Programming Cells: Towardsan automated “Genetic Compiler”

I. Summary

The increasing scale and sophistication of genetic engineering will necessitate a new generation of computer-aided design (CAD). For large genetic programs, keeping track of the DNA on the level of nucleotides becomes tedious and error prone. To push the size of projects, it is important to abstract the designer from the process of part selection and optimization. The vision is to specify genetic programs in a higher-level language, which a genetic compiler could automatically convert into a DNA sequence. Steps towards this goal include: defining the semantics of the higher-level language, algorithms to select and assemble parts, and biophysical methods to link DNA sequence to function. These will be coupled to graphic design interfaces and simulation packages to aid in the prediction of program dynamics, optimize genes, and scan projects for errors.

II. Introduction

Numerous genetic circuits have been built that encode functions that are analogous to electronic circuits [1–3]. Genetic programs have been built by combining multiple circuits. For example, we constructed an “edge detector” program that combines a ANDN gate, light sensor, and cell-cell communication that give bacteria the ability to draw the edge between light and dark regions of an image projected onto a plate [4]. Other genetic programs have been built that combine circuits to produce a push-on/push-off circuit [5], implement a counter [6], and reproduce predator-prey dynamics [7]. These represent toysystems, but the implementation of such programs in applications for industrial biotechnology is inevitable.

Automated DNA synthesis gives genetic engineers an unprecedented design capacity [8]. This technology enables the specification of every basepair for long sequences, without having to be concerned about the path to construction. Together with methods to rapidly combine genetic parts [9, 10] and assembly methods that scale to whole genomes [11–13], the problem of DNA construction has far outpaced our capacity for design [14]. A good example of this is the 2006 UCSF iGEM team to build a “remote-controlled bacterium.” DNA synthesis was used to build the first construct (requiring a few weeks), but after four years of additional tinkering, the paper will be submitted in 2010.

Our ability to design programs has been hampered by three problems. First, there is a lack of good, robust genetic circuits that can be easily connected. Second, there are few design rules that are sufficiently quantitative to be carried out by a design algorithm. Modeling can be helpful before the experiments to determine the topologies and parameter regimes required to obtain a particular function. However, simulations cannot be used to “reach down” to the DNA and suggest a specific mutation or select a part. Third, mistakes in the DNA sequence scale quickly with size. Currently, to scan for potential errors (e.g., transposon insertion sites or putative internal promoters), it requires the running of multiple,(usually web-based) programs. There is no unified software package to date that addresses all of these issues.

The creation of a simulation environment for genetic engineering is complicated by the diversity of cellular functions. When studying natural networks, there is a feeling of “peeling an onion,” where there are seemingly endless redundancies and classes of biochemical interactions. Even within the Registry of Standard Biological Parts (www.partsregistry.org), there are a wide variety of cellular functions: from enzymes and transcription factors to multi-gene gas vesicles and secretion chaperones. Each specific problem requires its own style of simulation; a dynamic program may be well satisfied by sets of differential equations, pattern formation by cellular automata, and enzymes by metabolic flux analysis. It would be daunting to create a simulation package that could encompass all of this.

To reduce the problem complexity and to frame recent computational work, we introduce the concept of a “Genetic Compiler,” whose inputs are high-level instructions (equivalent to VHDL or Verilog) and whose output is a DNA sequence. The sequence can be sent to a company for DNA synthesis or a robot for automated assembly. The problem is constrained by focusing on genetic programs that encode a desired logical or dynamical function, which can be integrated into many applications in biotechnology (Figure 1). This avoids the application-specific portions of the problem; for example, building a butanol sensor a particular metabolic pathway. It is distinct from tools for protein or metabolic engineering [15].

Figure 1. The compiler is focused on assembling the circuitry that links the inputs and outputs of a larger project.

