📄 systems biology

ChatGEM: An Agentic Architecture Enabling Interactive Simulation of Genome-Scale Metabolic Models

This paper introduces ChatGEM, an agentic platform that leverages natural language and retrieval-augmented generation to democratize genome-scale metabolic modeling, enabling researchers without computational expertise to perform complex simulations and successfully identify optimal microbial strains for bioproduction.

Original authors: Chowdhury, N., George, A., Purohit, S., Contolesi, A., Bredeweg, E. L., Czajka, J., Stratton, K. G., Gao, Y., Stephenson, M., Elmore, J. R., Scott, A., Leach, D. T., Jerger, A., Lemmon, T., Piehowski

Published 2026-07-21

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Chowdhury, N., George, A., Purohit, S., Contolesi, A., Bredeweg, E. L., Czajka, J., Stratton, K. G., Gao, Y., Stephenson, M., Elmore, J. R., Scott, A., Leach, D. T., Jerger, A., Lemmon, T., Piehowski, P., Tate, K., Fulcher, J. M., Beliaev, A., Burnum-Johnson, K., Rigor, P., Bardhan, J.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine a world where the tiny factories inside our cells—bacteria, yeast, or even our own human cells—could be reprogrammed to produce life-saving medicines, clean fuels, or biodegradable plastics. Scientists have been trying to do this for years using "Genome-Scale Metabolic Models" (GEMs). Think of a GEM as a massive, incredibly detailed blueprint or a video game map of a cell's internal chemistry. It lists every single chemical reaction the cell can perform, like a recipe book with thousands of ingredients and steps.

However, there's a catch: reading and using this blueprint has traditionally required a PhD in computer science. You had to know how to write complex code (programming) to ask the blueprint questions like, "What happens if we remove this ingredient?" or "How can we make the cell produce more of this chemical?" This created a huge wall between the biologists who understand the cell and the computers that could simulate it. It's like having a Ferrari but only being able to drive it if you can speak fluent C++. The big question has always been: Can we build a bridge so that anyone, just by talking to a computer, can drive this Ferrari?

This paper introduces ChatGEM, a new tool that acts as a universal translator between human language and these complex cellular blueprints. Instead of writing lines of code, a researcher can simply type, "I want to see what happens if we block this chemical pathway," and ChatGEM does the heavy lifting.

The Magic of the Talking Robot

The authors built ChatGEM on a smart system called ADEPT, which acts like a team of specialized robots working together. Imagine you are a chef who wants to cook a complicated dish but doesn't know how to use the stove or chop the vegetables. You ask a smart assistant, "Make me a lasagna." The assistant doesn't just guess; it has a library of verified recipes (this is the RAG or "Retrieval-Augmented Generation" part). It looks up the exact steps, grabs the right tools, and even checks if you have the right ingredients before it starts.

In the world of ChatGEM, the "chef" is the scientist, the "lasagna" is a complex biological simulation, and the "verified recipes" are thousands of lines of correct code that the system has learned from. The system uses a team of agents: one to understand your question, one to find the right code recipe, one to write the code, and one to run it safely in a digital sandbox.

The Big Test: From Simple to Super Hard

To see if this talking robot actually works, the researchers put it through three increasingly difficult tests using a bacterium called Pseudomonas putida (a tough little bug often used in biotechnology).

The Warm-up (Simple): They asked the system to simulate a basic chemical balance and turn off one specific reaction. Without the "recipe library" (RAG), the robot made a few clumsy mistakes, like using the wrong tool for the job. But with the library, it got it perfect, writing clean, professional code that followed all the rules.
The Intermediate Challenge: They asked it to design a new strain of bacteria to produce a specific chemical (a task called "OptKnock"). Without the library, the robot picked the wrong software tool entirely, like trying to fix a car engine with a hammer. With the library, it chose the right tool, set up the complex math correctly, and gave a clear answer.
The Boss Level (Hard): They asked it to combine enzyme limits (how much protein the cell can make) with thermodynamics (the laws of physics that stop reactions from happening). This is the "super hard" level. Without the library, the robot's code crashed immediately because the math was wrong. With the library, it successfully built a complex simulation that ran perfectly, producing detailed reports on how the cell would behave.

The results were striking. When the system used its "recipe library," its performance score jumped from a mediocre 2.63 to a strong 4.20 (on a 5-point scale). Even more impressive was the speed. A task that took a human expert over 254 seconds to code and debug took the robot just 3.9 seconds to do alone. For the hardest tasks, the robot was up to 65 times faster than a human doing it manually, saving hours of "debugging" (fixing errors) time.

