Project 18: Mining the potential of knowledge graphs for metadata on training

Abstract

Knowledge graphs (KGs) can greatly increase the potential of data by revealing hidden relationships and turning it into useful information. A KG is a graph-based representation of data that stores relations between subjects, predicates and objects in triplestores. These entities are typically described in pre-defined ontologies, which increase interoperability and connect data that would otherwise remain isolated in siloed databases. This structured data representation can greatly facilitate complex querying and applications to deep learning approaches like generative AI.

ELIXIR and its Nodes are making a major effort to make the wealth of open training materials on the computational life sciences reusable, amongst others by guidelines and support for annotating training materials with standardized metadata. One major step in standardizing metadata is the use of the Bioschemas training profile, which became a standard for representing training metadata. Despite being standardized and interoperable, there is still a lot of potential to turn these resources into valuable information, linking training data across various databases.

In this project, we aim to create queryable KGs derived from training metadata in the Bioschemas format available from platforms like TeSS and glittr.org. In a subsequent step, we will investigate the potential of such KGs for several use cases, including construction of custom learning paths, creation of detailed trainer profiles, and connection of training metadata to other databases. These use-cases will also shed light on the limits on the currently available metadata, and will help to make future choices on richer metadata and standards.

Leads

Geert van Geest, Harshita Gupta, Vincent Emonet

💬 MCP server

A Model Context Protocol (MCP) server to access and search through the training materials of multiple Elixir repositories, such as TeSS and Glittr.

⚡️ Usage

Use with STDIO transport:

uv run elixir-training-mcp

Use with Deploy as Streamable HTTP server:

uv run elixir-training-mcp --http

🧰 Available MCP tools

Once the server is running you can call the following tools from your MCP-compatible client:

🔌 Connect client to MCP server

Follow the instructions of your favorite chat client.

To add a new MCP server to VSCode GitHub Copilot:

• Install the GitHub.copilot extension
• Open the Command Palette (ctrl+shift+p or cmd+shift+p)
• Search for MCP: Add Server...
  • Choose STDIO, and provide the command: uvx elixir-training-mcp
  • Or choose HTTP, and provide the MCP server URL, e.g. http://localhost:8000/mcp

To use it with STDIO transport, your VSCode mcp.json should look like:

{
   "servers": {
      "elixir-training-mcp": {
         "type": "stdio",
         "command": "uvx",
         "args": ["elixir-training-mcp"]
      }
   }
}

You can also connect to a running server using Streamable HTTP:

{
    "servers": {
        "elixir-training-mcp-http": {
            "url": "http://localhost:8000/mcp",
            "type": "http"
        }
    }
}

Harvesting

The data/tess_harvest.ttl files is included in the repository (7MB), you can run the script to harvest JSON-LD data and build this ttl file, but it takes ~30min due to parsing JSON-LD being expensive:

uv run src/elixir_training_mcp/harvest/harvest_tess.py

Harvest GTN:

uv run src/elixir_training_mcp/harvest/harvest_gtn.py

Deploy a SPARQL endpoint on http://localhost:8000:

uv run rdflib-endpoint serve src/elixir_training_mcp/data/*_harvest.ttl