Deep model predictive control of gene expression in thousands of single cells.

Gene expression is inherently dynamic, due to complex regulation and stochastic biochemical events. However, the effects of these dynamics on cell phenotypes can be difficult to determine. Researchers have historically been limited to passive observations of natural dynamics, which can preclude studies of elusive and noisy cellular events where large amounts of data are required to reveal statistically significant effects. Here, using recent advances in the fields of machine learning and control theory, we train a deep neural network to accurately predict the response of an optogenetic system in Escherichia coli cells. We then use the network in a deep model predictive control framework to impose arbitrary and cell-specific gene expression dynamics on thousands of single cells in real time, applying the framework to generate complex time-varying patterns. We also showcase the framework's ability to link expression patterns to dynamic functional outcomes by controlling expression of the tetA antibiotic resistance gene. This study highlights how deep learning-enabled feedback control can be used to tailor distributions of gene expression dynamics with high accuracy and throughput without expert knowledge of the biological system.

Deep Neural Networks for Predicting Single-Cell Responses and Probability Landscapes.

Engineering biology relies on the accurate prediction of cell responses. However, making these predictions is challenging for a variety of reasons, including the stochasticity of biochemical reactions, variability between cells, and incomplete information about underlying biological processes. Machine learning methods, which can model diverse input-output relationships without requiring a priori mechanistic knowledge, are an ideal tool for this task. For example, such approaches can be used to predict gene expression dynamics given time-series data of past expression history. To explore this application, we computationally simulated single-cell responses, incorporating different sources of noise and alternative genetic circuit designs. We showed that deep neural networks trained on these simulated data were able to correctly infer the underlying dynamics of a cell response even in the presence of measurement noise and stochasticity in the biochemical reactions. The training set size and the amount of past data provided as inputs both affected prediction quality, with cascaded genetic circuits that introduce delays requiring more past data. We also tested prediction performance on a bistable auto-activation circuit, finding that our initial method for predicting a single trajectory was fundamentally ill-suited for multimodal dynamics. To address this, we updated the network architecture to predict the entire distribution of future states, showing it could accurately predict bimodal expression distributions. Overall, these methods can be readily applied to the diverse prediction tasks necessary to predict and control a variety of biological circuits, a key aspect of many synthetic biology applications.

Longitudinal Single-Cell Imaging of Engineered Strains with Stimulated Raman Scattering to Characterize Heterogeneity in Fatty Acid Production.

Understanding metabolic heterogeneity is critical for optimizing microbial production of valuable chemicals, but requires tools that can quantify metabolites at the single-cell level over time. Here, longitudinal hyperspectral stimulated Raman scattering (SRS) chemical imaging is developed to directly visualize free fatty acids in engineered Escherichia coli over many cell cycles. Compositional analysis is also developed to estimate the chain length and unsaturation of the fatty acids in living cells. This method reveals substantial heterogeneity in fatty acid production among and within colonies that emerges over the course of many generations. Interestingly, the strains display distinct types of production heterogeneity in an enzyme-dependent manner. By pairing time-lapse and SRS imaging, the relationship between growth and production at the single-cell level are examined. The results demonstrate that cell-to-cell production heterogeneity is pervasive and provides a means to link single-cell and population-level production.

Dynamic gene expression and growth underlie cell-to-cell heterogeneity in Escherichia coli stress response.

Cell-to-cell heterogeneity in gene expression and growth can have critical functional consequences, such as determining whether individual bacteria survive or die following stress. Although phenotypic variability is well documented, the dynamics that underlie it are often unknown. This information is important because dramatically different outcomes can arise from gradual versus rapid changes in expression and growth. Using single-cell time-lapse microscopy, we measured the temporal expression of a suite of stress-response reporters in Escherichia coli, while simultaneously monitoring growth rate. In conditions without stress, we found several examples of pulsatile expression. Single-cell growth rates were often anticorrelated with reporter levels, with changes in growth preceding changes in expression. These dynamics have functional consequences, which we demonstrate by measuring survival after challenging cells with the antibiotic ciprofloxacin. Our results suggest that fluctuations in both gene expression and growth dynamics in stress-response networks have direct consequences on survival.

Anticipating antibiotic resistance.

Machine learning can use clinical history to lower the risk of infection recurrence.