The Real-World Proof: Making Succinate

To prove this wasn't just a computer game, the researchers used ChatGEM to solve a real biological problem: figuring out which version of the Pseudomonas bacteria was best at making succinate (a chemical used to make plastics and other products).

They had four different versions of the bacteria, each tweaked in a slightly different way. They fed the system data about the proteins inside these bacteria and asked, "Which one is the best factory for succinate?"

ChatGEM analyzed the data and predicted that the "Constitutive" strain (a version where a specific gene is always "on") would be the winner. It figured this out by calculating a "leakage index"—a fancy way of saying, "How much of the cell's energy is wasted on eating the succinate instead of making it?" The robot predicted the Constitutive strain wasted the least energy.

Then, the scientists went into the lab and actually grew these bacteria. The experiment matched the robot's prediction perfectly. The Constitutive strain really did produce the most succinate. This confirmed that ChatGEM didn't just write code; it understood the biology well enough to make a correct, real-world prediction without a human ever writing a single line of code.

Why This Matters

The paper suggests that ChatGEM is a game-changer because it democratizes science. It suggests that you no longer need to be a coding wizard to use these powerful simulation tools. By turning complex math and biology into a simple conversation, it allows biologists, engineers, and curious minds to explore "what-if" scenarios instantly.

The authors are careful to note that this is a simulation tool that has been tested and validated in specific cases, not a magic wand that solves every problem in biology. However, by showing that a conversational AI can handle everything from simple checks to complex, physics-based modeling—and do it faster and more accurately than humans can alone—they have opened a door. They suggest that the future of metabolic engineering might not be about who can write the best code, but about who can ask the best questions.

Technical Summary: ChatGEM – An Agentic Architecture for Interactive Genome-Scale Metabolic Model Simulation

Problem Statement

Genome-scale metabolic models (GEMs) are foundational tools in systems biology for predicting cellular phenotypes and guiding strain engineering. However, their widespread adoption is hindered by a significant barrier to entry: the requirement for specialized computational expertise. Executing standard workflows—such as Flux Balance Analysis (FBA), strain design (e.g., OptKnock), or reconstructing enzyme-constrained GEMs (ecGEMs)—demands proficiency in the COBRA framework (Python/MATLAB), familiarity with biochemical databases (KEGG, ModelSEED), and the ability to debug complex code. This expertise gap restricts GEM-based analysis primarily to computational biologists, creating a disconnect between the analytical potential of these models and their utility for experimentalists and metabolic engineers. While Large Language Models (LLMs) have shown promise in code generation, existing tools lack a conversational interface capable of handling the full stack of quantitative GEM simulations across diverse modeling tasks.

Methodology

The authors present ChatGEM, an agentic platform designed to enable interactive GEM simulation through natural language. The system is built upon the ADEPT (Agentic Development Environment for Precision Tools) multi-agent framework, which provides a secure, scalable, and model-agnostic architecture.

System Architecture

Core Framework: ChatGEM integrates COBRApy, the StrainDesign toolbox, and commercial solvers (Gurobi) into the ADEPT ecosystem.
Multi-Agent Orchestration: The system utilizes specialized agents to coordinate code generation, execution, and interpretation. It employs a Retrieval-Augmented Generation (RAG) architecture to ground the LLM in domain-specific knowledge.
Knowledge Base: A curated repository of 52 human-verified Python scripts, ranging from simple FBA to complex ecOptMDFPathway workflows, is ingested into a vector database (ChromaDB) to support RAG.
Execution Environment: Code generated by the agents is executed in a secure, sandboxed container environment. The system supports observability via Langfuse/LangSmith and utilizes a "LLM-as-a-judge" (Claude Sonnet 4.6) for automated performance scoring.
Model Selection: The default reasoning model is o4-mini, chosen for its balance of reasoning capability and inference cost, though the system supports other providers (OpenAI, Anthropic, Ollama).

Evaluation Protocol

ChatGEM was benchmarked across three tasks of increasing biological and computational complexity using the Pseudomonas putida KT2440 GEM (iJN1463):

FBA with Reaction Knockouts: Routine flux analysis with specific reaction deletions.
OptKnock Strain Design: Formulating a bilevel optimization problem to identify gene knockouts that couple growth with target metabolite production.
ecOptMDFPathway: A fully integrated enzyme- and thermodynamic-constrained pathway optimization task.

