Control of mammalian cell-based devices with genetic programming

Abstract

Synthetic biology increasingly enables the construction of sophisticated functions in mammalian cells. A particularly promising frontier combines concepts drawn from industrial process control engineering—which is used to confer and balance properties such as stability and efficiency—with understanding as to how living systems have evolved to perform similar tasks with biological components. In this review, we first survey the state-of-the-art for both technologies and strategies available for genetic programming in mammalian cells. We then discuss recent progress in implementing programming objectives inspired by engineered and natural control mechanisms. Finally, we consider the transformative role of model-guided design in the present and future construction of customized mammalian cell functions for applications in biotechnology, medicine, and fundamental research.

Introduction

Engineered cell therapies comprise one of the most rapidly expanding frontiers in medicine, with mounting clinical successes in the field of cancer immunotherapy in particular [1,2]. A key step in the development of novel cell therapies is translating a clinical need into a design goal that can be implemented using genetic engineering. Since many medical needs can be described as a restoration of homeostasis via delivery of a therapeutic mediator in a manner that is regulated by environmental inputs, it is often possible to frame this need as a process controls objective. Thus, process control theory is increasingly employed to inform and guide cellular engineering [3-6]. This review provides a framework for posing and solving cellular engineering challenges using concepts derived from the field of process control engineering, with an emphasis on the use of genetic programs and associated mathematical models to achieve these goals.

Programming cells to meet control objectives

Composition of elements into modules and programs.

To describe the process of creating genetic programs, we define a hierarchical organization in which elements (Figure 1a) are combined into functional modules that each perform one specific function, and modules are combined into programs that cause an engineered cell to exhibit a desired behavior (Figure 1b). These terms functionally distinguish the components of genetic programs [7,8].

Figure 1. Fundamentals of mammalian genetic programs.

(a) Elements of genetic programming: red coloring highlights the element named, and grayscale coloring indicates related species associated with each element’s function. In this review, we use the term elements to describe DNA, RNA, or protein regulatory components that can be utilized in the construction of genetic programs. This usage is intended to disambiguate our discussion from existing definitions of “parts” [7,8]. Elements may be grouped into one of several categories based upon their function. Receptors [58-60] perform two key operations: sensing, in which the receptor interacts with an input and changes the receptor state, and actuation, in which the receptor in its new state effects a change in the cell (which is often termed the receptor output). Transcription factors (TFs) [14,15,60-62] (DV Israni, et al., bioRxiv doi: 10.1101/2021.02.22.432371) are proteins that bind to specific DNA sequences to modulate the rate of gene transcription. Promoters [61,63,64] are regulatory DNA sequences that define the location and efficiency of transcription initiation, determine which DNA strand will be transcribed, and influence the duration of transcription (i.e., transcript elongation). TFs bind to promoters to recruit RNA polymerase and co-factors to initiate transcription. DNA modifiers [20,65,66] are enzymes that either directly alter nucleotide sequences in DNA or change the state of chromatin. Protein regulators [16,23,67] control the state of specific proteins via directly or indirectly modulating stability, changing primary protein sequence (e.g., via cleavage or splicing), or modulating protein-protein interactions. Noncoding RNAs [17-19,31,33,50] modulate mRNA stability (either stabilizing or destabilizing) and can either inhibit or enhance translation of a target mRNA. In Supplementary Table 1, we present illustrative (but not comprehensive) examples of each category of element used for genetic programming. (b) A hierarchical organization of genetic programs is depicted. Elements are biomolecular components (DNA, RNA, or protein) that are combined to form functional modules. Each module is an assembly of elements that performs a specific function. Programs are assemblies of connected modules that together cause a cell to exhibit a desired behavior (i.e., exhibit a predetermined input-output response). A representative example of this hierarchy is shown in the context of genetic elements, modules, and programs based upon transcriptional regulation.

