Modular, programmable RNA sensing using ADAR editing in living cells

Proof-of-principle with synthetic trigger-sensor pairs

A RADAR sensor is inspired by recent advances in RNA editing^14–18, and consists of three parts (Fig. 1a): an optional marker coding sequence; a sensor sequence reverse complementary to the target RNA of interest (“trigger”), but with an editing-enhancing C:A mismatch at a central UAG stop codon (alternative pairings could be used); and an output coding sequence. The stop codon prevents the translation of the output CDS so that only the marker (e.g., mCherry) is expressed. In the presence of the trigger, double-stranded RNA (dsRNA) is formed, which recruits ADAR to edit the adenosine (A) in the UAG to an inosine (I), enabling the translation of the downstream CDS (e.g., EGFP). “Self-cleaving” 2A sequences¹⁹ insulate the flanking CDSs from the variable sensing sequence. We focused our initial efforts on identifying potential trigger sequences in 3’ untranslated regions (UTRs) for two reasons. First, ADAR editing is more efficient in 3’ UTR than CDS²⁰, likely because translating ribosomes may destabilize dsRNA. Second, RADAR is less likely to interfere with the production of the protein encoded by the trigger, because dsRNA formation in the 3’ UTR rather than the CDS will not affect the translation ribosome, and because any bystander editing that occur in the 3’ UTR of the trigger is less likely to cause detrimental outcomes. Most human (57%) and mouse (73%) genes have a 90 bp trigger candidate for at least one reported 3’ UTR variant.

We tested and optimized RADAR in human embryonic kidney (HEK) 293 cells. In our initial optimization, we concluded that the flanking 2A sequences improve the consistency of sensor behavior across different sensor sequences. We also discovered that trigger-independent sensor output (“baseline”) is minimized through the choice of the constitutive promoter driving sensor expression, with the SFFV promoter showing particularly low levels of “baseline” expression (Supplementary Fig. 1c), almost indistinguishable from autofluorescence (Supplementary Fig. 1b). Therefore, SFFV promoter is used for the rest of the paper. To explore the possible source of the baseline in the other promoters, we performed additional tests against CMV. We initially hypothesized that the RADAR stop codon could be edited due to the sensor region forming dsRNA with the rest of the sensor molecule, or translational readthrough/re-initiation. Translational re-initiation is unlikely, because methionine is absent in a long stretch of N terminal region in our reporter EGFP, and re-initiation at the next available methionine will likely result in a nonfluorescent protein. To test the rest of the hypotheses, we created two constructs, one with an additional stop codon after the RADAR stop codon and the other with a frameshifting single nucleotide insertion before the stop codon in the sensor sequence. If either hypothesis was true, these constructs would exhibit reduced baseline. Our observation was contrary to the prediction (Supplementary Fig. 1d), suggesting that the baseline is unlikely caused by “leaky editing” or readthrough. Our current speculation, pending future experimental validation, is that the baseline is caused by promoter-dependent splicing at cryptic junctions that skips over the stop codon.

We created a de novo designed trigger, T1, embedded in the 3’ UTR of a co-transfected gene (Fig. 1b). We validated that the output from the corresponding sensor S1 depends on T1 expression (Fig. 1b). This T1-dependent output is abolished in the absence of endogenous ADAR1 (Supplementary Fig. 1f), enhanced by over-expressing various forms of ADAR (Fig. 1b, Supplementary Fig. 1f–g), and increases with increasing levels of ADAR overexpression (Supplementary Fig. 1h), indicating that RADAR operates through the intended ADAR-dependent mechanism. Because we observed the highest output enhancement by the p150 isoform of ADAR1 (Supplementary Fig. 1g), this isoform was used throughout the paper to maximize outputs. In addition to the fluorescent output for editing (Fig. 1b), we directly assessed editing at the intended A using reverse transcription, PCR amplification, Sanger sequencing (Supplementary Fig. 1e), and NGS (Supplementary Fig. 2g), and obtained comparable results supporting that RADAR is operating through the intended mechanism.

