Skip to content

Applications

kgpBench

kgpBench is an open-source harness that evaluates the reasoning steps language-model agents take when they investigate variants in rare-disease genes.

Each kgpBench task poses an open research question about variants in the 1000 Genomes Project dataset, which the agent answers by querying 138 million variants and ~442 billion individual-level genotypes in real time through the OneKGPd MCP server. Language-model judges read the agent's full trajectory and score each reasoning step from a catalogue of checks, such as statistical power, population stratification, pedigree relatedness or shared haplotypes. Each step is scored on its own, and where a step admits several defensible choices, each anticipated choice is scored against its own expected value. The agent is told nothing about method or grading, so the benchmark measures what it does unprompted.

The harness enables contamination resistance that preserves fidelity. Rewriting the questions of a published benchmark has been found to raise its resistance to contamination only at a cost in fidelity (Sun et al., ICML 2025). kgpBench renews a retired question by asking it again about other data in the cohort: the renewed task calls for the same reasoning steps, judged against the same definitions, while its expected answers are computed anew from the cohort.

The harness is built on Inspect AI.

OneKGPd MCP Server

OneKGPd is a Model Context Protocol server that provides LLMs with natural-language access to the 1000 Genomes Project dataset hosted in Dnaerys. Implemented as a Java service on the Quarkus MCP Server framework, it exposes 14 MCP tools: variant, sample and genotype selection by coordinates, annotations, zygosity and population, kinship degree, AlphaMissense score averages and variant burden, along with population and sample metadata served from an embedded DuckDB instance. The server is available as a remote service over Streamable HTTP at https://db.dnaerys.org/mcp, and can run locally over stdio or HTTP transports.

IGV Genome Browser

IGV-Dnaerys is a web-based genome browser built on igv.js, the embeddable JavaScript implementation of the Integrative Genomics Viewer (IGV). A modified igv.js adds a gRPC-web track source, which fetches variants from Dnaerys in real time for the genomic region in view. Variants are rendered as features colour-coded by zygosity: homozygous, heterozygous or mixed across samples. The demo connects to the public 1000 Genomes Project dataset and displays a random subset of 10 samples in each session.

GA4GH Beacon v2

Beacon v2 is a GA4GH standard for discovery of genomic variants and associated biomedical data across single or federated resources. The National Computational Infrastructure implementation is a fork of the Barcelona Supercomputing Center's Jakarta EE Beacon v2 implementation, with Dnaerys serving all genomic variant queries over gRPC and MongoDB holding metadata for individuals, biosamples, analyses, runs, cohorts and datasets.