We introduce module types by considering selected examples; comprehensive overviews are provided elsewhere [9,10]. Some of the most widely used modules implement Boolean logic to relate inputs to outputs. Often called logic gates, these modules are conceptually comparable to their electronic counterparts, and this analogy is useful for describing individual gates as well as for defining rules by which such gates may be combined [11]. There now exist multiple examples of mammalian logic gates comprised of transcriptional activators and inhibitors including chromatin regulators [12-16], noncoding RNAs [17-19], or DNA recombinases [20]. This concept can be extended beyond digital logic (wherein the output states are discrete and depend on binary presence or absence of inputs) to implement analog logic (wherein the circuit implements a continuous relationship between input and output magnitudes) [16]. Similarly, modules can perform thresholding (e.g., suppressing output unless a minimum input magnitude is surpassed) or filtering (e.g., band-pass filters produce output only when input magnitudes fall within a specified range) [16,21-23]. Other modules amplify an input signal using positive feedback, often through a cascade wherein a transcription factor enhances its own expression [24]. Conversely, modules can also negatively regulate sensitivity to an input using negative feedback, such as a transcription factor repressing its own expression [25]. Finally, switches are a broad class of module that induce a change in gene expression state (i.e., ON or OFF) in response to an input. In some switches, sustaining the output state requires sustaining the input state (like a car horn) [10]. In other switches, the system exhibits memory such that a transient input signal can induce a temporally stable change in output state (like a light switch) [26-29]. In the following subsections, we focus on notable genetic program examples that achieve important control objectives in living cells.

Programs for mitigating intercellular variability and resource competition.

Intercellular heterogeneity poses a challenge for the design and characterization of genetic programs. Many transient and stable (i.e., genomic integration) methods for delivering the DNA or RNA that encodes genetic programs result in heterogeneity in expression (and sometimes copy number) across a population of cells. The magnitude of this heterogeneity varies with delivery method. Even within a genetically identical population of cells, expression levels vary due to differences such as ribosome copy number, cell cycle state, access to nutrients, and other factors [30]. All of these sources of heterogeneity can limit the extent to which each cell in a population executes an engineered function as desired. Some modules specifically address this challenge of making a genetic program robust to expression variability. Three recent examples employ incoherent feedforward loops (IFFLs) to achieve this objective, whereby expression of one element (a protein or miRNA) scales with delivery efficiency, and this element negatively regulates expression of other program elements [25,31,32]. A related challenge is that overexpression of transgenic components imposes burden because all elements are produced from a limited, common cellular resource pool. Resource competition inadvertently couples modules and can impair genetic program performance (C McBride, et al., bioRxiv doi: 10.1101/2021.05.26.445862). Modules that employ IFFLs [33,34], negative feedback topologies [25,35], or both [25] can mitigate this burden. In each of these IFFL examples, increasing the input drives expression of both the output and a mitigator element (inversely proportional to resource load) which negatively regulates production of the output (Figure 2a). In another recent example in bacteria, negative feedback was employed to maintain dCas9 concentrations at a level that is independent of sgRNA-load, mitigating competition between sgRNAs in genetic circuits [36]. These examples illustrate how genetic programming can be employed to address practical challenges associated with intercellular variation, enabling both improved characterization and ultimately enhanced performance of engineered cells.

Figure 2. Notable examples of control programs implemented in living cells.

(a-e) The upper part of each panel depicts a program topology, and the lower part of each panel depicts a scenario that illustrates the utility of each program. In (a), the system experiences an increase in genetic load as input increases, and the IFFL program maintains a consistent relationship between input and output, which does not occur without the IFFL program due to resource competition. In (b), the set point is increased (via inducing an increase in the reference species), and in response the program increases a manipulated input (to a degree that minimizes the error between output and reference levels), eventually causing the output to increase and track the change in the reference. This program is also able to reject disturbances, preventing unwanted sustained changes in output. In (c), the system rejects a disturbance that occurs within a network connecting input to output using a mitigator (which is produced proportionally to the output) to implement negative feedback. In (d), the set point is increased (via inducing an increase in the reference species), and the output tracks this change (rather than increasing too much, for example) because the output negatively regulates its own production by competing with the reference for binding to the mitigator. In (e), control is implemented using electronic equipment, wherein computer-based software measures output levels (e.g., fluorescence measured by microscopy), compares this value to an electronically defined set point (reference), and in response adjusts a manipulated input (e.g., via administration of a chemical inducer). In the scenario shown, the set point is increased and the output tracks this change.

Programs for disturbance rejection.

