Main<\/h2>\n\nThe efficient insertion of short DNA sequences into genomes could change the course of biotechnology and medicine1<\/a>,2<\/a><\/sup>. Small insertions can encode protein tags for purification and visualization, or manipulate protein function by altering protein localization, half-life or interaction profiles. Integrating sequences for transcription factor binding sites and splicing modulators provides control over gene expression while introducing structural elements or recombinase sites can change DNA conformation and provide a substrate for large-scale engineering. For therapeutic opportunities, over 16,000 small deletion variants have been causally linked to disease3<\/a>,4<\/a><\/sup>, and could in principle be restored by inserting the missing sequence5<\/a>,6<\/a><\/sup>. A prominent example is cystic fibrosis, where 70% of cases are caused by a three-nucleotide (nt) deletion7<\/a>,8<\/a><\/sup>. To enable reversing these mutations in practice, a technology must integrate insertions efficiently, accurately and safely, avoiding the unintended outcomes and double-strand break stress that hampers existing Cas9-based therapies9<\/a>,10<\/a>,11<\/a><\/sup>.<\/p>\nPrime editors can insert short DNA sequences without generating double-strand breaks or requiring an external template. They consist of a nicking version of Cas9 fused to a reverse transcriptase domain, which is complexed with a prime editing guide RNA (pegRNA)12<\/a><\/sup>. The pegRNA comprises a primer binding site homologous to the sequence in the target, and a reverse transcriptase template that includes the intended edit, all in the 3\u2032 extension of a standard CRISPR\u2013Cas9 guide RNA. At the target site, Cas9 nicks one strand of the genomic DNA, which then anneals to the primer binding site on the pegRNA, and is extended by the Cas9-fused reverse transcriptase using the pegRNA-encoded template sequence. Next, DNA repair mechanisms resolve the conflicting sequences on the two DNA strands, ultimately writing the intended edit into the genome. Where CRISPR\u2013Cas9 was compared with molecular scissors capable of disrupting target genes, and base editors were seen as molecular pencils for their ability to substitute single nucleotides, prime editors can be described as molecular word processors, able to perform search and replace operations directly on the genome13<\/a>,14<\/a>,15<\/a>,16<\/a><\/sup>.<\/p>\nThe prime editing system is complex, and the determinants of its efficiency are not fully understood. Several partly independent steps, including three DNA binding events and successful DNA repair, are needed to produce an edit, each potentially influenced by the introduced sequence. In the largest study so far to understand these biases, Kim et al. comprehensively tested the consequences of varying the reverse transcription templates and primer binding site lengths using a library of 55,000 pegRNAs. The editing rate increased with Cas9 guide RNA activity, as well as GC content and melting temperature of the primer binding site. While further optimization of sequences was possible, primer binding sites of 11\u201313\u2009nt and reverse transcriptase templates of 10\u201312\u2009nt had the highest average editing efficiencies17<\/a><\/sup>.<\/p>\nThe majority of libraries used by Kim et al. contained the same single-nucleotide substitution 5\u2009nt upstream of the nick site. Similarly, nearly all investigations of prime editing efficacy to date have predominantly focused on single-nucleotide substitutions12<\/a>,17<\/a>,18<\/a>,19<\/a>,20<\/a>,21<\/a><\/sup>. Of the many possible useful sequences in molecular biology, only a handful have been introduced with prime editing12<\/a><\/sup>. Therefore, in contrast to a relatively deep understanding of Cas9 mutagenesis10<\/a>,22<\/a>,23<\/a>,24<\/a><\/sup> and base editing outcomes25<\/a>,26<\/a>,27<\/a><\/sup>, very little is known about how the inserted sequence affects efficiency, and the length range of insertions feasible by prime editing has not been defined.<\/p>\nHere, we systematically measure the insertion efficiency of 3,604 sequences in several target sites and a variety of cellular and repair pathway contexts. We find that insertion sequence length, nucleotide composition and secondary structure all affect insertion efficiency. Moreover, we define the precise effect of mismatch repair (MMR) on thousands of insertion sequences and discover that overexpression of the 3\u2032 flap nucleases TREX1 and TREX2 abolished the insertion of longer sequences. Together, sequence features and repair pathway activity explain most of the variation in insertion rate. We then use these insights to train a sequence-based prediction model informed by MMR efficiency that predicts editing outcomes for novel sequences with high accuracy and demonstrate the model\u2019s usefulness for the selection of optimal reagents for new insertions.<\/p>\n<\/div>\n<\/div>\n
\nResults<\/h2>\n\nWe sought to systematically characterize how the length and composition of inserted sequence, as well as cell line, target site and the version of the prime editor system, affect insertion rates. To do so, we designed 3,604 pegRNAs encoding insertions immediately upstream of the nick site. These comprise 270 sequences useful for molecular biology (for example, His-6 tag, recombinase sites and mNeonGreen-11 (ref. 28<\/a><\/sup>)); 1,957 eukaryotic linear motifs29<\/a>,30<\/a>,31<\/a><\/sup>; 439 sequences with variable secondary structure; all single nucleotides, dinucleotides, trinucleotides and tetranucleotides; and 100 random sequences of each length between 5 and 10\u2009nt (Fig. 1a<\/a>). Insertions ranged from the length of 1 to 69\u2009nt, and varied in GC content (Fig. 1b<\/a>), while the primer binding site and homology arm lengths in the pegRNA were fixed to 13 and 34\u2009nt, respectively. We used lentiviruses to deliver the libraries against four target sites (three previously tested: HEK3<\/i>, EMX1<\/i>, FANCF<\/i>12<\/a><\/sup>, and the safe-harbor CLYBL<\/i> locus32<\/a><\/sup>) in two cell lines (HEK293T and HAP1), followed by transient transfection of the prime editor 2 plasmid (HEK293T cells) or doxycycline induction of PiggyBac transposase integrated prime editor (HAP1 cells), five\u2009d of selection and sequencing of two amplicons from the cell pool, one of the targeted locus and one of the pegRNA locus (Fig. 1c<\/a>). We calculated insertion efficiencies as the fraction of reads in the target site amplicon with a given insertion divided by the fraction of reads for the pegRNA encoding it in the pegRNA amplicon, and analyzed them as the main statistic in the rest of the study.<\/p>\n\nFig. 1: High-throughput measurement of prime insertion efficiencies.<\/b><\/figcaption>\n![\"Science](\"http:\/\/media.springernature.com\/lw685\/springer-static\/image\/art%3A10.1038%2Fs41587-023-01678-y\/MediaObjects\/41587_2023_1678_Fig1_HTML.png\")
<\/picture><\/a><\/div>\na<\/b>, Screen setup. Set 1 and Set 2 libraries were screened separately and data merged (Methods<\/a>); panels d<\/b>\u2013f<\/b> reflect Set 1 results only. b<\/b>, Library composition. The number of sequences in the library (y<\/i> axis) with different insert sequence lengths (x<\/i> axis, top panel) and %GC content (x<\/i> axis, bottom panel). c<\/b>, Experimental design. NGS, next generation sequencing. d<\/b>, Editing frequencies. Average mutation frequency (y<\/i> axis) for different screens (x<\/i> axis) stratified by mutation type (blue, insertions; gray, unintended outcomes). Markers represent one replicate and bars the average across n<\/i>\u2009=\u20093 biological replicates. e<\/b>, Replicate concordance. Pearson\u2019s R<\/i> between insertion rates in two screens (x<\/i> axis) for different comparisons (y<\/i> axis, colors). Markers, correlation value of one pair of screens (for replicate correlations, mean of pairwise comparison across n<\/i>\u2009=\u20093 biological replicates); line and whiskers, mean and s.e.m. f<\/b>, Representative examples of categories from e<\/b>. Percentage insertion in the HEK3<\/i> locus in HEK293T cells (y<\/i> axis) compared with values (x<\/i> axis) in other contexts (panels, colors) for insertion sequences (markers). Left panel, comparison of biological replicates; other panels, comparison of replicate averages. Label, R<\/i> of values in linear scale. Colors as in e<\/b>.<\/p>\n<\/div>\nFull size image<\/span><\/a><\/p>\n<\/figure>\n<\/div>\nInsertion efficiencies of sequences varied widely. The top 5% of templates were inserted 27\u2013134 times more efficiently than the bottom 5% across the various target site and cell line combinations (Supplementary Fig. 1a,b<\/a>), indicating substantial sequence-dependent variation. The insertion rates were highly consistent across biological replicates (median R<\/i>\u2009=\u20090.70; Supplementary Fig. 1c\u2013i<\/a>), but differed in magnitude across screens (average across pegRNAs, 0.18% for the CLYBL<\/i> locus in HEK293T to 6.7% for the HEK3<\/i> locus in HEK293T cells; Fig. 1d<\/a>). Unintended editing outcomes we observed included single-base mutations, small insertions and deletions around the nicking site, deletions overlapping primer binding site and reverse transcription template, insertion of mutated library sequences, duplications of the reverse transcription template, as well as partial scaffold integrations (Fig. 1d<\/a> and Supplementary Fig. 2a\u2013c<\/a>). These outcomes were rare overall (0.06\u20130.45%). Base changes at the target site were infrequent in reads with and without insertions (0.038% versus 0.030%), but slightly elevated upon insertion immediately downstream of the nick site and for the first nucleotides after the end of the homology arm (Supplementary Fig. 2d\u2013f<\/a>). Overall, the intended insertions were the dominant mutations generated, and we do not consider the unintended edits further.<\/p>\nTo understand the consistency of insertion efficiencies across contexts, we next compared them between replicates, cell lines and target sites. Insertion rates into the same target site in different cell lines were more correlated (mean R<\/i>\u2009=\u20090.52) than into different target sites in the same line (mean R<\/i>\u2009=\u20090.38). The correlation was weakest when both the target site and cell line were different (mean R<\/i>\u2009=\u20090.17; Fig. 1e,f<\/a>), demonstrating both target sequence-specific and cell line-dependent biases on insertion.<\/p>\nInsert size and MMR activity effects<\/h3>\n
