, David Osumi-SutherlandiD, Gregory S.X.E. JefferisiD, J. Douglas ArmstrongiD
Abstract
Biological databases store curated knowledge that researchers traditionally access through web interfaces or APIs. To move beyond casual browsing requires domain-specific knowledge and expertise to frame the queries necessary to explore this data. This generates barriers to both experienced users trying to integrate wide-ranging sources of information and for new users entering scientific fields undergoing paradigm shifts such as connectomics. A potentially powerful method to lower these barriers is to facilitate natural-language access to these databases by exposing them to large language models (LLMs) via the Model Context Protocol (MCP). Here we implement this for Virtual Fly Brain (VFB), an expert-curated and ontology-backed knowledgebase of Drosophila neuroscience, providing the precision needed to make recently integrated connectome and molecular data accessible. Benchmarked on 30 neuroscience tasks against a bare LLM and a web-search-assisted LLM, the VFB-MCP-equipped LLM produces precise, verifiable and appropriately quantified answers on 25/30 tasks vs 14/30 for web and 2/30 for bare. The MCP advantage is largest for tasks where data quantification is required (89% vs 11% web). This work establishes MCP layered over ontology-backed knowledge graphs as an effective method to improve LLM response quality for neuroscience and connectomics data, promoting accessibility and accelerating the reliable analysis of complex integrated datasets.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 1
- Organism
- Assay
- Research Focus
- Approach
In-depth summary
Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (14 Jul 2026) · CC BY 4.0
Model Context Protocol tools over Virtual Fly Brain's graph let Claude answer 25/30 Drosophila neuroscience tasks precisely, vs 14/30 with web search, 2/30 bare.
Querying curated knowledgebases such as Virtual Fly Brain (VFB) beyond browsing requires knowing their data models or coding against APIs, a barrier worsened by Drosophila connectomes with hundreds of thousands of neurons. The authors present, to their knowledge, the first Model Context Protocol (MCP) server for neuroscience and connectomics, letting large language models (LLMs) run graph queries at runtime instead of receiving pre-retrieved text, and benchmark it, which many biological MCP servers lack.
Key results
ShownWith VFB-MCP, Claude Opus 4.7 gave precisely correct, quantified and verifiable answers (score 2) on more tasks than the same model with web search or with no tools.
Score-2 answers: 25/30 MCP vs 14/30 web vs 2/30 bare. Friedman χ²(2) = 36.10, p = 1.4 × 10^-8. Holm-corrected pairwise Wilcoxon p-values: 7 × 10^-6 for MCP over bare, 1.57 × 10^-3 for web over bare, 3.00 × 10^-3 for MCP over web (30 tasks). MCP beat bare on 25 tasks and tied on 5; no task favoured bare. · Abstract; LLM + VFB-MCP produced the highest proportion of…; QUANTIFICATION AND STATISTICAL ANALYSIS
SuggestedThe gap between MCP and web grew as tasks demanded more quantification and integration: the two arms tied on simple-knowledge and survey tasks, but MCP led on multi-hop and graph-traversal/quantitative tasks.
Precise answers by category (MCP / web / bare): simple knowledge 6/8, 6/8, 0/8; integrative survey 5/5, 5/5, 1/5; multi-hop 6/8, 2/8, 1/8; graph traversal/quantitative 8/9, 1/9, 0/9. Pre-specified MCP vs web test on Category 4: Wilcoxon W = 5.0, p = 0.039. · The MCP advantage is greatest for multi-hop and…; QUANTIFICATION AND STATISTICAL ANALYSIS; Figure legends; Abstract
SuggestedNo arm produced a mostly fabricated answer, and partial-fabrication rates were similar. Error types differed: MCP answers kept VFB-derived facts correct and went wrong only when adding unsupported details or picking the wrong ontology class level.
No answer scored -1. Tasks with any partial fabrication: bare 6/30, web 9/30, MCP 6/30 (all pairwise McNemar p > 0.5). Bare errors were citations that did not support the claim, citations with wrong metadata, or facts attributed to the wrong entity. Web errors mismatched real papers and URLs, misread papers or bound data to the wrong entity. MCP errors added recalled details onto grounded results. · Entirely or majority fabricated answers were not…; Failure-modes vary by condition; Entity binding is a core advantage of structured…; Figure legends
ShownThree generic tools plus per-entity query introspection were enough for a cross-modal question. The model went from Kenyon cell classes to linked single-cell clusters to expression tables, then compared dopamine receptor expression between γ and α/β Kenyon cells.
Task T4.9 took five logical steps (search_terms, then get_term_info and run_query twice). MCP named Dop1R1, Dop1R2, Dop2R and DopEcR with per-cluster expression level and extent. Web named them without cluster-level numbers; bare declined to rank them. The actual trace had 43 tool rounds, nine of them VFB-MCP calls. · Example workflow: cross-modal query in five MCP…; Architecture: a small generic MCP surface with…
ShownMCP access cost more and ran longer than the bare model and, by a smaller margin, than web search. Large connectivity payloads also added file-handling overhead in Claude Code.
