1000 Genomes Project Dataset MCP Server

Natural language access to 1000 Genomes Project dataset, hosted online in Dnaerys variant store

Sequenced & aligned by New York Genome Center (GRCh38). 3202 samples: 2504 unrelated samples from phase three panel + 698 samples from 602 family trios - dataset details

Key Features

• real-time access to 138 044 723 unique variants and ~442 billion individual genotypes

• variant, sample and genotype selection based on coordinates, annotations, zygosity

• filtering by VEP (impact, biotype, feature type, variant class, consequences), ClinVar Clinical Significance (202502), gnomADe + gnomADg 4.1, AlphaMissense Score & AlphaMissense Class annotations

  • full annotation composition

• returned variants annotated with HGVSp, gnomADe + gnomADg, AlphaMissense score + cohort-wide statistics

Online Service

Remote MCP service via Streamable HTTP:

• http://db.dnaerys.org:80/mcp
• https://db.dnaerys.org:443/mcp

Examples

Macromolecular structural complexes

The MCM2-7 Complex (The "DNA Helicase Motor") is a molecular masterpiece. It’s a heterohexameric ring where each subunit is a distinct "gear" in the DNA-unzipping motor. Unlike homomeric rings (where every subunit is the same), this complex is asymmetric. Each interface between subunits is unique, and they don't all burn ATP at the same rate. The MCM2/5 interface is the "gate" that must physically open to allow DNA to enter the ring and then snap shut. This is a high-stress mechanical point.

Identify individuals in the KGP cohort carrying missense variants at the MCM2/5 interface. Specifically, look for 'charge-reversal' variants (e.g., Aspartate to Lysine). In these specific samples, analyze the 'compensatory coupling': do they carry a secondary, reciprocal charge-reversal variant on the opposing subunit interface that restores the electrostatic 'latch' ?

Identify individuals in the KGP cohort who carry high-pathogenicity variants in the Walker A or Walker B motifs (the ATP-burning heart) of any MCM subunit in MCM2-7 Complex. For these individuals, perform a 'Systemic Flux' analysis: look at their variants in the leading-strand polymerase (POLE) and the sliding clamp (PCNA). Do you detect a signature of 'Coordinated Deceleration' where the motor, the clamp, and the polymerase all carry variants that suggest a slower but highly-accurate replication fork ?

Macromolecular structural complexes

The human RNA Exosome (Exo-9 core) is a "dead machine" that acts as a scaffold. In lower organisms the ring itself can degrade RNA. In humans, the 9-subunit ring has lost all its catalytic teeth and is purely a structural tunnel that guides RNA into the catalytic subunits (DIS3 or EXOSC10) attached at the bottom. Since RNA is a highly negatively charged polymer, the residues lining this pore are typically positively charged (Lysine, Arginine), but not too "sticky" or RNA will jam. So, to reach the "shredder" at the bottom it must slide through a narrow pore formed by the Exo-9 ring.

The task: analyse all missense variants in the KGP cohort that map to the internal pore-lining residues of the Exo-9 ring. Look for 'charge-swap' variants where a positive residue (K, R) is replaced by a negative one (D, E). If an individual is healthy despite having a 'negative patch' in the tunnel that should repel RNA, do they carry a compensatory variant in the cap subunits (EXOSC1, 2, 3) that widens the entrance? Use a 3D electrostatic surface map to determine if the 'healthy' cohort maintains a specific electrostatic gradient.

Synergistic Epistasis in Redox Homeostasis

Cellular redox homeostasis is maintained by two parallel antioxidant systems: the glutathione system and the thioredoxin system. Complete loss of either GSR or TXNRD1 is incompatible with mammalian development, yet population databases contain individuals carrying variants predicted to impair enzyme function.

Identify clusters of individuals in the KGP cohort who carry multiple 'Moderate' impact VEP variants across both systems. Reasoning through the AlphaMissense structural implications, can you detect a 'balancing act' where a loss of efficiency in Glutathione reductase is consistently paired with high-confidence benign or potentially activating variants in the Thioredoxin system ? Synthesize a model of 'Redox Robustness' based on the co-occurrence of these variants across the cohort.