Given the repeatable sequence-dependent variation in insertion rates that spans over three orders of magnitude, we sought to understand the responsible features, starting with insert length. Insertion frequency did not decrease monotonically with insert length in HEK293T cells, but instead, had two modes of high values. First, sequences of 3 and 4\u2009nt were inserted on average 2.0\u20134.1 times more efficiently than others across the four targeted sites (Fig. 2a<\/a>). Second, sequences between 15 and 21\u2009nt were inserted on average 1.3\u20131.6 times more efficiently than 10\u201314-nt ones, and 1.5\u20132.0 times more efficiently than sequences longer than 21\u2009nt (Fig. 2a<\/a>). These relative biases in efficiency were shared between all target sites, despite a 20-fold range of their average insertion rates. Inserts longer than 45\u2009nt were incorporated less frequently, at a screen average rate that is 4\u20138 times lower than that of sequences shorter than 45\u2009nt. The longest sequence that was inserted at >1% frequency (1.4%, HEK3<\/i> site in HEK293T cells) was 66\u2009nt, demonstrating that integration of moderately long sequences is feasible with prime editing.<\/p>\n\nFig. 2: Prime insertion efficiency depends on insert length and MMR.<\/b><\/figcaption>\n![\"Science](\"http:\/\/media.springernature.com\/lw685\/springer-static\/image\/art%3A10.1038%2Fs41587-023-01678-y\/MediaObjects\/41587_2023_1678_Fig2_HTML.png\")
<\/picture><\/a><\/div>\na<\/b>, Insertion rate in HEK293T cells. Percentage of reads with insertion (y<\/i> axis, cut-off at 3\u2009s.d. above mean) for different insert sizes (x<\/i> axis) of individual sequences (blue markers) and averages for lengths with at least 30 measured sequences (dark blue line and markers) at different target sites (panels). Data represent the average of n<\/i>\u2009=\u20093 biological replicates. b<\/b>, As a<\/b>, but for HAP1 cells. c<\/b>, As a<\/b>, but for HAP1 \u2206MLH1<\/i> cells. d<\/b>, Insertion rate in one cell context (y<\/i> axis) compared with in another context (x<\/i> axis) at the HEK3<\/i> target of individual sequences (markers), comparing HEK293T with HAP1 cells (left panel) and HEK293T cells with HAP1 \u2206MLH1<\/i> cells (middle panel). Red, short sequences (up to 4\u2009nt); blue, medium sequences (5\u201313\u2009nt); teal, longer sequences (>13\u2009nt). Label, R<\/i> between rates. The data are an average from n<\/i>\u2009=\u20093 biological replicates (HEK293T) or n<\/i>\u2009=\u20092 biological replicates (HAP1). e<\/b>, Average insertion rates (y<\/i> axis) across insert lengths (x<\/i> axis) with at least 30 measured sequences in various cell line contexts (colors). Data are presented as mean\u2009\u00b1\u2009s.e.m. n<\/i>\u2009=\u20093 biological replicates (HEK293T) or n<\/i>\u2009=\u20092 biological replicates (HAP1). f<\/b>, The ratio of relative insertion rates (Methods<\/a>) at the HEK3<\/i> locus between HAP1 \u2206MLH1<\/i> and HAP1 cells (y<\/i> axis) for different lengths (x<\/i> axis) stratified by colors as in d<\/b>. Box, median and quartiles; whiskers, least extreme of 1.5 times the interquartile range from the quartile and most extreme values. Line, fit from an exponential model (ratio\u2009\u2248\u2009a<\/i>\u2009\u00d7 exp(\u2212b<\/i>\u2009\u00d7\u2009length)\u2009+\u20091). n<\/i>\u2009=\u20092 biological replicates.<\/p>\n<\/div>\nFull size image<\/span><\/a><\/p>\n<\/figure>\n<\/div>\nIn contrast to HEK293T cells, the insertion frequency of the short 1\u20134-nt sequences was not substantially higher than that of longer ones in HAP1 cells (0.60\u20131.27 times; Fig. 2b<\/a>). This reduced the concordance of insertion rates in the two cell lines at the same site (R<\/i>\u2009=\u20090.41 for FANCF<\/i> and 0.54 for HEK3<\/i>; Fig. 2d<\/a> and Supplementary Fig. 3a<\/a>) compared with replicates (median R<\/i>\u2009=\u20090.78; Fig. 1e<\/a>). One possible explanation is MMR proficiency, since HEK293T cells are partly MMR deficient due to promoter methylation of MLH1<\/i> (ref. 33<\/a><\/sup>), while HAP1 cells are not. The MMR pathway recognizes and excises short mismatches of less than 13\u2009nt and could therefore remove short insertions in HAP1 cells before the nicked strand is re-ligated34<\/a><\/sup>. Indeed, MMR antagonizes prime editing for substitutions and short insertions20<\/a>,35<\/a><\/sup>. Consistent with this explanation, we observed strong correlations between insertion rates in HAP1 and HEK293 cells for sequences longer than 13\u2009nt that are not affected by MMR (R<\/i>\u2009=\u20090.78 for the FANCF<\/i> locus and 0.91 for the HEK3<\/i> locus; Fig. 2c<\/a> and Supplementary Fig. 3a<\/a>).<\/p>\nTo experimentally test the hypothesis that rates of insertion of short sequences differ between cell lines due to MMR activity, we screened the HEK3<\/i>– and FANCF<\/i>-targeted libraries in HAP1 cells that are knocked out for MLH1<\/i> (HAP1 \u2206MLH1<\/i>; Fig. 2d<\/a> and Supplementary Fig. 3b,c<\/a>). We found that the average insertion rates of 1\u20134-nt sequences were most affected by the knockout, increasing by 7.2\u201311-fold, while the rates of 5\u201313-nt sequences increased 2.1\u20132.7-fold (Fig. 2e<\/a> and Supplementary Fig. 3d<\/a>). Overall, 66% (HEK3<\/i>) and 67% (FANCF<\/i>) of the variance in the fold changes (Fig. 2f<\/a> and Supplementary Fig. 3d<\/a>) was explained by a model where the loss of MMR increases the insertion rate of 1-nt sequences by 23\u201328-fold, with the increase in insertion efficiency dropping 40\u201348% for every additional nucleotide. The low correlations of insertion rates between HEK293T and wild-type HAP1 cells (R<\/i>\u2009=\u20090.41\u20130.54) also improved to close to replicate concordance when matching MMR status (R<\/i>\u2009=\u20090.73\u20130.96 between HEK293T and HAP1 \u2206MLH1<\/i> cell lines; Fig. 2c<\/a> and Supplementary Fig. 3a,e<\/a>). In summary, our findings highlight that MMR proficiency is the major source of independent variation between the tested cellular contexts for prime insertion of short sequences.<\/p>\nEffects of prime editing steps<\/h3>\n
Having confirmed MMR as a length-dependent determinant of insertion efficiency, we next sought to understand how different steps of prime editing affect insertion rates of our library sequences. Specifically, we dissected the contributions of (1) pegRNA expression, (2) reverse transcription by two different reverse transcriptases, (3) presence of a nicking guide and (4) overexpression of 3\u2032 and 5\u2032 flap nucleases (Fig. 3a<\/a>).<\/p>\n\nFig. 3: Effects of prime editing steps.<\/b><\/figcaption>\n![\"Science](\"http:\/\/media.springernature.com\/lw685\/springer-static\/image\/art%3A10.1038%2Fs41587-023-01678-y\/MediaObjects\/41587_2023_1678_Fig3_HTML.png\")
<\/picture><\/a><\/div>\na<\/b>, Schematic of molecular steps involved in prime editing. b<\/b>, Normalized pegRNA count derived from sequencing of PCR amplicons from genomic DNA (x<\/i> axis) or PCR amplicons from RNA (y<\/i> axis) for the HEK3<\/i> site in HEK293T cells for individual pegRNAs (markers). Pink, inserts with four or more consecutive adenines. Data represent the average of n<\/i>\u2009=\u20093 biological replicates. c<\/b>, Top panel, average insertion rate relative to length bin median (y<\/i> axis) for inserts stratified by the longest consecutive run of adenines (x<\/i> axis). Bottom panel, instead showing transcription rate (read counts from RNA\/read counts from DNA) on the y<\/i> axis. Data are presented as mean\u2009\u00b1\u2009s.e.m. n<\/i>\u2009=\u20093 biological replicates. d<\/b>, Insertion frequencies at the HEK3<\/i> site in HEK293T using the standard MMLV reverse transcriptase (PE2, x<\/i> axis) and the FeLV reverse transcriptase (PE-FeLV, y<\/i> axis) for different insertion sequences (markers). Colors, number of neighboring points. n<\/i>\u2009=\u20093 biological replicates. e<\/b>, As d<\/b>, but comparing PE3 and PE2 at the EMX1<\/i> site. f<\/b>, Schematic of screens with overexpression constructs. g<\/b>, Insertion frequencies for different overexpressions (y<\/i> axis and panels) compared with no overexpression (x<\/i> axis) for three biological replicate screens (markers) stratified by insertion sequence lengths (colors). h<\/b>, Average insertion rates (y<\/i> axis) across insert lengths (x<\/i> axis) with at least 30 measured sequences for overexpression constructs (colors). Data are presented as mean\u2009\u00b1\u2009s.e.m. n<\/i>\u2009=\u20093 biological replicates. i<\/b>, As h<\/b>, but instead displaying the insertion rate fold changes of screens with overexpressions compared with no overexpression (y<\/i> axis), calculated from the ratio of sums of all sequences (lines) or of ten randomly sampled sequences. j<\/b>, Top, average insertion frequency (grayscale) of four sequences with varying lengths (x<\/i> axis) when overexpressing eGFP stratified by homology arm lengths (panels). Bottom, insertion rate fold changes compared with eGFP (y<\/i> axis) when overexpressing TREX1 and TREX2 (colors). n<\/i>\u2009=\u20092 biological replicates. k<\/b>, Fraction of the nontemplated adenine allele at the +9 position (y<\/i> axis) for cells with overexpression constructs (x<\/i> axis and colors) stratified by experiment and homology arm lengths (panels). Markers show screen averages from three biological replicates for the pooled screen or from separate pegRNAs for the individual validation experiment.<\/p>\n<\/div>\n
