Synthetic gene circuits enable systems-level biosensor discovery at the host-microbe interface

The composition and function of the gut microbiota are strongly associated with human health, and dysbiosis is linked to an array of diseases, ranging from obesity and diabetes to infection and inflammation. Engineering synthetic circuits into gut bacteria to sense, record and respond to in vivo signals is a promising new approach for the diagnosis, treatment and prevention of disease. Here, we repurpose a synthetic bacterial memory circuit to rapidly screen for and discover new in vivo-responsive biosensors in commensal gut Escherichia coli. We develop a pipeline for rapid systems-level library construction and screening, using next-generation sequencing and computational analysis, which demonstrates remarkably robust identification of responsive biosensors from pooled libraries. By testing both genome-wide and curated libraries of potential biosensor triggers—each consisting of a promoter and ribosome binding site (RBS)—and using RBS variation to augment the range of trigger sensitivity, we identify and validate triggers that selectively activate our synthetic memory circuit during transit through the gut. We further identify biosensors with increased response in the inflamed gut through comparative screening of our libraries in healthy mice and those with intestinal inflammation. Our results demonstrate the power of systems-level screening for the identification of novel biosensors in the gut and provide a platform for disease-specific screening using synthetic circuits, capable of contributing to both the understanding and clinical management of intestinal illness. IMPORTANCE The gut is a largely obscure and inaccessible environment. The use of live, engineered probiotics to detect and respond to disease signals in vivo represents a new frontier in the management of gut diseases. Engineered probiotics have also shown promise as a novel mechanism for drug delivery. However, the design and construction of effective strains that respond to the in vivo environment is hindered by our limited understanding of bacterial behavior in the gut. Our work expands the pool of potential biosensors for the healthy and diseased gut, providing insight into host-microbe interactions and enabling future development of increasingly complex synthetic circuits. This method also provides a framework for rapid prototyping of engineered systems and for application across bacterial strains and disease models.


IMPORTANCE
The gut is a largely obscure and inaccessible environment. The use of live, engineered probiotics to detect and respond to disease signals in vivo represents a new frontier in the management of gut diseases. Engineered probiotics have also shown promise as a novel mechanism for drug delivery. However, the design and construction of effective strains that respond to the in vivo environment is hindered by our limited understanding of bacterial behavior in the gut. Our work expands the pool of potential biosensors for the healthy and diseased gut, providing insight into host-microbe interactions and enabling future development of increasingly complex synthetic circuits.
This method also provides a framework for rapid prototyping of engineered systems and for application across bacterial strains and disease models.

INTRODUCTION
Recent advances in our understanding of both the human microbiota and biological engineering techniques have created myriad possibilities for the development of synthetic microbes for in vivo clinical applications (1,2). Bacteria residing in the gut are uniquely positioned to monitor a variety of host, microbial, and environmental factors and to respond to changes in intestinal homeostasis. Engineered gut bacteria also offer the potential for in vivo production and delivery of therapeutics (2).
Environment-and disease-responsive functions, which could minimize both the metabolic burden of engineered systems on the bacteria and off-target effects on the patient, offer exciting prospects for clinical applications. To this end, recent in vivo approaches have developed sensors responding to inflammation (3,4), intestinal bleeding (5), and pathogen quorum-sensing systems (6,7). However, the construction of disease-responsive circuits in bacteria has been hindered by the limited number of characterized bacterial systems that can be reliably employed as sensors.
Mining the genomes of native gut bacteria is a promising approach for discovering new sensors that respond under conditions of interest, such as in the healthy or diseased gut. To date, these efforts have largely relied on transcriptome sequencing and proteomics of fecal samples. However, to obtain an instantaneous snapshot of bacterial behavior inside the gut using these techniques, invasive sampling is required (i.e., colonoscopy and biopsy). Furthermore, transient or low-abundance signals may not be detected, and any responsive genetic elements identified with these techniques may not function predictably when employed in synthetic circuits.
Approaches such as in vivo expression technology (IVET) and recombinase-based IVET (RIVET) have also been used to track in vivo-expressed genes non-invasively, but detect only constitutive expression (for IVET) and may have high false-positive rates (8).
Nevertheless, these technologies show the potential for systems-level approaches to interrogate the behavior of the microbiota.
We have previously developed an approach for non-invasive measurement of bacterial responses in the gut, based on a robust synthetic memory circuit, which records environmental stimuli via a transcriptional trigger (3,9). When activated, the trigger turns on a memory switch, which can retain the memory-on state for over a week in the gut (9). After the bacteria pass through the host, their memory state can be determined via reporter gene expression, enabling non-invasive readout of transient signals within the gut. The circuit can maintain functional and genetic stability during six months' colonization of the mouse gut, demonstrating its suitability for longitudinal studies and its potential to support the development of stable, engineered biosensors for in vivo deployment (3).
Here, we adapt this memory circuit for parallel, high-throughput screening of hundreds of potential triggers. We apply this method to identify new biosensors responding specifically to the in vivo gut environment. Through comparison between healthy mice and those suffering from inflammation, we also identify triggers that respond differentially during disease. Together, these results provide a platform for in vivo non-invasive biosensor discovery and longitudinal testing.

