⚡ electrical engineering

Host-Aware Control of Gene Expression using Data-Enabled Predictive Control

This paper demonstrates that Data-enabled Predictive Control (DeePC), enhanced with basis functions to handle nonlinearities, achieves robust, sample-efficient, and high-performance real-time regulation of gene expression and host growth in bacteria using minimal data compared to traditional AI-based controllers.

Original authors: Liam Perreault, Idris Kempf, Kirill Sechkar, Jean-Baptiste Lugagne, Antonis Papachristodoulou

Published 2026-05-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Liam Perreault, Idris Kempf, Kirill Sechkar, Jean-Baptiste Lugagne, Antonis Papachristodoulou

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to conduct an orchestra where the musicians are living bacteria. Your goal is to get them to play two specific notes perfectly at the same time: one note is how much of a special "glow-in-the-dark" protein they produce (gene expression), and the other is how fast the whole band grows (cell growth rate).

The tricky part is that these bacteria are stubborn. If you push them to grow faster, they might stop glowing. If you make them glow too much, they might get tired and stop growing. It's a delicate balancing act.

This paper introduces a new "conductor" (a computer controller) that learns how to lead this bacterial orchestra without needing a full score or a long rehearsal. Here is how they did it, broken down into simple concepts:

1. The Old Way vs. The New Way

The Old Way (Model-Based): Traditionally, scientists tried to write a perfect mathematical textbook describing exactly how every bacterium works. They would then use this textbook to predict what to do. The problem? Biology is messy. If the bacteria change even a little bit, the textbook becomes wrong, and the conductor fails.
The AI Way (Deep Learning): Another approach is to let a computer "learn" by trial and error, like a video game character grinding for experience points. This works well but requires a massive amount of data (thousands of hours of practice) and needs to be re-trained from scratch for every new type of bacteria.
The New Way (DeePC): The authors used a method called Data-enabled Predictive Control (DeePC). Think of this as a conductor who doesn't need a textbook or thousands of hours of practice. Instead, they just listen to the last few minutes of the music the band played and use that to guess what to do next. It's "plug-and-play."

2. The Two Knobs and Two Meters

The system has two controls (inputs) and two things to measure (outputs):

The Controls:
1. Light: Shining specific colors of light on the bacteria to turn genes on or off (like a dimmer switch).
2. Food: Changing the nutrient density in their food (like adjusting the speed of the band's tempo).
The Meters:
1. Glow: How bright the bacteria are.
2. Growth: How fast the bacteria are multiplying.

3. The Secret Sauce: "Basis Functions"

Bacteria aren't simple machines; they react in curves, not straight lines. If you double the light, you don't necessarily double the glow. The paper adds a special "translator" called basis functions.

Imagine trying to draw a smooth curve using only straight Lego bricks. It looks jagged and ugly. But if you have special curved Lego bricks (the basis functions), you can build a smooth, perfect curve. The authors added these "curved bricks" to their controller so it could understand the bacteria's natural, non-linear behavior without needing a complex math model.

4. The "Cheat Code" for Speed (Model Reduction)

Usually, to predict the future of a complex system, you need a lot of data. The authors realized that while the bacteria have 18 different internal "states" (like 18 different instruments playing), only 5 of them actually matter for the big picture.

They used a trick called model reduction to ignore the 13 "noise" instruments and focus only on the 5 main ones. This is like tuning a radio to cut out the static and only hear the main melody. Because they focused on the important parts, they needed significantly less data to teach the controller how to work.

5. The Results: Who Won the Contest?

The authors tested their new "DeePC with Curved Bricks" controller against four other methods:

PI Control: A simple, old-school rule-following method. (It struggled to keep up with changing targets).
Deep MPC: A heavy AI method that needs a massive amount of training data. (It performed well but was data-hungry).
DDPG: A reinforcement learning method (like a video game AI). (It worked well but also needed a lot of practice).
SLMPC: A method that uses a perfect mathematical model (the "textbook" approach). (It performed the best, but only because it had perfect knowledge of the bacteria, which is rare in real life).

The Winner: The new DeePC method performed almost as well as the "perfect knowledge" method (SLMPC) and the advanced AI methods, but it used the least amount of data.

Why This Matters

The paper claims this is a "proof-of-concept." It shows that you can control complex living systems in real-time without needing to know every single detail of how the biology works or spending months collecting data.

The authors specifically mention that this setup is designed for future experiments using microscopes and microfluidic devices (tiny chips that hold cells) to control bacteria one by one. The goal is to study how individual cells behave differently from each other under changing conditions, which is crucial for understanding biology and engineering new biological tools.

In short: They built a smart, data-efficient conductor that can lead a bacterial orchestra to play the right notes, even if the musicians are unpredictable, using very little rehearsal time.

