Provide established Metabolic Systems Biology tooling for Reconstruction and Analysis with LLMs

[Schmoho]

Background

Reconstruction and subsequent network analysis are the fundamental processes of the systems biology of genome-scale metabolic models (GEMs). A plethora of tools exist in this space, and some of them emerged as the de facto standards (specifically COBRApy, CarveMe, MEMOTE, Cytoscape). The skillful and creative usage of these tools is a prerequisite for reliable knowledge extraction and fruitful hypothesis generation using GEMs.

GEMs are graph data structures that are well-suited for loading to and querying from a graph database. Consequently, tooling to load SBML into Neo4J (Neo4JSBML) has been developed.

The dominant open standards for LLM tool usage currently are the Model Context Protocol (MCP) and the SKILLS.md mechanism(an intro can be found here). These standards are designed as lightweight mechanisms to allow LLM agents to call external software tools in pursuit of a given goal. MCP server-based tools suffer from a context bloat problem that can be mitigated using dynamic tool detection. The most important software tools in systems biology currently have no integrations available to make them available to LLM agents. Neo4J has an official MCP server that allows an LLM to query and manipulate graph databases.
Access to API tokens for open-weight models hosted in compliance with the GDPR can be made available for the duration of the project.

GEMs, like all formal models, have large potentials as curated, focused organism databases, as well as hypothesis generating oracles. Unlike with LLMs, their "responses" are fundamentally deterministic and mechanistic, grounded in knowledge. Their accessibility and usability however is severely limited by the sheer complexity and number of tools and databases involved. Furthermore, the idiosyncratic nature of lines of inquiry into a model usually means each analysis is (at least partially) as unique a gem as the GEM it investigates.

Goal

In this project, LLMs are not supposed to replace systems biology modeling, but to lower the cost of doing it correctly.

1. MCP Servers and/or SKILLS.md documentation for at least these reconstruction and analysis tools are made available:

• CarveMe
• COBRApy
• MEMOTE
• refineGEMs
• Cytoscape

2. A portable project setup (docker-compose.yml in case of MCP, clonable repo in case of skills + Neo4J database and MCP container setup) is made available, to minimize friction for users. The ideal is to have a single "install button" that would require little more than an API token from the user.

3. This setup is used in conjunction with the OpenCode LLM agent framework that supports both MCP and SKILLS.md. A proof of concept reconstruction and analysis of a bacterial metabolic network are performed.

Difficulty Level: Medium

Medium difficulty. The implementation itself is easy, but there is a large number of concepts involved in understanding the different parts that have to play together.

MCP servers are fundamentally HTTP wrappers around given functionality and their implementation is usually easy. SKILLS.md fundamentally uses documentation of tools, which can to a large degree be extracted and adapted from existing documentation. Both processes lend themselves exceedingly well to LLM-assistance.

The greatest difficulty in the project lies in understanding and evaluating the LLM output and acquiring sufficient understanding of the involved tools and data structures.

Size and Length of Project

• medium: 175h
• 12 weeks

Skills

Essential skills:

• Programming proficiency (pref. Python)
• HTTP APIs (i.e. fundamental HTTP concepts (methods, routes), JSON, some experience programming a server, for example using Flask)

Nice to have skills:

• Docker
• Proficiency with at least one visualization library (pref. ggplot or Vega-Lite)
• OpenAPI/Swagger (for Cytoscape REST)

Public Repository

https://github.com/Schmoho/gsoc-sysbio-llm-tools

Potential Mentors

Prof. Dr. Andreas Dräger (andreas.draeger@informatik.uni-halle.de)
Dario Eltzner (dario.eltzner@informatik.uni-halle.de)

AI Usage Policy

We welcome the use of AI tools by users who are capable and have proven expertise in the field they are using AIs for. We care first and foremost about the characteristics of the software artifact. That is, code needs to do exactly what it is supposed to, not "be written by a human". It needs to be maintainable, not "be pretty".

We therefore explicitly encourage participants to use AI tools, as long as they are capable and prepared to responsibly review and integrate the generated code.