RESULTS
Bacterial memory as a biosensor screening tool. To enable screening of new potential biosensors in parallel, we modified our previously-developed E. coli memory circuit, which is based on the λ phage lysis-lysogeny switch (Fig. S1A) (9). This modified circuit is referred to as the high-throughput memory system (HTMS) (Fig. 1A).
Both the original memory circuit and the HTMS consist of a trigger-based on a transcriptional promoter activated in the presence of a certain stimulus-and a memory switch. The memory-on and memory-off states of the switch correspond to the mutuallyrepressive proteins Cro and CI, respectively. Additionally, a b-galactosidase (LacZ) reporter is produced in the memory-on state.
One key modification for the HTMS is the triggering of memory using a dominantnegative mutant of the cI gene (cIDN), instead of cro used in the original trigger. When the trigger promoter is activated by a stimulus, CIDN monomers, which have an N55K mutation in the DNA binding region (10) dimerize with wild-type (WT) CI monomers expressed in the memory-off state, creating heterodimers that are deficient in DNAbinding. This leads to derepression of PR and transition to the memory-on state. As with the CI used in the memory element, CIDN carries a mutation to prevent RecA-mediated cleavage (ind-) (11).
Use of CIDN in the trigger ensures that there is no delay of switching to the memory-on state in the case of high, or constant, expression of the trigger promoter. To test this, a Ptet trigger driving cIDN or cro was integrated into E. coli K-12 MG1655 and NGF-1 strains containing a memory element. When grown in the presence of a high concentration (100 ng/ml) of anhydrotetracycline (aTc) inducer, cIDN-triggered strains showed switching to the memory-on state, while cro-triggered strains switched only after a subsequent period of growth in the absence of aTc (Fig. 1B).
The original memory circuit expresses a lacZ reporter gene for screening on indicator plates (9). To analyze pooled libraries containing many strains with varied trigger promoters, the HTMS also expresses a spectinomycin-selectable resistance gene (aadA) in the memory-on state.
The HTMS allows faithful selection of memory-on colonies with spectinomycin treatment. Plating of fully memory-off, fully memory-on, and 50-50 mixed cultures of PAS810, on indicator plates with and without spectinomycin further demonstrated that all spectinomycin-selected colonies were also lacZ positive (Fig. 1C). Spectinomycin did not yield false-positive results by inducing memory switching (fully memory-off: 0% +/-0% SE, n = 8), nor excessive false-negative results through inhibition of memory-on bacterial growth (fully memory-on: 93.0% +/-4.2% SE, n = 8; 50-50 mix: 49.7% +/-1.5% SE, n = 5) (Fig. 1C). Together these results demonstrate the ability of the HTMS to measure biosensor response and allow selection for downstream pooled analyses.