We then further characterized RADAR performance using the synthetic sensor-trigger pair. To estimate the editing efficiency, we compared trigger-induced output to a positive control sensor construct where the stop codon is replaced by the tryptophan codon, and determined that ~3% of the sensor RNAs are edited in response to the trigger (Fig. 1b). NGS from an independently performed experiment reported ~7% editing of the intended A in the trigger (Supplementary Fig. 2g), and both percentages are qualitatively consistent with the Sanger sequencing peaks (Supplementary Fig. 1e). We also examined the NGS signals in the dsRNA-forming region of the trigger, and observed low-level yet significant sensor-dependent bystander editing (Supplementary Fig. 2j), with 1% of reads showing evidence of bystander editing. Although these editing rates are much lower than the intended editing event, their prevalence is consistent with what has been reported¹⁷ and one future mitigation strategy would be to skip over these A sites on the trigger using discontinuous dsRNA¹⁷. Our prioritization of trigger regions in 3’ UTR would also make such bystander editing less likely deleterious than bystander editing in CDS. To assess the potential of using RADAR for quantitative measurements, we determined that S1 output is a linear function of T1 plasmid input across two orders of magnitude (Fig. 1c). RADAR output is modular, as we could utilize the Cre recombinase, a useful tool in neurobiology²¹, as an alternative output (Fig. 1d). Cre can serve as an output amplifier, although the increase in signal is at the cost of increased baseline EYFP (final output) expression (Fig. 1d), as a small amount of leaky Cre expression is likely also amplified by the EYFP reporter.

Potential to detect endogenous transcripts

We then validated RADAR in scenarios closer to eventual use cases. First, in addition to transiently delivered trigger plasmids, we saw increased output in response to doxycycline-induced expression of T1 embedded in the 3’ UTR of a genomically integrated EGFP (Fig. 1e, Supplementary Note 1). EGFP levels were not negatively impacted by the presence of S1 (Supplementary Fig. 3b). We also took advantage of this system to estimate RADAR’s input sensitivity to a genomically expressed trigger. We first established a calibration curve by directly adding different copy numbers of trigger-encoding plasmids as the template for qPCR reaction, and determined the range in which the Ct values and copy numbers are linearly correlated (Supplementary Fig. 2b). To derive the equivalent copy number per cell, we divide the total copy numbers by the number of cells we typically collect for RNA extraction for each condition. We then used reverse transcription and qPCR to measure the Ct values for the genomically expressed trigger mRNA under uninduced and dox-induced conditions, and used the calibration curve to read out the corresponding copy numbers for the trigger transcript (Supplementary Fig. 2b, c). We assumed 100% transcript recovery during RNA extraction and reverse transcription, and identical qPCR efficiencies on DNA plasmids used in calibration and cDNA from reverse transcription. The baseline uninduced trigger expression is in the range of a well expressed transcript, and the induced expression is much higher than a typical abundant transcript such as GAPDH (Supplementary Fig. 2c). Second, to test whether our design functions in the complex context of a natural 3’ UTR, we designed a sensor S_Bdnf for a subsequence within the 2.9 kb 3’ UTR of murine Bdnf, taking advantage of the mouse-human orthogonality for unambiguous testing. We observed a significant response from S_Bdnf to the Bdnf 3’ UTR expressed in HEK293 cells through transient transfection (Fig. 1f).

Finally, we assessed RADAR’s capability of detecting endogenous triggers in HEK293 cells. Based on the estimates from the previous paragraph, we designed sensors for six transcripts expressed at levels comparable to GAPDH, and as negative control we used sensors for which no endogenous triggers exist (Supplementary Fig. 2e, f). Only a subset of sensor outputs are significantly higher than the negative controls (Supplementary Fig. 2e), suggesting that RADAR performance might be affected by other uncharacterized factors, e.g., trigger region overlapping with other functional motifs. To pinpoint RADAR input sensitivity, we used siRNA to knock GAPDH down to different levels, and showed that the outputs are indistinguishable between 24% and 34% of the wildtype level, and is significantly elevated at 59% of the wildtype level and beyond (Fig. 1g), suggesting an input sensitivity between 34% and 59%. For the GAPDH sensor, the fluorescent output and the NGS editing rate of the intended A show comparable fold changes between 0 and 100 nM of siRNA (Supplementary Fig. 2h), suggesting that the dynamic range is limited by trigger-independent editing of A in the stop codon. This interpretation is also consistent with our observation that control sensor constructs with either an additional stop codon after the RADAR stop codon or a frameshifting single nucleotide insertion before the RADAR stop codon reduces the signal to a level below that of 100 nM siRNA (Supplementary Fig. 2d). As a second test case, we investigated a sensor for the 3’ UTR of DNAJB1, a member of the hsp40 heat shock response protein family. Although heat shock non-specifically increases RADAR output slightly, which we controlled for using the synthetic S1 sensor unmatched with DNAJB1 (Fig. 1h), there is statistically significant interaction between sensor type, heat shock, and siRNA against DNAJB1, indicating that the DNAJB1 sensor specifically responds to a signal sensitive to both induction by heat shock and knockdown by siRNA against DNAJB1, i.e., most likely DNAJB1 itself as intended (Fig. 1h). Heat shock induced DNAJB1 expression is at ~7% of GAPDH (Fig. 1h), providing an estimate for the input sensitivity in this second case. Here the output exhibits a larger fold change when measured by NGS editing rate of the intended A than by the fluorescent output (Supplementary Fig. 2i), suggesting that the dynamic range is limited by a “floor” on the fluorescent output, consistent with our observation that the baseline is close to cellular autofluorescence, in which case the key optimization goal would be to generally improve the output amplitude to elevate the operation range of the DNAJB1 sensor further above autofluorescence. Significant sensor-dependent bystander editing was observed in both endogenous GAPDH and DNAJB1 transcripts (Supplementary Fig. 2k, l), with 1–2.5% of reads having bystander edits mostly concentrated in the dsRNA regions, again indicating the need for future engineering efforts to mitigate RADAR’s potential effects on the host cells.