The inputs include genetic sensors that can respond to diverse signals, such as temperature, light, stress, or metabolites. The circuitry encodes the logic and dynamics. The output of the circuits control actuators, such as a metabolic pathway, the activation of cell-cell communication, or a stress response. The inputs and outputs tend to be problem-specific and the diversity of biological applications makes this difficult to encompass in a single simulation package. In contrast the circuitry can be reconfigured to build programs relevant to diverse problems in biotechnology.

The scope of this review is on the underlying algorithms and biophysical methods that would power such a compiler (Figure 2). Realizing this goal will require: 1. Libraries of reliable genetic circuits designed specifically to be part of a CAD program, 2. the definition of a higher-level language, 3. algorithms to assemble circuits according to a specified program, 3. biophysical methods to connect and optimize circuits, 4. simulation programs to debug the program dynamics, 5. algorithms for DNA assembly and experimental design. The scope has been limited to exclude several topics that are critical to synthetic biology, but have been well-reviewed elsewhere, notably codon optimization and tools from systems biology and metabolic engineering [15–17].

Figure 2. A Genetic Compiler.

The compiler automatically converts a higher-level language to a DNA sequence. Ideally, the designer would be completely blind to these steps (gray box). Simulations aid the debugging of the program, but the debugging would occur within the higher-level language.

III. Main Text

Robust Combinatorial Logic

Combinatorial logic is implemented by Boolean circuits and is the basis for digital computing. It is used to build circuits that apply Boolean algebra on a set of inputs to transform them into a set of desired outputs. Simple circuits can be layered in different configurations in order to achieve a computational operation. This has enabled the automated design that underlies VLSI. The ability for digital circuits to be flexibly used and easily captured by CAD comes at a cost of speed, design size, and power [18]. There are ongoing efforts to design the core biochemistry that would act as genetic logic gates [4, 19–29]. There are several important considerations in designing genetic logic gates to be used in CAD:

• Scalability. Each circuit in a design needs to be orthogonal because all of the circuits operate within the cell based on biochemical interactions. A circuit is scalable when the underlying genetic parts can be swapped to create orthogonal circuits. For example, orthogonal NOT gates could be built based on new repressor-operator pairs [30]. In contrast, an AND gate that we constructed is not scalable because its underlying parts do not lend themselves to making multiple gates [27].

• Extensibility. In order to layer genetic circuits, the inputs and outputs of the circuit have to have the same signal carrier [31]. Put simply, for circuits to be connected, the inputs and the outputs need to be the same biochemical form (phosphorylation, transcription, etc). For transcriptional circuits, it has been proposed that the flux of RNAPs on DNA could serve this purpose [31]. In practice, this can be implemented by making the circuit inputs and the outputs be promoters. Several examples of this are a NOT gate [30] and an AND gate [27]. This is in contrast to circuits where, for example, the inputs are small molecules and the output is the activation of a riboswitch [25].

• Modularity. The inputs and outputs of the circuits need to be able to be easily changed. Considering CAD, the ease has to be sufficient for a computer to be able to reliably perform this function. Promoters, protein-protein interaction domains, and RNA are sufficiently modular for this purpose [25]. Some biological systems that are considered modular – such as the protein domains of bacterial two-component systems – are not sufficiently modular for CAD. This is because small libraries are required to identify a functional chimera, as opposed to design rules that can be implemented by a machine.

• Robustness. The circuits must remain functional over large regions of parameter space. This is particularly important to be able to connect circuits in series, where robust circuits will allow for more uncertainty while still producing a functional connection [4, 32]. The circuits also have to be non-toxic and have minimal impact on host processes [33].

• Speed. The circuits also need to be reasonably fast. For example, a three-layer cascade of NOT gates required 20 minutes for each layer to compute [34], but a DNA inversion switch requires 4hours [2]. Even 20 minutes may be too slow for some applications, thus requiring logic to be implemented by phosphorelays or protein-protein interactions, which can occur in milliseconds [35].