Biosensor library construction. To build biosensor libraries for genomic
integration, we adapted a Tn7 transposon genome insertion plasmid (12) for rapid Golden Gate assembly (13) of bacterial promoters upstream of the cIDN trigger and insertion into the genome of memory bacteria ( Fig. 2A and S2). The modularity of this cloning strategy allows for adjustment of trigger sensitivity through incorporation of ribosomal binding site (RBS) variants, which vary the translation rate of mRNA transcripts ( Fig. 2A). To test this concept, triggers consisting of a Ptet promoter combined with nine synthetic RBS sequences-previously demonstrated to vary widely in their translation strength (14) (Fig. 2A)-were constructed and inserted into the genome of HTMS bacteria, and the HTMS response to varying concentrations of aTc (0-100 ng/ml) was characterized (Fig. 2B). The RBS variants differed in their extent of memory induction at 0.1-10 ng/ml aTc (EC50 ranging from 0.5 to 4.1 ng/ml for responsive strains), illustrating our ability to tune trigger sensitivity.
We explored two approaches for generating biosensor libraries: 1) a genomewide collection of trigger promoters that would enable screening of a bacterium's entire range of transcriptional responses (MG1655 library), and 2) a curated collection of promoters with sensitivity variants aimed at detecting inflammation (Nissle 1917 library).
Both libraries were assembled into an HTMS-containing E. coli NGF-1 strain (PAS811), as NGF-1 has proven to be an efficient and persistent colonizer in the mouse gut (3,9,15). The genome-wide library was sourced from a previously published collection of 1600 unique promoters from E. coli K-12 MG1655 (16). Promoters and their wild-type RBSs were amplified by PCR from this collection, assembled into our transposon plasmid, and integrated as triggers into the genome of PAS811. Because our method focuses on detecting off-to-on sensor transitions, the resultant library was further subsampled by pooling 500 colonies that were LacZ-negative under routine in vitro culture. Sequencing confirmed the presence of 155 unique strains in this final genomewide library.
Our second library was constructed with a subset of promoters sourced from the human probiotic E. coli Nissle 1917, which are involved in anaerobic respiration of sulfur-or nitrogen-oxides or nitrate, produced by the gut epithelium during inflammation (17,18). For each promoter, a trigger with its wild-type RBS, as well as with five different synthetic RBSs (MCD5, MCD10, MCD15, MCD17 and MCD23) (14) were included to tune sensitivity. Sequencing confirmed that the assembled library contained 61 unique strains out of 66 total designed constructs.

Parallel analysis faithfully reports biosensor response.
To screen for biosensor response, HTMS libraries are exposed to a condition of interest (Fig. 3A), and put through a processing, sequencing and analysis pipeline (Fig. 3B). After exposure, HTMS bacteria are recovered and cultured. The initial culture is split into two and backdiluted, and one of the two new cultures is subjected to spectinomycin selection.
Following selection, the trigger regions of both cultures are sequenced and analyzed to produce an odds ratio for each trigger promoter in the library, corresponding to that trigger's memory state. To calculate odds ratios, results are normalized to a positive normalization strain (PAS812) which remained in a memory-on state (Fig. 3B).
Pooled library analysis is predictive of the on/off state of HTMS bacteria. The Nissle 1917 library was cultured aerobically in liquid media and analyzed to obtain odds ratios as described above. Concurrently, individual strains from this library were grown on indicator plates to assess each strain's in vitro memory state directly. Both tests showed strong agreement, with strains that were LacZ-positive also displaying higher odds ratios (Fig. 3C). Receiver operating characteristic analysis confirmed efficient distinction between memory-on and memory-off states, with an odds ratio of approximately 0.02 delimiting the boundary (Fig. 3D). This confirmed our sequencing method as a reliable indicator of biosensor memory state.
Differential biosensor response in the healthy mouse gut. To screen for biosensor response to growth within the murine gut, the MG1655 library was administered to specific-pathogen free (SPF) mice by oral gavage (~10 7 bacteria/mouse), and fecal samples were collected over one (n=2) or seven (n=3) days.
High library diversity was maintained in both experiments (92% and 82% of strains identified in gavage samples present at experiment endpoint, respectively). Analysis of HTMS strains recovered from gavage suspension and fecal samples identified 23 unique strains that responded specifically to growth within the gut (gavage: odds ratio < 0.02; fecal samples: ³ 1 timepoint odds ratio ³ 0.02 and p < 0.05) ( Fig. 4A and 4B; Data Set S1 and S2). Five strains (containing ydiL, ydjL, gatY, gcvA and ubiG triggers) were detected in the memory-on state in at least 4 of 5 mice. The two most consistent responders (ydjL and ydiL triggers) were selected for follow-up testing.
To validate the response of the ydiL and ydjL triggers during gut transit, memory bacteria containing these triggers (ydiL: PAS813; ydjL: PAS814) or a promoterless cIDN gene (negative control: PAS815) were administered to SPF mice as monocultures.
Together these results demonstrate the ability for HTMS analysis to rapidly identify biosensors in vivo and the power of varying trigger sensitivity to tune the strength of biosensor response.