Expanding input repertoire

We then explored strategies for expanding the coverage of the human/mouse transcriptome. First, using the 3’ UTR of Bdnf, we observed that RADAR performance was largely maintained down to 72 bp dsRNA (Fig. 2a). Second, inspired by work showing that ADAR enzymes are tolerant of or even benefited by disruptions in the target dsRNA^17,22, we validated that we could rescue the signal from the non-functional 36 bp sensor if we introduced an additional 54 bp sequence complementary to another part of the long Bdnf 3’ UTR, constituting a “split” design (Fig. 2b). Such a split design offers flexibility for several scenarios (Supplementary Note 1): to skip over undesired reverse-complement stop codons, to skip over complex secondary structures, miRNA binding sites, or other functional features of the trigger RNA, to distinguish between homologs where the sequence containing CCA is shared, but unique regions exist elsewhere, and to detect gene fusions and alternatively spliced transcripts where the unique junction is not amenable to direct sensing (no CCA within the proximity of the junction). Finally, we validated that a sensor for the CDS of EGFP showed a significant response to trigger induction (Fig. 2c). Sensing CDS is less ideal than 3’ UTR, because its output in response to the transiently transfected trigger is weaker than the 3’ UTR sensor (Supplementary Fig. 3a), and it slightly reduces the GFP expressed from the genomically integrated trigger (Supplementary Fig. 3c), suggesting interference with the trigger through dsRNA formation or alteration of the GFP CDS through bystander editing. Nevertheless, the CDS sensor results offer an alternative when necessary but also reaffirming our initial prioritization of 3’ UTRs.

Combining the ability of shortening the sensor, using a split design, and sensing CDSs, > 85% of human (Fig. 2d, Supplementary Fig. 3d) and mouse (Supplementary Fig. 3e) genes have at least one candidate trigger sequence compatible with RADAR.

Chimeric ADAR as an alternative to ADAR1

While ADAR over-expression greatly improves RADAR’s dynamic range, it may have detrimental side effects. One alternative is to bias the editing towards our RADAR sensor using an engineered version of ADAR that contains only an editing-enhancing mutant deaminase domain of ADAR2 and the MS2 RNA binding protein MCP (“ADAR2(DD)E488Q-MCP”)^14,23, and inserting the MS2 sequence in the corresponding sensor. Combining this chimeric ADAR with MS2-bearing sensors, we achieved a dynamic range comparable to ADAR1^p150 over-expression (Fig. 2e, Supplementary Fig. 3f). MS2 insertion in the 3’ UTR of the sensor (Fig. 2e) appears to generate a higher output dynamic range than insertion next to the sensor sequence (Supplementary Fig. 3f), probably due to the latter’s conflict with translating ribosomes. ADAR2(DD)E488Q-MCP did not affect the original MS2-lacking sensor (Fig. 2e), consistent with our expectation that this design biases editing towards MS-bearing sensors.

Demonstrating unique features

RADAR has several unique features and potential applications. For example, RADAR can be used for cell classification (Fig. 3a, Supplementary Note 1). RADAR can also integrate multiple inputs using OR and AND logic (Fig. 3b, 3c); for the latter, two sensor sequences can be straightforwardly concatenated such that both stop codons have to be edited for output expression.