Higher-Level Languages and their Deconstruction

A higher-level language abstracts the designer from the details of the machine. In electronics, it enables the CPU operations to be hidden. In synthetic biology, it would hide the details of the molecular biology and DNA sequence, and possibly even the choice of circuits. Impeding the development of a higher-level language in synthetic biology is the diversity of cellular functions. This can be partially overcome by limiting the language to describe problems that are common to many applications, while leaving the application-specific aspects unabstracted (Figure 1). There are two areas that have shown promise in moving towards a higher-level language in synthetic biology. First, vocabularies and grammars are being developed that implement design rules and can be used to describe genetic parts and their combination. This will form the backbone of the higher-level language. Second, there are a number of algorithms that can be borrowed from electrical engineering, notably logic minimization, which can convert the language into integrated circuits.

The Semantics of Genetic Programs

A higher-level language in synthetic biology needs a vocabulary and grammar to represent genetic parts and the rules underlying how they are combined to build circuits and programs [36]. A vocabulary captures how genetic functions are described and how the data underlying a part is stored and accessed. A rule is an enforced relationship between parts. For example, an expression cassette is defined as needing a promoter, cistron, and terminator and a cistron requires a RBS sequence and gene [37].

An example of a program that describes a well-known genetic oscillator (the “repressilator”) is shown in Figure 3. At first glance, this appears unnecessarily complex compared to the more familiar plasmid visualization. However, it becomes advantageous as the designs get larger and more complex. Plasmid construction programs (like Vector NTI or ApE) toggle between the plasmid map for visualization and the DNA sequence for design. Both are difficult for debugging large programs, whereas formal grammars give an intermediate level of abstraction for writing programs and the rules enable automated troubleshooting [37]. Rules also allow for the permutation of parts in order to create multiple designs [38].

Figure 3. Semantics of genetic programs.

The plasmid map (A) and Eugene code (B) for a genetic oscillator [94] is shown (B). The author of the Eugene code is Adam Liu [41].

An emerging group of tools (GenoCAD, Eugene, and GEC) propose formalisms for vocabularies and provide a means to embed domain expertise in software allowing less experienced users to benefit from it [39–41]. Domain experts can express problem-specific design strategies as formal languages using a controlled vocabulary to represent the types of genetic parts usable in a particular organism or even for a particular application [40]. Formal languages provide a means to develop a knowledge base on a collection of defined terms organized with respect to their structural relationships to each other in the DNA sequence [36]. The advantage of this approach is that the relationship of the individual parts to each other are defined in the software and can provide a means of indexing and retrieving knowledge about parts and designs associated with defined terms. When expressing design strategies with formal languages it is possible to use parsers to verify that parts used in a design are properly positioned with respect to each other [42]. Thus, users have a level of in silico validation through the use of such systems. Beyond this syntactic or structural layer, languages can be augmented by adding a semantic layer to derive dynamics encoded in complex DNA sequences composed of interacting parts using algorithms originally developed to compile the source code of computer programs [37].

Instead of simulating the function of a DNA sequence, it is possible to perform a semantic analysis of DNA sequences using expert systems capable of reasoning using the relationships between defined terms to infer many aspects of their use, assembly based on data present in the part or associated with the biology from which the part was derived. An expert system would scan in silico designs for logical errors in construction, identify constraints that would preclude use of particular combinations of parts. Such an expert system can be constructed using ontologies – a rigorous organization of knowledge concerning a particular domain of knowledge [43]. Ontologies are particularly good examples of expert systems in that they permit the development of a model of the entities and their relationships within the domain [44]. Such definitions are logically consistent and can be used by reasoner softwares to verify known information on entities tagged by the ontology and to infer new information based upon these definitions. Softwares developed to use ontologies have the advantage that as data models change in the ontology, the software simply needs to adapt to the ontology, simplifying many aspects of software data model development and design [44]. The synthetic biology community has started the Synthetic Biology Ontology Language to represent concepts tied to synthetic biology [45]. Ongoing efforts aim clarifying how the SBOL approach and the languages used in GenoCAD complement each other.

