Off target V(D)J recombination drives lymphomagenesis and is escalated by loss of the Rag2 C-terminus

SUMMARY

Genome wide analysis of thymic lymphomas from Tp53^−/− mice with wild-type or C-terminally truncated Rag2 revealed numerous off target, RAG-mediated DNA rearrangements. A significantly higher fraction of these errors mutated known and suspected oncogenes/tumor suppressor genes than did sporadic rearrangements (p<0.0001). This tractable mouse model recapitulates recent findings in human pre-B ALL and allows comparison of wild-type and mutant RAG2. Recurrent, RAG-mediated deletions affected Notch1, Pten, Ikzf1, Jak1, Phlda1, Trat1, and Agpat9. Rag2 truncation substantially increased the frequency of off target V(D)J recombination. The data suggest that interactions between Rag2 and a specific chromatin modification, H3K4me3, support V(D)J recombination fidelity. Oncogenic effects of off target rearrangements created by this highly regulated recombinase may need to be considered in design of site-specific nucleases engineered for genome modification.

INTRODUCTION

Antigen receptor genes are assembled by controlled, large-scale deletions and inversions (Little et al., 2014). Errors in this process, known as V(D)J recombination, produce aberrant genomic rearrangements, which can fuel the development of T and B cell malignancies (Onozawa and Aplan, 2012). Normal V(D)J recombination entails double-strand DNA cleavage at pairs of gene segments by the heteromeric Rag1 and Rag2 endonuclease (hereafter referred to as RAG). RAG recognizes conserved Recombination Signal Sequences (RSSs) positioned adjacent to coding segments that are subsequently joined to form immunoglobulin (Ig) and T cell receptor (TCR) variable region exons (Fig. 1A). Enzymatic steps in joining are carried out by general DNA repair functions.

Fig. 1. Normal and Aberrant V(D)J Recombination.

A. Normal V(D)J recombination. RAG binds Recombination Signal Sequence (RSS) motifs, mediates synapsis and cleaves the DNA target. Double strand breaks occur 5' of the RSS. Two coding ends (CE) lose and gain nucleotides in interim processing events and then join as a coding joint (CJ). Signal ends are ligated into a signal joint (SJ) with no or minimal interim modification. Products may contain junctional inserts (not diagrammed). GC-rich, “N regions", are nucleotide additions that may occur in either joint. Palindromic “P nucleotides” are found in CJs only.

B–D. V(D)J recombination errors. B: A cryptic RSS-like element (cRSS; orange symbols) may join to an RSS or to a second cRSS. Off target cleavage gives rise to otherwise normal recombination products, yielding a cCJ and cSJ. C: Alternatively, errors can occur at the joining stage. Coding and signal ends may be swapped in error to form ‘hybrid joints’ (HJ). Alternatively, RAG-cleaved ends can be erroneously connected to a third incidental break (end-donation; gapped red line). D: Joining in trans (between two chromosomes) is a third distinguishable type of aberration and can occur along with recognition or joining errors. “Trans cut?” and “trans join?” reflect that trans events could occur at either stage. A reciprocal ‘type I’ translocation is diagrammed (reviewed in Marculescu et al, 2006).

Mechanisms underlying aberrant V(D)J recombination comprise three general categories according to outcome and/or genetic dependencies. One is off target cleavage (Fig. 1B) at sequences fortuitously resembling an RSS (i.e. a cryptic RSS or “cRSS”). A second type of mistake may occur at the point at which DNA ends are joined (Fig. 1C). A third involves interchromosomal (trans) recombination (Fig. 1D). Such errors, which can be compounded, leave distinctive traces in the structure of the junctions they produce in the genomes of lymphoid neoplasms (Marculescu et al., 2006; Onozawa and Aplan, 2012). A stunning example of V(D)J recombination gone wrong was provided by recent genomewide sequence analysis of pediatric acute lymphocytic leukemia (ALL), which revealed tumor genomes peppered with structural rearrangements caused by off target V(D)J recombination (Papaemmanuil et al., 2014).

Studies in mice and in cultured cells indicate that multiple strategies direct RAG to appropriate loci in appropriate cells, but the degree to which these also serve to prevent aberrant V(D)J recombination genome wide has not been determined. Possible failsafes include regulation of Rag1 and Rag2 expression, attenuation of RAG activity during S-phase, chromatin features that define possible recombination sites (such as transcription-associated accessibility, locus contraction, histone H3 methylation, etc.), a vigilant DNA damage response, and the RSS specificity of RAG endonuclease itself (Little et al., 2014). Two features in the Rag2 C-terminal domain have been explored at the molecular level. One is a specific phosphorylation site, targeted by cyclinA/CDK2, that confines RAG activity to the G0/G1 phase of the cell cycle. Another, also located in the C-terminus, is a plant homeodomain-like domain (PHD) that specifically interacts with H3K4me3 inducing a structural change in RAG that relieves a negative autoregulation of its activity (Lu et al., 2015).