Full size image<\/span><\/a><\/p>\n<\/figure>\n<\/div>\nWe first assessed expression levels of pegRNAs targeting the HEK3<\/i> site in HEK293T cells using deep sequencing. Abundance in the transcriptome was well correlated between replicates (median R<\/i>\u2009=\u20090.97; Supplementary Fig. 4a<\/a>) and with the DNA-derived read count frequency (R<\/i>\u2009=\u20090.56; Fig. 4b<\/a>). The exceptions were sequences that resulted in four or more consecutive thymines on the pegRNA cassette (adenines in the inserted DNA), which act as transcription terminators for RNA polymerase III (refs. 36<\/a>,37<\/a><\/sup>). Upon removing pegRNAs with terminator motifs, the correlation between measured DNA and RNA sequence coverage increased to 0.59 (Fig. 3b<\/a>). Sequences with four or more consecutive adenines were 4.8-fold less expressed and, accordingly, their average insertion rate was 4.8-fold lower compared with other sequences (Fig. 3c<\/a> and Supplementary Fig. 4b<\/a>). Overall, 23 of the 24 inserts (96%) that were not observed in any screen contained at least one run of four or more adenines, highlighting this feature as a useful filter in pegRNA design.<\/p>\n\nFig. 4: Cytosine content and secondary structure of the insert sequence are positively correlated with the insertion rate.<\/b><\/figcaption>\n![\"Science](\"http:\/\/media.springernature.com\/lw685\/springer-static\/image\/art%3A10.1038%2Fs41587-023-01678-y\/MediaObjects\/41587_2023_1678_Fig4_HTML.png\")
<\/picture><\/a><\/div>\na<\/b>, Correlation of length-normalized insertion rate with nucleotide frequency in the insert (colors) for each nucleotide (y<\/i> axis) in each screen (x<\/i> axis). Data represent the average of n<\/i>\u2009=\u20093 (HEK293T) or n<\/i>\u2009=\u20092 (HAP1) biological replicates. b<\/b>, As a<\/b>, but for a new set of screens with 18-nt inserts and 15-nt homology arms targeting five novel sites within 1\u2009kb of the HEK3<\/i> site. c<\/b>, Insertion rate at the HEK3<\/i> site in HEK293T cells relative to length bin median (y<\/i> axis) for inserts (markers) with different cytosine content (x<\/i> axis). Line, linear regression fit; shaded area, 95% posterior confidence interval of the fit. Data represent the average of n<\/i>\u2009=\u20093 biological replicates. d<\/b>, Insertion rates at the HEK3<\/i> site in HEK293T cells relative to length bin median (y<\/i> axis) for inserts (markers) with calculated Gibbs free energy (\u2206G<\/i>) from ViennaFold (x<\/i> axis). Line, linear regression fit; shaded area, 95% posterior confidence interval of the fit. Data represent the average of n<\/i>\u2009=\u20093 biological replicates. e<\/b>, Correlation (x<\/i> axis) between insertion rates and insert sequence free energy calculated from different parts of the 3\u2032 extension (y<\/i> axis). Box, median and quartiles; whiskers, least extreme of 1.5 times the interquartile range from the quartile and most extreme values. n<\/i>\u2009=\u20093 (HEK293T) or n<\/i>\u2009=\u20092 (HAP1) biological replicates. f<\/b>, Insertion rates for sequences (markers) at the HEK3<\/i> site in HEK293T for pegRNAs (x<\/i> axis) and epegRNAs (y<\/i> axis). Data represent the average of n<\/i>\u2009=\u20093 biological replicates. g<\/b>, Percentage increase in insertion rate with each standard deviation increase in structure strength (colors) for different overexpression constructs (x<\/i> axis) and insertion sequence lengths (y<\/i> axis). h<\/b>, Insertion rates relative to length bin median (y<\/i> axis) for sequences that disrupt or preserve (x<\/i> axis) scaffold loops (panels). Colored lines show screen medians and the thicker black lines and dots show the median across all screens. i<\/b>, The predicted secondary structure of a 66-nt insert sequence (ELMI003108) with the HEK3<\/i> homology arm.<\/p>\n<\/div>\n
Full size image<\/span><\/a><\/p>\n<\/figure>\n<\/div>\nSecond, to disentangle the contribution of the reverse transcription step, we made a prime editor construct with the nicking Cas9 fused to an engineered feline leukemia virus reverse transcriptase (MashUp RT: pipettejockey.com) with similar fidelity to the murine leukemia virus one used in prime editor 2. The average insertion rates observed using this construct were 6.7-fold lower compared with the standard PE2 (0.72% and 4.86%, respectively; Supplementary Fig. 5a\u2013d<\/a>), but highly correlated to PE2 (R<\/i>\u2009=\u20090.80; Fig. 3d<\/a>). Therefore, the effects of the insert sequence on insertion are not specific to the murine reverse transcriptase used in PE2 and highlight the possibility to perform prime editing experiments with alternative constructs.<\/p>\nThe PE3 system includes an additional guide RNA to nick the nonedited strand, which increases editing efficiency as well as indel formation rate12<\/a><\/sup>. We explored how the addition of this extra sgRNA affects the insertion frequencies of our library. We chose the EMX1<\/i> locus in HEK293T cells where we observed poor insertion efficiencies of 0.28% on average without the nicking guide RNA and cotransfected a nicking guide RNA that targets 77\u2009nt downstream of the pegRNA target38<\/a><\/sup>. We found that the extra nick increased the average insertion rate by 5.6-fold to 1.5% (Supplementary Fig. 5d\u2013g<\/a>), and increased the indel rate by 2.3-fold to 0.31%, including deletions between the nick sites of the pegRNA and sgRNA that were not observed for PE2 (Supplementary Fig. 5h<\/a>). Importantly, the relative insertion rates for sequences in the library were highly concordant between PE2 and PE3 in HEK293T cells (R<\/i>\u2009=\u20090.84; Fig. 4f<\/a>).<\/p>\nAn important step in prime editing is to resolve between the intermediates with a 5\u2032 flap (containing the wild-type sequence) or a 3\u2032 flap (containing the insertion) that compete. We speculated that the activity of the respective flap nucleases can steer the balance between the two outcomes. To test this, we overexpressed the 5\u2032 flap nuclease FEN1 and the 3\u2032 flap nucleases TREX1 and TREX2 in the context of the HEK3<\/i> site-targeting screen in HEK293T cells. As a control, we overexpressed eGFP in the same backbone used for the nucleases (Fig. 3f<\/a>). The insertion rates after FEN1 or eGFP overexpression were highly correlated to those measured in screens without overexpression (R<\/i>\u2009=\u20090.93 and 0.97; Fig. 3g<\/a>) with similar length dependence (Fig. 3h<\/a> and Supplementary Fig. 6a\u2013d<\/a>). Intriguingly, TREX1 and TREX2 overexpression abolished the insertion of longer sequences. For cells that did not overexpress nucleases or overexpressed eGFP, the average insertion rate for sequences longer than 4\u2009nt was 4.4\u20136.0% which is 4.4\u20135.8 times less than for shorter sequences. This is in contrast to cells overexpressing TREX1 and TREX2, where the average insertion rate for sequences >4\u2009nt was only 0.66% or 0.97%, 25.3\u201326.7-fold lower than that of shorter ones (Fig. 3h,i<\/a>).<\/p>\nWe confirmed that TREX1 and TREX2 antagonize prime insertions in a length-dependent manner. To do so, we cotransfected HEK293T cells with overexpression constructs encoding eGFP, TREX1 or TREX2 (Fig. 3f<\/a> and Supplementary Fig. 6e<\/a>) and individual pegRNAs targeting the HEK3<\/i> site encoding a 1-, 3-, 9- or 30-nt insertion (C, CAG, BCL6 binding site and Myc-tag) in the context of 25- or 34-nt homology arms (Fig. 3j<\/a>). Overexpressing TREX1 and TREX2 decreased editing rates across all insert and homology arm lengths, but disproportionately more for longer inserts (1.6\u20133.0-fold for the 1-nt insertion compared with 20\u2013108-fold for the 30-nt insertion; Fig. 3k<\/a>). This effect could be driven by the length of the insert sequence alone or of the entire 3\u2032 flap (corresponding to insertion\u2009+\u2009homology arm). In line with the results from our pooled screens (Fig. 3i<\/a>), we observed a strong correlation between the log fold change of insertion rates for TREX1\/2 over eGFP and the insert sequence length (R<\/i>\u2009=\u20090.97) which decreased when considering the total extension length (R<\/i>\u2009=\u20090.86\u20130.92; Supplementary Fig. 6f<\/a>), suggesting a more important role for the insertion length than the overall flap length.<\/p>\nThe HEK3<\/i> locus in HEK293T contains a single-nucleotide variation at position 9 after the prime editor nick site. The pegRNA homology arm encodes a G for this position, while one of the three chromosome copies encodes an A. If a 3\u2032 flap containing the edit and at least 9\u2009nt of the homology arm was fixed into the genome, we would expect a decreased frequency of the A allele. Indeed, for both pooled and validation screen conditions without TREX1\/2 overexpression, we only observed 0.95\u20131.6% (screen averages) of reads with library insertions containing A in the +9 position compared with 33\u201336% for unedited reads (Fig. 3k<\/a>). This is in contrast to screens overexpressing TREX1\/2 where the percentage of the A allele increased to 3.4\u20136.9%, suggesting a higher proportion of flaps where the homology arm was digested to below 9\u2009nt (Fig. 3k<\/a>). Taken together, our data demonstrate that TREX1\/2 antagonize the insertion of longer sequences with prime editing, presumably by digesting the 3\u2032 flap intermediate containing the edit.<\/p>\nSequence content effects on insertion efficiency<\/h3>\n