A larger model that the synthetic biology community can take advantage of is the National Center for Biomedical Ontologies [46]. The NCBO supports researchers by providing resources to access, review and integrate many biomedical ontologies. The Ontology for Biomedical Research (OBI) is especially interesting as it has been developed to describe the disparate elements and their relations that are necessary and sufficient to describe an experiment [47]. Many of the concepts from OBI are very applicable to the description of synthetic biology workflows, protocols, parts and device usage. NCBO also promotes standards for sharing data across ontologies, most notably the Minimum Information to Reference an External Ontology [48]. Use of MIREOT guidelines during the development of an ontology will support the consistent inclusion of terms from ontologies as diverse as Gene, Chemical Information and Pathway Information and Phenotypic Information.

Logic Minimization

Several programs developed in electrical engineering have the ability to take a truth table as an input and then outputa wiring diagram from simpler circuits (Figure 4). The ESPRESSO program, developed in the early 1980s, has been used extensively for logic minimization in VLSI design [49] and it is embedded along with other tools in the abc program that is currently maintained by UC-Berkeley [50]. The output of the logic minimization tools feeds into programs, such as Logic Friday [51], which both act as a visualization tool and enable constraints to be applied to the construction of a circuit diagram. The minimum identified by these programs is not guaranteed to be the global minimum. The logic minimization programs will have to be modified to include biological constraints, including the size of the DNA, and availability and orthogonality of the gates. Two additional considerations in choosing a wiring diagram are:

Figure 4. Automated program design using logic minimization algorithms.

An example of a multi-input single-output truth table is shown. The truth table is converted to an equation F, which is a function of the four inputs (a,b,c,d). Each term corresponds to each row of the truth table where the output is 1 and the prime (′) is shorthand for the NOT function. Logic minimization algorithms, such as ESPRESSO [49], can be used to simplify the full equation to its minimal form (simplest sum of products). This equation is then converted to circuit diagrams using programs such as Logic Friday [51], which can also implement constraints. For example, the large wiring diagram consists of only 2-input NOR gates (top), whereas the smaller wiring diagram was built allowing for multiple-input OR/NOR/AND/NAND gates. Biological constraints, such as gate availability and DNA size can then be applied to search for the optimal diagram.

• Fault tolerance with respect to asynchronous computing. Genetic circuits are asynchronous because there is no clock controlling when each layer of the computation is performed [52]. Delays in the progression of the signal can lead to faults in the calculation. One approach to this problem is to build a genetic clock, and progress towards a robust oscillator [1, 53] and counter [6] indicate that this may be possible. The design of asynchronous logic blocks that are robust to faults is also an active area of research, especially with respect to minimizing power requirements [54–56]. The principles from this work could be applied to the construction of integrated genetic logic, although unlike simple logic minimization, there has been historically a lack of CAD tools for asynchronous circuit design [9, 57]. Recently, a system for asynchronous system design (RALA) has been developed and this framework is particularly applicable for biological systems [58].

• Robustness to perturbations. The wiring diagrams need to be robust with respect to perturbations due to fluctuations in molecules and environmental conditions [59]. In the design of electronic circuits, it has been noted that asynchronous circuits are also more robust [9, 56]. Sole and co-workers enumerated digital systems composed of layered NAND gates and found that fault tolerance resulted from designs that contained a high degree of degeneracy, defined as occurring when multiple subsets of a circuit diagram that are structurally different perform the same function [60].

Biophysical Models of Part Function

The automated selection of genetic parts is one of the most difficult aspects of synthetic biology CAD [61]. Biophysical models can map the DNA sequence of a genetic part to its function. This enables the impact of an individual mutation to calculated for a part and then coupled to a dynamical model to determine the effect on circuit function. In other words, they facilitate linking a simulation of reaction kinetics to physical DNA. Biophysical models are also useful in screening parts selected from a registry for context effects and can scan the DNA sequence of programs to identify unintended functional sequences (e.g., internal promoters or terminators) and correct potential errors.

