Multi-arm RNA junctions encoding molecular logic unconstrained by input sequence for versatile cell-free diagnostics

Multi-arm junctions for controlling gene expression

Our general strategy for regulating gene expression using multi-arm junctions is illustrated in Fig. 1. A multi-arm structure is placed at the 5’ end of an mRNA and is followed by the coding sequence of the regulated gene. This configuration establishes strong secondary structures in the mRNA that conceal the ribosome binding site (RBS) and start codon (AUG) necessary to initiate translation. As a result, translation is repressed unless complementary input RNAs bind and unwind the structure. The stem-loop arms of the structure act as sensors that provide binding sites for each of the input RNAs. The form of molecular logic evaluated by the multi-arm assembly is programmed by controlling the length and number of the sensor arms. OR logic is implemented using short sensor arms that remain ‘unlocked’ and available for input RNA binding. Hybridization of either input RNA unwinds the cognate sensor arm through to the base stem to activate translation. To implement AND logic, all but one sensor arm is lengthened to establish a ‘locked’ configuration that prevents binding by the corresponding input RNA. Binding of successive inputs unlocks additional sensor arms for input binding and ultimately allows the RNA strand to be fully unwound, activating translation of the downstream gene. Compared to other approaches for implementing logic-gated gene expression with RNAs, this strategy abolishes the need for any sequence correlations between input RNAs for AND operations. Moreover, it does not require extended N-terminal residues to be added to the output protein for OR operations, which can interfere with protein folding. Implementing this strategy, however, first required that we develop loop-based RNA–RNA interactions that functioned reliably in vivo to enable unwinding of the junction structure through binding to the sensor arms.

Loop-initiated translational activation

We thus developed a set of riboregulators designed to be integrated into the stem-loop regions of multi-arm junctions. While many recent high-performance riboregulators, such as toehold switches⁷, have relied on single-stranded toehold domains to initiate reactions^7,36,38, we hypothesized that long loop domains could be utilized to provide similar performance. Such long loops would provide a strong thermodynamic driving force to initiate RNA–RNA interactions and provide an input RNA binding site that is sufficiently labile and unconstrained to offer good reaction kinetics. Figure 2a shows the resulting loop-initiated RNA activators (LIRAs) that feature hairpins with extended loop domains regulating the expression of a downstream output gene. In the absence of the input RNA, translation by the LIRA is strongly repressed by sequestering both the RBS and the start codon of the output gene within the RNA duplex of the hairpin structure. A long loop domain a* of 21 nt is incorporated into the hairpin structure to promote the initial interaction between the LIRA and the activating RNA. After binding to the loop through the complementary a* sequence, the input RNA is designed to bind into the b* domain at the top of the hairpin stem, disrupting the existing base pairs and driving apart those located lower in the hairpin stem. Importantly, this effect enables the release of base pairs, in this case the RBS and the start codon that are completely unrelated to the sequence of the cognate input RNA. Thus, LIRAs can accommodate input RNAs without imposing any sequence constraints, and they can regulate a variety of different proteins without requiring modifications to the N-terminal sequence. In comparison, the design of toehold-switch riboregulators causes three to four N-terminal residues in the output protein to be defined by the sequence of the input RNA (Extended Data Fig. 2a), which prevents recognition of input RNAs that would generate stop codons upstream of the output gene⁷. To enable efficient expression and testing in vivo, four bulges were incorporated into the LIRA hairpin structure to reduce the likelihood of premature rho-independent transcriptional termination and to increase the thermodynamics driving the input–LIRA reaction. A 6 bp clamp domain was also added immediately after the start codon in the LIRA stem to reduce the likelihood of translational leakage (Fig. 2a and Extended Data Fig. 2b,c).

A library of 24 different LIRA sequences were designed de novo using the NUPACK software package³⁹ (Methods) and plasmids were constructed to express the input and LIRA transcripts using T7 RNA polymerase in E. coli BL21 Star DE3 cells (Supplementary Tables 1 and 2 for primer and LIRA sequence information). These experiments employed green fluorescent protein (GFP) as the reporter protein and measured fluorescence from the cells using flow cytometry. ON/OFF ratios for the LIRAs were determined by measuring the ON-state GFP expression in the presence of the cognate input and dividing it by the GFP expression measured in the OFF state where a non-cognate input was expressed in the cell (Fig. 2b, LIRA ON and OFF values are shown in Supplementary Fig. 1). We found that 16 out of 24 of the LIRAs provided ON/OFF ratios over 50-fold, with the highest one yielding an ON/OFF ratio of ~350-fold. Additional tests using input RNAs complementary to different regions of the LIRA hairpin and different loop domain lengths revealed that the base LIRA design with a loop of 21 nt and a 31 nt input RNA provides the best overall performance (Supplementary Figs. 2 and 3, and Tables 3 and 4 for sequence information).

