Specs MCP Server

An MCP server built with FastMCP that serves similarity search and analytics over a compound dataset (compound_aggregate_with_annotations.csv). The server exposes tools for compound-to-compound and MOA-to-MOA similarity, gene co-occurrence, disease rollups, and related lookups.

Repository layout

• main.py — MCP server definition and all registered tools.
• data/compound_aggregate_with_annotations.csv — PCA embeddings (150 dims) plus annotations per compound.
• requirements.txt — Python dependencies.
• Dockerfile — container recipe for running the server.

Prerequisites

• Python 3.10+ recommended.
• pip for dependency installation.

Remote connection

If you host this server remotely, point clients to the deployed base URL:

• Remote MCP endpoint: https://specs-mcp.serve.scilifelab.se/mcp

Running the MCP server locally with Docker

Build and run with Docker:

docker build -t specs-mcp .
docker run --rm -p 8000:8000 specs-mcp

After the server starts, point your MCP-compatible client at the URL it prints (e.g., http://localhost:8000).

Available tools

• query_compound_similarity — global embedding neighbors for a compound.
• compound_neighbors — scoped embedding neighbors by MOA/disease/shared_genes (no global mode).
• compound_metadata — core fields plus optional disease-matched neighbors.
• get_compound_details — detailed metadata for one compound.
• list_available_compounds — list compounds with optional substring filter.
• match_closest_compounds — fuzzy-match free text to compound names.
• compare_two_compounds — pairwise similarity and shared annotations.
• compound_targets — gene-overlap neighbors for a compound.
• gene_compounds — compounds targeting a specific gene.
• gene_context — co-genes, MOAs, diseases for a gene.
• search_by_target_gene — quick gene-to-compound lookup.
• query_moa_similarity — intra-MOA stats and nearest compounds outside the MOA.
• list_available_moas — list MOAs with optional filters.
• match_closest_moa — fuzzy-match free text to MOA strings.
• list_compounds_by_moa — compounds within a MOA.
• moa_summary — MOA stats plus optional compound rows.
• moa_relations — nearest MOA centroids to a given MOA.
• phase_summary_for_moa — clinical phase distribution within an MOA.
• disease_summary_for_moa — disease distribution within an MOA.
• moa_summary_for_phase — MOA distribution within a clinical phase.
• top_viability_in_group — viability ranking within an MOA or phase.
• disease_summary — disease-level stats plus optional compounds.
• disease_relations_by_phase — top compounds in a phase and disease mix.
• moa_summary_for_disease — MOA distribution within a disease area.
• genes_for_group — top genes within an MOA or disease group.
• compound_neighbors_shared_genes — see compound_neighbors (scope shared_genes).
• compound_neighbors_disease — see compound_neighbors (scope disease).
• compound_neighbors_moa — see compound_neighbors (scope moa).
• top_viability_in_group_moa — see top_viability_in_group (group_type moa).
• list_available_compounds_filtered — see list_available_compounds with filter.
• list_available_moas_filtered — see list_available_moas with filter.
• get_dataset_statistics — dataset summary stats.
• reload_dataset — reload the CSV into memory after updates.

All tools operate on the CSV loaded at startup; call reload_dataset after replacing compound_aggregate_with_annotations.csv to refresh the in-memory cache.