Cellular regulation is highly redundant and there are often many ways to control the same parameter. For example, there are many options in controlling the concentration of a protein, from plasmid copy number to protein stability. This gives flexibility to the designer to take whatever route is convenient. In terms of design automation, it will be important to choose those routes that are the most reliable when modeled on the computer. For example, the interactions between RNA basepairs is relatively straightforward to calculate and efficient algorithms have been developed to calculate the folding of RNA secondary structure [62, 63]. This has formed the core of predictive methods to capture how changes in the sequence affect the strength of ribosome binding sites [32, 64, 65], the efficiency of terminators [66–68], the shape of the transfer function of an shRNA switch controlled by a small molecule [69], and to create orthogonal rRNA-mRNA binding sequences [70]. In each of these models, the effect of arbitrary mutations on part function can be quantitatively calculated.

The activity of promoters have been successfully modeled using position weight matrixes (PWMs) [71, 72]. These models predict the free energy by which a transcription factor binds to its operator sequence in a promoter. They make an additive assumption, where each base in the operator is assumed to be independent and a relatively small data set is used to parameterize a model. PWMs have suffered from high false positive rates, which has been a challenge in using them to scan genomes for promoters. However, they have been very successful at modeling how mutations affect the strength of a given promoter [71]. In this sense, they are ideal for applications where it is necessary to link how mutations will affect circuit dynamics by impacting promoter function.

Biophysical models are particularly applicable to a common problem that occurs when connecting genetic circuits in series (Figure 5). If the output of the first circuit is not in the correct dynamic range to trigger the second circuit, then the combined circuits will not function properly. One way to accomplish this is to vary the ribosome binding site (RBS) linking the circuits, typically by substituting different sites from a part library [73, 74] or through random mutagenesis [27, 75]. More recently, a biophysical model of RBS function was developed and shown to be able to identify de novo RBS sequences that can connect two previously-characterized circuits [32]. This approach could accelerate the automated connection and optimization of many genetic circuits.

Figure 5. Connecting genetic circuits.

(A) A simple program is shown consisting of a sensor and circuit. The sensor turns on a promoter in response to an INPUT. A RBS controls a translation initiation rate S that then serves as an input to the circuit. The circuit consists of an activator that turns on a promoter that serves as the output. (B) If the transfer function of the sensor is too low (blue) or the basal level is too high (black), then it will not cross the dynamic range to turn on the circuit (dotted lines). When the sensor output crosses that which is required to trigger the circuit, then the program is functional (red). (C) Experimental data is shown where the RBS is modified in the connection of a sensor to an AND gate. The solid line is the predicted fitness curve derived from the transfer functions of the sensor and circuit measured in isolation of the program.

Simulation and Design Environments

Kinetic simulations are commonly used to study the dynamics of molecules in the cells and there are a number of packages from systems biology that aid the modeling of signaling and regulatory networks [76–79]. A limitation in applying these tools to synthetic biology is the need to connect the output of a simulation to a specific design choice that is on the level of the DNA sequence, whether it be a mutation or the choice of a part. In the context of a compiler, the role of simulations would be to predict the dynamics of the assembled genetic circuits and serve as a debugging tool that would aid the user in writing the higher-level language (Figure 2).

A number of software tools have been developed to facilitate the selection and combination of parts gleaned from the Registry of Standard Biological Parts. For example, BioJade and SynBioSS enable the import of BioBricks along with descriptive parameters sets that can form the basis for simulations or algorithms for part selection [80, 81]. Several algorithms enable the selection of parts to match an objective function that describes a desired circuit, such as an oscillator or toggle switch or logic gate [59, 82]. All of these simulation packages suffer from a lack of parts and circuits whose kinetics are well-documented and easily imported.

An underlying assumption of this approach is that parts will quantitatively identical kinetic parameters when measured individually or in the context of different genetic programs. There are several elegant examples where this assumption has been shown to hold, including in the construction of AND gates [4], toggle switches, and feedforward loops [83] through the assembly of different permutations of underlying parts. In our work with the edge detector, we found that the behavior of the program could be explained based on the component circuits [4]. However, it is not clear how often this is true and there are many published (as well as unpublished!) reports where the final behavior could not be predicted based on the components [33, 84, 85]. How to quantitatively measure and report this information is still an open question.