DeLTA 2.0: A deep learning pipeline for quantifying single-cell spatial and temporal dynamics.

Improvements in microscopy software and hardware have dramatically increased the pace of image acquisition, making analysis a major bottleneck in generating quantitative, single-cell data. Although tools for segmenting and tracking bacteria within time-lapse images exist, most require human input, are specialized to the experimental set up, or lack accuracy. Here, we introduce DeLTA 2.0, a purely Python workflow that can rapidly and accurately analyze images of single cells on two-dimensional surfaces to quantify gene expression and cell growth. The algorithm uses deep convolutional neural networks to extract single-cell information from time-lapse images, requiring no human input after training. DeLTA 2.0 retains all the functionality of the original version, which was optimized for bacteria growing in the mother machine microfluidic device, but extends results to two-dimensional growth environments. Two-dimensional environments represent an important class of data because they are more straightforward to implement experimentally, they offer the potential for studies using co-cultures of cells, and they can be used to quantify spatial effects and multi-generational phenomena. However, segmentation and tracking are significantly more challenging tasks in two-dimensions due to exponential increases in the number of cells. To showcase this new functionality, we analyze mixed populations of antibiotic resistant and susceptible cells, and also track pole age and growth rate across generations. In addition to the two-dimensional capabilities, we also introduce several major improvements to the code that increase accessibility, including the ability to accept many standard microscopy file formats as inputs and the introduction of a Google Colab notebook so users can try the software without installing the code on their local machine. DeLTA 2.0 is rapid, with run times of less than 10 minutes for complete movies with hundreds of cells, and is highly accurate, with error rates around 1%, making it a powerful tool for analyzing time-lapse microscopy data.

Microsecond fingerprint stimulated Raman spectroscopic imaging by ultrafast tuning and spatial-spectral learning.

Label-free vibrational imaging by stimulated Raman scattering (SRS) provides unprecedented insight into real-time chemical distributions. Specifically, SRS in the fingerprint region (400-1800 cm-1) can resolve multiple chemicals in a complex bio-environment. However, due to the intrinsic weak Raman cross-sections and the lack of ultrafast spectral acquisition schemes with high spectral fidelity, SRS in the fingerprint region is not viable for studying living cells or large-scale tissue samples. Here, we report a fingerprint spectroscopic SRS platform that acquires a distortion-free SRS spectrum at 10 cm-1 spectral resolution within 20 µs using a polygon scanner. Meanwhile, we significantly improve the signal-to-noise ratio by employing a spatial-spectral residual learning network, reaching a level comparable to that with 100 times integration. Collectively, our system enables high-speed vibrational spectroscopic imaging of multiple biomolecules in samples ranging from a single live microbe to a tissue slice.

DeLTA: Automated cell segmentation, tracking, and lineage reconstruction using deep learning.

Microscopy image analysis is a major bottleneck in quantification of single-cell microscopy data, typically requiring human oversight and curation, which limit both accuracy and throughput. To address this, we developed a deep learning-based image analysis pipeline that performs segmentation, tracking, and lineage reconstruction. Our analysis focuses on time-lapse movies of Escherichia coli cells trapped in a "mother machine" microfluidic device, a scalable platform for long-term single-cell analysis that is widely used in the field. While deep learning has been applied to cell segmentation problems before, our approach is fundamentally innovative in that it also uses machine learning to perform cell tracking and lineage reconstruction. With this framework we are able to get high fidelity results (1% error rate), without human intervention. Further, the algorithm is fast, with complete analysis of a typical frame containing ~150 cells taking <700msec. The framework is not constrained to a particular experimental set up and has the potential to generalize to time-lapse images of other organisms or different experimental configurations. These advances open the door to a myriad of applications including real-time tracking of gene expression and high throughput analysis of strain libraries at single-cell resolution.

Cell-machine interfaces for characterizing gene regulatory network dynamics.

Gene regulatory networks and the dynamic responses they produce offer a wealth of information about how biological systems process information about their environment. Recently, researchers interested in dissecting these networks have been outsourcing various parts of their experimental workflow to computers. Here we review how, using microfluidic or optogenetic tools coupled with fluorescence imaging, it is now possible to interface cells and computers. These platforms enable scientists to perform informative dynamic stimulations of genetic pathways and monitor their reaction. It is also possible to close the loop and regulate genes in real time, providing an unprecedented view of how signals propagate through the network. Finally, we outline new tools that can be used within the framework of cell-machine interfaces.