Engineered mouse models bearing mutant Rag2 alleles implicate the Rag2 C-terminal domain in curtailing the oncogenic potential of V(D)J recombination. A T490A mutation at the phosphorylation site as well as either partial (Rag2^FS361) or complete (Rag2³⁵²) deletion of the Rag2 C-terminus (to amino acid 361 and 352 respectively), contributes to lymphomagenesis when assayed in a p53-deficient background (Deriano et al., 2011; Gigi et al., 2014; Zhang et al., 2011). Rag2^T490A Rag2^FS361 and Rag2³⁵², Tp53^−/− lymphomas exhibit chromosomal translocations cytogenetically mapping near Ig and TCR antigen receptor loci (Deriano et al., 2011; Gigi et al., 2014; Zhang et al., 2011). Such translocations are rarely, if ever, seen in lymphomas from RAG2^wt/wt Tp53^−/− mice [(Deriano et al., 2011; Zhang et al., 2011) and cited therein].

To investigate the role of Rag2’s C-terminus in suppressing oncogenic V(D)J recombination events, we analyzed structural genome variations in lymphomas from either Rag2^wt/wt, Tp53^−/− or Rag2^352/352, Tp53^−/− mice. The Rag2³⁵² mutation causes a significantly more rapid onset of lymphomagenesis than any of the other above-mentioned Rag2 mutations in the Tp53^−/− background. This experimental model has several advantages over genome wide characterizations of human tumors. It is possible to collect independent tumors from animals of the same genotype. Irrelevant “passenger” mutations are disfavored by the very short interval - within the first 14 weeks of life - in which tumors arise. Mouse studies avoid potentially confounding effects of therapy, which can be very difficult to control when examining human tumors. Additionally, there is a large body of genome wide information on chromatin modification, Rag2 binding, and DNA accessibility as applied to purified subsets of primary developing T cells, a level of characterization not yet available for humans.

Analysis of 9 thymic lymphoma genomes yielded 275 validated somatic structural rearrangements. Most chromosome translocations near Ig and TCR loci lacked evidence of V(D)J recombination error (Fig.1). Furthermore, the rearrangements showed no apparent connection to lymphomagenesis. In contrast, there were numerous deletions arising from off target V(D)J recombination, many of which mutated genes with a known or strongly supported role in oncogenesis. It was evident that this type of aberrant V(D)J recombination is an inherent risk of the antigen receptor gene rearrangement process, being observed in both Rag2³⁵² as well as wild-type Rag2-expressing mice. Results with the experimentally tractable model broadly recapitulate observations in human B-ALL. Further, the distribution of off target breakpoints relative to histone H3K4me3-differentiated chromatin suggests that a known Rag2-mediated negative autoregulation of RAG activity deters the endonuclease from generating oncogenic rearrangements.

RESULTS

Deletion is the predominant structural variant class in p53-deficient lymphomas

Through genome wide analyses we compared the somatic structural variants (SVs) in T cell lymphomas from Tp53^−/− mice to those from Rag2^352/352, Tp53^−/− double mutants (Experimental Procedures, tumor phenotypes are given in Fig. S1A,B). Tumors were sequenced with an average coverage of 14X. 57 somatic SVs (range: 8–20) were identified in four Tp53^−/− lymphomas and 218 SVs (range: 23–68) were detected in five Rag2^352/352, Tp53^−/− lymphomas (Fig. 2). Breakpoints mapped to scattered positions throughout the genome (Fig. 2). Over half of the junctions were deletions, the remaining being divided among inversions, translocations, and apparent duplications (Fig. S1D).

Fig. 2. Structural alterations in thymic lymphoma genomes.

Circos plot representation of all rearrangements (including normal Ig/TCR junctions) detected in the genomewide analyses. Ig and TCR loci are marked by symbols. Translocations (seen as arcs) are more prevalent and show a tendency to localize in the region of Ig/TCR genes in Rag2^352/352, Tp53^−/− tumors. See also Fig. S1 and S4.

Normal and abnormal rearrangement at TCR and IGH loci

Among the 275 validated somatic rearrangements, 76 authentic V(D)J junctions were recovered (Fig. S1, Table S1). Some junctions were supported by as few as two paired reads, confirming the high sensitivity of our analysis pipeline (Fig. S1, Table S2). About 15% of the Ig or TCR locus recombinants were comprised of abnormal J-to-J junctions, junctions between V gene segments and “locus-deleting elements”, incomplete (out-of-order) V-D recombination, directly joined Vδ to Jδ junctions (skipping D), and, in the case of the only detected signal joint, an atypical loss of sequence from the RSS of both signal ends (Table S1). Each type of abnormality is documented in wild-type, untransformed thymocytes. Rearrangements involving authentic V(D)J gene segments provided an internal control for various comparisons made in this study.

Off target V(D)J recombination

To date, available systematic methods for identifying cRSS are either of unproved accuracy, or use an algorithm of limited availability, nonetheless, an ability to identify (and quantify) off target V(D)J recombination errors is a key necessity for the present study. We formulated and tested two simple criteria. First, a CAC must exist to the right (or GTG to the left) of both breakpoints (a point addressed more fully below), and second, must occur within a specified distance from the breakpoint. The conserved CAC of an RSS heptamer demarcates the position at which RAG makes its cut, and the distance between the CAC motif and a breakpoint reflects the limited number of base pairs that may be lost when an end is processed prior to ligation. For coding joints, the CAC distance-to-breakpoint was set at 21 bp, the maximum observed for authentic RSS junctions in the data (Fig. 3A). When each is plotted according to their two distance-to-breakpoint values, authentic V(D)J junctions fell into a tight cluster averaging 3 nucleotides from either CAC motif. When the remaining 198 SVs from our dataset were similarly plotted, two distinct distributions emerged: junctions with CAC distance-to-breakpoint values like authentic recombination products and those with a wider, randomly distributed set of distances. The presence of two populations in our experimental data was confirmed by analyzing a set of 180 simulated junctions (see Supplemental Experimental Procedures), which reconstructed only the broadly scattered distribution (Fig. 3A).