Genetic circuits usually consist of small numbers of molecules, and stochastic effects can dominate. There has been much work performed in the area of stochastic modeling and this has been applied to understanding natural and synthetic genetic networks [4]. A challenge is to convert the results of a stochastic simulation to a design choice; for example, the selection of a promoter. Noise can be controlled through the position of an operator in a promoter [85], and it is influenced by the number of layers in a genetic program [86]. In a theoretical study, Hasty and co-workers demonstrated that the variability of the circuit response can be tuned by independently controlling the transcription and translation rates and DNA copy number [87]. Cytometry data is the most common form of data produced when characterized by a genetic circuit and this contains significant and often unutilized information about the variability in the circuits. Munsky et al. have developed a method that enables the full cytometry distributions to be used to parameterize a genetic circuit, including its stochastic behavior [88].

Optimizing DNA Assembly and Experimental Design

Once the DNA sequence is debugged on the computer, it needs to be constructed. This can either be by automated DNA synthesis or by the robotic assembly of parts. Despite the declining cost of DNA synthesis, the ability to automatically assemble parts using BioBricks and related cloning strategies will be an advantage when it is desired to make many combinatorial variants of a genetic system. This is useful when there is potential degeneracy in the design, allowing multiple constructs to be tested for function. Often, only a subset of the permutations function as expected.

Grammars have been modified to include information for assembly [40]. The inclusion of rules can significantly reduce the construction cost by reducing the number of variants that need to be assembled [38]. This can be further coupled to kinetic simulations to reduce the permutations to those that are likely to perform a given circuit function. For example, Peccoud and co-workers have applied grammars to permutations of the toggle switch in order to identify variants that are likely to perform the correct function [37]. After rules and simulations constrain the set of permutations, a path to fabrication needs to be determined. Densmore and co-workers have developed algorithmic methods to reduce the path length in the number of assembly steps when constructing a genetic system [8]. The same approach can be applied to reducing the cloning steps when using the same set of underlying parts to assemble many systems.

Methods from the statistical design of experiments (DOE) could also be integrated into a CAD program [89, 90]. Of particular interest is the design of sequential experiments, where the design of the next round of DNA constructs is influenced by the results from the previous round. This has been applied to highly-dimensional problems, such as the optimization of fermentation, where there are many variables [pH, feed rate, oxygen, media, etc] that need to be optimized. In the context of genetic programs, the rigorous application of DOE methods would randomize an initial set of constructs. After the constructs were measured for the desired function, this would be fed back to the algorithm to identify the next set of constructs. The process is iterated until the function is optimized. This is a particularly powerful approach when coupled with optimization algorithm, such as Nelder-Mead or ant colonization [91].

IV. Conclusions

Genetic engineering is moving towards becoming an information science. The model of storing and distributing genetic material is slowly loosing relevance. It is routine to outsource the task of constructing DNA from the designer to synthesis facilities. This has created a strong need for computer aided design programs that are able to facilitate the organization and construction of large projects. One the parts are experimentally characterized, it is unnecessary to distribute the DNA. Rather, the CAD program can access the sequence and function information to design a genetic program. The sequence information corresponding to the desired program is sent to a synthesis/assembly facility for construction.

Existing genetic circuits have not been designed to be sufficiently robust to be automatically connected using CAD. These circuits and the genetic parts on which they are based need to be constructed specifically with the intent of coupling with a CAD program. One of the key problems is that there simply are not enough robust circuits to feed into a CAD program. There are significant efforts underway to develop computational methodologies specifically to create the orthogonal parts that would underlie such programs [70, 92, 93]. What is particularly needed are sets of orthogonal transcription factors that bind to different operator sequences. It is the co-development of fundamental circuits with the computational algorithms to assemble them that will allow us to move towards the vision of being able to program living cells.