The ability of living systems to maintain a steady state despite disturbances in the environment, which is known as homeostasis, is critical for survival. In industrial processes and everyday mechanical devices, this capability is implemented thorough process control engineering in order to maintain a steady system output despite disturbances (i.e., to reject disturbances). In general, an error signal describes the difference between the desired output level (set point) and the actual output level, and the controller takes action to reduce the error. These concepts are increasingly being extended to implement synthetic disturbance rejection (sometimes termed robust perfect adaptation [37]) in living cells. One of the first such demonstrations was implemented using an architecture called the antithetic motif, which integrates error over time [38,39]. Just as electronic systems compute error by subtraction, the biological antithetic motif comprises two components that interact to annihilate each other in a stoichiometric manner, such that only the component that is in excess will remain (representing both the magnitude and sign of the error). This error is employed in a program (typically by adjusting a manipulated input) to bring the output back to the set point after a disturbance impacting either antithetic component (Figure 2b). This motif has been mathematically analyzed [40] (M Filo, et al., bioRxiv doi: 10.1101/2021.03.21.436342) (M Gomez-Schiavon, et al., bioRxiv: 10.1101/2020.10.09.334078) (E Hancock, et al., bioRxiv doi: 10.1101/2021.04.18.440372) and implemented in bacteria (using sigma factors) [39] and mammalian cells (using miRNA) (T Frei, et al., bioRxiv doi: 10.1101/2020.12.06.412304). Other motifs capable of integrating error have also been proposed but not yet implemented [40,41]. Similarly, genetic programs that employ negative feedback modules can also mitigate the effect of disturbances on the final output, and this strategy has proven to be useful in studying endogenous signaling pathways and creating synthetic pathways in yeast [42] (Figure 2c).

Programs for set point tracking.

Another useful application of process control is ensuring that the output will track dynamic changes made to the set point. This function enables a cell to dynamically adjust production of an output in response to changing magnitudes of a reference signal that serves as the set point (e.g., by adjusting the concentration of a chemical inducer). In bacteria, this strategy was implemented using synthetic scaffold and anti-scaffold proteins to perform subtraction and make the output track changes in the reference [43] (Figure 2d). The antithetic motif, previously discussed for accomplishing disturbance rejection, is also useful for implementing set point tracking in mammalian cells (T Frei, et al., bioRxiv doi: 10.1101/2020.12.06.412304). Finally, set point tracking can be implemented using in silico computation (as opposed to biomolecular computation) by integrating imaging and microfluidic systems to precisely regulate protein concentrations in yeast [44] and mammalian cells [45] (Figure 2e).

Opportunities and challenges for genetic control programs.

With the development of new elements for mammalian genetic programming, including essential quantitative characterizations, genetic programs can now be designed and employed to achieve sophisticated control objectives. Mitigating resource competition facilitates element characterization [33,34] and may benefit biomanufacturing, cell-based therapies, and fundamental research wherein targeted perturbations that avoid global impacts on cell state are desired. Programmed disturbance rejection and set point tracking could enable safe, reliable, and effective engineered cell-based therapies that diagnose and treat disease. However, important challenges still exist. First, the fact that cells divide and dilute their components poses challenges for current cellular computation strategies (extensively reviewed elsewhere [46]). Second, modules cannot be fully decoupled from one another or other processes in the cell, because all natural and engineered elements draw from a shared resource pool. Third, heterogeneity amongst cells in a population is an intrinsic feature of such systems, and genetic programs must be designed to operate within (or even employ) this fact. Given our ever-increasing appreciation of the role that heterogeneity plays in the function of natural biological systems [30], developing engineering design principles for explicitly extending these insights to constructing novel systems is an exciting frontier. An increasingly essential strategy for understanding these phenomena, and ultimately overcoming related challenges, is the use of computational models.

Model-guided design of genetic programs enables improved understanding and prediction of performance

Mathematical modeling can streamline the canonical engineering “design, build, test” (DBT) cycle. When building genetic programs, an empirical “DBT” cycle is often iterated until a program that performs sufficiently well is identified (often through laborious trial-and-error and tuning). While this approach has yielded some success, it quickly becomes intractable when considering the high dimensional problem of building multi-component genetic programs and/or programs which are intended to perform complex functions. Mathematical modeling can greatly accelerate this process by improving understanding of the mechanisms that influence genetic program performance and by prioritizing promising genetic program designs for testing in the wet lab. A model is essentially a formal, quantitative statement of a hypothesis; evaluating explanatory models is often the only way that a researcher can systematically test whether many distinct observations can be consistent with a single mechanistic hypothesis. The rigorous hypothesis testing procedure that is necessary to build and analyze a mathematical model is one of the key advantages of employing model-guided design rather than solely iterative tuning that relies more heavily upon insight and intuition. Dynamic models, such as those incorporating ordinary differential equations (ODEs) with reaction rates described by algebraic equations, are well-suited to this task, as such models can describe and predict the behavior of gene regulatory systems by simulating the dynamic trajectories of biologically relevant states and comparing these simulations to experimental data. Thus, our discussion focuses on this class of models, although some considerations are extensible to other modeling frameworks. In Supplementary Table 2, we present illustrative examples of models employed to address different modeling goals.

Employing models for explanation.