One 30-task sweep cost about $1.36 (~14 min) bare, $11.80 (~43 min) web and $16.30 (~61 min) MCP: MCP/bare ≈ 12x, MCP/web ≈ 1.4x, under a $3 per-prompt cap. Single run_query responses reached 1.2 MB and 40,000+ rows, adding 10-20 file-handling rounds on heavy connectivity questions. · Observed cost and runtime; Minimising cost; Oversized MCP responses and CLI spillover
Methods, data and reproducibility
A TypeScript server exposes VFB's Neo4j knowledge graph (built on the Drosophila Anatomy Ontology) through three generic tools: search_terms, get_term_info (which lists the query types valid for each entity) and run_query. Claude Opus 4.7 answered 30 tasks in four categories with no tools, VFB-MCP only, or web search/fetch only; one VFB curator scored the answers from -1 to 2.
- 30-task Drosophila neuroscience benchmark (8 simple knowledge, 5 integrative/survey, 8 multi-hop, 9 graph traversal/quantitative) with mark schemes and expert scoring sheets · Drosophila · n = 30 tasks x 3 conditions, 1 run each · 10.5281/zenodo.21143107
- Virtual Fly Brain knowledgebase: Neo4j graph combining connectome, single-cell transcriptomic and transgenic driver data under the Drosophila Anatomy Ontology · Drosophila
- Drosophila Anatomy Ontology (fbbt.owl) · Drosophila
- Methods & tools
- VFB-MCP v1.8.1, VFB-Query v1.9.0, Claude Opus 4.7, Claude Code CLI, WebSearch / WebFetch (Claude Code), Neo4j, Solr, MCP SDK v1.26.0, Express.js, Docker, Drosophila Anatomy Ontology, scipy
- Code
- https://github.com/VirtualFlyBrain/VFB3-MCPhttps://github.com/VirtualFlyBrain/vfb-mcp-paper-supplementhttps://vfb3-mcp.virtualflybrain.org/
- Data & accessions
- https://doi.org/10.5281/zenodo.21143107https://github.com/VirtualFlyBrain/vfb-mcp-paper-supplementhttp://purl.obolibrary.org/obo/fbbt.owlhttps://virtualflybrain.org/
- Software & environment
- Server written in TypeScript v5.9.3 with MCP SDK v1.26.0 and Express.js, shipped as a Docker container (Node.js 18 Alpine). Benchmarks used server v1.8.1 and VFB-Query v1.9.0, run through the Claude Code CLI (claude -p) with a $3 budget cap and a 600 s timeout.
- Experimental conditions
- Runs billed through the API. One 30-task sweep on Opus 4.7 cost about $1.36 (~14 min) bare, $16.30 (~61 min) with VFB-MCP and $11.80 (~43 min) with web.
- Reproducibility gaps
- One run per task, so run-to-run variation is not captured. Ground truth is VFB's database state at the time of scoring, which may change. Reproducing the results needs access to Claude Opus 4.7 through Claude Code.
Limitations (6)
- Only Claude Opus 4.7 was evaluated, with one run per task, so other models and the run-to-run randomness of LLM output were not measured. · stated by the authors
- The scorer was not blinded to condition, because VFB query citations and web URLs reveal which arm produced an answer. · stated by the authors
- The 30 tasks do not cover the full range of neuroscience questions. Bare and web performance also depend on whether published papers cover the cell class in question. · stated by the authors
- Broad run_query connectivity calls can return payloads above 1 MB, which the client harness must handle. MCP answers still layered unsupported training-data details over grounded results; the planned PubMed MCP addition is untested. · stated by the authors
- The VFB team designed the battery, and one VFB curator (an author) scored it with VFB's own database as primary ground truth. Where the literature and the database disagree, this rule gives MCP answers a 2 and literature-based answers a 1. · noted in this summary
- Each category comparison rests on 5 to 9 tasks, and only the Category 4 MCP-vs-web contrast was formally tested (p = 0.039). The other category differences are descriptive. · noted in this summary
Who should read:Read this if you build or evaluate LLM access to curated biological knowledge graphs, or want to query Drosophila connectome, ontology and single-cell data in natural language through Virtual Fly Brain.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- NeuroNLP: a natural language portal for aggregated fruit fly brain data
Nikul H. Ukani, Adam Tomkins, … Aurel A. Lazar·bioRxiv·Dec 14, 2016
- FlyBrainLab: Accelerating the Discovery of the Functional Logic of the Fruit Fly Brain in the Connectomic/Synaptomic Era
Aurel A. Lazar, Tingkai Liu, … Yiyin Zhou·eLifePeer-reviewed·Feb 22, 2021
- The Fruit Fly Brain Observatory: from structure to function
Nikul H. Ukani, Chung-Heng Yeh, … Aurel A. Lazar·bioRxiv·Dec 14, 2016
- A Programmable Ontology Encompassing the Functional Logic of the Drosophila Brain
Aurel A. Lazar, Mehmet Kerem Türkcan, Yiyin Zhou·Frontiers in NeuroinformaticsPeer-reviewed·Jun 20, 2022
- Tools for comprehensive reconstruction and analysis of a female Drosophila ventral nerve cord
Anthony W. Azevedo, Ellen Lesser, … John C Tuthill·NaturePeer-reviewed·Jun 26, 2024
- NeuroGraphBench: Interacting with Drosophila Connectomes at Scale for Exploring the Functional Logic of Neural Circuits
Aurel A. Lazar, Shashwat Shukla, Yiyin Zhou·bioRxiv·Oct 11, 2026