A molecular recorder that tracks events over time by sequentially editing DNA

How do we learn the order of molecular events in living systems? A first approach is direct observation, for example, live-cell fluorescence microscopy to order interactions in real time. A second approach is time series experiments, for example, destructively sampling and transcriptionally profiling a system at different time points, followed by pseudotemporal ordering. A third approach is epistatic analysis, for example, ordering the actions of genes by comparing the phenotypes of single and double mutants. Although these and other approaches have important strengths, they are also limited in key ways. For example, live imaging is largely restricted to in vitro models. For time series experiments, resolution and accuracy are constrained by the frequency of sampling and the reproducibility of the biological process under investigation. Epistatic analysis is confounded by pleiotropy, particularly in multicellular organisms.

Another approach, theoretically promising but methodologically underdeveloped relative to the aforementioned alternatives, is a DNA memory device³, which we define here as an engineered system for digitally recording molecular events through permanent changes to a cell’s genome that can be read out post factum. Thus far, several proof-of-concept DNA memory devices have been described that leverage diverse approaches for the ‘write’ operation, including site-specific recombinases (SSRs)^4,5, CRISPR–Cas9 genome editing^6–9, CRISPR integrases^10,11, terminal deoxynucleotidyl transferases (TdTs)¹², base-pair misincorporation¹³, base editing¹⁴ and others¹.

The nature of the write operation in such DNA memory devices shapes their performance in terms of channel capacity for encoding and decoding signals, temporal resolution, interpretability and portability¹. For example, SSRs record molecular signals with high efficiency, but the number of distinct signals that can be concurrently recorded is limited by the number of available SSRs. DNA memory devices relying on CRISPR–Cas9 can potentially overcome this limitation, for example, if each signal of interest were coupled to the expression of a different guide RNA (gRNA), but in that case each signal would also require its own target(s). Furthermore, the CRISPR–Cas9 molecular recorders described thus far rely on double-stranded breaks (DSBs) and non-homologous end joining (NHEJ) to ‘scar’ target sites¹. In addition to being toxic, frequent DSBs often excise or corrupt consecutively located target sites, the molecular equivalent of accidental data deletion.

A further handicap of nearly all DNA memory devices described thus far is that, while recordings might stochastically accumulate at unordered target sites, the order in which they occurred is not explicitly captured. CRISPR spacer acquisition systems, which rely on signal-induced, unidirectional incorporation of DNA spacers or transcript-derived tags to an expanding CRISPR array, overcome this limitation^{10,11,15–17}. However, at least thus far, their reliance on accessory integration host factors has restricted such recorders to prokaryotic systems. Another approach, CHYRON, enables directional writing of information to DNA by combining self-targeting CRISPR gRNAs with the expression of TdT, whose presence shifts the most likely outcome of NHEJ from short deletions to short insertions¹⁸. While this approach unidirectionally inserts nucleotides in a signal-responsive manner, it continues to rely on NHEJ-mediated repair of DSBs. Furthermore, because each gRNA/target yields a homogenous signal (TdT-mediated insertions of variable length), it is not clear how this approach could be used to explicitly record the precise order of more than a handful of distinct signals. Finally, at least two groups have independently developed ‘logic-circuit architectures’ that use sequential base editing to record the order and identity of biological signals in both bacterial and mammalian cells (DOMINO¹⁹ and CAMERA¹⁴). However, because base editors are currently limited to writing single-base substitutions to predefined targets, the order of signals can only be recorded via preprogrammed circuits, rendering multiplex recording challenging.

Here we describe a DNA memory device that is (1) highly multiplexable, that is, compatible with the concurrent recording of at least thousands of distinct symbols or event types; (2) sequential and unidirectional in recording events to DNA and therefore able to explicitly capture the precise order of recorded events; and (3) active in mammalian cells. This system, which we call DNA Typewriter, begins with a tandem array of partial CRISPR–Cas9 target sites (DNA Tape), all but the first of which are truncated at their 5′ ends and are therefore inactive (Fig. 1a–c). Each of many prime editing gRNAs (pegRNAs), together with the prime editing enzyme², is designed to mediate the insertion of a k-mer within the sole active site of the tandem array, which is initially its 5′-most target site. In the simplest implementation, all pegRNAs target the same 20-bp spacer but each encodes a unique ‘symbol’ in the form of a k-mer insertion. Specifically, the 5′ portion of the k-mer insertion is the variable and encodes the identity of the pegRNA, while its 3′ portion is constant and activates the subsequent target site in the tandem array by restoring its 5′ end. Thus, each successive edit records the identity of the pegRNA mediating the edit while also shifting the position of the active target site by one unit along the array. At any moment, an intact spacer and protospacer adjacent motif (PAM) are present at only one location along the array, analogous to the ‘write-head’ of a disk drive or the ‘type guide’ of a typewriter.

To test this idea, we designed a DNA Tape (TAPE-1) by modifying a spacer sequence previously shown to be highly amenable to prime editing by the PE2 enzyme (HEK293 target 3, or HEK3)². In TAPE-1, a 3-bp key (GGA) is followed by a tandem array of a 14-bp monomer (TGATGGTGAGCACG) that includes the PAM sequence (TGG) at positions 4–6. At the 5′-most end of the TAPE-1 array, the key sequence, the first 14-bp monomer and the first 6 bases of the subsequent 14-bp monomer collectively make up an intact 20-bp spacer and PAM (Fig. 1a). We further designed a set of 16 pegRNAs to target TAPE-1, with each pegRNA programming a distinct 5-bp insertion (Fig. 1b). The first 2 bp of the insertion is unique to each of the 16 pegRNAs. The remaining 3 bp of the insertion corresponds to the key (GGA). We reasoned that, when a pegRNA/PE2-mediated insertion occurred at the active TAPE-1 site, it would (1) record the identity of the pegRNA via the 2-bp portion of the insertion; (2) inactivate the current active site by disrupting its sequence; and (3) activate the next monomer along the array, as the newly inserted GGA key, together with the subsequent 20 bp, creates an intact 20-bp spacer and PAM. In the next iteration of genome editing, a pegRNA-mediated insertion to the second monomer would be recorded while also moving the type guide to the third monomer and then to the fourth, the fifth and so on (Fig. 1c).