Fig. 3. Application of a ‘CAC-distance-to-breakpoint’ method to identify off target, RAG-mediated “coding joint-like” junctions.

A. CAC-distance-to-breakpoint values distinguish two populations of junctions in T cell lymphomas.

Left: Tumor V(D)J coding joints are plotted according to CAC distance-to-breakpoint values. Middle: A plot of CAC distance-to-breakpoint values for all junctions other than authentic V(D)J coding joints shows a clustered and dispersed pattern. Right: A CAC-distance-to-breakpoint plot for a set of simulated junctions. The boxed area in each panel indicates the 21bp cut-off used to discriminate between cRSS and nonRSS junctions as tested in B and C. Numbers at the top refer to the number of junctions plotted within each set (junctions with distance to breakpoint values exceeding 100 bp on one or both sides are not shown). The 74 RSS-to-RSS junctions represent all the coding joints generated by V(D)J recombination, excluding1 hybrid joint and 1 signal joint. The “all except RSS-to-RSS junctions” set includes all other junctions (275 total junctions minus 74 coding joints, 1 hybrid joint, and 1 signal joint).

B. The CAC-distance-to-breakpoint method identifies junctions with and without extended matches to the canonical RSS sequence.

Top: V(D)J junctions in tumors (seen in part A, left panel) occur near RSS motifs that, as expected, exhibit canonical heptamer/nonamer sequences. The group of provisionally identified cRSS junctions (red boxed area middle panel part A) reveals a consensus sequence that matches 11 of 13 canonical heptamer and nonamer positions. Putative cRSSs from nonRSS and simulated junctions (outside the red boxed area, including those not shown) do not resemble a canonical RSS. Some bias toward nonamer matches is introduced by the method of nonamer assignment (Supplemental Experimental Procedures) but would not artifactually enhance identities at functionally important positions. The CAC defining the heptamer start in all categories is shown for clarity, not for purposes of comparison.

C. Cryptic RSS junctions as defined by the CAC-distance-to-breakpoint method have other, independent properties in common with V(D)J junctions.

Junctional insertions are seen in RSS, cRSS and nonRSS junctions. cRSS junction inserts have restricted length and high G:C composition, and P nucleotides, a unique feature of V(D)J recombination - similar to inserts observed in RSS junctions. Inserts in nonRSS junctions do not recapitulate these properties. Values in parentheses exclude an extreme outlier (length exceeding Q3_cRSS + (MAX_RSS − Q3_RSS) \ IQR_RSS) * IQR_cRSS). IQR = interquartile range.

Sequence logos (Crooks et al., 2004) were used to compare visually RSS sequences as used in the tumor Ig/TCR rearrangements to cRSS sequences derived from cryptic junctions (identified according to CAC-to-breakpoint values), to analogous CAC-defined “nonRSS” sequences from junctions falling outside the 21 bp range and finally, to “simRSS” sequences from the simulated junctions (Fig. 3B). Authentic, targeted RSS elements gave the best match to the canonical RSS [CACAGTG (12/23 spacer) ACAAAAACC], most consistently at functionally important positions (Little et al., 2014). Similarly, the consensus sequence of the presumptive cRSSs reconstructed the canonical RSS sequence except for the last two (CC) of the nonamer. There was a strong signal for the heptamer sequence beyond CAC and a signal at the functionally important nonamer positions 5,6, and 7 (Little et al, 2014). The logos of the nonRSS and simRSS groups lacked these significant features.

A second test of the CAC distance-to-breakpoint method evaluates the cRSS junction group for two hallmark features of V(D)J recombination. These are “N regions”, short G/C- rich insertions generated by terminal deoxynucleotidyl transferase and palindromic “P nucleotides”. These arise when hairpin-terminated RAG cleavage intermediates are nicked a few bp from the terminus and the opened ends happen to escape trimming before becoming joined. In both respects, cRSS junctions compared very well to the authentic V(D)J junctions (Fig. 3C). In contrast, the insertions seen in nonRSS junctions were AT-rich and more variable in size. P nucleotides were identified at the same frequency among cRSS as authentic V(D)J junctions, while almost no qualifying inserts were seen in the nonRSS group. These data indicate that the CAC distance-to-breakpoint criteria indeed identify off target V(D)J recombination products.

The CAC distance-to-breakpoint method was largely in agreement with the computational approach (Papaemmanuil et al. 2014). In their study, 44 junctions were assigned with high confidence to off target V(D)J recombination, 39 of which were also identified by our criteria (Table S3). For 214 junctions determined not to be cRSS-related in the same study, there was agreement on 185, again a good concordance (86%). The cryptic junctions detected by the distance-to-breakpoint method in the nonRSS group appear to be properly assigned, because the sequence logo was well matched to the canonical nonamer sequence. The third cRSS discovery method, based upon “RIC” (Recombination Information Content) scores (Cowell et al, 2002), did not perform as robustly. Of the 39 concordant junctions (i.e. off target V(D)J recombination products with strong support), only 12 (31%) bore a cRSS at both breakpoints according to RIC scores and individually over half of the 39 cRSSs failed the RIC test (Table S3). Overall, our test and that of Papaemmanuil et al. (2014) appear equally discriminatory, given the percent coincidence for both positive and negative assignments is about the same.