We next examined sequence content-dependent variation in insertion rate. To address this in a length-independent way, we calculated the insertion rate of each insert relative to sequences with the same or similar length (Methods<\/a>) and then measured its correlation with sequence features, computed from the perspective of the written sequence (that is, the reverse complement of the pegRNA molecule sequence). We observed a consistent cytosine preference across all four target sites and cell lines (Fig. 4a<\/a> and Supplementary Fig. 7a<\/a>), with each extra percentage of cytosine in the insert increasing the relative insertion rate by an average of 2.2%. Conversely, the percentages of adenine and thymine decreased insertion rates for all loci and cell lines (Fig. 4a<\/a> and Supplementary Fig. 7a<\/a>).<\/p>\nOur observations of nucleotide content effect were limited to four target sites, and moderately variable. To confirm whether the sequence influences hold more broadly, we performed an additional set of screens in HEK293T cells, targeting the original HEK3<\/i> site and five novel sites within 1\u2009kilobase (kb) of the HEK3<\/i> site (dubbed HEK3-S2<\/i> to HEK3-S6<\/i>) with pegRNA libraries encoding 356\u2013388 18-nt inserts on pegRNAs with 15-nt homology arms (average insertion rate 3.2%, median R<\/i> between replicates 0.81; Supplementary Fig. 7b<\/a>). Reassuringly, the sequence preferences were recapitulated in this experiment, with a strong preference for cytosines (average R<\/i> between insertion rate and cytosine fraction\u2009=\u20090.47; Fig. 4b<\/a> and Supplementary Fig. 7c<\/a>).<\/p>\nWe next sought to understand how pegRNA secondary structure affects insertion rates. As the strength of the structure depends on the length of the insert, we calculated the secondary structure\u2019s free energy relative to a large sample of sequences of the same length (Methods<\/a>). We observed that sequences with relatively stronger structures were more efficiently inserted (R<\/i>\u2009=\u20090.46; Fig. 4d<\/a>). To better understand this effect, we considered which combination of the pegRNA parts (primer binding site, insert and homology arm) gives predicted free energies that best reflect insertion efficacy. We observed the strongest correlation when the structure was calculated from the reverse transcribed portion of the extension (that is, the combination of insert sequence and homology arm; average R<\/i> across screens\u2009=\u20090.38), and the additional inclusion of the primer binding site sequence decreased correlation (Fig. 4e<\/a> and Supplementary Fig. 8a,b<\/a>). Further, the free energies of pegRNA extensions designed for one target site always predicted insertion efficiency better at the same site than other target sites (Supplementary Fig. 8c<\/a>). Since the homology arm is specific to the target, this also explains some of the differences in insertion rates we observed across the target sites.<\/p>\nStructure in the insert and homology arm could increase prime editing efficiency by protecting the pegRNA itself from nuclease degradation, a strategy explored in engineered pegRNAs (epegRNAs) which contain structured RNA elements to the 3\u2032 of the primer binding site21<\/a>,39<\/a>,40<\/a><\/sup>. However, we did not observe an increased abundance of more structured pegRNAs in the transcriptome (Supplementary Fig. 8d<\/a>), suggesting an alternative mechanism. To better understand the interplay of structure in various parts of the pegRNA and how it affects insertion rates, we screened 439 inserts of varying free energy from the original pegRNA library in the epegRNA construct, targeting the HEK3<\/i> site in HEK293T cells (Supplementary Fig. 8e\u2013g<\/a>). We found that the additional structure in the insert and homology arm also increased insertion rates for epegRNAs (R<\/i>\u2009=\u2009\u22120.34) but to a lesser extent than for regular pegRNAs (R<\/i>\u2009=\u2009\u22120.53; Supplementary Fig. 8h<\/a>), and that the insertion rates between regular and epegRNAs were highly correlated (R<\/i>\u2009=\u20090.79; Fig. 4f<\/a>). Together, this implies that structure past the protective cap still influences insertion rates via ways beyond transcript abundance, and that our results on insertion efficiencies are relevant for epegRNAs as well.<\/p>\nWe further noticed that structure in the reverse transcribed portion of the pegRNA was not correlated to the insertion rates of sequences <5\u2009nt, but was well correlated for longer sequences (Fig. 4g<\/a>). Since insertion rates of longer sequences are more impacted by overexpression of TREX1 and TREX2, we speculated that the structure protects the reverse transcribed 3\u2032 DNA flap containing the edit from degradation. Indeed, we observed that structure has a 2.4\u20132.6-fold stronger effect for cells overexpressing TREX1 or TREX2 compared with cells overexpressing FEN1, eGFP or nothing (Fig. 4g<\/a> and Supplementary Fig. 9a\u2013d<\/a>).<\/p>\nStructure plays a role in other parts of the pegRNA molecule as well. For instance, the 13\u2009nt of the primer binding site are perfectly complementary to the protospacer (positions 5\u201317) and can therefore hybridize with each other. If the first nucleotides of the insert create further base pairing with the protospacer and scaffold, the strength of this structure is enhanced, and the protospacer could be sequestered from base pairing with the target site or ribonucleoprotein complex formation with Cas9 could be impaired. To test if this additional pairing affects insertion rates, we predicted minimum free energy configurations of the primer binding site and the first three insert nucleotides with the spacer and the first guanine of the scaffold and observed 27% lower editing rates for inserts with extended base pairing 3\u2009nt into the protospacer compared with no extension (Supplementary Fig. 10a<\/a>). Finally, we tested if the disruption of the structural scaffold loops, which are required for association with Cas9, by the insert sequence reduces insertion rates. We calculated the minimum free energy configuration of the insert with the scaffold and observed 26% lower average editing for the pegRNAs with the first scaffold loop disrupted (screen range 10\u201343%) and 20% with the second and third loops (screen range 11\u201335%) compared with other inserts of the same length (Fig. 4h<\/a>). This loop dependence is in agreement with recent findings that scaffold variants with additional point mutations to stabilize the stem-loops can increase prime editing efficiencies41<\/a><\/sup>.<\/p>\nCombining effects of insert sequence length, cytosine content and structure explained why some sequences are inserted much better than others. For example, the long 66-nt ELMI003108 sequence that was inserted in the HEK3<\/i> locus at 1.39% insertion frequency (0.66% on average for the other 10 sequences >66\u2009nt) formed a strong structure together with the HEK3<\/i> homology arm (minimum free energy\u2009=\u2009\u221235.2\u2009kcal\u2009mol\u22121<\/sup>; 1.5\u2009s.d. lower than the average free energy of 66-nt sequences; Fig. 4i<\/a>). Other longer sequences that inserted frequently relative to their size were recombinase sites which are often near-palindromic and therefore form strong structures (Supplementary Fig. 10b,c<\/a>). Finally, our library included eight codon variations of the His-6 tag in forward and reverse orientations. The average insertion difference between the best codon variant and the worst was 13.3-fold, with the highest insertion rate for the cytosine-richest CAC histidine codons (Supplementary Fig. 10d<\/a>). This directly demonstrates the practical utility of this new understanding for guiding the codon choice for tags to insert (see the Supplementary Note<\/a> for a more thorough discussion).<\/p>\nPredicting insertion rates<\/h3>\n
Given our improved understanding of prime insertion rates, we next aimed to predict the relative efficiencies of inserting different sequences into the same site. We extracted 53 salient features such as insert length, nucleotide composition and folding energy for each pegRNA in eight screens (Fig. 5a<\/a>, Supplementary Table 1<\/a> and Supplementary Fig. 11<\/a>), and used tenfold cross-validation to select an accurate model (Methods<\/a> and Supplementary Fig. 12<\/a>). Based on feature correlations, their marginal effect we uncovered above and interpretability, we manually picked a final set of ten features, such that adding the remaining 43 extracted features did not improve the model performance further on the training data (Fig. 5b<\/a> and Supplementary Fig. 12<\/a>). The contribution of individual features to prediction reflected the understanding developed above: insert sequence length, the secondary structure of the pegRNA and reverse transcribed sequence, sequence composition and MMR each had a substantial impact, and the direction of these effects was consistent with expectations (Fig. 5c<\/a> and Supplementary Table 1<\/a>). The final model trained on the full training set achieved a correlation of 0.68 on held-out sequences, with performance ranging from R<\/i>\u2009=\u20090.44 to 0.92 when restricted to individual screens, exceeding correlation of individual biological replicates in noisier ones (Fig. 5d<\/a> and Supplementary Fig. 12<\/a>). We call this method MinsePIE (Modeling insertion efficiency for Prime Insertion Experiments) and incorporated it into a package available at https:\/\/github.com\/julianeweller\/MinsePIE<\/a>, and produced a web application to predict prime editing insertion rates at https:\/\/elixir.ut.ee\/minsepie\/<\/a>.<\/p>\n\nFig. 5: Predicting prime insertion efficiencies.<\/b><\/figcaption>