We synthesized and cloned TAPE-1 arrays with varying numbers of monomer units (2×TAPE-1, 3×TAPE-1, 5×TAPE-1) and stably integrated these arrays into the genome of HEK293T cells via the piggyBac system. We then transiently transfected the resulting cells with a pool of plasmids designed to express PE2 (pCMV-PE2-P2A-GFP; Addgene, 132776) and 16 pegRNAs, each programmed to insert an NNGGA barcode to TAPE-1, and collected the cells after 4 days. The TAPE-1 region was PCR amplified from genomic DNA and sequenced.

For each TAPE-1 array, we categorized sequencing reads into those in which (1) no editing occurred; (2) the observed pattern was consistent with sequential, directional editing; and (3) the observed pattern was inconsistent with sequential, directional editing (Fig. 1d–f and Supplementary Table 1). Overall editing rates were modest, as only 4.7% ± 0.5%, 5.2% ± 0.6% and 5.9% ± 0.8% of all reads for 2×TAPE-1, 3×TAPE-1 and 5×TAPE-1, respectively, exhibited any editing. However, within the set of reads showing edits, the data were overwhelmingly consistent with sequential, directional editing. For example, with 2×TAPE-1, the second monomer was edited in 22.8% ± 1.7% of reads in which the first monomer was also edited (Fig. 1d). By contrast, the second monomer was only edited in 0.6% of reads in which the first monomer was not edited. This observation strongly suggests that edits of the second monomer were dependent on an edit of the first monomer having already occurred. Furthermore, this confirms that the 3-bp mismatch at the PAM-distal end of ‘inactive’ spacers of the TAPE-1 design is sufficient to inhibit prime editing. Data obtained from 3×TAPE-1 and 5×TAPE-1 were also consistent with sequential genome editing. For example, 98.5% (3×TAPE-1) and 99.0% (5×TAPE-1) of reads that were edited at the second monomer were also edited at the first monomer, while 97.6% (3×TAPE-1) and 98.8% (5×TAPE-1) of reads that were edited at the third monomer were also edited at the first and second monomers (Fig. 1e,f). These results were consistent across three transfection replicates (Supplementary Table 1).

An interesting phenomenon was that, while the observed editing rate of the first TAPE-1 monomer was ~6%, the editing rates of the second or third TAPE-1 monomer, conditional on the preceding monomers already being edited, were ~20% (Extended Data Fig. 1a). A simple explanation for this ~14% greater ‘elongation’ than ‘initiation’ of editing is that some integrated tapes are more amenable to prime editing than others, resulting in an excess of fully unedited tapes. However, we also observed a similar pattern with episomal tapes, as well as following multiple sequential transfections with pegRNA- and PE2-expressing plasmids to edit integrated tapes (7–15% increase in the conditional editing efficiency of the second site). Factors that might contribute to the observed ‘pseudo-processivity’ include heterogeneous susceptibility of cells to transfection, chromatin context^20,21 and cell cycle phase, but the primary explanation remains unclear. We also observed modest reductions in the conditional editing efficacy after the second site (1–10% decreases), which might be explained simply by each site being ‘active’ for less time than its predecessor.

We next analysed the distribution of the 16 NNGGA barcode insertions, focusing on 5×TAPE-1. Their frequencies were correlated across three replicates as well as between the first and second target sites (Pearson’s r = 0.97–0.99; Extended Data Fig. 1b,c). The observed variation was partly explained by the relative abundances of the individual pegRNAs in the plasmid pool (Pearson’s r = 0.87; Extended Data Fig. 1d). To explore whether the sequence of the insertion itself influences editing efficiency, we repeated the experiment with an equimolar pool of 16 pegRNA-expressing plasmids that had been individually cloned and purified (rather than cloned as a pool). For each of the NNGGA insertions in each experiment, we calculated ‘edit scores’ as the log₂-scaled insertion frequencies normalized by the abundances of pegRNAs in the corresponding plasmid pools (Fig. 1g). The maximal edit score difference between the best barcode (CCGGA with an edit score of 0.98) and the worst barcode (TGGGA with an edit score of −2.38) was 3.36, that is, a nearly ten-fold difference in editing efficiency. However, 10 of 16 barcodes exhibited efficiencies within a two-fold range. Edit scores were well correlated between 5×TAPE-1 edited by the 16 pegRNA plasmids pooled before versus after cloning (Spearman’s p = 0.97; Extended Data Fig. 1e), in line with an insertion sequence-dependent bias. Indeed, when we used the relative efficiencies observed in the ‘post-cloning pooling’ experiment to correct the TAPE-1 unigram barcode frequencies measured in the ‘pre-cloning pooling’ experiment, the correlation of the frequencies with the abundance of the corresponding pegRNAs in the plasmid pool improved (Pearson’s r = 0.87→0.94; Extended Data Fig. 1d) and vice versa (Pearson’s r = 0.27→0.67; Extended Data Fig. 1f).

Several strategies to improve the efficiency of prime editing through modular engineering were recently reported: (1) addition of degradation-resistant secondary structure to the 3′ end of the pegRNA²² (resulting in enhanced pegRNAs, or epegRNAs); (2) introduction of human MLH1 dominant-negative peptide (hMLH1dn) to favour the intended edit²³; and (3) modifications to the primary sequence of the prime editing enzyme²³ (resulting in PEmax). Combined deployment of these strategies has been reported to improve editing efficiency by ~3.5-fold in HEK293T cells and ~72-fold in HeLa cells, relative to PE2 and pegRNAs²³.