Cell states, especially in medical contexts, are often defined by not only the expression level of RNAs but also the presence of new RNA sequences. We leveraged RADAR’s unique features to sense the latter. First, as ADAR is sensitive to the base identities surrounding the edited adenosine¹⁵, RADAR is uniquely suitable for distinguishing certain short genetic variants. As a demonstration, we validated a sensor that distinguishes two common oncogenic mutations of TP53 associated with different invasive traits in cancer cells²⁴ (Fig. 3d) as well as a single base distinguisher (Supplementary Fig. 3g). ~5% of known or likely pathogenic variants reported in ClinVar can potentially be distinguished from their wild type alleles using RADAR. Second, as gene fusions drive many cancers²⁵, we validated that the split design can potentially serve as a fusion-specific sensor (Supplementary Fig. 3h).

Finally, to show that RADAR is a self-contained module, we assessed RADAR in plants that lack endogenous ADAR. Without any optimization other than switching to plant promoters and vectors, we observed elevated fluorescent output in response to the matching T1 trigger compared to a non-matching trigger in Nicotiana benthamiana (Fig. 3e), indicating that the core RADAR mechanism is functional in a heterologous context, and is therefore potentially useful in a wide range of species, even those lacking native ADAR machinery.

Plasmid generation

Plasmids were generated using standard molecular cloning practices, including InFusion of linearized plasmids and PCR fragments and restriction-ligation of linearized fragments and annealed phosphorylated oligonucleotides. A complete list of plasmids and associated maps is found in Supplementary Table 1. Plasmids are available upon request from the corresponding author and are deposited to Addgene. Human ADAR plasmids as well as the ADAR1 knockout cell line were generously provided by prof. Billy Li. Cre and Cre reporter plasmids were kindly gifted by prof. Liqun Luo. pUBC_stdMCP_serinemod_E488QADAR_p2A_yGFP (“ADAR2(DD)E488Q-MCP”) was a gift from Robert Singer (Addgene plasmid # 154787; http://n2t.net/addgene:154787; RRID:Addgene_154787).

Tissue culture

Flp-In T-REx Human Embryonic Kidney (HEK) 293 cells were purchased from Thermo Scientific (catalog #R78007). Cells were cultured in a humidity-controlled incubator under standard culture conditions (37°C with 5% CO2) in Dulbecco’s Modified Eagle Medium, supplemented with 10% fetal bovine serum (Fisher Scientific catalog #FB12999102), 1 mM sodium pyruvate (EMD Millipore catalog # TMS-005-C), 1x penicillin-streptomycin (Genesee catalog #25–512), 2 mM L-glutamine (Genesee catalog #25–509) and 1x MEM non-essential amino acids (Genesee catalog #25–536). To induce expression of certain constructs, 100 ng/mL of doxycycline was added at the time of transfection. The inducible-trigger containing cell line was generated by transfecting the construct in a PiggyBac backbone along with PiggyBac integrase (4:1), with 50 ug/mL hygromycin added for selection when reseeding into 10 cm dishes two days after transfection.

Transient transfections

HEK293 cells were cultured in either 24-well or 96-well tissue culture-treated plates under standard culture conditions. When cells were 70–90% confluent, the cells were transiently transfected with plasmid constructs using the jetOPTIMUS^® DNA transfection Reagent (Polyplus catalog # 117–15), as per manufacturer’s instructions using 0.375 uL of reagent per 50 uL of jetOPTIMUS buffer for 500 ng total DNA transfections in the 24-well format and 0.13 uL of reagent per 12.5 uL of buffer for 130 ng total DNA in the 96-well format. All transfections are detailed in Supplementary Table 2.

Flow cytometry and data analysis

Cells were harvested approximately 48 hours after transfection by trypsinization and resuspended in flow buffer (HBSS + 2.5 mg/mL bovine serum albumin). Post 40 um straining, cells were analyzed by flow cytometry (Biorad ZE5 Cell Analyzer), and data was processed with the cytoflow Python package. An overview of gating and analysis is given in Supplementary Fig. 1a, 1b.

siRNA knockdown experiments

ON-TARGETplus siRNA SMARTpool reagents were purchased from Dharmacon (DNAJB1: L-012735–01-0005; GAPDH: D-001830–10-05; VCL: L-009288–00-0005 (used as control for DNAJB1 knockdown); PPIB: D-001820–01-05 (used as control for GAPDH siRNA titration)) and resuspended according to manufacturer’s instructions. Transfection was performed according to manufacturer’s instructions using 100 nM total siRNA, antibiotic free serum, and the DharmaFECT 1 transfection reagent. DNA transfection was performed immediately after as described above, without an additional media change.