We observed that multiple LIRAs provided very low translational leakage in the absence of the input RNA. Figure 2c shows the OFF-state GFP fluorescence of eight different LIRAs with ON/OFF ratios greater than 50 compared with the autofluorescence of cells lacking GFP and a set of previously reported toehold-switch riboregulators with wide dynamic range⁷ (Extended Data Fig. 3a). We found that the toehold switches in the OFF state yielded 2- to 3-fold higher fluorescence than the cells lacking GFP plasmids. In contrast, all eight of the LIRAs examined provided fluorescence leakage that was statistically indistinguishable from the background cellular fluorescence, with P > 0.067 for all the LIRA devices in the plot compared with cell autofluorescence. Comparison of ON-state signals showed that three out of the four toehold switches provided higher signal output than the LIRAs (Extended Data Fig. 3b). To explain the very low OFF-state signals, we hypothesized that low-leakage LIRAs could be making use of a combined translational and transcriptional regulation mechanism to yield virtually undetectable leakage (Extended Data Fig. 3c), assisted by the strong secondary structure of the LIRA hairpin (see Extended Data Fig. 3d for comparison of minimum free energies for LIRA and toehold-switch hairpins). To verify this hypothesis, we performed reverse transcription quantitative polymerase chain reaction (RT-qPCR) experiments to measure the concentration of LIRA RNAs and cognate and non-cognate input RNAs expressed from cells in the ON and OFF states, respectively (Supplementary Fig. 4). We found that expression of the LIRA transcript with a non-cognate input was only 25% of that measured for the transcript with a cognate input, confirming that part of the LIRA regulation is due to transcriptional control. We also studied a set of four LIRA variants that contained different sequences in the stem below the start codon (see Supplementary Table 5 for sequence information), which in turn modify the N-terminal residues in the output protein. We found that these clamp sequence changes did not impact the OFF-state signal of the LIRAs (Extended Data Fig. 2c), but they did cause variations in ON-state expression levels ranging from 40% to 230% of the parent LIRA (Extended Data Fig. 2d). Despite these variations, all of the devices with clamp modifications displayed ON/OFF ratios of at least 50-fold (Extended Data Fig. 2e), indicating that changes in clamp sequence and the N-terminal residues are well tolerated by the riboregulators.

Foreshadowing their use in multi-arm junctions, we evaluated LIRA orthogonality by measuring the crosstalk observed between the 16 devices providing the widest dynamic range. A 16 × 16 matrix of pairwise LIRA–input RNA interactions was measured by transforming cells with different combinations of plasmids. Figure 2d shows the measured crosstalk between the devices. Cognate interactions along the diagonal are normalized to 1 for the riboregulators in their ON states, while off-diagonal, non-cognate interactions reflect the percent activation with respect to the ON state. We found that crosstalk from the non-cognate inputs was very low, less than 4% in nearly all cases, with a single strong off-target interaction observed between LIRA 22 and input 18 showing 5.6% crosstalk. Thus, the LIRAs provided a set of 15 orthogonal devices for regulation of gene expression in vivo.

On the basis of their low crosstalk and lack of sequence constraints, we also investigated whether LIRAs could be designed to detect mRNAs within the cell. A set of LIRAs targeting regions of low secondary structure in the mRNAs for mCherry and the antibiotic resistance genes aadA, ampR and cat, conferring resistance to spectinomycin, ampicillin and chloramphenicol, respectively, was investigated (see Supplementary Table 6 for sequence information). All LIRAs were based on a high-performance design identified during library screening and were generated simply by replacing the original target-binding site with the reverse complement of the mRNA target site. We found that all four mRNAs could be readily detected using the LIRAs and provided ON/OFF GFP levels ranging from 22- to 38-fold (Fig. 2e).