We note, however, that because the sensitivity of our method is based upon the presence of a cRSS on both sides of the junction, it is not suited to the detection of “one-sided” events (which are predicted for end-donation errors, Fig.1B). Nevertheless, we found no evidence that end-donations could comprise more than a minor fraction of aberrant V(D)J recombination products. End-donations suffering extensive resection after cleavage would not have been detected by our analysis. Moreover many one-sided junctions in the collection of Papaemmanuil et al. (2014) qualified as two-sided products according to the CAC distance-to-breakpoint test (Table S3).

Off target V(D)J recombination causes recurrent deletions in Notch1 and Jak1

Mutations causing constitutive activation of the Notch1 oncogene have been demonstrated in both human and mouse T cell neoplasms. In the latter, recurrent cRSS-mediated deletions are found that produce a Notch1 protein activated via N-terminal truncation (Ashworth et al., 2010; Jeannet et al., 2010; Van Vlierberghe and Ferrando, 2012). The two most commonly observed Notch1 deletions (Tsuji et al., 2004) were also seen in our dataset, along with additional examples (Fig. 4A, Fig. S2). Our results, along with data from Tsuji et al. (2004) identify 14 functional cRSSs in a 30kb span of the Notch1 locus (Fig. 4A). This high density of functional cRSSs agrees with predictions based on artificial recombination substrates (Lewis et al., 1997).

Fig. 4. Recurrent cRSS-mediated deletions.

A. Notch1 activation via cRSS-mediated recombination.

Diagram of the Notch1 protein indicating relevant domains. Middle: representation of the Notch1 locus with lines locating the corresponding translated regions in the protein diagram. Triangles indicate cRSSs identified in the present study as well as candidate cRSSs reported in Tsuji et al., 2004 (see Supplemental Experimental Procedures for inclusion details). The most commonly observed deletion removes the 5’ promoter (blue arrow) causing an internal promoter to be used (red arrow). The resulting protein lacks the extracellular EGF1-like repeats, and the negative regulatory region (NRR) but retains most of the transmembrane (TM) domain. The truncated form is thought to insert into the membrane and become cleaved in a ligand-independent fashion, releasing constitutively activated, intracellular Notch1 (ICN). All other Notch1 deletions likewise deleted the 5’ promoter. See Table S2 for junction sequences.

B. Internal Jak1 deletions suggest another instance of oncogene activation through cRSS-mediated recombination.

A diagram of domains within the Jak1 protein is displayed above a representation of the locus. Individual cRSS-mediated deletions are shown as horizontal arrows. cRSS coordinates (corresponding to the heptamer border, mm9 genome assembly) are given for each deletion. All parts of the diagram in red indicate the most frequent cRSS deletion in Jak1. Clustering of the deletions suggests that deletions in this region may be activating (see Discussion; for sequences, Fig. S2).

C. Predicted reciprocal cryptic signal joints can be detected in tumor DNA samples by PCR.

Sequences of independently isolated cryptic signal joints from recurrent Jak1, Agpat9, and Trat1 rearrangements. See also Fig. S4. For evidence of ongoing recombination in thymoma, see Fig. 3.

PCR primers specific for each Notch1 cRSS defined in the present study were used to assay for additional deletions that, due to coverage, may not have been detected in the genomewide analysis. These were indeed seen, giving a total of 8 distinct Notch1 deletions (Fig. S2), distributed between all five RAG2^352/352, Tp53^−/− tumors and two of the Tp53^−/− -only isolates (Fig. 5). All deletions encompassed the 5’ region of the gene, deleting the 5’ promoter, as is the case for cRSS-mediated mutations previously demonstrated to result in constitutive Notch1 activation (Ashworth et al., 2010; Jeannet et al., 2010).

Fig 5. Recurrently disrupted genes and cancer-associated genes are found in tumors of both genotypes.

Genes mutated in tumors from different mice are listed above the thick line. The tumor harboring the detected mutations is indicated at the top of each column. Check marks and other symbols are defined in the Key. There were no clustered, recurrent breakpoints shared between tumors occurring outside a gene. A comprehensive list of cancer gene mutations, whether multiply or singly observed in this study is given. Yellow boxes indicate genes that occur among the 572 currently listed in the Cosmic Cancer Gene Census (http://cancer.sanger.ac.uk/census). Green boxes indicate genes with a likely cancer association (Table S4).