Identification of individual cells from z-stacks of bright-field microscopy images.

Obtaining single cell data from time-lapse microscopy images is critical for quantitative biology, but bottlenecks in cell identification and segmentation must be overcome. We propose a novel, versatile method that uses machine learning classifiers to identify cell morphologies from z-stack bright-field microscopy images. We show that axial information is enough to successfully classify the pixels of an image, without the need to consider in focus morphological features. This fast, robust method can be used to identify different cell morphologies, including the features of E. coli, S. cerevisiae and epithelial cells, even in mixed cultures. Our method demonstrates the potential of acquiring and processing Z-stacks for single-layer, single-cell imaging and segmentation.

Balancing a genetic toggle switch by real-time feedback control and periodic forcing.

Cybergenetics is a novel field of research aiming at remotely pilot cellular processes in real-time with to leverage the biotechnological potential of synthetic biology. Yet, the control of only a small number of genetic circuits has been tested so far. Here we investigate the control of multistable gene regulatory networks, which are ubiquitously found in nature and play critical roles in cell differentiation and decision-making. Using an in silico feedback control loop, we demonstrate that a bistable genetic toggle switch can be dynamically maintained near its unstable equilibrium position for extended periods of time. Importantly, we show that a direct method based on dual periodic forcing is sufficient to simultaneously maintain many cells in this undecided state. These findings pave the way for the control of more complex cell decision-making systems at both the single cell and the population levels, with vast fundamental and biotechnological applications.

Extracellular Vesicle Production Loaded with Nanoparticles and Drugs in a Trade-off between Loading, Yield and Purity: Towards a Personalized Drug Delivery System.

Extracellular vesicles (EVs) released by cells and circulating in body fluids are recognized as potent vectors of intercellular self-communication. Due to their cellular origin, EVs hold promise as naturally targeted "personalized" drug delivery system insofar as they can be engineered with drugs or theranostic nanoparticles. However, technical hurdles related to their production, drug loading, purification, and characterization restrain the translation of self-derived EVs into a clinical drug delivery system. Herein, different methods are compared to generate and to purify EVs encapsulating iron oxide nanoparticles and a clinical photosensitizer drug (Foscan) as biocamouflaged agents for photodynamic therapy, magnetic resonance imaging, magnetic manipulation, and hyperthermia. Theranostic EVs are produced from drug- and nanoparticle-loaded endothelial cells either by spontaneous release in complete medium, by starvation in serum-free medium or by mechanical stress in a microfluidic chip mimicking vessel shear stress, and purified by ultracentrifugation or magnetic sorting. The impact of the production and purification protocols is investigated on EV yield and size, nanoparticle and drug cargo, and finally on their therapeutic efficacy. EV production by starvation combined with purification by ultracentrifugation may be considered a reasonable trade-off between loading, yield, and purity for biogeneration of theranostic EVs.

Noise propagation in synthetic gene circuits for metabolic control.

Dynamic control of enzyme expression can be an effective strategy to engineer robust metabolic pathways. It allows a synthetic pathway to self-regulate in response to changes in bioreactor conditions or the metabolic state of the host. The implementation of this regulatory strategy requires gene circuits that couple metabolic signals with the genetic machinery, which is known to be noisy and one of the main sources of cell-to-cell variability. One of the unexplored design aspects of these circuits is the propagation of biochemical noise between enzyme expression and pathway activity. In this article, we quantify the impact of a synthetic feedback circuit on the noise in a metabolic product in order to propose design criteria to reduce cell-to-cell variability. We consider a stochastic model of a catalytic reaction under negative feedback from the product to enzyme expression. On the basis of stochastic simulations and analysis, we show that, depending on the repression strength and promoter strength, transcriptional repression of enzyme expression can amplify or attenuate the noise in the number of product molecules. We obtain analytic estimates for the metabolic noise as a function of the model parameters and show that noise amplification/attenuation is a structural property of the model. We derive an analytic condition on the parameters that lead to attenuation of metabolic noise, suggesting that a higher promoter sensitivity enlarges the parameter design space. In the theoretical case of a switch-like promoter, our analysis reveals that the ability of the circuit to attenuate noise is subject to a trade-off between the repression strength and promoter strength.