Performance was evaluated under two conditions: without RAG (relying solely on the LLM's internal knowledge) and with RAG (utilizing the retrieved domain-specific scripts). Code quality was assessed by an LLM judge on two metrics:

Coding Accuracy (1–5): Correctness of biological outputs (fluxes, knockouts) and adherence to COBRApy best practices.
Code Completeness (1–5): Structural quality, API usage, modularity, and documentation.
Overall Performance Score (OPS): A weighted metric ( $0.7 \times \text{Accuracy} + 0.3 \times \text{Completeness}$ ).

Key Results

Performance and Efficiency

Impact of RAG: The inclusion of RAG significantly improved code generation quality. The mean OPS across all three tasks increased from 2.63 (without RAG) to 4.20 (with RAG).
- FBA Task: OPS improved from 3.60 to 4.30. Without RAG, the model often used generic string searches and non-standard API calls; with RAG, it correctly identified model identifiers and used native methods.
- OptKnock Task: OPS improved from 2.30 to 4.00. Without RAG, the model selected an incorrect library (Cameo) violating the protocol; with RAG, it correctly formulated the bilevel problem using the StrainDesign toolbox.
- ecOptMDFPathway Task: OPS improved from 2.00 to 4.30. Without RAG, the code contained critical logical errors in stoichiometric mass balance and thermodynamic constraints, rendering it non-executable. With RAG, it successfully generated a functional Mixed-Integer Linear Programming (MILP) formulation.
Model Comparison: The reasoning-optimized o4-mini outperformed the general-purpose GPT-4.1 in the absence of RAG (2.63 vs. 2.18 mean OPS). However, with RAG, both models converged to an identical mean OPS of 4.20, suggesting that domain-specific context retrieval can compensate for differences in raw reasoning capacity.
Time Efficiency: ChatGEM drastically reduced execution time. For the FBA task, time dropped from ~254 seconds (manual) to ~3.9 seconds (ChatGEM, no inspection), a 65-fold reduction. For complex tasks like OptKnock and ecOptMDFPathway, ChatGEM completed workflows in minutes (e.g., 774s and 41.8s respectively) compared to over a day for manual coding.

Biological Case Study: Succinate Overproduction in P. putida

ChatGEM was applied to reconstruct eight ecGEMs for four engineered P. putida KT2440 strains (Wild-type, Landing Pad, Constitutive dCas12a, Inducible dCas12a) at two time points (12h and 24h).

Model Validation: The ecGEMs showed strong agreement between predicted enzyme requirements and experimentally measured iBAQ protein abundances ( $r = 0.50$ ). Predicted growth rates were consistent with biological expectations.
Metabolic Insights: The analysis identified EDD, GLCDpp, and ACONTa as the dominant protein sinks in the central carbon pathway, consistent with the known dominance of the Entner-Doudoroff pathway in P. putida.
Strain Prioritization: A new metric, the Succinate Leakage Index (SLI), was defined to quantify the ratio of enzyme usage for succinate consumption versus production.
- The Constitutive strain exhibited the lowest SLI (0.12 at 12h, 0.13 at 24h), predicting it as the optimal chassis for succinate overproduction.
- Experimental validation confirmed this prediction: the Constitutive strain produced the highest succinate titers.
- The Induced strain showed an increasing SLI over time, correlating with a decline in succinate production observed in simulations and experiments.

Significance and Claims

The paper claims that ChatGEM represents a qualitative shift in the practice of metabolic engineering by democratizing access to GEM-based analysis. Its primary contributions are:

Lowering the Barrier to Entry: It enables researchers without computational expertise to perform sophisticated, multi-step GEM simulations (including ecGEM reconstruction and strain design) through natural language interaction.
Workflow Compression: It compresses workflows that previously required specialized coding and debugging into single conversational sessions, reducing time requirements by orders of magnitude.
Scientific Validity: The system successfully generated experimentally validated predictions (identifying the optimal chassis for succinate production) without the user writing any code, demonstrating that agentic architectures can handle complex, quantitative mechanistic simulations.
RAG as a Critical Enabler: The study highlights that for domain-specific scientific coding, Retrieval-Augmented Generation is essential for achieving high accuracy and correctness, effectively equalizing the performance of different LLM backbones when provided with verified domain knowledge.

The authors position ChatGEM not merely as an automation tool, but as an "executable interface for knowledge," amplifying the capabilities of experimental biologists by making the full stack of quantitative metabolic modeling accessible through conversation.