We also identified recurrent, cRSS-mediated rearrangements at a second known oncogene, Jak1. One specific recurrent deletion (8.2 kb) eliminated exons 6 through 8 (Fig. 4B, Fig. S2), and was found in every Rag2^352/352 thymic lymphoma as well as in one Rag2^w/wt tumor by PCR. These Jak1 deletions share several properties with the activating Notch1 deletions. They recur in independent tumors, involve multiple sets of cRSSs, overlap one another over a specific area of the locus and may provide for production of a modified (rather than absent) protein product. The similarities raise the possibility that, as for Notch1, these cRSS-mediated mutations might activate Jak1.

cRSS-mediated deletions and candidate driver mutations

Cryptic RSS junctions affected genes with a potential oncogenic impact in that they govern cell proliferation, differentiation or survival. Of these, 21 specific deletions were selected for PCR analysis in the hope of discovering additional recurrently mutated genes. We examined the nine original tumors and six additional Rag2^352/352 p53^−/− tumors, revealing recurrent, cRSS-mediated deletions at 3 more loci: Trat1, Phlda1 and Agpat9 (Fig. S2). All recurrent mutations, except the Phlda1 deletion, were verified in both Tp53^−/−, RAG2^352/352 and Tp53^−/− -only lymphomas (Fig.5). Recurring examples of the same cRSS junction in independent tumors suggest the possibility that, like in Notch1, these specific gene disruptions are driver mutations.

If off target V(D)J recombination is a force in the oncogenic process, it should be possible to detect enrichment of cRSS junctions among SVs that interrupt genes of significance in cancer. “Cancer genes” were defined by a metric based on the co-occurrence of search terms in the Web of Science Core Collection (http://apps.webofknowledge.com/) (Experimental Procedures, Table S4). Almost 90% of cRSS junctions in the collection interrupted a gene and well over half of these genes (65%) were associated with cancer (for additional details see Supplemental Experimental procedures and Table S4). By comparison, 70% of nonRSS junctions had breakpoints located within a gene, and only 24% of these disrupted cancer genes, a highly significant difference (p<0.0001, Table S5). These data, summarized in Fig. 6, support the possibility that off target V(D)J recombination contributes to oncogenesis.

Fig. 6. Overview and role of off target V(D)J recombination.

A. SVs generated by either RSS, cRSS or nonRSS mechanisms in RAG2³⁵² vs. RAG2^wt tumors.

Off -target V(D)J joints (cRSS-cRSS and cRSS-RSS), on-target RSS-RSS V(D)J and non-RSS junctions are indicated by colors given in the key. cRSS-mediated events account for 46% of the structural genomic alterations in the RAG2^352/352, Tp53^−/− tumors, but this is the case for only 15% of rearrangements in cRSS-mediated in thymic tumors (Table S5).

B. The role of cRSS recombination in tumor formation.

The Y-axis in each plot represents junction numbers. Black sections, stippled and white section indicate junctions in cancer genes, non-cancer genes and non-gene for all junctions except those involving an RSS (RSS-RSS and RSS-cRSS). Assignment of cancer gene association is given in Table S4. There is a highly significant difference between cRSS vs. nonRSS mechanisms with respect to observed mutations in cancer-associated genes (p <0.0001 - see Table S5). A further breakdown of the data in part B, right panels, represents the impact of the Rag2 C-terminus RAG2 carboxyl terminal domain with respect to three types of rearrangements. RAG2^352/352 tumors show a bias toward cRSS-mediated cancer gene mutations relative to non-RSS mechanisms that is absent in RAG2 wt tumors (compare the four black sections in each of the 4 bars, right panel; p value for the difference is 0.0044, Table S5).

To investigate whether the Rag2 C-terminus reduces oncogenic mutation, the degree to which cRSS versus nonRSS junctions interrupt cancer genes was further compared between Tp53^−/−, RAG2^352/352 and Tp53^−/−-only tumors. The fraction of cancer genes that were “hit” by cRSS recombination with Rag2³⁵² was significantly different than that observed with intact Rag2 (p=0.0044, Fig. 6B, Table S5C).

Excision products containing cRSS signal joints (cTrecs) indicate recent recombination activity

V(D)J recombination co-generates reciprocal products, a coding joint and a signal joint. When a coding joint is created by deletion, the signal joint will occur on an excised circular DNA molecule ((Fujimoto and Yamagishi, 1987), Fig. 1A). The circular excision products, dubbed “Trecs” (TCR rearrangement excision circles) do not replicate and are progressively diluted by cell division. We sought evidence that cryptic Trecs (“cTrecs”) might be generated by off target recombination. To maximize the possibility of detection, we focused the effort on recurrent cRSS rearrangements. Excised signal joints reciprocal to the Jak1, Phlda and Trat1 cryptic coding joints were recovered by PCR (Fig. 4C). The demonstration of cTrecs corresponding to chromosomal cRSS rearrangement is a new observation. By analogy to Trecs, it appears that off target V(D)J rearrangements occurred relatively recently in the history of the rapidly dividing tumor, indicating ongoing recombination activity. (Identification of cTrecs corresponding to potential oncogenic events does not, however, mean they are related to founder mutations, but rather reflects the evolving, multiclonal nature of the lymphomas, as noted in the Discussion.) Consistent with this possibility, RAG transcripts were detected in several tumors (Fig. S3). Furthermore, a high level of V(D)J recombination was detected with a standard extrachromosomal recombination assay in a cell line established from a Rag2^352/352, Tp53^−/− tumor (Fig. S3). The presence of specific cTrecs could conceivably provide an additional parameter for phenotyping human tumors as well as being an indicator of recent mutation.