Models can help identify important contributors to genetic program performance. This explanatory use focuses on mathematically describing a set of experimental training data and does not aim to predict any novel behavior. Such models can identify mechanistic assumptions that are necessary (or not necessary) to explain a dataset or observation, and they can help one understand important interactions between components in a genetic program [15,25,33,47] (T Frei, et al., bioRxiv doi: 10.1101/2020.12.06.412304). The process of generating and refining a model for explanation is particularly useful for generating potential mechanistic explanations for previously unintuitive experimental results (i.e., for generating new hypotheses).

Employing models for prediction.

Models can help to identify or prioritize genetic program designs for experimental investigation. The starting point for building a model capable of prediction is often a model used for explanation, which is then directly validated with additional experimental data that were not used to develop the original model. Predictive models can be used to consider the performance of a genetic program in a previously untested experimental condition, to efficiently evaluate many as-of-yet untested design choices in silico, and to facilitate genetic program design automation [16,32,34,42,48,49].

Guiding design using in silico data.

In order to consider designs based upon elements that might be created, but do not yet exist, predictive models can employ in silico data rather than (or in addition to) exclusively experimental data. [40,46,50,51]. This strategy assumes that each biomolecular implementation considered could be developed in the wet lab, and thus such analyses are most appropriately guided by informed biophysical estimates as to what is possible to build (e.g., via de novo protein design or tuning of existing genetic elements). In silico analyses help to manage the reality that a tiny fraction of the potential design space has been characterized experimentally, and this approach is readily updated to keep pace with improving estimates of technical risk and effort required to build (or find) new elements. As model-guided design continues to advance, in silico analyses may become indispensable for prioritizing experimental development and characterization.

Considering context.

Models that explicitly consider context are increasingly necessary to facilitate design of modules that function reliably when implemented in different programs or under different conditions. The genetic context (e.g., DNA delivery method), cellular context (e.g., organism, cell type), cellular state (e.g., number of RNA polymerase or ribosomes in a cell), and environmental context (e.g., nutrient concentration) of a genetic program substantially impacts its performance [52-54]. Many genetic programs are first prototyped in workhorse cell lines using efficient, convenient DNA delivery methods, which may differ from the cell type(s) and delivery method(s) that might be used in a final application (e.g., a clinical product). This strategy is often motivated by the need to (at least initially) rigorously evaluate multiple genetic program designs via many independent experiments, which may not be feasible with less efficient methods. Moreover, a given genetic program might ultimately be used in a variety of final applications, such as in primary cells which differ from donor to donor, such that simply starting with a near-final implementation strategy only partially circumvents the challenge of dealing with context. In general, a shift in context at some point in the design process is virtually inescapable. Therefore, developing genetic elements, modules, and programs that are robust to cellular context—or that can be described by design rules that enable mapping of existing knowledge to new contexts of interest—would be widely useful for design-guided engineering. Modeling may help achieve this goal by facilitating identification and improved understanding of how changes in context impact genetic program performance on a mechanistic level. Early steps toward meeting this need include developing a resource-aware framework for ODE modeling of genetic programs in mammalian cells [33]. This framework aims to provide a generalizable modification to any ODE model of genetic elements that are either transcriptionally or translationally limited by cellular burden [33]. Related work in bacterial cells employs a mechanistic model of resource competition to characterize genetic circuit modules and predict their behavior when combined with additional modules (C McBride, et al., bioRxiv doi: 10.1101/2021.05.26.445862). Other work focuses on mapping data collected (or in this case simulated) in one context to another context (specifically, mapping from the transient context to the stable context), using statistical methods to understand how the input/output (I/O) relationship obtained in one context can be appropriately compared to the I/O relationship in another context [55].

Opportunities for automated genetic program design.

There now exists great opportunity for extending the modeling concepts described above towards automating genetic program design. The state-of-the-art in this space is Cello—a suite of software employed for engineering bacteria [56] and yeast [57]. Cello employs mathematical models (algebraic transfer functions for bacteria and ODE-based dynamic models for yeast) of genetic elements described by an electrical engineering-derived hardware design language. However, adapting this approach to engineer mammalian cells (and different types of genetic elements) remains challenging. For example, a key feature that enables Cello-guided design is insulation of elements from genetic context, ensuring that elements function in the same way even when used in different circuits [57]; this property is difficult to achieve in mammalian cells. Predictive design of genetic programs in mammalian cells was recently reported using an alternative model-driven design framework that is not based upon analogy with electronic components [16], but this approach has not yet been extended to achieve automated design. As synthetic biology continues to evolve into a true engineering discipline, model-guided design of elements, modules, and programs will prove to be as transformative for the engineering of biology as it has been for all other technical fields.