As our initial experiments with PE2 and pegRNAs resulted in only modest editing of the first site of TAPE-1 (~6%), we sought to incorporate these new strategies. We cloned a pool of U6-driven epegRNAs, each programmed to insert an NNGGA barcode to TAPE-1, and transfected them into HEK293T cells in which 5×TAPE-1 was integrated (5×TAPE-1(+) HEK293T) along with a plasmid expressing PEmax and hMLH1dn (pCMV-PEmax-P2A-hMLH1dn; Addgene, 174828). After 4 days, we collected genomic DNA and then PCR amplified and sequenced TAPE-1. We observed 18.1% ± 0.5% editing of the first site (Extended Data Fig. 2a), a nearly three-fold increase relative to PE2 and pegRNAs, while editing remained overwhelmingly sequential (>99.5%). We then cloned four more pools, encoding 6-bp (NNNGGA) to 9-bp (NNNNNNGGA) barcodes. The epegRNA–PEmax–hMLH1dn prime editing system achieved reasonably high efficiencies for longer insertions (for example, 10.6% ± 0.5% for 9-bp insertions; Extended Data Fig. 2a). Edit scores for pegRNA–PE2 versus epegRNA–PEmax–hMLH1dn were highly correlated (Spearman’s p = 0.96 for NNGGA and Spearman’s p = 0.88 for NNNGGA; Extended Data Fig. 2b–e). The edit scores for epegRNAs were more uniform than those for standard pegRNAs, as 14 of 16 NNGGA barcodes exhibited efficiencies within a two-fold range (Extended Data Fig. 2c) and 59 of 64 NNNGGA barcodes exhibited efficiencies within a four-fold range (Extended Data Fig. 2e). We also calculated edit scores for more than 1,900 barcodes in NNNNNNGGA (or 6N+GGA) TAPE-1-targeting epegRNAs in a single experiment (Extended Data Fig. 2f–i), markedly expanding the number of unique symbols that can be encoded and deployed to write to a shared DNA Tape by two orders of magnitude relative to our original NNGGA experiment. Overall, 1,509 of the 1,908 6N+GGA barcodes exhibited efficiencies with edit scores between –1 and 1, that is, a four-fold range (Extended Data Fig. 2h).

To evaluate the compatibility of DNA Typewriter with cell types other than HEK293T cells, we integrated the 5×TAPE-1 target into mouse embryonic fibroblasts (MEFs) and mouse embryonic stem (mES) cells using the piggyBac transposase system and transfected cells with either a pool of 16 NNGGA epegRNAs or a pool of 64 NNNGGA epegRNAs with PEmax- and hMLH1dn-expressing plasmids through electroporation with DNA plasmids. After 4 days, we collected genomic DNA and then amplified and sequenced TAPE-1. We observed 7.0–18.1% editing of the first site after 4 days (Extended Data Fig. 2j). In mESCs, where prolonged culturing was permitted compared with MEFs, we performed a second transfection with the same set of epegRNA-, PEmax- and hMLH1dn-expressing plasmids, 4 days after the first transfection. The cumulative editing of the first site increased to 28.7% ± 2.8% when the sample was collected another 4 days after the second transfection. Of note, the edit scores for NNGGA and NNNGGA pegRNAs in mESCs were reasonably well correlated with those measured in HEK293T cells (Extended Data Fig. 2k,l), suggesting that measurements of relative pegRNA efficiencies made in HEK293T cells are applicable to other cell types. Collectively, these results demonstrate that the performance of DNA Typewriter can be improved using methods that enhance prime editing and, furthermore, that the method can be used in primary and stem cells. Overall, we suspect that the range and efficiency of DNA Typewriter will be tightly coupled to that of prime editing, which has also been demonstrated to work in human induced pluripotent stem cells (iPSCs) and primary human T cells²³.

Our TAPE-1 construct exhibited sequential, directional editing, wherein the editing of any given site along the array was strongly dependent on all preceding sites having already been edited. This behaviour is consistent with DNA Typewriter’s design, as the key sequence must be inserted 5′ to any given monomer within the DNA Tape to complete the spacer that is recognized by any of the gRNAs used. However, performance would presumably be corrupted by non-specific editing, for example, if a guide were able to mediate edits to a non-type-guide monomer despite several mismatches at the 5′ end of the spacer^24,25.

Although TAPE-1 exhibited reasonable efficiency and specificity, we sought to explore whether this would be the case for other spacers. To this end, we designed and synthesized 48 TAPE constructs (TAPE-1 through TAPE-48), each derived from one of eight basal spacers that previously demonstrated reasonable efficiency for prime editing^2,26,27 and one of six design rules that vary monomer sequence, key sequence and key/monomer length (Extended Data Fig. 3a). In each of these 48 constructs, a 3×TAPE region was accompanied by a pegRNA-expressing cassette designed to target it with a 4- to 6-bp insertion (16 possible 2-bp barcodes followed by a 2- to 4-bp key sequence). We then transiently transfected HEK293T cells with PE2-encoding plasmid and a pool of 48 pegRNA-by-3×TAPE constructs and collected cells after 4 days. The 3×TAPE region was PCR amplified from genomic DNA and sequenced.

We calculated two quantities for each 3×TAPE array: (1) efficiency, calculated by summing all edited reads and dividing by the total number of reads, and (2) sequential error rate, calculated by summing all edited reads inconsistent with sequential, directional editing and dividing by the total number of edited reads (Extended Data Fig. 3b). Of note, our initial TAPE-1 construct had one of the lowest sequential error rates among the 48 tested DNA Tapes. The only construct that had a lower sequential error rate than TAPE-1 was TAPE-6, which was derived from the same basal spacer (HEK3) but had a 4-bp rather than a 3-bp key sequence. Indeed, across the full experiment, a longer key sequence was associated with a lower sequential error rate (Extended Data Fig. 3c). Performance differences between basal spacers were modest, with DNA Tapes based on the HEK3 and FANCF spacers exhibiting the best combination of efficiency and specificity (Extended Data Fig. 3d). Among FANCF-based spacers, TAPE-27 exhibited over 50% greater efficiency than TAPE-1 but also had a two-fold-higher sequential error rate (Extended Data Fig. 3b). Performance characteristics were highly consistent when we repeated the experiment with integration rather than transient transfection of the constructs (Extended Data Fig. 3e).

