Neurosynth Compose MCP Server for agentic review

[adelavega]

Proposal: Neuroimaging review-to-meta-analysis agent

(drafted using GPT)

Goal

Build an MCP-backed agent that turns a research question into a reproducible, executable neuroimaging meta-analysis:

1. Define a versioned review protocol.
2. Search NeuroStore and external literature sources.
3. Screen publications for relevance and eligibility.
4. Select eligible Analyses, not merely papers.
5. Create an immutable StudySet.
6. Generate and validate a NiMARE/Compose meta-analysis specification.
7. Execute it locally or in the cloud and return provenance-linked results.

flowchart LR
    Q[Research question] --> P[Review protocol]
    P --> S[Study screening]
    S --> A[Analysis selection]
    A --> SS[StudySet]
    SS --> M[MA specification]
    M --> R[Execution and results]

Loading

Key architectural decisions

• NeuroStore is the authoritative state store. The chat coordinates the workflow, but protocols, decisions, StudySets, specifications, jobs, and results are persisted and versioned server-side.
• Separate Study and Analysis screening. An eligible publication does not imply that every reported contrast or coordinate set is eligible.
• Use study_schema as the evidence layer. Screening should use structured fields, extraction provenance, evidence spans, missingness, and confidence—not untraceable generated summaries.
• Use external sources through MCP adapters. PubMed expands bibliographic recall, PMC supplies permitted full text, and Semantic Scholar supports citation/reference expansion.
• GPT is the interactive reviewer and planner. It applies review criteria, explains decisions, resolves straightforward cases, and escalates ambiguity. It is not the database or final scientific authority.
• Generate typed executable specifications. GPT may propose algorithms and parameters, but the server validates them against supported NiMARE/Compose schemas before execution.
• Make every artifact reproducible. StudySets and executions should pin exact Study/Analysis versions, coordinate hashes, protocol versions, extraction versions, software environments, parameters, and random seeds.

Critical scientific distinction

Included Studies !- Included Analyses

Analysis selection must consider the structured relationship between Analysis, Effect, Cell, ModelTerm, FactorLevel, groups, tasks, inference settings, coordinate space, and coordinate tables.

Coordinates alone are insufficient to determine whether a contrast is appropriate.

Recommended product modes

1. NeuroStore synthesis

   • Fast and highly structured.
   • Limited to NeuroStore coverage.
   • Appropriate for exploratory or corpus-bounded analyses.

2. Extended systematic review

   • NeuroStore + PubMed + Semantic Scholar expansion.
   • PMC or user-authorized full text for eligibility and extraction.
   • Complete search provenance and PRISMA-compatible decisions.
   • Appropriate for comprehensive reviews.

Detailed review workflow

1. Convert the research question into a structured, versioned protocol.
2. Search NeuroStore and PubMed independently.
3. Deduplicate by DOI, PMID, PMCID, Semantic Scholar ID, and bibliographic similarity.
4. Expand promising candidates through citation/reference traversal.
5. Screen titles, abstracts, and compact schema-based “screening cards.”
6. Retrieve source evidence for uncertain or provisionally included papers.
7. Determine Study eligibility.
8. Extract or validate Study, model, Analysis, and coordinate information.
9. Screen individual Analyses using the frozen protocol.
10. Human-review ambiguous Analyses and low-confidence coordinate extraction.
11. Create an immutable StudySet snapshot.
12. Generate, validate, approve, and execute the meta-analysis specification.
13. Register outputs, logs, maps, environment information, and provenance.

Proposed MCP tool groups

Discovery

• search_literature
• expand_citation_graph
• get_study_screening_card
• get_study_evidence
• get_study_analyses
• get_analysis_evidence

Review state

• create_review_protocol
• amend_review_protocol
• record_study_decisions
• record_analysis_decisions
• get_review_progress
• get_disagreements

StudySet

• preview_studyset
• validate_studyset
• create_studyset_snapshot
• create_analysis_annotation

Meta-analysis

• list_supported_estimators
• draft_ma_specification
• validate_ma_specification
• estimate_execution
• submit_ma
• get_ma_status
• register_ma_results

Write operations and execution should require explicit user confirmation.

Proposed skills

• neuroimaging-systematic-review

  • Protocol formulation, searching, deduplication, screening, eligibility, and PRISMA accounting.

• neuroimaging-analysis-selection

  • Effect interpretation, Analysis selection, coordinate validation, and ambiguity handling.

• neuroimaging-meta-analysis

  • StudySet materialization, algorithm/corrector selection, specification generation, execution, and result interpretation.

The MCP provides authenticated data and controlled actions. Skills provide the reusable scientific workflow around those tools.

Screening and validation safeguards

Use four screening outcomes:

• include
• exclude
• uncertain
• duplicate

Every exclusion should have a standardized reason and supporting evidence. Missing information should produce uncertain, not automatic exclusion.

Flag Analyses for human review when there are:

• Potentially merged or duplicated contrasts
• Missing or inconsistent coordinate spaces
• ROI/whole-brain ambiguity
• Unclear contrast direction
• Missing model/effect relationships
• Potentially duplicated participant samples
• Low-confidence coordinate-table extraction
• Incompatible statistic or estimator inputs

Validate Study screening, Analysis selection, and coordinate correctness separately. Downstream stability of the resulting meta-analytic maps should also be evaluated.

External source and licensing considerations

• Use PubMed for broad biomedical bibliographic discovery.
• Use PMC’s approved APIs for permitted automated full-text retrieval.
• Use Semantic Scholar primarily for metadata, citations, references, and related-paper discovery.
• Do not assume that a PMC record, Semantic Scholar PDF link, or locally held paper is licensed for model processing or redistribution.
• Store license and retrieval metadata alongside every imported document.
• Enforce document permissions server-side; do not rely on skill instructions.
• Return structured evidence or minimal necessary excerpts rather than complete restricted papers.

Suggested implementation sequence

1. MVP: NeuroStore search → Study screening → Analysis screening → StudySet.
2. Add versioned review protocols, evidence-backed decisions, and resumable screening.
3. Add PubMed discovery and deduplication.
4. Add Semantic Scholar citation expansion.
5. Add permitted PMC full-text retrieval and external-paper ingestion.
6. Add typed meta-analysis specification generation and server validation.
7. Add cloud/local execution and versioned result registration.
8. Validate against published meta-analyses and expert screening decisions.

Product positioning

This should be positioned as an evidence-to-executable-meta-analysis system, not merely an AI paper screener.

The primary advantages are:

• Neuroimaging-specific structured evidence
• Analysis-level selection
• Traceable screening decisions
• Versioned StudySets
• Executable NiMARE specifications
• End-to-end provenance and reproducibility

[adelavega]

Counterpoint: start with Autonima + the NeuroStore evidence layer

A more attainable near-term approach is to keep Autonima as the execution engine and add NeuroStore as a reusable evidence source.

Before screening or extraction, Autonima would check NeuroStore:

• Reuse complete Study, Analysis, and coordinate records.
• Augment incomplete records.
• Screen and extract from scratch only when necessary.

This preserves Autonima’s advantages—batch execution, reproducibility, BYOK/local models, and external literature support—while addressing its largest inefficiency: repeatedly processing papers already represented in NeuroStore.

The immediate API savings from avoiding abstract screening are likely small. The greater value comes from reusing validated Analysis structures and coordinates, reducing extraction retries and producing more consistent meta-analysis inputs.

This hybrid should be implemented and benchmarked first. An MCP/ChatGPT interface can then be added over the same workflow once the underlying evidence-reuse path is proven.