The efficient insertion of short DNA sequences into genomes could change the course of biotechnology and medicine1<\/a>,2<\/a><\/sup>. Small insertions can encode protein tags for purification and visualization, or manipulate protein function by altering protein localization, half-life or interaction profiles. Integrating sequences for transcription factor binding sites and splicing modulators provides control over gene expression while introducing structural elements or recombinase sites can change DNA conformation and provide a substrate for large-scale engineering. For therapeutic opportunities, over 16,000 small deletion variants have been causally linked to disease3<\/a>,4<\/a><\/sup>, and could in principle be restored by inserting the missing sequence5<\/a>,6<\/a><\/sup>. A prominent example is cystic fibrosis, where 70% of cases are caused by a three-nucleotide (nt) deletion7<\/a>,8<\/a><\/sup>. To enable reversing these mutations in practice, a technology must integrate insertions efficiently, accurately and safely, avoiding the unintended outcomes and double-strand break stress that hampers existing Cas9-based therapies9<\/a>,10<\/a>,11<\/a><\/sup>.<\/p>\n Prime editors can insert short DNA sequences without generating double-strand breaks or requiring an external template. They consist of a nicking version of Cas9 fused to a reverse transcriptase domain, which is complexed with a prime editing guide RNA (pegRNA)12<\/a><\/sup>. The pegRNA comprises a primer binding site homologous to the sequence in the target, and a reverse transcriptase template that includes the intended edit, all in the 3\u2032 extension of a standard CRISPR\u2013Cas9 guide RNA. At the target site, Cas9 nicks one strand of the genomic DNA, which then anneals to the primer binding site on the pegRNA, and is extended by the Cas9-fused reverse transcriptase using the pegRNA-encoded template sequence. Next, DNA repair mechanisms resolve the conflicting sequences on the two DNA strands, ultimately writing the intended edit into the genome. Where CRISPR\u2013Cas9 was compared with molecular scissors capable of disrupting target genes, and base editors were seen as molecular pencils for their ability to substitute single nucleotides, prime editors can be described as molecular word processors, able to perform search and replace operations directly on the genome13<\/a>,14<\/a>,15<\/a>,16<\/a><\/sup>.<\/p>\n The prime editing system is complex, and the determinants of its efficiency are not fully understood. Several partly independent steps, including three DNA binding events and successful DNA repair, are needed to produce an edit, each potentially influenced by the introduced sequence. In the largest study so far to understand these biases, Kim et al. comprehensively tested the consequences of varying the reverse transcription templates and primer binding site lengths using a library of 55,000 pegRNAs. The editing rate increased with Cas9 guide RNA activity, as well as GC content and melting temperature of the primer binding site. While further optimization of sequences was possible, primer binding sites of 11\u201313\u2009nt and reverse transcriptase templates of 10\u201312\u2009nt had the highest average editing efficiencies17<\/a><\/sup>.<\/p>\n The majority of libraries used by Kim et al. contained the same single-nucleotide substitution 5\u2009nt upstream of the nick site. Similarly, nearly all investigations of prime editing efficacy to date have predominantly focused on single-nucleotide substitutions12<\/a>,17<\/a>,18<\/a>,19<\/a>,20<\/a>,21<\/a><\/sup>. Of the many possible useful sequences in molecular biology, only a handful have been introduced with prime editing12<\/a><\/sup>. Therefore, in contrast to a relatively deep understanding of Cas9 mutagenesis10<\/a>,22<\/a>,23<\/a>,24<\/a><\/sup> and base editing outcomes25<\/a>,26<\/a>,27<\/a><\/sup>, very little is known about how the inserted sequence affects efficiency, and the length range of insertions feasible by prime editing has not been defined.<\/p>\n Here, we systematically measure the insertion efficiency of 3,604 sequences in several target sites and a variety of cellular and repair pathway contexts. We find that insertion sequence length, nucleotide composition and secondary structure all affect insertion efficiency. Moreover, we define the precise effect of mismatch repair (MMR) on thousands of insertion sequences and discover that overexpression of the 3\u2032 flap nucleases TREX1 and TREX2 abolished the insertion of longer sequences. Together, sequence features and repair pathway activity explain most of the variation in insertion rate. We then use these insights to train a sequence-based prediction model informed by MMR efficiency that predicts editing outcomes for novel sequences with high accuracy and demonstrate the model\u2019s usefulness for the selection of optimal reagents for new insertions.<\/p>\n<\/div>\n<\/div>\n We sought to systematically characterize how the length and composition of inserted sequence, as well as cell line, target site and the version of the prime editor system, affect insertion rates. To do so, we designed 3,604 pegRNAs encoding insertions immediately upstream of the nick site. These comprise 270 sequences useful for molecular biology (for example, His-6 tag, recombinase sites and mNeonGreen-11 (ref. 28<\/a><\/sup>)); 1,957 eukaryotic linear motifs29<\/a>,30<\/a>,31<\/a><\/sup>; 439 sequences with variable secondary structure; all single nucleotides, dinucleotides, trinucleotides and tetranucleotides; and 100 random sequences of each length between 5 and 10\u2009nt (Fig. 1a<\/a>). Insertions ranged from the length of 1 to 69\u2009nt, and varied in GC content (Fig. 1b<\/a>), while the primer binding site and homology arm lengths in the pegRNA were fixed to 13 and 34\u2009nt, respectively. We used lentiviruses to deliver the libraries against four target sites (three previously tested: HEK3<\/i>, EMX1<\/i>, FANCF<\/i>12<\/a><\/sup>, and the safe-harbor CLYBL<\/i> locus32<\/a><\/sup>) in two cell lines (HEK293T and HAP1), followed by transient transfection of the prime editor 2 plasmid (HEK293T cells) or doxycycline induction of PiggyBac transposase integrated prime editor (HAP1 cells), five\u2009d of selection and sequencing of two amplicons from the cell pool, one of the targeted locus and one of the pegRNA locus (Fig. 1c<\/a>). We calculated insertion efficiencies as the fraction of reads in the target site amplicon with a given insertion divided by the fraction of reads for the pegRNA encoding it in the pegRNA amplicon, and analyzed them as the main statistic in the rest of the study.<\/p>\n a<\/b>, Screen setup. Set 1 and Set 2 libraries were screened separately and data merged (Methods<\/a>); panels d<\/b>\u2013f<\/b> reflect Set 1 results only. b<\/b>, Library composition. The number of sequences in the library (y<\/i> axis) with different insert sequence lengths (x<\/i> axis, top panel) and %GC content (x<\/i> axis, bottom panel). c<\/b>, Experimental design. NGS, next generation sequencing. d<\/b>, Editing frequencies. Average mutation frequency (y<\/i> axis) for different screens (x<\/i> axis) stratified by mutation type (blue, insertions; gray, unintended outcomes). Markers represent one replicate and bars the average across n<\/i>\u2009=\u20093 biological replicates. e<\/b>, Replicate concordance. Pearson\u2019s R<\/i> between insertion rates in two screens (x<\/i> axis) for different comparisons (y<\/i> axis, colors). Markers, correlation value of one pair of screens (for replicate correlations, mean of pairwise comparison across n<\/i>\u2009=\u20093 biological replicates); line and whiskers, mean and s.e.m. f<\/b>, Representative examples of categories from e<\/b>. Percentage insertion in the HEK3<\/i> locus in HEK293T cells (y<\/i> axis) compared with values (x<\/i> axis) in other contexts (panels, colors) for insertion sequences (markers). Left panel, comparison of biological replicates; other panels, comparison of replicate averages. Label, R<\/i> of values in linear scale. Colors as in e<\/b>.<\/p>\n<\/div>\n Full size image<\/span><\/a><\/p>\n<\/figure>\n<\/div>\n Insertion efficiencies of sequences varied widely. The top 5% of templates were inserted 27\u2013134 times more efficiently than the bottom 5% across the various target site and cell line combinations (Supplementary Fig. 1a,b<\/a>), indicating substantial sequence-dependent variation. The insertion rates were highly consistent across biological replicates (median R<\/i>\u2009=\u20090.70; Supplementary Fig. 1c\u2013i<\/a>), but differed in magnitude across screens (average across pegRNAs, 0.18% for the CLYBL<\/i> locus in HEK293T to 6.7% for the HEK3<\/i> locus in HEK293T cells; Fig. 1d<\/a>). Unintended editing outcomes we observed included single-base mutations, small insertions and deletions around the nicking site, deletions overlapping primer binding site and reverse transcription template, insertion of mutated library sequences, duplications of the reverse transcription template, as well as partial scaffold integrations (Fig. 1d<\/a> and Supplementary Fig. 2a\u2013c<\/a>). These outcomes were rare overall (0.06\u20130.45%). Base changes at the target site were infrequent in reads with and without insertions (0.038% versus 0.030%), but slightly elevated upon insertion immediately downstream of the nick site and for the first nucleotides after the end of the homology arm (Supplementary Fig. 2d\u2013f<\/a>). Overall, the intended insertions were the dominant mutations generated, and we do not consider the unintended edits further.