Multi-arm RNA junctions for in vivo molecular logic

Having developed a set of orthogonal LIRAs lacking sequence constraints, we next integrated them as sensor modules into the multi-arm RNA junction structures for computing intracellular OR and AND logic expressions. The sensor arms of the resulting logic gate RNA are each capped by different LIRA modules and designed to direct the unfolding of the structure as input RNAs bind to the gate RNA. Two-input OR logic devices were constructed upstream of a GFP reporter using a three-arm junction containing a pair of LIRA sensor arms (Fig. 3a). The base arm contains the RBS and start codon signals topped by the LIRA arms to provide binding sites A* and B* for interaction with the complementary input RNAs A and B. Binding of either input RNA disrupts the cognate stem-loop structure and further draws apart the base arm to reveal the RBS and start codon for translation initiation. To increase translational output for this input and reduce the likelihood of transcriptional regulation, we also incorporated a hairpin reconfiguration domain (indicated in dark blue in Fig. 3a) that generated an additional stem-loop upon binding of input A to the gate RNA. This newly formed stem disrupted the bottom grey portion of the input B LIRA module, providing a single-stranded region upstream of the RBS and greater space to better accommodate the ribosomal footprint. During transcription, the hairpin domain can also help delay formation of the strong LIRA stem-loop structures to discourage transcriptional termination.

We tested the two-input OR device by transcribing the input and gate RNAs off separate high- and medium-copy plasmids, respectively, in E. coli (see Supplementary Table 7 for sequence information). Using flow cytometry, we found that GFP expression increased by 38- to 84-fold when any combination of the two-input RNAs was expressed (Fig. 3b,c). We also constructed a three-input OR gate RNA using three orthogonal LIRA modules (Fig. 3d). This four-arm junction system contained a base arm with the RBS and the start codon and inserted the LIRA stem loop for input C between modules for inputs A and B. Similar to the two-input device, the two left LIRA stem loops also contained hairpin reconfiguration domains to enable increased translation upon binding of inputs A and C. This circuit also performed as expected in vivo, with low expression for the null-input logical FALSE case and 6- to 19-fold increases in expression when the input RNAs were expressed in any combination (Fig. 3e,f).

Multi-arm junctions for AND logic employ sensor arms of different strengths to implement locked and unlocked LIRA sites (Fig. 4a). The gate RNA contains a base arm topped by a weak, unlocked sensor arm for LIRA module A* and a strong, locked sensor arm for LIRA module B*. The locked arm also conceals the RBS and the start codon translation initiation signals within an RNA duplex. For the logical FALSE case when only input B is expressed, the locked stem-loop structure and the base stem are designed to be too thermodynamically stable to be disrupted by input RNA B, preventing system activation. However, if input A interacts with the gate RNA first, its binding energy is sufficiently strong to disrupt both the left stem loop and the base stem of the gate RNA. Unwinding the base stem in turn unlocks the LIRA B* module, making it sufficiently weak to interact with input B. Thus, when input B is also expressed, the B* module is completely disrupted and the RBS and start codon are exposed for translation of the GFP reporter gene. Unlike the LIRA OR gates, use of locked sensor arms for LIRA AND gates does add multiple N-terminal residues to the output protein (Fig. 4a and Extended Data Fig. 1c).

We tested the two-input AND device in E. coli using different combinations of input RNAs (see Supplementary Table 7 for sequence information). We found that only strong GFP reporter expression was observed for the logical TRUE case with both inputs expressed. GFP expression increased by 79-fold for the TRUE case compared with the case with neither input transcribed (Fig. 4b,c). In addition, we found that translational leakage in the presence of input RNA B was low, 43-fold lower than the TRUE state, indicating that the extended stem-loop structure effectively blocked access of the transcript to the gate RNA. We also extended the AND ribocomputing strategy to three inputs using the four-arm junction structure shown in Fig. 4d. This device incorporated the binding site for input C to lock modules A* and B* and prevent them from interacting with their corresponding input RNAs without expression of input C. To increase translational output and encourage stem-loop disruption, hairpin reconfiguration domains were added to the arms for inputs A and C. This device also functioned properly in E. coli, providing a 36-fold increase in GFP expression in the logical TRUE case with all three inputs expressed compared with the null-input case (Fig. 4e,f). Leakage in all logical FALSE conditions was low, with the TRUE state providing at least 16-fold higher GFP output in all cases.

