A bioreactor-based architecture for in vivo model-based and sim-to-real learning control of microbial consortium composition

TL;DR

The paper tackles the challenge of stably regulating density and composition in two-strain microbial consortia without genetic modifications. It introduces a bioreactor-based two-chamber architecture (mixing chamber and reservoir) with three tunable flow-rate inputs and both model-based and sim-to-real learning controllers, validated in vivo on an Escherichia coli system. The results show precise, robust regulation of total biomass and strain composition, successful tracking of time-varying references, and resilience to perturbations, achieved with minimal experimental calibration data. While limitations arise from aggregate sensing and hardware backflow, the work demonstrates a flexible framework that can be extended to more strains and adaptive strategies, potentially enabling scalable, non-genetic bioproduction workflows.

Abstract

Microbial consortia offer significant biotechnological advantages over monocultures for bioproduction. However, industrial deployment is hampered by the lack of scalable architectures to ensure stable coexistence between populations. Existing strategies rely on genetic modifications, which impose metabolic load, or environmental changes, which can reduce production. We present a versatile control architecture to regulate density and composition of a two-strain consortium without genetic engineering or drastic environmental changes. Our bioreactor-based control architecture comprises a mixing chamber where both strains are co-cultured and a reservoir sustaining the slower-growing strain. For both chambers we develop model-based and sim-to-real learning controllers. The control architecture is then validated in vivo on a two-strain Escherichia coli consortium, achieving precise and robust regulation of consortium density and composition, including tracking of time-varying references and recovery from perturbations.

Figures (9)

• Figure 1: a) A schematic representation of the control architecture, comprising a mixing chamber where strains 1 (in green) and 2 (in pink) are co-cultured, and a reservoir where the strain with a lower growth rate (i.e. strain 2) is grown independently. The mixing chamber and the reservoir are supplied with fresh growth medium at rates $D_1$ and $D_R$, respectively. Additionally, the reservoir feeds strain 2 to the mixing chamber at a rate $D_2$. Both reactors are connected to waste containers to ensure the volume is kept constant. b) Implementation of the control architecture with two interconnected Chi.Bio. reactors. The pumps that regulate the exchanges of fresh media and waste are highlighted with the teal and yellow boxes, while the tubes that interconnect the reservoir and the mixing chamber are highlighted with pink arrows.
• Figure 2: Block diagram of the closed-loop control architecture. The controller regulates the concentrations of the biomasses $x_1$ and $x_2$ in the mixing chamber to the desired levels $x_{1,\mathrm{d}}$ and $x_{2,\mathrm{d}}$, respectively. The current values of the states are reconstructed using an extended Kalman filter from the aggregate measure $x_1+x_2$. An independently designed controller regulates $x_2^R$ in the reservoir to the desired level $x_{2,\mathrm{d}}^R$.
• Figure 3: Sim-to-real pipeline. Step 1) System Identification: the parameters of the dynamical system are identified with an open-loop experiment; Step 2) In silico training: the mathematical model is employed to produce synthetic data for the training of the neural network. Step 3) In vivo deployment: The trained network is employed to regulate the cell population density within the bioreactor.
• Figure 4: Open loop experiments to identify the parameters $\mu_1^*$ and $\mu_2^*$. Time evolution $y_1$ when growing fast-growing strain (top) and slow-growing strain (bottom) monocultures in the mixing chamber, while imposing $D_2=0$. A qualitative comparison between the experimental data (in blue) and the model predictions (in grey) is shown. For each experiment the dilution rate $D_1$ administered during the experiment is shown.
• Figure 5: Experimental protocol to measure ex-post the relative numbers between the strains comprising the consortium in the mixing chamber. a) Every 10 minutes, samples are taken from the mixing chamber. b) Example measurements collected from the Chi.Bio. Top panel: time evolution of $y_1$. The red arrows indicate the first four time instants in which the samples are collected; bottom panel: dilution rates. c) Gating strategy used to distinguish the strains using a flow cytometer. After excluding debris and cell agglomerates using the FSC and SSC channels, cells were classified by their fluorochrome (FITC). Specifically, cells expressing high levels of green fluorescent protein (FL1-H $\geq 2\times 10^3$ on log scale), were classified as belonging to strain 2. The remaining cells were classified as belonging to strain 1. d) Example of $x_2/x_1$ ratio computed using the data classified using the flow cytometer.

Theorems & Definitions (2)

• Remark 1
• Remark 2