Context-Aware Biosensor Design Through Biology-Guided Machine Learning and Dynamical Modeling

Abstract

Addressing the challenge of achieving a global circular bioeconomy requires efficient and robust bio-based processes operating at different scales. These processes should also be competitive replacements for the production of chemicals currently obtained from fossil resources, as well as for the production of new-to-nature compounds. To that end, genetic circuits can be used to control cellular behavior and are instrumental in developing efficient cell factories. Whole-cell biosensors harbor circuits that can be based on allosteric transcription factors (TFs) to detect and elicit a response depending on the target molecule concentrations. By modifying regulatory elements and testing various genetic components, the responsive behavior of genetic biosensors can be finely tuned and engineered. While previous models have described and characterized the behavior of naringenin biosensors, additional data and resources are required to predict their dynamic response and performance in different contexts, such as under various gene expression regulatory elements, media, carbon sources, or media supplements. Tuning these conditions is pivotal in optimizing biosensor design for applications operating in varying conditions, such as fermentation processes. In this study, we assembled a library of FdeR biosensors, characterized their performance under different conditions, and developed a mechanistic model to describe their dynamic behavior under reference conditions, which guided a machine learning-based predictive model that accounts for context-dependent dynamic parameters. Such a Design-Build-Test-Learn (DBTL) pipeline allowed us to determine optimal condition combinations for the desired biosensor specifications, both for automated screening and dynamic regulation. The findings of this work contribute to a deeper understanding of whole-cell biosensors and their potential for precise measurement, screening, and dynamic regulation of engineered production pathways for valuable molecules.

Keywords: biosensor; context dependence; dynamical modeling; genetic circuit; scientific machine learning; synthetic biology.

1

Scheme of the cloning process and final biosensor circuit.

2

Normalized fluorescence profile of the 17 Level-2 biosensor constructs. The experiments used 400 μM naringenin induction, M9 as growth media, and glucose (0.4%) as carbon source. The line color determines the promoter used. The RBS used in each case is shown at the end of each fluorescence curve.

3

Media and supplements have a significant effect on the dynamics of the reference construct. (a) Normalized fluorescence over 16-h experiments, comparing media (color) and supplements (shape). (b) Normalized fluorescence heat map (media vs supplement) of the mean of the maximum values in each experiment.

4

Isolated effect of the four DoE variables on the normalized fluorescence of the naringenin biosensor. (a) Effect of promoter selection. (b) Effect of ribosome binding site (RBS) selection. (c) Effect of media selection. (d) Effect of supplement selection.

5

Heat map pairwise comparisons of the four factor combinations. The color intensity describes the normalized fluorescence value. In the case where a combination appears in more than one DoE experiment, the mean was calculated. White squares indicate combinations that do not appear in the 32 DoE experiments. (a) Promoter vs RBS. (b) Promoter vs media. (c) Promoter vs supplement. (d) RBS vs media. (e) RBS vs supplement. (f) Media vs supplement.

6

Fitting and prediction results obtained in the cross-validation for Experiment 46. (a) Experimental growth curves (OD) vs simulated and predicted response. (b) Experimental reporter curve (GFP/OD) vs simulated and predicted responses in the ensemble modeling obtained by bagging. (c) Comparison of RMSE performance for OD curves vs simulated and predicted responses. (d) Comparison of RMSE performance for GFP/OD vs simulated and predicted responses.