Overall, these results show considerable variation in efficiencies and sequential error rates, specific to particular 13- to 15-bp TAPE sequences. Although a single well-performing monomer such as either TAPE-1 or TAPE-27 is sufficient to construct a generic substrate to which thousands of distinct symbols can be written, additional screening might yield monomers with even better performance characteristics and would also facilitate modelling of the sequence determinants of monomer performance^24–26,28.

We next asked whether we could apply DNA Typewriter to record, recover and decode complex event histories. We prepared a set of synthetic signals by individually cloning 16 individual pegRNA-expressing plasmids, each encoding a unique 2-bp barcode insertion to TAPE-1. We also prepared a polyclonal population of HEK293T cells with integrated 5×TAPE-1 to serve as the substrate for recording. Finally, we designed a set of five ‘transfection programmes’−complex event histories that we could attempt to record and then subsequently decode (Fig. 2a).

At the beginning of each epoch of each transfection programme, one or more pegRNA plasmids were introduced to a population of HEK293T cells with integrated 5×TAPE-1 (5×TAPE-1(+) HEK293T) via transient transfection of plasmids expressing the corresponding pegRNA(s) and PE2. After each transfection, cells were passaged the next day into a new plate and excess cells were collected for genomic DNA. 5×TAPE-1 from each epoch of each programme was amplified and sequenced. Successive epochs occurred at 3-day intervals.

Programmes 1 and 2 each consisted of a distinct, non-repeating sequence of transfection of the 16 pegRNAs, that is, one per epoch. The specific orders aimed to maximize (programme 1) or minimize (programme 2) the edit distances between temporally adjacent signals. On the basis of sequencing of 5×TAPE-1 after epoch 16, we observed that barcodes introduced in the early epochs were more frequent at the first target site (site 1) than barcodes introduced at late epochs (Fig. 2b). This is expected, as each editing round shifts more of the type guides to site 2 (and subsequently to site 3 to site 5) (Extended Data Fig. 4a), with minimal effects on the integrity of the 5×TAPE-1 array (Extended Data Fig. 4b). A trivial decoding approach would be to simply arrange barcodes in order of decreasing site 1 unigram frequency, but for both programmes 1 and 2 this resulted in an incorrect order (Extended Data Fig. 4c).

However, inference can be improved by leveraging the sequential aspect of DNA Typewriter, for instance, by analysing bigram frequencies or pairwise appearance of events as used to infer orders from the CRISPR–Cas spacer acquisition process (Cas1–Cas2 system used in bacteria)^11,29. For example, if signal B preceded signal A, then we expect many more B–A bigrams than A–B bigrams at adjacent edited sites in 5×TAPE-1. In Fig. 2c,d, we show heatmaps of bigram frequencies measured from all four pairs of adjacent editing sites on 5×TAPE-1, arranged by the true order in which the signals were introduced for programmes 1 and 2. Indeed, the bigram frequencies appear to capture event order information, as evidenced by the gross excess of observations immediately above versus immediately below the diagonal (for example, in programme 1, CA-GC >> GC-CA). One way to leverage this information is by enumerating ‘ordering rules’ among all events for possible permutations and then checking which the observed data best match^11,29. However, the number of ordering rules for n events increases to the order of n² (for ordering of 16 events, there are 136 ordering rules, or (n² + n)/2 in general), while the number of possible permutations increases to n factorial. As a more computationally efficient approach, we implemented the following algorithm: (1) initialize with the event order inferred from site 1 unigram frequencies; (2) iterate through adjacent epochs from beginning to end and swap signals A and B if the bigram frequency of B–A is greater than that for A–B; and (3) repeat step 2 until no additional swaps are necessary. For both programmes 1 and 2, this algorithm resulted in correct ordering of the 16 signals, out of 16 factorial or 21 trillion possibilities (Extended Data Fig. 4c). This inference was robust to the sequencing depth, as the correct order could be reconstructed from as few as 2,500 reads of the 5×TAPE-1 amplicon (Extended Data Fig. 4d).

The dearth of bigrams inconsistent with the true order, illustrated by the lack of signal below the diagonal in the programme 1 and programme 2 heatmaps (Fig. 2c,d), indicates minimal interference between adjacent epochs; that is, transfected pegRNAs from adjacent epochs did not overlap in their activities. To evaluate performance in the presence of such overlap, we designed programme 3, in which two barcodes are introduced in each epoch but adjacent epochs always share one barcode (Fig. 2a). Concurrent transfection of two pegRNAs with distinct barcodes is evident in the resulting bigram frequency matrix, specifically by the signal both immediately above and below the diagonal (Fig. 2e). Our aforementioned decoding algorithm performed slightly worse on programme 3, with a single swap between epochs 4 and 5 required to revise the inferred order to the correct order (Extended Data Fig. 4c).

Finally, we asked whether the relative strength of signals could be inferred from symbols recorded to DNA Tape. For this, we designed programmes 4 and 5, which have the same order of barcodes−a pair in each epoch−but with each pair at different ratios in the two programmes. In programme 4, pegRNAs encoding each pair of barcodes were always mixed at a 1:3 ratio, whereas in programme 5 the same pairs for each epoch were mixed at a 1:1, 1:2, 1:4 or 1:8 ratio (Fig. 2a). For both programmes, the resulting bigram frequency matrix was consistent with expectation and the order of events was accurately inferred (Fig. 2f,g and Extended Data Fig. 4c). However, in addition, we were able to compare the relative ratios at which each pair of barcodes was introduced within each epoch between programmes 4 and 5 and found these to be well correlated with expectation (Fig. 2h and Extended Data Fig. 4e,f). Taken together, these results show that DNA Typewriter can record, recover and decode complex event histories including the order, overlap and relative strength of signals.