Identification of disease-specific biosensors.
To look for sensors responding differentially to disease, we compared the response of the Nissle 1917 library in healthy mice (n = 4; as previously displayed in Fig. 4D, 4E and S3, and Data Set S3) to a murine intestinal inflammation model (Fig. 5A, 5B and S4, and Data Set S3). After library gavage, SPF mice were provided water containing 4% w/v dextran sulfate sodium (DSS) ad libitum for five days, and HTMS analysis was performed on fecal samples throughout. Weight loss (Fig. S4A), colon length reduction at endpoint (Fig.   S4B), and increased E. coli CFU counts (Fig. S4C) were all consistent with increasing inflammation throughout the experiment. Six strains registered memory-on at more timepoints in the DSS-treated group than in the control group ( Fig. 5A and 5B). In particular, the ynfE17 trigger strain (PAS819) responded specifically in DSS-treated mice (control: no response; DSS-treated: 93% of timepoints with odds ratio ³ 0.02 and p < 0.05) ( Fig. 5A and 5B, and Data Set S3).
To validate ynfE17 (PAS819) response to inflammation, the strain was administered to SPF mice as a monoculture, after which a subset of the mice was provided water containing 4% DSS. Fecal samples were cultured for memory bacteria on indicator plates for 7 days following gavage. As above, body weights (Fig. S4D), post-dissection colon lengths (Fig. S4E) and CFU counts of HTMS bacteria (Fig. S4F) reflected increased inflammation in DSS-treated mice. Confirming the screen results, ynfE17 showed increased response in DSS-treated mice compared to untreated controls, with the greatest difference between groups at days 6 and 7 (at Day 6, +DSS: 24% +/-9% SE, n= 4; control: 5% +/-2% SE, n = 8) (Fig. 5C). The strong response of PAS819 in the absence of DSS in one of the control group mice indicates that in vivo 13 conditions other than DSS treatment can activate the ynfE17 trigger. In vitro anaerobic growth both in rich media and in cecal media did not induce ynfE17 (0 +/-0% SE; n = 7; 0 +/-0% SE, n = 3, respectively), in contrast to ynfE15 (Fig. 4G), suggesting a lower nitrate threshold for ynfE17 activation and that individual bacteria may experience low nitrate conditions within the inflamed mouse gut.