More examples

---

Architecture

Implemented as a Java EE service, accessing KGP dataset via gRPC calls to public Dnaerys variant store service.

• provides MCP over Streamable HTTP, HTTP/SSE and STDIO transports
• service implementation is based on Quarkus MCP Server framework
• tools: computeAlphaMissenseAvg, computeVariantBurden, countSamples, countSamplesHomozygousReference, countVariants, countVariantsInSamples, getDatasetInfo, getKinshipDegree, selectSamples, selectSamplesHomozygousReference, selectVariants, selectVariantsInSamples
  • implementation

Installation

Project can be run locally with MCP over stdio and/or http transports

Option A - build & run locally

• build the project and package it as a single über-jar:
  • jar is located in target/onekgpd-mcp-runner.jar and includes all dependencies

./mvnw clean package -DskipTests -Dquarkus.package.jar.type=uber-jar

with skipping test compilation

./mvnw  clean package -Dmaven.test.skip=true -Dquarkus.package.jar.type=uber-jar

• run it locally with dev profile
  • both stdio and http transports are enabled
  • http transport is on quarkus http.port
  • project expects JRE 21 to be available at runtime

java -Dquarkus.profile=dev -jar <full path>/onekgpd-mcp-runner.jar

Option B - build & run in docker

• in order to run in docker, stdio transport needs to be disabled to prevent application from stopping itself due to closed stdio in containers

  • it's already configured in prod profile
  • it's the default configuration overall

• build with prod profile

docker build -f Dockerfile -t onekgpd-mcp .

• run as you prefer, e.g.

docker run -p 9000:9000 --name onekgpd-mcp --rm onekgpd-mcp

Option C - pull from Docker Hub

• pull prebuilt image; stdio transport disabled, http transport on port 9000

docker pull dnaerys/onekgpd-mcp:latest

• run

docker run -p 9000:9000 --name onekgpd-mcp --rm onekgpd-mcp

---

Connecting with MCP clients

• to connect via http transport, remote or local, simply direct the client to a destination, e.g. http://localhost:9000/mcp or https://db.dnaerys.org:443/mcp

  • NB: Claude Desktop won't work with http://localhost:9000/mcp option (e.g. when running MCP server in a docker container). This option is for clients like Goose.

• to connect via stdio transport, MCP client should start application with dev profile and with a full path to the jar file

  • e.g. for Claude Desktop add to config files (e.g. claude_desktop_config.json):

{
  "mcpServers": {
    "OneKGPd": {
      "command": "java",
      "args": ["-Dquarkus.profile=dev", "-jar", "/full/path/onekgpd-mcp-runner.jar"]
    }
  }
}

Verification

How many variants exist in 1000 Genome Project ?

Test Coverage Status

Component	Type	Tests	Status
Entity Mappers (9 classes)	Unit	314	✅ Complete
DnaerysClient	Unit	58 (7 disabled)	✅ Complete
DnaerysClient	Integration	5 (1 disabled)	✅ Complete
OneKGPdMCPServer	Unit	26	✅ Complete
OneKGPdMCPServer	Integration	5	✅ Complete
Other	Unit	1	✅ Complete
Other	Integration	1	✅ Complete
Total		410 tests	402 passing, 8 disabled

Test Breakdown:

• Unit tests: 399 (7 disabled, 392 passing)
• Integration tests: 11 (1 disabled, 10 passing)

Disabled Tests:

• 7 DnaerysClient unit tests (PaginationTests, streaming gRPC limitation - wiremock-grpc-extension:0.11.0 cannot mock streaming RPCs yet)
• 1 DnaerysClient integration test (PaginationLogicTests, streaming gRPC limitation - wiremock-grpc-extension:0.11.0 cannot mock streaming RPCs yet)

Running Tests

# Unit tests only (no server required)
./mvnw test

# Integration tests (requires db.dnaerys.org access)
./mvnw verify -DskipITs=false

# Update test baselines after data changes
./mvnw verify -DskipITs=false -DupdateBaseline=true

See TEST_SPECIFICATION.md for detailed test documentation.

---

Test part of this project is written by Claude. Fun part is written by humans.

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.