We next designed a strategy to record and decode short text messages in populations of cells with DNA Typewriter. In brief, we modified the Base64 binary-to-text encoding scheme by assigning each of the 64 possible 3-mers to 6-bit binaries. The Base64 scheme encodes uppercase and lowercase English characters, numbers from 0 to 9, and two symbols. In our TAPE64 scheme, we encoded uppercase English characters, four symbols and a whitespace, with two-fold or four-fold redundancy (Fig. 3a and Supplementary Table 2).

We selected three messages to encode: (1) “WHAT HATH GOD WROUGHT?”, the first long-distance message transmitted by Morse code in 1844; (2) “MR. WATSON, COME HERE!”, the first message transmitted by telephone in 1876; and (3) “BOUND FOREVER, DNA”, a translation of a lyric from the 2017 song DNA by the K-pop music group BTS. Each message was split into sets of four characters. Plasmids encoding a given set of pegRNAs were concurrently transfected along with a plasmid encoding PE2 into 5×TAPE-1(+) HEK293T cells at a ratio of 7:5:3:1, such that the ratio encoded the order of the four characters within each set (Fig. 3b). As such, each full message could be recorded by five to six consecutive transfections spaced by 3-day intervals.

To recover and decode the recorded messages, we collected populations of cells corresponding to each message and amplified and sequenced the DNA Tape region. From the resulting reads, we first identified all characters in the message by examining NNNGGA insertions at site 1 of 5×TAPE-1. We then grouped these characters into sets by hierarchical clustering (Extended Data Fig. 5a), while also ordering the sets relative to one another by applying the algorithm used for the previous experiment to the bigram transition matrix (Fig. 3c–e). Finally, we arranged the four characters within each set by decreasing the order of their edit score-corrected frequencies, as within each set earlier characters were encoded at a higher plasmid concentration.

For all three messages, our reconstructions of the original text were reasonable but imperfect. From the first message, 17 out of 22 characters were correctly recovered and ordered, with 3 deletion errors and 1 swap between adjacent characters to yield “WA HATH GOD WRUOGT?” (Fig. 3c). Of note, the deletion errors were due to repeated use of pegRNA barcodes ACT, CAT and GAC to encode multiple ‘H’ or ‘T’ characters that, as such, were not expected to be recovered separately. These deletion errors are the result of our encoding scheme, which used only 64 unique pegRNAs. We anticipate that greater information content per edit can be achieved with pegRNAs with longer barcodes; for example, 6-bp barcodes would have allowed each instance of repeated characters to be represented by different insertions, thereby avoiding this kind of error. In line with our previous analysis on decoding complex event histories, this inference was robust to sequencing depth, as undersampling did not appreciably add more errors to decoded messages (Extended Data Fig. 5b). From the second message, 20 out of 22 characters were correctly recovered and ordered, with two deletions and one insertion to yield “MR. WATSON, COMI HEE!” (Fig. 3d). From the third message, 16 out of 18 characters were correctly recovered and ordered, with a single swap between adjacent characters to yield “BOUND FOREVE,R DNA” (Fig. 3e). Despite these errors, our experiment demonstrates the potential of DNA Typewriter to digitally record the content and order of information to the genomes of populations of mammalian cells.

Beginning with genome editing of synthetic target arrays for lineage tracing (GESTALT), several approaches have been developed that leverage stochastic genome editing to generate a combinatorial diversity of mutations that irreversibly accumulate to a compact DNA barcode during in vivo development^6,30. Such stochastically evolving barcodes mark cells and enable inference of their lineage relationships on the basis of patterns of shared mutations. However, despite their promise, GESTALT and similar recorders remain sharply limited by several technical challenges, including (1) a failure to explicitly record the order of editing events, which renders phylogenetic reconstruction of cell lineage highly challenging^31,32; (2) a reliance on DSBs and NHEJ to introduce edits (DSBs frequently delete or corrupt consecutively located targets within a barcode); and (3) the decreasing number of target sites available to CRISPR–Cas9 as sites are irreversibly edited, which effectively makes it impossible to sustain continuous lineage recording over long periods of time without sacrificing resolution.

The ordered manner in which edits accrue with DNA Typewriter, the use of a prime editor with a Cas9 nickase to insert one of many possible symbols at the type guide, the predefined sequences and locations of potential edits, and the fact that one and only one monomer is an active type guide at any given moment have the potential to address all of these limitations at once. To demonstrate this potential, we sought to record cell lineage during the expansion of a monoclonal cell line, leveraging DNA Typewriter in combination with single-cell RNA-seq (scRNA-seq). First, we constructed a HEK293T cell line with doxycycline (Dox)-inducible PE2 expression (iPE2(+) HEK293T). We also designed and cloned a lentiviral construct that includes (1) the 5×TAPE-1 sequence, associated with a random 8-bp barcode region (TargetBC) at its 5′ end; (2) a transcription cassette for TargetBC-5×TAPE-1 with a reverse transcription capture sequence for enrichment during scRNA-seq; and (3) a constitutive pegRNA expression cassette that targets TAPE-1 for a 6-bp insertion (NNNGGA, referred to below as InsertBC; GGA is the key sequence for TAPE-1) (Fig. 4a). Lentiviral transduction of this construct into the cell line at a high multiplicity of infection (MOI) was followed by serial dilution to isolate a monoclonal cell line that grew from 1 cell to ~1.2 million cells via ~20 doublings over 25 days in the presence of Dox (Fig. 4b and Extended Data Fig. 6a). After collection, we used scRNA-seq to recover and sequence multiple TargetBC-5×TAPE-1 arrays from each of ~12,000 cells.