DISCUSSION
Here, we have expanded the use of a robust genetic memory circuit to assess the in vivo responses of multiple bacterial sensors in parallel. The original memory circuit (9) was altered to allow off-to-on transitions in the presence of constant induction, and to enable selection of memory-on strains from pooled libraries using spectinomycin.
We developed a screening, sequencing and analysis pipeline to efficiently identify in vivo-responding trigger-RBS combinations. Tests conducted with both genome-wide and curated libraries containing hundreds of sensors demonstrated that our method is an effective, non-invasive way to identify new biosensors responding in the gut. We identified and validated biosensors responding to growth in the healthy mouse gut and preferentially in inflamed conditions. Together, these results demonstrate the power of tuning trigger sensitivity to physiological responses and for the HTMS to assess unique features of the mammalian gut environment in vivo.
One advantage of our method is its ability to discover sensors that could not be rationally designed based on existing knowledge, presenting an opportunity to apply the rapidly increasing but largely uncharacterized genetic diversity identified through microbiome sequencing. For instance, the two validated healthy-gut sensors from our E. coli MG1655 library (PAS813 and PAS814) are derived from operons with largely uncharacterized function and regulation. PAS814 is triggered by the promoter of the ydjLKJIHG operon, which putatively includes a kinase, a transporter protein, two dehydrogenases, an aldolase and an aldo-keto reductase. Only the activity of the aldoketo reductase, YdjG, has been confirmed through reduction of methylglyoxal (20,21).
Interestingly, a previous analysis of E.coli protein expression in germ-free mice showed that YdjG was expressed 3.5-fold higher in the cecum than in vitro (22). Another gene which has been studied in this operon, ydjK, may play a role in osmotolerance, showing a 50% increased growth rate in high-salt media when overexpressed in E. coli (23). It is not known whether methylglyoxal or osmotic stress can directly trigger transcription of the ydjLKJIHG operon. However, methylglyoxal occurs in many foods and is also produced by intestinal bacteria (24); it can also inhibit bacterial growth, suggesting a possible motivation for expression of ydjLKJIHG in the gut.
Promoters derived from three unique Nissle 1917 operons (ynfEFGH, torCAD, and yeaR-yoaG) showed robust memory response in the healthy mouse gut (Fig. 4D and 4E). The ynfEFGH operon encodes a DMSO reductase which has also been shown to reduce selenate (25,26). It is activated by FNR in anaerobic conditions and repressed by phosphorylated NarL in the presence of nitrate (19), which was further confirmed by our in vitro tests (Fig. 4G).
Tuning of trigger sensitivity (e.g., by RBS modulation) is important for generating responses to physiological conditions of interest and for successful application in synthetic engineered circuits. As we observed, RBS tuning can be used to increase the response of the ynfE promoter to as high as 100% in healthy mice (PAS816; Fig. 4F), and to adjust the response to distinguish the inflamed gut state (PAS819; Fig. 5C).
Importantly, the sensors we identify can be used directly in downstream applications with the memory circuit. This provides an engineering advantage over any responsive genetic elements identified through analysis in their native context, for which incorporation into synthetic circuits would routinely require additional optimization.
Inflammation leads to an increase in nitric oxide produced by the host, which generates nitrate in the intestine (27). However, because the ynfE promoter is activated by a decrease in nitrate, our results suggest that DSS-induced inflammation may lead to lower levels of free nitrate available to E. coli NGF-1, possibly due to increased local competition for nitrate via respiration by NGF-1 and other Enterobacteriaceae. This idea is supported by our observation of increasingly higher NGF-1 bacterial loads in fecal samples of DSS-treated mice ( Fig. S4C and S4F), suggesting a bloom of E. coli-and potentially other Enterobacteriaceae capable of nitrate respiration-correlated with increasing duration of DSS treatment. This is consistent with previous descriptions of E.
coli experiencing a growth advantage due to anaerobic respiration of host-derived nitrate (27). Thus, we hypothesize that PAS819 responds in DSS-treated mice specifically through sensing inflammation-induced changes in its own microenvironment. The potential to engineer synthetic circuits into commensal gut bacteria is a promising new approach to the management of intestinal disease. Synthetic biology is just beginning to tap into the evolutionary breadth of capabilities found in natural systems, and our method represents a practical means for expanding the toolkit of useful sensors for in vivo application.

MATERIALS AND METHODS
Media, culture conditions. Unless otherwise mentioned, bacterial cultures were grown at 37°C in LB broth or agar (10 g/L NaCl, 5 g/L yeast extract, 10 g/L tryptone).
Mixed liquid cultures (i.e., libraries) were grown in LBPS, which contains Peptone Special (Sigma) instead of tryptone. To quantify memory response on indicator plates, agar was supplemented with 60 µg/ml X-gal.
High-throughput memory system (HTMS) construction. The spectinomycin resistance gene, aadA, was added downstream of the PL promoter in the original memory element (9) by overlap extension PCR and genomically integrated by λ Red recombineering (28) into E. coli TB10 (29) between mhpR and lacZ, driving endogenous lacZ as a memory-on reporter. From TB10, transfer into streptomycin-resistant E. coli

In vivo induction of strains and libraries. The Harvard Medical School Animal
Care and Use Committee approved all animal protocols. Experiments were conducted in female 7-to 14-week-old BALB/c (Charles River; MG1655 library) or C57BL/6J (Jackson; Nissle 1917 library) mice. Before experiments, all mice were confirmed to be absent of native streptomycin-and chloramphenicol-resistant flora. Food and water were removed ~4 h before each gavage; water was replaced immediately, and food was replaced < 2 h following gavage.
Concurrently, an overnight culture of the positive normalization strain, PAS812 was back-diluted 1:100 into LBPS-chloramphenicol. Cultures were grown 4 h, or until OD600 ~1. PAS812 OD600 was adjusted to match the library culture, then diluted 1:1000 into the library culture. The resulting mix was back-diluted 1:1000 into 50 ml LBPSchloramphenicol, and immediately split into two 25 ml volumes. Spectinomycin (50 µg/ml) was added to one culture, and both were grown overnight before centrifugation to collect bacterial pellets, which were stored at -80°C.