<\/p>\n To understand the consistency of insertion efficiencies across contexts, we next compared them between replicates, cell lines and target sites. Insertion rates into the same target site in different cell lines were more correlated (mean R<\/i>\u2009=\u20090.52) than into different target sites in the same line (mean R<\/i>\u2009=\u20090.38). The correlation was weakest when both the target site and cell line were different (mean R<\/i>\u2009=\u20090.17; Fig. 1e,f<\/a>), demonstrating both target sequence-specific and cell line-dependent biases on insertion.<\/p>\n Given the repeatable sequence-dependent variation in insertion rates that spans over three orders of magnitude, we sought to understand the responsible features, starting with insert length. Insertion frequency did not decrease monotonically with insert length in HEK293T cells, but instead, had two modes of high values. First, sequences of 3 and 4\u2009nt were inserted on average 2.0\u20134.1 times more efficiently than others across the four targeted sites (Fig. 2a<\/a>). Second, sequences between 15 and 21\u2009nt were inserted on average 1.3\u20131.6 times more efficiently than 10\u201314-nt ones, and 1.5\u20132.0 times more efficiently than sequences longer than 21\u2009nt (Fig. 2a<\/a>). These relative biases in efficiency were shared between all target sites, despite a 20-fold range of their average insertion rates. Inserts longer than 45\u2009nt were incorporated less frequently, at a screen average rate that is 4\u20138 times lower than that of sequences shorter than 45\u2009nt. The longest sequence that was inserted at >1% frequency (1.4%, HEK3<\/i> site in HEK293T cells) was 66\u2009nt, demonstrating that integration of moderately long sequences is feasible with prime editing.<\/p>\n a<\/b>, Insertion rate in HEK293T cells. Percentage of reads with insertion (y<\/i> axis, cut-off at 3\u2009s.d. above mean) for different insert sizes (x<\/i> axis) of individual sequences (blue markers) and averages for lengths with at least 30 measured sequences (dark blue line and markers) at different target sites (panels). Data represent the average of n<\/i>\u2009=\u20093 biological replicates. b<\/b>, As a<\/b>, but for HAP1 cells. c<\/b>, As a<\/b>, but for HAP1 \u2206MLH1<\/i> cells. d<\/b>, Insertion rate in one cell context (y<\/i> axis) compared with in another context (x<\/i> axis) at the HEK3<\/i> target of individual sequences (markers), comparing HEK293T with HAP1 cells (left panel) and HEK293T cells with HAP1 \u2206MLH1<\/i> cells (middle panel). Red, short sequences (up to 4\u2009nt); blue, medium sequences (5\u201313\u2009nt); teal, longer sequences (>13\u2009nt). Label, R<\/i> between rates. The data are an average from n<\/i>\u2009=\u20093 biological replicates (HEK293T) or n<\/i>\u2009=\u20092 biological replicates (HAP1). e<\/b>, Average insertion rates (y<\/i> axis) across insert lengths (x<\/i> axis) with at least 30 measured sequences in various cell line contexts (colors). Data are presented as mean\u2009\u00b1\u2009s.e.m. n<\/i>\u2009=\u20093 biological replicates (HEK293T) or n<\/i>\u2009=\u20092 biological replicates (HAP1). f<\/b>, The ratio of relative insertion rates (Methods<\/a>) at the HEK3<\/i> locus between HAP1 \u2206MLH1<\/i> and HAP1 cells (y<\/i> axis) for different lengths (x<\/i> axis) stratified by colors as in d<\/b>. Box, median and quartiles; whiskers, least extreme of 1.5 times the interquartile range from the quartile and most extreme values. Line, fit from an exponential model (ratio\u2009\u2248\u2009a<\/i>\u2009\u00d7 exp(\u2212b<\/i>\u2009\u00d7\u2009length)\u2009+\u20091). n<\/i>\u2009=\u20092 biological replicates.<\/p>\n<\/div>\n Full size image<\/span><\/a><\/p>\n<\/figure>\n<\/div>\n In contrast to HEK293T cells, the insertion frequency of the short 1\u20134-nt sequences was not substantially higher than that of longer ones in HAP1 cells (0.60\u20131.27 times; Fig. 2b<\/a>). This reduced the concordance of insertion rates in the two cell lines at the same site (R<\/i>\u2009=\u20090.41 for FANCF<\/i> and 0.54 for HEK3<\/i>; Fig. 2d<\/a> and Supplementary Fig. 3a<\/a>) compared with replicates (median R<\/i>\u2009=\u20090.78; Fig. 1e<\/a>). One possible explanation is MMR proficiency, since HEK293T cells are partly MMR deficient due to promoter methylation of MLH1<\/i> (ref. 33<\/a><\/sup>), while HAP1 cells are not. The MMR pathway recognizes and excises short mismatches of less than 13\u2009nt and could therefore remove short insertions in HAP1 cells before the nicked strand is re-ligated34<\/a><\/sup>. Indeed, MMR antagonizes prime editing for substitutions and short insertions20<\/a>,35<\/a><\/sup>. Consistent with this explanation, we observed strong correlations between insertion rates in HAP1 and HEK293 cells for sequences longer than 13\u2009nt that are not affected by MMR (R<\/i>\u2009=\u20090.78 for the FANCF<\/i> locus and 0.91 for the HEK3<\/i> locus; Fig. 2c<\/a> and Supplementary Fig. 3a<\/a>).<\/p>\n To experimentally test the hypothesis that rates of insertion of short sequences differ between cell lines due to MMR activity, we screened the HEK3<\/i>– and FANCF<\/i>-targeted libraries in HAP1 cells that are knocked out for MLH1<\/i> (HAP1 \u2206MLH1<\/i>; Fig. 2d<\/a> and Supplementary Fig. 3b,c<\/a>). We found that the average insertion rates of 1\u20134-nt sequences were most affected by the knockout, increasing by 7.2\u201311-fold, while the rates of 5\u201313-nt sequences increased 2.1\u20132.7-fold (Fig. 2e<\/a> and Supplementary Fig. 3d<\/a>). Overall, 66% (HEK3<\/i>) and 67% (FANCF<\/i>) of the variance in the fold changes (Fig. 2f<\/a> and Supplementary Fig. 3d<\/a>) was explained by a model where the loss of MMR increases the insertion rate of 1-nt sequences by 23\u201328-fold, with the increase in insertion efficiency dropping 40\u201348% for every additional nucleotide. The low correlations of insertion rates between HEK293T and wild-type HAP1 cells (R<\/i>\u2009=\u20090.41\u20130.54) also improved to close to replicate concordance when matching MMR status (R<\/i>\u2009=\u20090.73\u20130.96 between HEK293T and HAP1 \u2206MLH1<\/i> cell lines; Fig. 2c<\/a> and Supplementary Fig. 3a,e<\/a>). In summary, our findings highlight that MMR proficiency is the major source of independent variation between the tested cellular contexts for prime insertion of short sequences.<\/p>\n Having confirmed MMR as a length-dependent determinant of insertion efficiency, we next sought to understand how different steps of prime editing affect insertion rates of our library sequences. Specifically, we dissected the contributions of (1) pegRNA expression, (2) reverse transcription by two different reverse transcriptases, (3) presence of a nicking guide and (4) overexpression of 3\u2032 and 5\u2032 flap nucleases (Fig. 3a<\/a>).<\/p>\n a<\/b>, Schematic of molecular steps involved in prime editing. b<\/b>, Normalized pegRNA count derived from sequencing of PCR amplicons from genomic DNA (x<\/i> axis) or PCR amplicons from RNA (y<\/i> axis) for the HEK3<\/i> site in HEK293T cells for individual pegRNAs (markers). Pink, inserts with four or more consecutive adenines. Data represent the average of n<\/i>\u2009=\u20093 biological replicates. c<\/b>, Top panel, average insertion rate relative to length bin median (y<\/i> axis) for inserts stratified by the longest consecutive run of adenines (x<\/i> axis). Bottom panel, instead showing transcription rate (read counts from RNA\/read counts from DNA) on the y<\/i> axis. Data are presented as mean\u2009\u00b1\u2009s.e.m. n<\/i>\u2009=\u20093 biological replicates. d<\/b>, Insertion frequencies at the HEK3<\/i> site in HEK293T using the standard MMLV reverse transcriptase (PE2, x<\/i> axis) and the FeLV reverse transcriptase (PE-FeLV, y<\/i> axis) for different insertion sequences (markers). Colors, number of neighboring points. n<\/i>\u2009=\u20093 biological replicates. e<\/b>, As d<\/b>, but comparing PE3 and PE2 at the EMX1<\/i> site. f<\/b>, Schematic of screens with overexpression constructs. g<\/b>, Insertion frequencies for different overexpressions (y<\/i> axis and panels) compared with no overexpression (x<\/i> axis) for three biological replicate screens (markers) stratified by insertion sequence lengths (colors). h<\/b>, Average insertion rates (y<\/i> axis) across insert lengths (x<\/i> axis) with at least 30 measured sequences for overexpression constructs (colors). Data are presented as mean\u2009\u00b1\u2009s.e.m. n<\/i>\u2009=\u20093 biological replicates. i<\/b>, As h<\/b>, but instead displaying the insertion rate fold changes of screens with overexpressions compared with no overexpression (y<\/i> axis), calculated from the ratio of sums of all sequences (lines) or of ten randomly sampled sequences. j<\/b>, Top, average insertion frequency (grayscale) of four sequences with varying lengths (x<\/i> axis) when overexpressing eGFP stratified by homology arm lengths (panels). Bottom, insertion rate fold changes compared with eGFP (y<\/i> axis) when overexpressing TREX1 and TREX2 (colors). n<\/i>\u2009=\u20092 biological replicates. k<\/b>, Fraction of the nontemplated adenine allele at the +9 position (y<\/i> axis) for cells with overexpression constructs (x<\/i> axis and colors) stratified by experiment and homology arm lengths (panels). Markers show screen averages from three biological replicates for the pooled screen or from separate pegRNAs for the individual validation experiment.