The frequency distribution of recurrently observed TargetBCs and InsertBCs in these data suggested that the MOI for this monoclonal cell line was ~19 (Extended Data Fig. 6b,c and Methods). However, the DNA Tapes associated with some TargetBCs were recovered more effectively than others (Extended Data Fig. 6b), presumably owing to site-of-integration effects on expression. To minimize complications related to missing data, we focused our analysis on cells for which we recovered DNA Tape sequences from all of the 13 most frequently observed TargetBCs, excluding one DNA Tape sequence with a corrupted type guide (TargetBC ATAAGCGG). Although the sequencing error rate was estimated to be very low (Extended Data Fig. 6d,e), accumulation of errors across edited sites might affect lineage reconstruction. We therefore also required that all edits to these DNA Tapes be among the 19 most frequently observed InsertBCs.

Applying these filters left 3,257 cells, for each of which we recovered intact TAPE-1 sequences for each of the 13 prioritized TargetBCs. Although nine of these TAPE-1 sequences were the expected five monomers in length, three were four monomers in length (TargetBCs TGGACGAC, TTTCGTGA and TGGTTTTG) and one was two monomers in length (TargetBC TTCACGTA). Because of their consistent length across the dataset, we infer that these TargetBC-specific contractions are due to pre-existing heterogeneity in the TAPE-1 lentiviral library before integration, rather than having been caused by editing. Thus, the TAPE-1 arrays on which we focused our analyses included 13 active type guides and 59 editable sites. With 59 editable sites and 19 potential edits per site, the overall complement of assayed DNA Tape in each cell has on the order of 10⁷⁵ possible states.

During monoclonal expansion, the generation of lineage barcodes in each cell was efficient, such that the vast majority of the cells assayed contained a unique editing pattern across the 59 sites (3,236/3,257 or 99.4%; 9 patterns recurred in 2 cells and 1 pattern recurred in 3 cells). After 25 days of editing, the first sites of active TAPE-1 arrays were edited to near saturation (mean, 96.8%) while the fifth sites were only modestly edited (mean, 19.7%) (Fig. 4c). Across all 13 DNA Tape arrays, the number of edits accruing per cell resembled a Poisson distribution, with the mean number of discrete events per cell (μ = 39.4) roughly equalling the variance (σ² = 40.0) (Fig. 4d). Assuming 20 cell divisions, this corresponds to an average of two edits accruing per cell division. The mean number of pairwise differences between cells, including sites at which one cell was edited and the other was unedited, was 41.9 ± 5.3 (Fig. 4e).

We next sought to construct a cell lineage tree. In contrast to GESTALT and other CRISPR–Cas9-based lineage recording systems, edits accruing to the multicopy DNA Tape derive from a finite set of pegRNA-specified symbols, analogous to the finite set of nucleotides or amino acids used to build conventional phylogenetic trees. However, in further contrast to GESTALT but also to conventional phylogenetics, DNA Typewriter provides explicit information regarding the order in which differences accrued. To leverage this, we constructed a 3,257 × 3,257 similarity matrix by calculating, for all possible pairs of cells, the number of shared edits across the 59 sites. However, for shared edits at any given site to be counted, we required that all earlier sites along that DNA Tape also be identically edited (Methods). Across all 5.3 million pairwise comparisons of cells, 24 million of 33 million shared edits met this criterion; those that did not presumably correspond to coincident occurrences of the same edit at the same site in different cells and, as such, are appropriate to discount. After converting this similarity matrix to a distance matrix, we generated two phylogenetic trees, using either the unweighted pair group method with arithmetic mean (UPGMA) or the neighbour-joining hierarchical clustering method. When comparing these two methods, UPGMA resulted in a tree with a lower parsimony score of 123,625, compared with the score of 124,997 for the tree constructed using neighbour-joining hierarchical clustering. A compact representation of the UPGMA tree is shown in Fig. 5a, with the full tree in Supplementary Fig. 1.

To assess robustness, we first focused on two distantly related clades of 16 cells from the global UPGMA tree, merged them into a new set of 32 cells and then performed conventional bootstrapping, treating the sites associated with each of the 13 TargetBCs as independent groups, sampling 13 TargetBC groups with replacement, and then constructing and comparing UPGMA-based trees (Methods). Across 100 resamplings, all 31 branchings were observed multiple times, 20 with bootstrap values over 50%, with a bootstrap value of 100% for the separation between the two distantly related clades (Fig. 5b). Bootstrap analysis of an additional clade of 81 cells is shown in Extended Data Fig. 6f; for this clade, all 80 branchings were observed multiple times, 38 with bootstrap values over 50%. Finally, we performed bootstrap analysis of the entire matrix, resulting in a tree in which 76% of branches were seen multiple times and 25% had bootstrap values over 50% (Supplementary Fig. 1).

In summary, over the course of 25 days of expansion of a monoclonal cell line from 1 to ~1.2 million cells, we observed the ordered accumulation of 39.4 ± 6.3 edits to 59 sites located within 13 DNA Tape arrays. Although the number of active type guides at these arrays declined (from 13 in the founding cell to a mean of 8.6 active type guides per cell after 25 days), we did not exhaust the recording capacity of the system (only 1 of the 3,257 sampled cells was edited at all 59 sites). To further assess whether editing was maintained throughout the experiment, we examined the number of pairwise differences between each cell and its nearest neighbour within the sampled set of 3,257 cells (Fig. 4f). On average, cells were separated from their nearest neighbour by 22.8 edits (or, assuming a constant rate of ~2 edits per generation, 11 to 12 generations). We interpret this as strong support for the conclusion that editing of the DNA Tapes was maintained throughout clonal expansion.