[atulbraj]

Hi @Schmoho, @draeger, @khanspers

I’m very interested in this GSoC project and have read through the full description and goals.
My background is primarily in Python-based tooling and API-driven systems, with experience working on data- and infrastructure-heavy projects. I’m particularly interested in the idea of using LLMs as orchestrators around deterministic, domain-specific tools rather than replacements for them, which strongly resonates with this project’s motivation.

I’d like to start by familiarizing myself with the repository and one of the listed tools (e.g., COBRApy or MEMOTE) and make a small initial contribution—most likely in the form of SKILLS.md-style documentation or a minimal prototype—to better understand the workflow, interfaces, and expectations.

Please let me know if there is a preferred starting point or any guidance on which tool or aspect of the project would be most useful to explore first.

[dyrpsf]

Hi @Schmoho and @draeger ,

I’m very interested in contributing to this project as part of GSoC. I have a strong background in Python and experience with APIs, JSON, and backend development. I’m also familiar with Docker and enjoy working on research-oriented and interdisciplinary projects.

I have previously contributed to SBML and JSBML by making multiple pull requests using Java, and I received positive feedback on my contributions. This experience has helped me understand biological modeling tools and the importance of writing clean, maintainable, and reliable code.

I’m eager to learn more about metabolic modeling tools like COBRApy and CarveMe and contribute to building MCP/Skills-based integrations for LLM agents. I’m comfortable setting up reproducible environments and writing well-documented, production-quality code.

I would love to discuss how I can get started and contribute meaningfully to this project. Looking forward to your guidance.

Thank you!

— Deepak

[kushvinth]

Hi @Schmoho, @draeger,

I'm very interested in this project and find the goal of making established systems biology tools accessible to LLM agents particularly compelling if, there is a open role here.

The approach of using LLMs as orchestrators rather than replacements for deterministic, mechanistic tools strongly resonates with my experience. I've worked on similar projects like ainqa, torax, and roadgpt where LLM integration with domain-specific tools was key.

Background:

I'm proficient in Python, HTTP APIs (Flask/FastAPI), Docker, and JSON. Last summer, I contributed to Google DeepMind's Torax and Apple's ML Stable Diffusion:

• Fix: Correct usage of R vs R_major in geometry and transport modules google-deepmind/torax#1704
• Add configurable deployment target support for iOS 18 quantisation compatibility apple/ml-stable-diffusion#391

I've built several production systems including:

• shazam: Music recognition using audio fingerprinting from scratch
• homelab: Self-hosted infrastructure running 20+ containerized services with 99.9% uptime (check out https://status.kushvinth.com)
• lic: Cross-platform CLI tool on PyPI & Homebrew with 30+ stars, used by Oracle devs
• dosai: iOS app with AI-powered food analysis using Swift and AVFoundation
• freedle: Wordle clone with glassmorphic UI at https://freedle.kushvinth.com

I have experience with visualization libraries (matplotlib, plotly) and creating reproducible Docker-based workflows.

Why This Project Excites Me:

The challenge of bridging complex bioinformatics tooling (COBRApy, CarveMe, MEMOTE, Cytoscape) with LLM agents via MCP and SKILLS.md is exactly the type of integration work I enjoy. I appreciate the emphasis on maintainable, correct code over "human-written" code having the right characteristics of the software artifact is what matters.

I'm particularly interested in:

• Creating MCP servers as HTTP wrappers around the reconstruction/analysis tools
• Developing SKILLS.md documentation extracted from existing tool docs
• Building the docker-compose setup for minimal friction deployment
• Working with Neo4J's MCP server for graph database integration

Next Steps:

I'd like to start by familiarising myself with the gsoc-sysbio-llm-tools repository and one of the core tools (COBRApy or MEMOTE). I'm happy to make an initial contribution either SKILLS.md documentation or a minimal MCP server prototype to demonstrate understanding and gather feedback.

Would you recommend starting with a specific tool or aspect of the project? I'm also curious about the preferred balance between MCP servers vs SKILLS.md for different tools.

Looking forward to discussing this further!

Best regards,
Kushvinth Madhavan

GitHub: https://github.com/kushvinth
LinkedIn: https://www.linkedin.com/in/kushvinth-madhavan-97352a322/