<\/p>\n<\/div>\n Full size image<\/span><\/a><\/p>\n<\/figure>\n<\/div>\n We first assessed expression levels of pegRNAs targeting the HEK3<\/i> site in HEK293T cells using deep sequencing. Abundance in the transcriptome was well correlated between replicates (median R<\/i>\u2009=\u20090.97; Supplementary Fig. 4a<\/a>) and with the DNA-derived read count frequency (R<\/i>\u2009=\u20090.56; Fig. 4b<\/a>). The exceptions were sequences that resulted in four or more consecutive thymines on the pegRNA cassette (adenines in the inserted DNA), which act as transcription terminators for RNA polymerase III (refs. 36<\/a>,37<\/a><\/sup>). Upon removing pegRNAs with terminator motifs, the correlation between measured DNA and RNA sequence coverage increased to 0.59 (Fig. 3b<\/a>). Sequences with four or more consecutive adenines were 4.8-fold less expressed and, accordingly, their average insertion rate was 4.8-fold lower compared with other sequences (Fig. 3c<\/a> and Supplementary Fig. 4b<\/a>). Overall, 23 of the 24 inserts (96%) that were not observed in any screen contained at least one run of four or more adenines, highlighting this feature as a useful filter in pegRNA design.<\/p>\n a<\/b>, Correlation of length-normalized insertion rate with nucleotide frequency in the insert (colors) for each nucleotide (y<\/i> axis) in each screen (x<\/i> axis). Data represent the average of n<\/i>\u2009=\u20093 (HEK293T) or n<\/i>\u2009=\u20092 (HAP1) biological replicates. b<\/b>, As a<\/b>, but for a new set of screens with 18-nt inserts and 15-nt homology arms targeting five novel sites within 1\u2009kb of the HEK3<\/i> site. c<\/b>, Insertion rate at the HEK3<\/i> site in HEK293T cells relative to length bin median (y<\/i> axis) for inserts (markers) with different cytosine content (x<\/i> axis). Line, linear regression fit; shaded area, 95% posterior confidence interval of the fit. Data represent the average of n<\/i>\u2009=\u20093 biological replicates. d<\/b>, Insertion rates at the HEK3<\/i> site in HEK293T cells relative to length bin median (y<\/i> axis) for inserts (markers) with calculated Gibbs free energy (\u2206G<\/i>) from ViennaFold (x<\/i> axis). Line, linear regression fit; shaded area, 95% posterior confidence interval of the fit. Data represent the average of n<\/i>\u2009=\u20093 biological replicates. e<\/b>, Correlation (x<\/i> axis) between insertion rates and insert sequence free energy calculated from different parts of the 3\u2032 extension (y<\/i> axis). Box, median and quartiles; whiskers, least extreme of 1.5 times the interquartile range from the quartile and most extreme values. n<\/i>\u2009=\u20093 (HEK293T) or n<\/i>\u2009=\u20092 (HAP1) biological replicates. f<\/b>, Insertion rates for sequences (markers) at the HEK3<\/i> site in HEK293T for pegRNAs (x<\/i> axis) and epegRNAs (y<\/i> axis). Data represent the average of n<\/i>\u2009=\u20093 biological replicates. g<\/b>, Percentage increase in insertion rate with each standard deviation increase in structure strength (colors) for different overexpression constructs (x<\/i> axis) and insertion sequence lengths (y<\/i> axis). h<\/b>, Insertion rates relative to length bin median (y<\/i> axis) for sequences that disrupt or preserve (x<\/i> axis) scaffold loops (panels). Colored lines show screen medians and the thicker black lines and dots show the median across all screens. i<\/b>, The predicted secondary structure of a 66-nt insert sequence (ELMI003108) with the HEK3<\/i> homology arm.<\/p>\n<\/div>\n Full size image<\/span><\/a><\/p>\n<\/figure>\n<\/div>\n Second, to disentangle the contribution of the reverse transcription step, we made a prime editor construct with the nicking Cas9 fused to an engineered feline leukemia virus reverse transcriptase (MashUp RT: pipettejockey.com) with similar fidelity to the murine leukemia virus one used in prime editor 2. The average insertion rates observed using this construct were 6.7-fold lower compared with the standard PE2 (0.72% and 4.86%, respectively; Supplementary Fig. 5a\u2013d<\/a>), but highly correlated to PE2 (R<\/i>\u2009=\u20090.80; Fig. 3d<\/a>). Therefore, the effects of the insert sequence on insertion are not specific to the murine reverse transcriptase used in PE2 and highlight the possibility to perform prime editing experiments with alternative constructs.<\/p>\n The PE3 system includes an additional guide RNA to nick the nonedited strand, which increases editing efficiency as well as indel formation rate12<\/a><\/sup>. We explored how the addition of this extra sgRNA affects the insertion frequencies of our library. We chose the EMX1<\/i> locus in HEK293T cells where we observed poor insertion efficiencies of 0.28% on average without the nicking guide RNA and cotransfected a nicking guide RNA that targets 77\u2009nt downstream of the pegRNA target38<\/a><\/sup>. We found that the extra nick increased the average insertion rate by 5.6-fold to 1.5% (Supplementary Fig. 5d\u2013g<\/a>), and increased the indel rate by 2.3-fold to 0.31%, including deletions between the nick sites of the pegRNA and sgRNA that were not observed for PE2 (Supplementary Fig. 5h<\/a>). Importantly, the relative insertion rates for sequences in the library were highly concordant between PE2 and PE3 in HEK293T cells (R<\/i>\u2009=\u20090.84; Fig. 4f<\/a>).<\/p>\n An important step in prime editing is to resolve between the intermediates with a 5\u2032 flap (containing the wild-type sequence) or a 3\u2032 flap (containing the insertion) that compete. We speculated that the activity of the respective flap nucleases can steer the balance between the two outcomes. To test this, we overexpressed the 5\u2032 flap nuclease FEN1 and the 3\u2032 flap nucleases TREX1 and TREX2 in the context of the HEK3<\/i> site-targeting screen in HEK293T cells. As a control, we overexpressed eGFP in the same backbone used for the nucleases (Fig. 3f<\/a>). The insertion rates after FEN1 or eGFP overexpression were highly correlated to those measured in screens without overexpression (R<\/i>\u2009=\u20090.93 and 0.97; Fig. 3g<\/a>) with similar length dependence (Fig. 3h<\/a> and Supplementary Fig. 6a\u2013d<\/a>). Intriguingly, TREX1 and TREX2 overexpression abolished the insertion of longer sequences. For cells that did not overexpress nucleases or overexpressed eGFP, the average insertion rate for sequences longer than 4\u2009nt was 4.4\u20136.0% which is 4.4\u20135.8 times less than for shorter sequences. This is in contrast to cells overexpressing TREX1 and TREX2, where the average insertion rate for sequences >4\u2009nt was only 0.66% or 0.97%, 25.3\u201326.7-fold lower than that of shorter ones (Fig. 3h,i<\/a>).<\/p>\n We confirmed that TREX1 and TREX2 antagonize prime insertions in a length-dependent manner. To do so, we cotransfected HEK293T cells with overexpression constructs encoding eGFP, TREX1 or TREX2 (Fig. 3f<\/a> and Supplementary Fig. 6e<\/a>) and individual pegRNAs targeting the HEK3<\/i> site encoding a 1-, 3-, 9- or 30-nt insertion (C, CAG, BCL6 binding site and Myc-tag) in the context of 25- or 34-nt homology arms (Fig. 3j<\/a>). Overexpressing TREX1 and TREX2 decreased editing rates across all insert and homology arm lengths, but disproportionately more for longer inserts (1.6\u20133.0-fold for the 1-nt insertion compared with 20\u2013108-fold for the 30-nt insertion; Fig. 3k<\/a>). This effect could be driven by the length of the insert sequence alone or of the entire 3\u2032 flap (corresponding to insertion\u2009+\u2009homology arm). In line with the results from our pooled screens (Fig. 3i<\/a>), we observed a strong correlation between the log fold change of insertion rates for TREX1\/2 over eGFP and the insert sequence length (R<\/i>\u2009=\u20090.97) which decreased when considering the total extension length (R<\/i>\u2009=\u20090.86\u20130.92; Supplementary Fig. 6f<\/a>), suggesting a more important role for the insertion length than the overall flap length.<\/p>\n The HEK3<\/i> locus in HEK293T contains a single-nucleotide variation at position 9 after the prime editor nick site. The pegRNA homology arm encodes a G for this position, while one of the three chromosome copies encodes an A. If a 3\u2032 flap containing the edit and at least 9\u2009nt of the homology arm was fixed into the genome, we would expect a decreased frequency of the A allele. Indeed, for both pooled and validation screen conditions without TREX1\/2 overexpression, we only observed 0.95\u20131.6% (screen averages) of reads with library insertions containing A in the +9 position compared with 33\u201336% for unedited reads (Fig. 3k<\/a>). This is in contrast to screens overexpressing TREX1\/2 where the percentage of the A allele increased to 3.4\u20136.9%, suggesting a higher proportion of flaps where the homology arm was digested to below 9\u2009nt (Fig. 3k<\/a>). Taken together, our data demonstrate that TREX1\/2 antagonize the insertion of longer sequences with prime editing, presumably by digesting the 3\u2032 flap intermediate containing the edit.