Validation of LIRAs in paper-based diagnostics

The sensing and logic capabilities of LIRAs and multi-arm junction RNA structures also make them promising devices for use in paper-based cell-free systems, where they can be used as diagnostics without the need for expensive equipment and provide results that can be detected by the naked eye^14,15,16,17. Since RNA–RNA interactions differ in cell-free reactions compared with the cytoplasmic environment, we first tested LIRAs by using them as riboregulators in paper-based reactions (see Supplementary Table 8 for sequence information). These reactions employed freeze-dried cell-free transcription–translation reactions along with LIRA plasmids, the lacZ ω subunit and the lacZ colourimetric substrate CPRG (chlorophenol-red-β–d-galactopyranoside) deposited onto 2-mm-diameter paper discs (Fig. 5a). At the time of use, the paper discs were rehydrated with solutions containing RNAs for detection by the embedded LIRA riboregulators. We first tested the paper-based reactions with LIRAs that showed wide dynamic range during in vivo experiments. However, in the cell-free reactions, they were unable to be turned on by their cognate input RNAs. To increase translational output and encourage stem-loop disruption, a hairpin reconfiguration domain was added to the 5’ end of each LIRA sensor (Fig. 5b). Applying synthetic viral RNA targets to a final concentration of 5 µM, we found that the updated LIRA pathogen sensors provided strong increases in absorbance at 575 nm wavelength as the yellow-to-purple CPRG cleavage reaction was carried out by lacZ (Fig. 5c). Reactions with the pathogen RNAs turned to the expected pink or purple colour as the reactions proceeded, while those without the pathogen RNAs remained yellow to yellow-pink in colour depending on the sensor (Fig. 5c, bottom).

To enable detection of RNAs at the concentrations typically present in clinical samples, we used nucleic acid sequence-based amplification (NASBA) to amplify low-concentration pathogen RNAs before use in the paper-based assays. In NASBA, a combined reaction featuring reverse transcription, T7 RNA polymerase, RNase H and DNA primers that incorporate the T7 promoter sequence is used to generate multiple RNA copies from a starting RNA template. Synthetic RNA targets from norovirus and yellow fever virus (YFV) were supplied to NASBA reactions at an initial concentration of 200 aM and amplified over 2 h at 41 °C. We found that both pathogen RNAs could be detected in the colourimetric paper-based reactions following NASBA (Fig. 5d).

In addition, we applied the assay to clinical serum samples positive and negative for the dengue virus. The serum samples were first diluted by 10-fold into water and heated to 95 °C for 2 min to release the viral genome from the capsid. The RNA was then amplified using NASBA and applied to the paper-based LIRA sensors. We found that LIRAs could unambiguously identify the clinical dengue sample through the resulting purple colour. To determine the detection limit of the dengue assay, we carried out a series of NASBA/LIRA reactions with synthetic dengue target RNA concentrations ranging from 200 fM down to 0.2 aM. We found that the dengue transcript could be detected down to concentrations as low as 20 aM in the NASBA reaction, which corresponds to 12 RNA copies per µl of reaction (Fig. 5e).

Paper-based diagnostic with embedded molecular logic

Diagnostic devices that combine visible readouts with the ability to perform information processing on biomolecular inputs have the potential to improve assay capabilities by expanding the number of pathogens a single test can detect, reducing false positives, and lowering assay complexity and cost. To demonstrate the potential of such logic-enabled paper-based diagnostic devices, we carried out proof-of-concept studies exploiting the logic capabilities of multi-arm junction molecular logic for HIV and SARS-CoV-2 detection. HIV continues to be a major global health threat with HIV-1 group M being the predominant cause of infections worldwide^40,41. Within group M, there are nine different subtypes with genetic distances of 25% to 35% and prevalences that vary depending on the geographic region. HIV-1 subtype C causes >50% of infections worldwide and circulates mostly in India and regions of Africa, while HIV-1 subtype B predominates in Europe and the Americas⁴². We thus aimed to develop a logic system capable of detecting both HIV-1 subtype B and C using a single OR operation, which could be deployed in an area such as Southern Brazil where both subtypes are common⁴³.