RNA extraction, reverse-transcription, and sequencing

HEK293 cells grown in 24-plates were spun down and RNA was extracted using the following kits: RNAasy mini kit (Qiagen), RNase-Free DNase Set (Qiagen), and QIAshredder (Qiagen). After extraction RNA quality was assessed by running 500 ng on 1% agarose gel. 500 ng of purified RNA was then reverse-transcribed using iScript cDNA synthesis (Biorad). For Sanger sequencing, cDNA was PCR amplified and sent to Genewiz/Azenta with a matching primer. For next-generation sequencing, cDNA was PCR amplified using partial Illumina adapters, gel extracted, and sent to Genewiz/Azenta according to Amplicon-EZ sample preparation instructions. NGS reads were aligned to a synthetic reference genome containing all expected amplicons (all samples had at least 95% correct mapping rate and at least 350,000 mapped reads), and paired reads were combined into single “consensus” reads using a minimum base quality threshold of 15. Pileups of consensus reads were used to quantify editing efficiency and bystander editing rates. Raw reads have been deposited in the Sequence Read Archive (BioProject PRJNA874842).

qPCR measurements

qPCR was carried out on a QuantStudio3 (Applied Biosystems) using SYBR-Green. RNA estimation was calculated based on the calibration curve of purified plasmid and normalized by the Ct threshold. The following primer pair sequences for GFP (Signagen) and normalizing gene (β-actin) were used: GFP-F AAGCAGAAGAACGGCATCAA, GFP-r TCCAGCAGGACCATGTGATC, β-actin-F CGTCCACCGCAAATGCTT, β-actin-R GTTTTCTGCGCAAGTTAGGTTTTGT. For qPCR of endogenous transcripts, QuantStudio6 (Applied Biosystems) was used with primers listed in Supplementary Table 3 and the SensiMix SYBR HiROX reagent (Meridian QT605–05).

Leaf infiltration and analysis

Three leaves in three Nicotiana benthamiana plants were infiltrated with Agrobacterium separately transformed with RADAR components (sensor and human ADAR1^p150), and a matching or unrelated trigger, with two infiltrations of either trigger per leaf. Quantification based on averages of average fluorescence intensities from six rectangular regions per infiltration spot. EGFP (output) fluorescence was normalized using the mCherry (marker) fluorescent intensity.

Statistical analysis

Values are reported as the means from at least three biological replicates, representative of two independent biological experiments. For experiments comparing two groups, a Bonferroni-corrected two-tailed Student t-test was used to assess significance. For the analysis of the heatshock experiment, a linear regression on a full factorial experimental design was used to assess the interaction between siRNA knockdown, heatshock, and the sensor used. To compare the output of endogenous sensors, a z-score analysis was performed, comparing the sample mean of a sensor’s output to the means of five negative control sensors designed not to be triggered in human cells.

Bioinformatics

We analyzed all genes with annotated 3’ UTRs in the human genome in search of trigger sequences compatible with the RADAR design. Later, we also considered all genes with a CDS annotation. 3’ UTR and CDS sequences were obtained from Ensembl Biomart. Candidates must have a 5’ CCA 3’ sequence (which is paired with the 5’ UAG 3’ in the sensor) flanked by a sufficiently long stretch of sequence; the reverse complement of the trigger sequence should not create an in-frame stop codon in the sensor. Furthermore, the trigger sequences should be unique (assessed by mapping to the genome), and the sensor sequence easily synthesizable (lacking homopolymer runs and with 35–80% GC content). The hg38 genome build was used for human transcriptome analysis and the GRCm39 build for murine transcriptome analysis. The blastn tool version 2.9.0+ was used with the -task blastn -evalue 1 arguments, and further filtering was done on the alignment length (minimum 30 matches) and position (must overlap the central CCA or alternative sequence). Results from alternative chromosomes duplicative in nature were removed. A single hit -- the target sequence -- was allowed, with sensors having more hits discarded from analysis; this means that some designs were discarded due to pseudogenes and other repetitive regions that may or may not be expressed as RNA. As sequences other than the UAG:CCA pairing can efficiently be edited, we also analyzed candidate trigger sequences with a central GCA, UCA, or CAA sequence.

The 20211025 release of ClinVar was used to analyze the potential for detecting known and likely pathogenic variants, assuming that UAG can be edited when paired by one of CAA, CUA, CGA, ACA, UCA, GCA, CCA, CCU, CCC¹⁵. Alternative and reference alleles were padded based on the GRCh38 (hg38) reference genome by one base on either side. A variant was considered distinguishable by RADAR if the padded alternate allele contained one of the edit-triggering variants, but the padded reference did not contain any.