Chromosomal cRSS signal joints are surprisingly rare

If, as expected, cRSS targets are randomly oriented with respect to one another in the genome, recombination between any pair of cRSS should be as likely to generate a signal joint as a coding joint. Rearrangement of a pair of cRSSs that happen to be oriented “heptamer”-to-“heptamer” produces a cryptic signal joint that remains in the chromosome. Surprisingly, chromosomal coding joints outnumbered chromosomal signal joints 6- to 7-fold in the tumor genomes (Tables S1, S3). A marked deficit was also seen in data from human ETV6-Runx1 ALL (Papaemmanuil et al., 2014): we found only one cryptic signal joint among 140 candidate junctions (Table S3). Indeed, very few signal joints representing off target V(D)J recombination products have been reported thus far [e.g. (Mendes et al., 2014)]. The rarity of cryptic signal joints in the tumor genomes is not due to a failure in their production, because as described above, cTrecs are found corresponding to Trat1, Agpat9 and Jak1 rearrangements. It is not obvious what the chromosomal cSJ deficit means, but the consistency with which it is observed suggests further investigation could be revealing.

Genomic context of functional cRSSs

A potential for off target V(D)J recombination is a property of wild-type RAG [(Papaemmanuil et al., 2014) and this study], underscoring the importance of the mechanisms that limit or prevent recognition errors during V(D)J rearrangements. Such errors might be minimized by chromatin/RAG interactions. Three relevant chromatin properties have been measured at the whole-genome level in primary mouse cell populations corresponding to CD4+ CD8+ T cell precursors or to unfractionated thymocytes (most of which are CD4+ CD8+ pre-T cells). The properties measure accessibility, as reflected by DNase I sensitivity (Sabo et al., 2006), methylation of lysine 4 in histone H3 (Ji et al., 2010; Vigano et al., 2014) and the location of chromatin-bound Rag2 (Ji et al., 2010).

The location of each 39 bp interval (a length that accommodates both 12 and 23bp versions of a RAG target) corresponding to 162 RSSs, 163 cRSSs, and 224 nonRSS (Table S6) in the thymic tumors was determined. The relationship of the target motif and the position of H3K4me3-modified chromatin is not known, so this choice minimized imprecision introduced by taking a larger interval to both sides of the breakpoint. The percent of RSS targets overlapping DNAse I-sensitivity hotspots, H3K4me3-modified chromatin or Rag2 binding regions is shown in Fig. 7B. Even in the case of authentic RSSs, slightly over half lay in chromatin bearing all three features, and 14% of RSS breakpoints occupied positions that lacked any of the properties. This is likely due to the fact that the distribution of marks is measured in aggregate, and thus may not reflect minor subpopulations. Similarly, the degree to which these aggregate measurements reflect the situation in a tumor cell at the time of cRSS rearrangement cannot be determined.

Fig. 7. Truncation of the Rag2 C-terminus disrupts a correlation between cRSS breakpoints and H3K4me3.

A. Distribution of breakpoints (colors indicate RSS, cRSS and nonRSS) relative to published genomewide determinations of LEFT: DNase I hotspots, MIDDLE: H3K4me3 ChIP-seq analysis of a genetically-isolated DP thymocyte population and RIGHT: RAG2 ChIP-seq analysis, same population. Bar heights represent the percent of breakpoints (defined in the text) in DNase I hotspots, tri-methylated H3K4 peaks, or RAG2 binding peaks. Correlations between breakpoint categories and chromatin features are significantly different from a random distribution as indicated for the evaluated RSS and cRSSs. The correlation between nonRSSs breakpoints and DNase I hotspots reached significance, but nonRSS breakpoints were not markedly different from random with respect to H3K4me3 or RAG2 binding peaks (Table S5).

B. Comparisons made between breakpoints in p53-null tumors with WT (solid bars) versus C-terminally truncated RAG2. LEFT: Blue bar graphs give the percent of authentic RSS breakpoints co-localized with accessibility, H3K4me3 modification and RAG binding. No statistically significant differences are seen between genotypes. RIGHT: Orange bars indicate the percent of cRSS break points co-localized with accessibility, H3K4me3 modification and Rag2 binding. The cRSS breakpoints have a significantly reduced coincidence with H3K4me3 in RAG2³⁵² tumors when compared to the cRSS breakpoints in RAG2^wt tumors (two-tailed Fisher’s exact test). None of the other comparisons reach significance (Table S5).

Despite these caveats, cryptic RSS breakpoints exhibited a non-random distribution relative to DNAse I hotspots, H3K4me3-modified chromatin and RAG2 binding (p <0.0001, Table S5E). Localization of cRSS breakpoints to the above markers may define the state of genomic regions susceptible to off target V(D)J recombination. Alternatively, considering that all three properties including Rag2 binding (Ji et al., 2010) are directly or indirectly associated with active gene expression, a more trivial possibility is that the distribution of cRSS targets nonspecifically reflects DNA target accessibility.

Because Rag2³⁵² lacks the PHD domain implicated in recognition of H3K4me3 marks, the location of breakpoints in relation to this chromatin mark was further examined. Comparing cRSS sites in wild-type Rag2, Tp53^−/− tumors with those in arising in Rag2^352/352, Tp53^−/− revealed no significant differences with respect to colocalization of authentic RSS targets and H3K4me3 (Fig. 7). In contrast, the distribution of cRSS targets with respect to H3K4me3 was significantly different (p=0.0007). C-terminally truncated RAG2 thus appears markedly less constrained by the presence of H3K4me3 that its wild-type counterpart.