As illustrated by this lineage tracing experiment, we can deploy and recover at least a dozen DNA Tapes in each cell, which substantially increases information capacity. However, even with multiple DNA Tapes, the maximum potential recording duration of each DNA Tape remains directly proportional to the number of consecutive monomers on each DNA Tape. Although 5×TAPE-1 appears to be very stable within cells as well as throughout amplification and sequencing (Extended Data Fig. 4b), longer tandem arrays might introduce additional technical challenges, for example, by being difficult to synthesize, clone and maintain, prone to instability during in vivo DNA replication or repair as well as during in vitro PCR, and difficult to accurately and fully sequence.

To evaluate the extent to which such issues might be limiting in practice, we sought to generate a synthetic minisatellite in the form of 12 or 20 repeats of the 14-bp TAPE-1 monomer. 12×TAPE-1 was synthesized as single-stranded DNA (IDT), and 20×TAPE-1 was synthesized as a plasmid (GenScript). PCR amplicons of each array were cloned into the piggyBac vector through Gibson assembly. Of note, cloned constructs were used ‘as is’, even though it is possible that some degree of variation in repeat number was already present (Extended Data Fig. 7a,b). We integrated piggyBac vectors bearing ~12×TAPE-1 or ~20×TAPE-1 into HEK293T cells expressing both PE2 and pegRNAs targeting TAPE-1 for NNNGGA insertions (PE2(+) 3N-TAPE-1-pegRNA(+) HEK293T) in triplicate. We cultured these cell lines for 40 days before collecting genomic DNA. PCR amplification of TAPE-1 was followed by standard library construction and sequencing on the Pacific Biosciences Sequel platform to obtain circular consensus sequencing (CCS) reads. On average, we recovered 8.4 ± 3.3 repeats of TAPE-1 monomers from 12×TAPE-1 and 12.5 ± 4.3 repeats from 20×TAPE-1. In each case, there was a sharp drop-off after the intended length of 12 or 20 monomers, suggesting that, regardless of the mechanism, these longer arrays are more prone to contraction than expansion (Extended Data Fig. 7c). Of note, the editing rates were the same for the constructs (4.5 ± 1.3 edits and 4.5 ± 1.5 edits for the 12×TAPE-1 and 20×TAPE-1 arrays, respectively; Extended Data Fig. 7d). This is expected, as each DNA Tape has exactly one active type guide and, as such, the rates at which they are written to should be independent of their length.

We grouped CCS reads within each replicate on the basis of a degenerate 8-bp barcode (TargetBC), as these presumably derived from the same integration. On average, each TargetBC group had 3.1 ± 3.4 and 3.8 ± 5.7 reads for ~12×TAPE-1 and ~20×TAPE-1, respectively. Within TargetBC groups, shorter arrays appeared more stable, with a greater proportion matching the maximum length within that group (Extended Data Fig. 7e,f). Of representative CCS reads for 4,784 and 6,254 integrated arrays for 12×TAPE-1 and 20×TAPE-1, respectively, the overwhelming majority (>99.5%) exhibited clear patterns of sequential, directed editing (Extended Data Fig. 7g,h). In terms of the maximum extent to which any given DNA Tape was edited, we observed one TargetBC for which 14 distinct 3-bp insertion events were recorded along a 14-monomer DNA Tape.

This experiment illustrates that it is possible to construct and use synthetic minisatellites corresponding to at least 20 monomers as a DNA Tape and that sequential recording of at least 14 consecutive events with DNA Typewriter is possible. Nonetheless, further experiments are required to quantify the extent to which variation in synthetic minisatellite length is due to (1) piggyBac vector heterogeneity, that is, variation that existed before integration; (2) DNA replication and microsatellite instability in HEK293T cells; (3) DNA repair subsequent to prime editing-induced nicks; and/or (4) PCR amplification artefacts. Of note, the observed variation in array length tended to occur within the unedited portion of the DNA Tape (Extended Data Fig. 7g,h). We have yet to observe any clear examples of ‘information erasure’, possibly because the edits themselves disrupt the tandem repeats, inhibiting processes that might otherwise lead to erasure from spreading proximal to the type guide.

Digital systems represent information through both the content and order of discrete symbols, with each symbol drawn from a finite set. Digital systems are ancient and include written text, Morse code and binary data as well as, of course, genomic DNA. In this proof-of-concept study of DNA Typewriter, we demonstrate how sequential genome editing of a monomeric array constitutes an artificial digital system that is operational within living eukaryotic cells, capable of ‘writing’ thousands of discrete symbols to DNA in an ordered fashion.

DNA Typewriter improves on existing CRISPR-based molecular recorders in important ways (Supplementary Table 3). The sequential editing achieved by DNA Typewriter resembles Cas1–Cas2-based recording^10,11,16, which at present is limited to bacterial systems. In DOMINO¹⁹ and CAMERA¹⁴, base editors are used to record biological signals to ‘preprogrammed logic circuits’ composed of multiple targets for base editing. Although these methods are conceptual predecessors to DNA Typewriter, there are critical differences. In particular, with all three methods, a recording event creates a new target for further editing (that is, the type guide). However, with DOMINO and CAMERA, each logic circuit is designed to record a specific order. By contrast, a single DNA Typewriter construct can potentially record any order. For example, to distinguish pairwise orderings within a set of n events, DOMINO or CAMERA would require n-choose-2 recording logic circuits or a system that contains on the order of n² unique gRNAs and their targets. By contrast and as demonstrated here (Fig. 2), DNA Typewriter requires only a single target array such as 5×TAPE-1, along with n unique pegRNAs that encode different insertions but share the same target.