Technical Summary: Host-Aware Control of Gene Expression using Data-Enabled Predictive Control

Problem Statement
Cryogenetics aims to precisely regulate cellular processes, such as gene expression and growth rates, for applications in biomanufacturing and fundamental research. Current control strategies face a dichotomy: model-based controllers (e.g., MPC) rely on accurate linearized models which may lack validity in complex biological systems, while AI-based methods (e.g., Deep MPC, Reinforcement Learning) often require extensive training data and re-training for new systems. There is a need for a control strategy that offers high sample efficiency, robustness to parameter uncertainty, and the ability to handle nonlinearities without requiring extensive prior system identification or re-training for every new application.

Methodology
The authors propose a Data-Enabled Predictive Control (DeePC) framework applied to a two-input, two-output (2x2) biological system in E. coli.

System Model: The controlled system integrates a coarse-grained bacterial host model with a mechanistic model of the CcaS/CcaR optogenetic toggle switch. The inputs are nutrient concentration ( $u_s$ ) and the ratio of red/green light intensities ( $u_g$ ). The outputs are the host growth rate ( $y_\lambda$ ) and the expression level of a fluorescent reporter gene ( $y_g$ ). The system is inherently nonlinear, with 18 states, and exhibits coupling where growth rate affects gene expression via dilution and resource sharing.
Controller Design: The core algorithm is linear DeePC, which predicts future system behavior directly from measured input-output data using Hankel matrices, bypassing the need for an explicit parametric model. To address the system's nonlinearities (specifically sigmoidal responses to light and nutrients), the authors introduce basis functions ( $\phi_1, \phi_2$ ) that transform the inputs before they enter the DeePC optimization problem.
Data Efficiency: To minimize the data required for controller initialization, the authors employ balanced model reduction. By analyzing the observability and controllability Gramians of the linearized system, they demonstrate that 99.4% of the system's dynamics can be captured by reducing the state dimension from 18 to 5. This reduction significantly lowers the minimum number of samples required for the persistence of excitation condition.
Implementation: The controller operates in a receding horizon manner. To handle arbitrary reference signals and maintain the rank of the Hankel matrix, new input-output data is appended at each time step. The cost function penalizes input differences rather than absolute inputs to introduce integral action.

Key Contributions

Host-Aware Control Strategy: The paper presents a "plug-and-play" controller that simultaneously regulates synthetic gene expression and host growth by leveraging nutrient density, avoiding the need for additional synthetic circuits or external dilution methods used in prior work.
Nonlinearity Handling via Basis Functions: The integration of basis functions into the DeePC framework allows the controller to manage sigmoidal nonlinearities common in biological systems while maintaining the computational simplicity of a linear formulation.
Sample Efficiency: Through model reduction, the authors demonstrate that the controller can be implemented with significantly fewer data points (minimum $T \ge 89$ samples) compared to using the full-order model ( $T \ge 167$ ), making it suitable for time-constrained in vivo experiments.
Comprehensive Benchmarking: The study provides a rigorous comparison of DeePC against four other control paradigms: Model-Free (PI), Model-Based (Successive Linearisation MPC or SLMPC), and AI-based (Deep MPC and Deep Deterministic Policy Gradient/DDPG).

Results

Performance: In simulations with sinusoidal and step reference signals, the DeePC controller with basis functions (DeePC/BF) tracks references with performance comparable to SLMPC (which assumes full state feedback and perfect model knowledge) and significantly outperforms PI and Deep MPC.
Robustness: The controller demonstrates robustness to measurement noise, time delays in the optogenetic response, and parameter uncertainty in the basis functions. While performance degrades with increased noise or delay, the controller maintains stability and reaches steady states.
Sample Efficiency: DeePC/BF is shown to be significantly more sample-efficient than Deep MPC and DDPG. While Deep MPC and DDPG required hundreds of thousands of training samples to plateau in performance, DeePC/BF achieved comparable tracking performance with fewer than 200 samples.
Limitations of Comparators: Deep MPC struggled due to the binarization of inputs required by its implementation, and PI control failed to accurately track time-varying references. DDPG performed well on step responses but required extensive training episodes.

Significance and Claims
The paper positions this work as a proof-of-concept for future in vivo experiments on microscopy platforms capable of single-cell optogenetic control. The authors claim that their approach offers a viable path toward real-time, single-cell-level regulation in engineering biology without the labor-intensive training and re-training associated with AI-based methods. By minimizing the data acquisition period through model reduction and basis functions, the method addresses a critical bottleneck for experiments where duration is limited (e.g., cell viability). The study concludes that DeePC, augmented with basis functions and model reduction, provides a robust, data-efficient alternative for host-aware cybergenetic control, offering guidelines for optimal hardware and experimental design in future biological applications.