Most translocation junctions reveal no apparent link to V(D)J recombination

In 22 translocation SVs, we detected 23 translocation junctions (one SV contained 2 junctions). Genomewide SV analysis confirmed earlier findings that translocations with breakpoints near TCR loci were present in Rag2^352/352, Tp53^−/− lymphomas, and not the Tp53^−/− only counterparts [(Deriano et al., 2011), Fig. 2]. Between the two BAC probes used in our previous study (Deriano et al, 2011) we identified a total of 5 junctions in two RAG2^352/352, Tp53^−/− tumors (# 132, 12154). Of these, only one can be, without doubt, classified as an off target V(D)J rearrangement. This is the translocation between chr4 and chr14 creating a signal joint composed of the RSS of a TCR Vα-gene segment (chr14) and a cryptic RSS within Kdm1a locus (chr4; Fig. S4). In the case of a simple trans joining error, a coding joint should be present on a reciprocal derivative chromosome, and this was indeed demonstrated by PCR of the tumor DNA (Fig. S4). The 4:14 exchange illustrated that trans joining errors likely play a role in tumor formation, because there is abundant evidence that mutation of Kdm1a, a histone deacetylase gene, is significant in leukemia (Harris et al., 2012). This was in contrast to the other translocations, of which none interrupted a cancer-associated gene (Table S4),. Furthermore, we found no evidence of RSS or cRSS in these junctions, suggesting another mechanism is responsible. These results raise the question of whether or not apparent locus-specific structural rearrangements in other Tp53^−/− mouse models can be attributed to V(D)J recombination, as few recombinant junctions have been analyzed at the DNA sequence level.

DISCUSSION

The genomic lesions described here bear striking similarities to those recently reported in pediatric ETV6-Runx1 B-ALL (Papaemmanuil et al., 2014). Common to both are a preponderance of deletions, many resulting from off target V(D)J recombination. In both cases, RAG-mediated deletions between pairs of cRSS create known and suspected driver mutations. Trans events involving V(D)J recombination are much less frequent, and in both studies misrecognition of RSSs is seen to occur in tumors with wild-type RAG. In human ETV6-Runx1 B-ALL, initiating events are thought to occur prenatally, but tumors were not available for investigation until they reached, on average, 57 months of age (Papaemmanuil et al., 2014). By comparison, Rag2^352/352, Tp53-null mice die of thymic lymphomas at a median age of 13 weeks. The status of the genome in the murine system provides as close a reflection of the ‘smoking guns’ in the oncogenic process as is currently possible.

The paucity of chromosome translocations involving antigen receptor loci in our dataset is striking, given our previous identification of such translocations in thymic lymphomas from Rag2^352/352, Tp53-null mice (Deriano et al, 2011). This discrepancy likely arises because the translocations detected by cytogenetic techniques in our earlier work, while recurrent, were not uniformly present in all tumors, and in many cases were present in minor subpopulations (Deriano, 2011, see Table S1). Thus, those translocations, which based on their frequency are unlikely to represent driver mutations, might not have been present in the tumor samples analyzed here, or were at low enough abundance to have escaped detection by whole genome sequencing.

Rag2 C-terminus- H3K4me3 interactions may enhance the fidelity of V(D)J recombination

Comparison of genomic lesions in lymphomas from Rag2^wt/wt, Tp53^−/− and Rag2^352/352, Tp53^−/− mice revealed that although off target V(D)J recombination events are found in both genotypes, they are much more prevalent when Rag2 is C-terminally truncated. cRSSs identified in Rag2^352/352, Tp53^−/− tumors occur significantly more frequently outside H3K4me3-modified regions of the genome (as measured in normal pre T cells (Ji et al., 2010; Vigano et al., 2014) than those in Tp53^−/− mice. The significance of this bias is underscored by the fact that comparisons of a less specific “accessibility” (DNase I sensitivity and Rag2 binding) show inconsistent trends between genotypes and fail to reach statistical significance. These data establish a mechanistic link between the PHD domain located in RAG2’s C-terminus, H3K4me3-modified chromatin, and suppression of off target V(D)J recombination. The possibility of such a link has been suggested, most recently, in a study that biochemically demonstrated reversal of RAG autoinhibition when the PHD finger in the Rag2 C-terminus interacts with H3K4me3 peptides (Lu et al., 2015). The observations reported here support a role for H3K4me3-mediated reversal of autoinhibition in minimizing the oncogenic impact of faulty RSS recognition.

Numerous rearrangements lacking identifiable features of V(D)J recombination junctions in Rag2^352/352 tumors

Curiously, all classes of SV --authentic V(D)J recombination, cRSS rearrangements and nonRSS junctions subgroups (deletion, inversion, duplication) -- were quantitatively increased in Rag2^352/352, Tp53^−/− tumors. We considered several possible explanations for this observation. An overall increase in SVs might be caused by the loss the T490 phosphorylation site. Rag-mediated breaks arising inappropriately during S phase could induce DNA repair functions that are error-prone when handling other types of damage. A more speculative suggestion is that replication of coding ends bearing unresolved hairpin termini, a DNA intermediate normally excluded from S phase, generates DNA palindromes which are known to provoke large scale rearrangements in mice, humans and yeast [(Coté and Lewis, 2008) and cited therein]. The apparently non-specific junctions could reveal a new aspect of aberrant V(D)J recombination. For example, an uncontrolled resection of cleaved ends (at the kilobase-to-megabase scale), might obliterate any signature of off target RAG cleavage in junctions.

