Multiplexing microbial biosensors through machine learning

Introduction:: A major goal of synthetic biology is to engineer microbes to sense and respond to environmental cues [1–4]. To date, at least 50 microbial biosensors have been constructed to sense conditions ranging from pH to arsenic concentration [5,6]. Successful implementations of biosensors for real-world applications have been limited to simple systems that detect one signal, with few groups achieving more complex multi-signal sensors [7–10]. As synthetic biology shifts toward addressing more complex problems, like diagnosing gut diseases or measuring environmental contaminant concentrations, complex biosensors that can measure multiple signals in parallel will be necessary. This has been difficult to achieve due to crosstalk between biosensors. The typical approach is to optimize each sensor component to be orthogonal, but this takes extensive effort and may not reliably work outside of the conditions it was optimized for [11,12]. My objective is to develop a machine learning pipeline that computationally disentangles crosstalk of non-orthogonal microbial sensors to determine chemical concentrations. By enabling new and existing gene circuit sensors to skip the orthogonality optimization step, this computational approach will streamline the translation of individual biosensors from bench testing to using in combination with other biosensors in medical and environmental applications.

Materials and Methods:: Our computational approach to synthetic biology integrates mechanistic modeling and machine learning to determine chemical concentrations from fluorescence measurements of microbial communities of engineered biosensors. Experimentally, co-cultures of 2-3 microbial sensors are inoculated with combinations of their respective input chemicals at different controlled concentrations. The cultures’ optical density (OD600) and fluorescence are measured by a plate reader for 24 hours. The crosstalk of each sensor in the communities is also characterized by measuring their monoculture response to the other chemicals. The co-culture timecourse data is used to fit a set of ordinary differential equations (ODEs) by optimizing parameters. Using the optimized parameters and equations, we simulate thousands of timecourses to augment the experimental data for training the machine learning model. We have tested two machine learning pipelines which are trained to use culture fluorescence measurements to predict the chemical input concentrations, where the prediction accuracy is assessed via R². Support vector regression (SVR) is trained and tested on both simulated and experimental data separately. We have also developed a more advanced pipeline consisting of a variational autoencoder (VAE) to encode the timecourse data into 10 latent dimensions, which are then used as inputs to a multi-layer perceptron (MLP) to predict the chemical concentrations. This model is trained on a combined dataset of simulated and experimental data.


Results, Conclusions, and Discussions:: The two machine learning approaches successfully predict concentrations of multiple inputs from microbial biosensor communities regardless of crosstalk. Simulations and SVR defined parameter limits for prediction accuracy. Among factors affecting accuracy, crosstalk had minimal impact, decreasing R² by 0.11 when crosstalk increased from 0% to 100% for two sensors with noise and no growth differences. Increasing the sensor count from 2 to 10 decreases R² by 0.18. Noise and extreme growth rate differences had the greatest impact on predictions. This noise constraint is not unique to our method as noise decreases accuracy of single sensor measurements too. R² values only dropped below 0.8 for growth rate differences exceeding 40%, which is experimentally avoidable. SVR and VAE-MLP also interpreted experimental sensor responses accurately. For two orthogonal sensor populations, SVR predicted input concentrations for training and validation data with R²=0.96 and R²=0.87 respectively. VAE-MLP reached an R² of 0.93 for this experiment’s chemical concentration predictions. For two non-orthogonal sensor populations, SVR achieved an R² of 0.62, possibly due to limited experimental data. SVR predicted binary inputs from three non-orthogonal sensors with R²=0.77, which is low due to one sensor's inconsistent response. These prediction accuracies can improve by tuning data pre-processing, hyperparameter selection, and the models’ architectures. Future experiments will involve collecting more data from biosensors from literature with varying crosstalk levels. The future directions include demonstrating the concept’s viability by testing multiple sensors together in practical applications, such as measuring gut inflammation biomarkers from fecal samples, where inducer concentrations are uncontrolled. Our computational approach represents a pioneering method for determining input concentrations from microbial biosensor mixtures. It overcomes the limitations of traditional techniques like fitting a hill function to an individual dose response curve, which cannot comprehensively capture crosstalk dynamics. Our novel approach to engineering multiplexed microbial biosensors using machine learning will accelerate biosensor development for use in medical and environmental monitoring applications.


Acknowledgements (Optional): :

References (Optional): : 1. Kotula, J. W. et al. Programmable bacteria detect and record an environmental signal in the mammalian gut. Proc. Natl. Acad. Sci. U. S. A. 111, 4838–4843 (2014).

2. Lee, S. K. & Keasling, J. D. A Propionate-Inducible Expression System for Enteric Bacteria. Appl. Environ. Microbiol. 71, 6856–6862 (2005).

3. Ding, N., Zhou, S. & Deng, Y. Transcription-Factor-based Biosensor Engineering for Applications in Synthetic Biology. ACS Synth. Biol. 10, 911–922 (2021).

4. Rogers, J. K. & Church, G. M. Genetically encoded sensors enable real-time observation of metabolite production. Proc. Natl. Acad. Sci. 113, 2388–2393 (2016).

5. Wan, X. et al. Cascaded amplifying circuits enable ultrasensitive cellular sensors for toxic metals. Nat. Chem. Biol. 15, 540–548 (2019).

6. Schmidl, S. R. et al. Rewiring bacterial two-component systems by modular DNA-binding domain swapping. Nat. Chem. Biol. 15, 690–698 (2019).

7. Daeffler, K. N.-M. et al. Engineering bacterial thiosulfate and tetrathionate sensors for detecting gut inflammation. Mol. Syst. Biol. 13, 923 (2017).

8. Cartwright, I. M. et al. Mucosal acidosis elicits a unique molecular signature in epithelia and intestinal tissue mediated by GPR31-induced CREB phosphorylation. Proc. Natl. Acad. Sci. 118, e2023871118 (2021).

9. Wang, B., Barahona, M. & Buck, M. A modular cell-based biosensor using engineered genetic logic circuits to detect and integrate multiple environmental signals. Biosens. Bioelectron. 40, 368–376 (2013).

10. Cheng, Y.-Y. et al. Programming bacteria for multiplexed DNA detection. Nat. Commun. 14, 2001 (2023).

11. Meyer, A. J., Segall-Shapiro, T. H., Glassey, E., Zhang, J. & Voigt, C. A. Escherichia coli “Marionette” strains with 12 highly optimized small-molecule sensors. Nat. Chem. Biol. 15, 196–204 (2019).

12. Jung, J. K. et al. Cell-free biosensors for rapid detection of water contaminants. Nat. Biotechnol. 38, 1451–1459 (2020).