How might we write biologically specific information using DNA Typewriter? Here we use pegRNAs to encode symbols (that is, insertional barcodes), but these pegRNAs are introduced by artificial transduction or stochastic expression. However, several groups have engineered gRNAs whose activity is dependent on the binding of specific small molecules or ligands^33–35. Also, we recently developed ENGRAM, a prime editing-based system in which biological signals of interest such as NF-κB and Wnt signals are coupled to the production of specific pegRNAs³⁶. These pegRNAs mediate the insertion of signal-specific barcodes to a DNA-based recording site, providing quantitative information with respect to the strength and/or duration of the signal(s). At least in principle, such strategies are compatible with the current implementation of DNA Typewriter, potentially enabling the temporal dynamics of multiple biological signals or other cellular events to be recorded and resolved. In this context, the use of longer and therefore more diverse insertion barcodes could enable extensive multiplexing, although this might come at the expense of recording efficiency. A further caution is that we estimate the rate of prime editing to be on the order of days, such that DNA Typewriter may be most useful for recording information about biological processes that unfold over a timescale of days or weeks, rather than minutes or hours.

One such process is biological development, wherein the unfolding of a cell lineage tree is of fundamental interest. In a proof-of-concept experiment, we show how DNA Typewriter overcomes the major limitations of earlier editing-based lineage recorders such as GESTALT^6,30 by reducing ambiguity about the order in which editing events occurred, eschewing DSBs and thereby minimizing the risk of inter-target deletion, predefining the locations to which edits accrue, predefining the ‘symbol set’ from which edits are drawn and stabilizing the rate of editing by ensuring one and only one type guide per active DNA Tape. These attributes clearly paid off in our proof-of-concept experiment, as we were able to sustain a seemingly steady accumulation of edits to multicopy DNA Tape across 25 days of in vitro expansion, from a single cell to over 1 million cells. Although this is longer than the gestation period of a mouse, we do not exhaust the recording capacity of the system. Furthermore, the resulting data are sufficiently rich and complete that we can build and characterize cell lineage trees from these data with conventional phylogenetic algorithms (for example, UPGMA and NJ), with only minor modifications directed at leveraging information about the order of edits, not available in other contexts in which phylogenetics is applied. In this experiment, the number of edits accruing per cell resembled a Poisson distribution. Further experiments are needed to assess the extent to which this rate of accrual is a function of absolute time, the cell cycle or some combination thereof. However, as it has been shown that prime editing continues to take place in non-mitotic cells such as neurons², we suspect that it is primarily a function of time.

What are the limits of this approach? Under the assumption that we can achieve similar performance in vivo (multiple efficiently recovered DNA Tapes per cell; steady accrual of edits over several weeks; multiple edits per lineage per cell division), we can readily conceive of a technical path to Sulstonesque reconstructions³⁷ of the cell lineage histories of non-transparent model organisms, for example, flies, mice, zebrafish and macaques. We further envision that a single synthetic DNA construct that encodes a prime editing enzyme, multiple recording arrays and a combination of stochastic and signal-specific pegRNAs could be used to simultaneously record both lineage and biological signals in any multicellular system, that is, a molecular ‘flight recorder’ locus. A single-locus design would be less affected by site-of-integration effects, such as we have observed with multiple DNA Tape constructs integrated across the genome. Alternatively, if genomic sites with a high prime editing efficiency can be identified, such sites might be leveraged to boost information capture. A separate risk is that prime editing efficiency might vary substantially across cell types. However, any such variation could potentially be ameliorated by technical improvements to system components^22,23, by increasing recording capacity and/or by modelling it during tree reconstruction. Although challenging to engineer, a generic recorder locus would allow us to take full advantage of DNA as an in vivo digital recording medium, for example, not only to characterize wild-type development, but also to enable systematic comparison of the developmental histories of wild-type and mutant individuals.

Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41586-022-04922-8.

We thank the members of the Shendure laboratory, especially S. Srivatsan and J.-B. Lalanne, for helpful discussions and approaches for data analysis. We thank L. Starita for critical reading of the manuscript. We thank the members of the Allen Discovery Center for Cell Lineage Tracing, particularly A. Schier and M. Elowitz, for helpful discussions over several years. We thank D. Liu’s laboratory for sharing the prime editing plasmids. This work was supported by a grant from the Paul G. Allen Frontiers Group (Allen Discovery Center for Cell Lineage Tracing to J.S.), the National Human Genome Research Institute (UM1HG011586 to J.S.; R01HG010632 to J.S.) and the Brotman Baty Institute for Precision Medicine. J.C. is a Howard Hughes Medical Institute Fellow of the Damon Runyon Cancer Research Foundation (DRG-2403-20). J.S. is an Investigator of the Howard Hughes Medical Institute.

This project was initiated during discussions among J.C., W.C. and J.S. J.C. and J.S. designed experiments with input from W.C. and A.M. throughout the study. J.C., W.C., A.M., F.M.C., C.C.S., S.G.R. and S.D. contributed to reagent and material preparation. J.C. performed experiments with assistance from W.C., F.M.C., S.G.R., N.H., C.L., B.M. and R.M.D. Single-cell assays were performed by J.C. and S.G.R. with input from A.M. and F.M.C. J.C. and J.S. analysed the data with assistance from W.C., A.M., F.M.C. and S.G.R. J.C. and J.S. wrote the manuscript with input from all other authors.

Raw sequencing data have been uploaded to the Sequence Read Archive with associated BioProject ID PRJNA757179↗.

Custom analysis code for this project is available at GitHub (https://github.com/shendurelab/DNATickerTape↗) and Figshare (10.6084/m9.figshare.19607811).

The University of Washington has filed a patent application partially based on this work in which J.C., W.C. and J.S. are listed as inventors. The remaining authors declare no competing interests.

Junhong Choi, Email: junhongc@uw.edu.

Jay Shendure, Email: shendure@uw.edu.

is available for this paper at 10.1038/s41586-022-04922-8.

The online version contains supplementary material available at 10.1038/s41586-022-04922-8.