<\/p>\n We next examined sequence content-dependent variation in insertion rate. To address this in a length-independent way, we calculated the insertion rate of each insert relative to sequences with the same or similar length (Methods<\/a>) and then measured its correlation with sequence features, computed from the perspective of the written sequence (that is, the reverse complement of the pegRNA molecule sequence). We observed a consistent cytosine preference across all four target sites and cell lines (Fig. 4a<\/a> and Supplementary Fig. 7a<\/a>), with each extra percentage of cytosine in the insert increasing the relative insertion rate by an average of 2.2%. Conversely, the percentages of adenine and thymine decreased insertion rates for all loci and cell lines (Fig. 4a<\/a> and Supplementary Fig. 7a<\/a>).<\/p>\n Our observations of nucleotide content effect were limited to four target sites, and moderately variable. To confirm whether the sequence influences hold more broadly, we performed an additional set of screens in HEK293T cells, targeting the original HEK3<\/i> site and five novel sites within 1\u2009kilobase (kb) of the HEK3<\/i> site (dubbed HEK3-S2<\/i> to HEK3-S6<\/i>) with pegRNA libraries encoding 356\u2013388 18-nt inserts on pegRNAs with 15-nt homology arms (average insertion rate 3.2%, median R<\/i> between replicates 0.81; Supplementary Fig. 7b<\/a>). Reassuringly, the sequence preferences were recapitulated in this experiment, with a strong preference for cytosines (average R<\/i> between insertion rate and cytosine fraction\u2009=\u20090.47; Fig. 4b<\/a> and Supplementary Fig. 7c<\/a>).<\/p>\n We next sought to understand how pegRNA secondary structure affects insertion rates. As the strength of the structure depends on the length of the insert, we calculated the secondary structure\u2019s free energy relative to a large sample of sequences of the same length (Methods<\/a>). We observed that sequences with relatively stronger structures were more efficiently inserted (R<\/i>\u2009=\u20090.46; Fig. 4d<\/a>). To better understand this effect, we considered which combination of the pegRNA parts (primer binding site, insert and homology arm) gives predicted free energies that best reflect insertion efficacy. We observed the strongest correlation when the structure was calculated from the reverse transcribed portion of the extension (that is, the combination of insert sequence and homology arm; average R<\/i> across screens\u2009=\u20090.38), and the additional inclusion of the primer binding site sequence decreased correlation (Fig. 4e<\/a> and Supplementary Fig. 8a,b<\/a>). Further, the free energies of pegRNA extensions designed for one target site always predicted insertion efficiency better at the same site than other target sites (Supplementary Fig. 8c<\/a>). Since the homology arm is specific to the target, this also explains some of the differences in insertion rates we observed across the target sites.<\/p>\n Structure in the insert and homology arm could increase prime editing efficiency by protecting the pegRNA itself from nuclease degradation, a strategy explored in engineered pegRNAs (epegRNAs) which contain structured RNA elements to the 3\u2032 of the primer binding site21<\/a>,39<\/a>,40<\/a><\/sup>. However, we did not observe an increased abundance of more structured pegRNAs in the transcriptome (Supplementary Fig. 8d<\/a>), suggesting an alternative mechanism. To better understand the interplay of structure in various parts of the pegRNA and how it affects insertion rates, we screened 439 inserts of varying free energy from the original pegRNA library in the epegRNA construct, targeting the HEK3<\/i> site in HEK293T cells (Supplementary Fig. 8e\u2013g<\/a>). We found that the additional structure in the insert and homology arm also increased insertion rates for epegRNAs (R<\/i>\u2009=\u2009\u22120.34) but to a lesser extent than for regular pegRNAs (R<\/i>\u2009=\u2009\u22120.53; Supplementary Fig. 8h<\/a>), and that the insertion rates between regular and epegRNAs were highly correlated (R<\/i>\u2009=\u20090.79; Fig. 4f<\/a>). Together, this implies that structure past the protective cap still influences insertion rates via ways beyond transcript abundance, and that our results on insertion efficiencies are relevant for epegRNAs as well.<\/p>\n We further noticed that structure in the reverse transcribed portion of the pegRNA was not correlated to the insertion rates of sequences <5\u2009nt, but was well correlated for longer sequences (Fig. 4g<\/a>). Since insertion rates of longer sequences are more impacted by overexpression of TREX1 and TREX2, we speculated that the structure protects the reverse transcribed 3\u2032 DNA flap containing the edit from degradation. Indeed, we observed that structure has a 2.4\u20132.6-fold stronger effect for cells overexpressing TREX1 or TREX2 compared with cells overexpressing FEN1, eGFP or nothing (Fig. 4g<\/a> and Supplementary Fig. 9a\u2013d<\/a>).<\/p>\n Structure plays a role in other parts of the pegRNA molecule as well. For instance, the 13\u2009nt of the primer binding site are perfectly complementary to the protospacer (positions 5\u201317) and can therefore hybridize with each other. If the first nucleotides of the insert create further base pairing with the protospacer and scaffold, the strength of this structure is enhanced, and the protospacer could be sequestered from base pairing with the target site or ribonucleoprotein complex formation with Cas9 could be impaired. To test if this additional pairing affects insertion rates, we predicted minimum free energy configurations of the primer binding site and the first three insert nucleotides with the spacer and the first guanine of the scaffold and observed 27% lower editing rates for inserts with extended base pairing 3\u2009nt into the protospacer compared with no extension (Supplementary Fig. 10a<\/a>). Finally, we tested if the disruption of the structural scaffold loops, which are required for association with Cas9, by the insert sequence reduces insertion rates. We calculated the minimum free energy configuration of the insert with the scaffold and observed 26% lower average editing for the pegRNAs with the first scaffold loop disrupted (screen range 10\u201343%) and 20% with the second and third loops (screen range 11\u201335%) compared with other inserts of the same length (Fig. 4h<\/a>). This loop dependence is in agreement with recent findings that scaffold variants with additional point mutations to stabilize the stem-loops can increase prime editing efficiencies41<\/a><\/sup>.<\/p>\n Combining effects of insert sequence length, cytosine content and structure explained why some sequences are inserted much better than others. For example, the long 66-nt ELMI003108 sequence that was inserted in the HEK3<\/i> locus at 1.39% insertion frequency (0.66% on average for the other 10 sequences >66\u2009nt) formed a strong structure together with the HEK3<\/i> homology arm (minimum free energy\u2009=\u2009\u221235.2\u2009kcal\u2009mol\u22121<\/sup>; 1.5\u2009s.d. lower than the average free energy of 66-nt sequences; Fig. 4i<\/a>). Other longer sequences that inserted frequently relative to their size were recombinase sites which are often near-palindromic and therefore form strong structures (Supplementary Fig. 10b,c<\/a>). Finally, our library included eight codon variations of the His-6 tag in forward and reverse orientations. The average insertion difference between the best codon variant and the worst was 13.3-fold, with the highest insertion rate for the cytosine-richest CAC histidine codons (Supplementary Fig. 10d<\/a>). This directly demonstrates the practical utility of this new understanding for guiding the codon choice for tags to insert (see the Supplementary Note<\/a> for a more thorough discussion).<\/p>\n Given our improved understanding of prime insertion rates, we next aimed to predict the relative efficiencies of inserting different sequences into the same site. We extracted 53 salient features such as insert length, nucleotide composition and folding energy for each pegRNA in eight screens (Fig. 5a<\/a>, Supplementary Table 1<\/a> and Supplementary Fig. 11<\/a>), and used tenfold cross-validation to select an accurate model (Methods<\/a> and Supplementary Fig. 12<\/a>). Based on feature correlations, their marginal effect we uncovered above and interpretability, we manually picked a final set of ten features, such that adding the remaining 43 extracted features did not improve the model performance further on the training data (Fig. 5b<\/a> and Supplementary Fig. 12<\/a>). The contribution of individual features to prediction reflected the understanding developed above: insert sequence length, the secondary structure of the pegRNA and reverse transcribed sequence, sequence composition and MMR each had a substantial impact, and the direction of these effects was consistent with expectations (Fig. 5c<\/a> and Supplementary Table 1<\/a>). The final model trained on the full training set achieved a correlation of 0.68 on held-out sequences, with performance ranging from R<\/i>\u2009=\u20090.44 to 0.92 when restricted to individual screens, exceeding correlation of individual biological replicates in noisier ones (Fig. 5d<\/a> and Supplementary Fig. 12<\/a>). We call this method MinsePIE (Modeling insertion efficiency for Prime Insertion Experiments) and incorporated it into a package available at https:\/\/github.com\/julianeweller\/MinsePIE<\/a>, and produced a web application to predict prime editing insertion rates at https:\/\/elixir.ut.ee\/minsepie\/<\/a>.<\/p>\nResults<\/h2>\n
<\/picture><\/a><\/div>\nInsert size and MMR activity effects<\/h3>\n
<\/picture><\/a><\/div>\nEffects of prime editing steps<\/h3>\n
<\/picture><\/a><\/div>\n<\/picture><\/a><\/div>\nSequence content effects on insertion efficiency<\/h3>\n
Predicting insertion rates<\/h3>\n