Off target V(D)J recombination generates known and candidate driver mutations

Where mutation is repeatedly discovered in the same gene in independent tumors, it is usually taken as evidence of biological selection in the pathological process. If, however, V(D)J recombination is the mutation mechanism, a relationship is less certain. The repeated observation of a cryptic junction could be due to properties that fortuitously favor recombination rather than biological selection. Whereas there is little question that Notch1, Jak1, Ikzf1, Fhit, and Pten, all included in the short, expertly curated list of 572 COSMIC Cancer Gene Census (http://cancer.sanger.ac.uk/census/), are drivers in the oncogenic process, the four remaining recurrently mutated genes occupy a gray zone.

To address this, we developed a system for stratifying all the genes mutated in the present study according cancer relevance (Experimental Procedures and Table S4). By this approach, Nr3c1 and Phlda1 belong to the same group as the above-mentioned genes and can also be considered cancer-associated. Agpat9 (Table S4B) and Trat1 however have modest or no support for a cancer role in the literature (see Table S4C). Although Agpat9 and Trat1 are low-rated cRSS rearrangements, it may be significant that both operate in pathways that are connected to oncogenesis. Trat1 (T cell receptor associated protein) is upstream in the generation of PIP3, which has regulatory interactions with both Akt1 and Pten. The Pten pathway is evidently an Achilles heel, given that Pten itself is recurrently mutated by cRSS mechanisms here (Fig.1) and in human T-ALL (Mendes et al., 2014). Agpat9 is a 1-acylglycerol-3-phosphate o-acyltransferase involved in the production of phosphatidic acid, a cofactor in the AKT1/Tor pathway (Agarwal, 2012). It remains to be seen whether cRSS-mediated deletions pinpoint new cancer genes, but given the potential applications, this question bears further investigation.

Potential mechanism for Jak1 activation

Analyzing off target mutations may enhance understanding of cancer development mechanisms. It has been demonstrated that cRSS-rearrangements at Notch1 cause the production of a truncated, activating form of the protein (Ashworth et al., 2010; Jeannet et al., 2010). It would appear, given the number of similarities between the two, that the Jak1 mutations seen here, like the deletions at Notch1, may be activating. Activated Jak1, caused by point mutations or other small changes that increase protein stability is seen in both human and murine lymphoid tumors (Chen et al., 2012; Bellanger et al., 2014; Flex et al., 2008). The commonly observed 8.2 kb cRSS-mediated deletion here is compatible with the possibility of an in-frame splice between exons 5 and 9 in the mRNA. This would delete a large part of the FERM domain (Fig. 4), eliminating functions thought at present to be a pre-requisite for mediating the oncogenic effects of Jak1 activation.

Are RAG-mediated off target events relevant to nuclease-mediated genome editing?

Suppressing off target cleavage is an exceptionally challenging problem in the V(D)J recombination system, and as such may inform strategies for increasing the accuracy of genome editing in clinical applications. During V(D)J recombination, several layers of regulation converge to limit RAG action at inappropriate target sites. Such measures include the tempo with which T cell intermediates traverse the vulnerable developmental stage (Haines et al., 2006), the destruction of Rag2 upon entry into S phase of the cell cycle, and the physical accessibility of the underlying cRSS sequences upon transcription, histone modification and chromatin looping (Little et al., 2014). Additionally, highlighted here is the possible importance of chromatin-modulated Rag2 auto-inhibition in the suppression of off target activity. Finally, it goes without saying that a highly important requirement for fidelity is an intact p53 pathway. These safety nets, the product of millions of years of evolution, counteract the potential for disaster inherent in a process that occurs in cells with high proliferative ability, throughout life. In the short term, it may not be a simple task to computationally predict the unintended targets of man-made genome editing nucleases, nor to reduce their numbers to a safe level by engineering ever-more stringent enzymes. The strategies employed in the V(DJ recombination system may inform these efforts.

Experimental Procedures

Whole genome tumor and normal paired-end libraries were sequenced and analyzed with a custom analysis pipeline (Mijuskovic et al., 2012; Mijušković et al., 2012). Candidate SVs that passed all filters were tested for tumor specificity by PCR (Table S2). The DNA sequence of all 275 confirmed junctions was then determined by Sanger sequencing. Detailed methods, including downstream analysis, identification of cRSS junctions, creation of arbitrary junctions, and stratification of cancer genes are presented in Supplemental Experimental Procedures.

Analysis of the genomic context of breakpoints in Figure 7 was based on publicly accessible genomewide DNAse I hypersensitivity data (wgEncodeUwDgfThymusC57bl6MAdult8wksHotspotsRep1), H3K4me3 peaks, (GSM530317) and RAG2 binding peaks (GSM530316) (Ji et al., 2010; Sabo et al., 2006).

To visually confirm overlaps relative to H3K4me3 peaks (Ji et al., 2010), the track was uploaded to the UCSC browser along with WIG files of genomewide data from early T cell precursors (Vigano et al., 2014)(GSM1164638, GSM1164639, GSM1164643, GSM1164644). There was close agreement at the monitored breakpoints, despite the use of different cell purification methods and different sources for H3K4me3 antibodies. p values were calculated using a two-tailed Fisher’s exact test.