To implement the system, we first identified conserved regions in the genomes of HIV-1 subtypes B and C to use as circuit input RNAs. Complementary sequences for these inputs were then incorporated into a two-input OR three-arm junction gate RNA (Fig. 6a, see Supplementary Table 9 for sequence information). To ensure the system functioned properly in paper-based cell-free reactions, the binding site for the input RNAs was extended so that it included both the loop domain and the entire stem of the LIRA module. The gate RNA was transcribed in the cell-free reactions and supplied with the HIV-1 subtypes B and C input RNAs. For both inputs, output of the lacZ α subunit was produced as evidenced by increased production of the purple cleavage product in the paper-based reactions (Fig. 6b). Reactions lacking either input RNA remained yellow.

We next made use of AND logic operations to implement RNA devices for SARS-CoV-2 detection. SARS-CoV-2, which was first reported in 2019 in Wuhan, China, has now become a global pandemic with over 100 million reported cases and over 3 million deaths worldwide according to data from the Johns Hopkins Coronavirus Resource Center. SARS-CoV-2 can be transmissible even before any symptoms have developed^44,45 and studies have shown that many patients who test positive for the virus do not show any symptoms⁴⁶. These factors have allowed the pandemic to take hold and emphasize the importance of developing diagnostic assays that can be widely deployed to detect SARS-CoV-2, even in carriers who do not have any signs of illness.

Following the US Centers for Disease Control and Prevention (CDC) recommendations⁴⁷, SARS-CoV-2 infections are often identified by amplification of two selected regions of the virus nucleocapsid (N) gene, 2019-nCoV_N1 and 2019-nCoV_N2. RT-qPCR is the most common method of detection of SARS-CoV-2 given its excellent specificity and sensitivity. However, it requires well-trained personnel and expensive equipment, which makes virus detection more challenging in rural areas with limited medical resources and requires additional time to ship samples to centralized facilities. Previous paper-based cell-free assays have been limited to detecting only a single pathogen target sequence at a time, and parallel assays that detect target RNAs in separate reactions can suffer as a result of differences in riboregulator activation speeds and lead to increased assay cost.

To overcome these issues, we combined AND logic multi-arm junctions with isothermal amplification reactions to simultaneously detect two different SARS-CoV-2 N gene sequences using a single paper-based readout reaction. The resulting two-input AND gate RNAs contained a hairpin reconfiguration domain to encourage binding between the gate RNA and the input viral RNAs (Fig. 6c, see Supplementary Table 9 for sequence information). We first evaluated several devices using synthetic targets and identified two with the best performance. Gate RNA N1*N2* recognizes the antisense sequences in regions N1 and N2 of the SARS-CoV-2 N gene, with the left and right sensor arms targeting N1* and N2*, respectively (Fig. 6d). Similarly, gate N2*N1 targets the antisense sequence of the N2 region with the left sensor arm and the sense sequence of the N1 region with the right sensor arm (Fig. 6e). Both devices show clear colour changes in the presence of the two-input RNAs, but did not activate when one or both inputs were absent, thus carrying out AND logic (Fig. 6d,e).

We then designed specific NASBA primer pairs for each of the two devices to amplify the input RNAs from the SARS-CoV-2 genome. NASBA reactions were performed using heat-inactivated SARS-CoV-2 virus particles at different concentrations. We found that gate N1*N2* performed better than gate N2*N1 and enabled detection of SARS-CoV-2 down to concentrations of 20 aM in the NASBA reactions when viewed by the naked eye (Fig. 6f,g), a concentration that is within the range necessary for detecting the virus in clinical samples⁴⁸. Using a plate reader, SARS-CoV-2 down to a concentration of 2 aM in the NASBA reaction could be distinguished. We then tested six positive saliva samples from SARS-CoV-2 patients together with six negative ones. Figure 6h illustrates the process from sample treatment to paper-based reactions. Diluted saliva samples were subjected to a brief 95 °C heating step for 2 min to release the viral RNA and then added to NASBA reactions for amplification of each input RNA. The resulting amplicons were then applied to paper-based cell-free reactions for testing with the SARS-CoV-2 AND gate N1*N2* RNA. As shown in Fig. 6i, the gate RNA detected the six positive samples, generating a clearly visible purple colour, while the six negative samples remained yellow in colour. A similar strategy was also applied to differentiate influenza A subtypes and distinguished H1N1, H5N1 and H1N2 from closely related virus subtypes (Extended Data Fig. 4, see Supplementary Table 10 for sequence information).