Hot spots of retroviral integration in human CD34⁺ hematopoietic cells

Gene therapy of genetic blood disorders requires stable genetic modification of hematopoietic stem cells. Gene transfer vectors derived from murine gamma-retroviruses, such as the Moloney murine leukemia virus (MLV), have been used for more than a decade to transduce human bone marrow (BM)–derived or mobilized hematopoietic stem/progenitor cells (HSCs) in a clinical context. Retroviral vector–mediated gene transfer has recently achieved therapeutic efficacy, allowing correction of life-threatening diseases such as severe combined immunodeficiencies (SCIDs)^1–3  or chronic granulomatous disease (CGD).⁴  MLV-derived vectors, however, have also raised significant safety concerns for the genotoxic risk potentially associated with their uncontrolled integration into the human genome.^5–7  Indeed, insertional activation of a T-cell proto-oncogene has been correlated with the occurrence of lymphoproliferative disorders in 3 patients treated with retrovirally transduced hematopoietic cells for X-linked SCID (X-SCID).¹  Recent studies have shown that gamma-retroviral vectors integrate preferentially within transcribed genes and around promoters and CpG islands,⁸  where insertion of the viral long terminal repeat (LTR) transcriptional enhancer has a high probability to interfere with gene regulation.⁹  Nevertheless, no adverse event related to viral insertion was reported in other clinical trials for X-SCID,³  adenosine deaminase-deficient SCID (ADA-SCID),²  CGD,⁴  or graft-versus-host disease (GVHD),⁹  suggesting the existence of specific risk factors that are incompletely understood.¹⁰ 

Analysis of MLV integration patterns in natural or experimentally induced leukemias/lymphomas showed the existence of insertion sites recurrently associated with a malignant phenotype. These “common insertion sites” (CISs) include proto-oncogenes or other genes associated with cell growth and proliferation, the activation or deregulation of which has a causal relationship with the establishment and/or progression of neoplasia.¹¹  Some of these sites, such as the EVI1-MDS1 locus, have been identified at relatively high frequency also in the nonmalignant progeny of transduced hematopoietic cells in mice,¹²  nonhuman primates,¹³  and humans,⁴  indicating that insertion into certain genes may cause clonal amplification of transduced progenitors in vivo. From these studies, however, it is not clear whether clonal dominance is entirely the result of in vivo selection, or is favored by the existence of highly preferred regions of retroviral integration that make clonal amplification more likely to occur. This issue is highly relevant in understanding the different outcomes of different gene therapy clinical trials, in assessing the relative safety of using MLV-derived vectors in specific clinical applications, and in comparing the safety profile of alternative vectors (eg, HIV-derived lentiviral vectors) or vector designs.

We report an analysis of gamma-retroviral (RV) and lentiviral (LV) vector integration hot spots from large collections of integration sites obtained from human cord blood (CB)– and BM–derived CD34⁺ HSCs transduced in vitro and analyzed without selection. Hot spots account for more than 20% of the MLV integration sites, while they are significantly less frequent in the case of HIV-derived vectors. Integration sites associated with clonal dominance and neoplasia in both mice and humans, including LMO2, are hot spots of gamma-RV but not LV integration in human hematopoietic cells.

CB–derived CD34⁺ cells were transduced with the previously described LGSΔN and LGSΔN-ΔCAAT RV vectors,¹⁴  driving the expression of green fluorescent protein (GFP) under an intact or a U3-deleted MLV LTR, and of ΔLNGFR under an internal SV40 promoter. CB-derived CD34⁺ cells were also transduced with the self-inactivating (SIN) pRRLsin-18.pptCMV-GFPwpre LV vector¹⁵  containing a U3-deleted HIV-1 LTR and a cytomegalovirus (CMV)–driven GFP cassette,¹⁶  or with the pHR2pptCMV-GFPwpre or the pHR2pptGSΔN LV vectors, retaining HIV-1 wild-type LTRs and driving the expression of GFP or ΔLNGFR under internal CMV or SV40 promoters. To generate the pHR2pptCMVGFPwpre construct, a pptCMVGFPwpre fragment from the pRRLsin-18.pptCMVGFPwpre vector was cloned into ClaI-EcoRI sites of pHR2MD-NGFR.¹⁷  To obtain the pHR2pptGSΔN LV construct, the pHR2pptCMVGFPwpre vector was digested with BamHI/EcoRI and ligated to a GFP-SV40ΔLNGFR cassette. BM–derived CD34⁺ cells were transduced with previously described RV vectors expressing either the ADA (GIADA1² ) or the γ_c receptor³  cDNA.

RV vector supernatants were produced by transient transfection of the amphotropic Phoenix packaging cell line. Infectious particle titer was determined on K562 cells. Vesicular stomatitis virus–G protein (VSV-G) pseudotyped LV particles were prepared by transient cotransfection of 293T cells, collected and concentrated as already described,¹⁷  and titrated on 293T cells. Transduction efficiency was evaluated by flow cytometry. Amphotropic or GaLV-pseudotyped ADA and γ_c receptor RV vectors were titered as previously described.^2,3 

CD34⁺ HSCs were purified from CB Ficoll fractions by magnetic sorting (MiniMACS; Miltenyi, Auburn, CA) and prestimulated for 24 to 48 hours in serum-free Iscove modified Dulbecco medium (IMDM) supplemented with 20% BIT (Stem Cell Technologies; Vancouver, BC, Canada), 20 ng/mL human thrombopoietin (TPO), 100 ng/mL Flt3-ligand (PeproTech, Rocky Hill, NJ), 20 ng/mL IL-6, and 100 ng/mL stem-cell factor (SCR; R&D Systems, Minneapolis, MN). RV transduction was performed by spinoculation (3 rounds at 1500 rpm for 45 minutes) in the presence of 4 μg/mL polybrene. LV transduction was performed by overnight incubation at a MOI of 200 in the presence of 4 μg/mL polybrene. Transduction efficiency was evaluated by analysis of enhanced GFP (EGFP) and/or ΔLNGFR expression by flow cytometry using a mouse anti–human NGFR antibody (Becton Dickinson, San Jose, CA).

BM- or peripheral blood (PB)–derived CD34⁺ cells were purified from healthy donors or patients with SCID by magnetic sorting, prestimulated for 24 hours in IMDM containing human serum or serum-free X-Vivo-10 medium and a cytokine cocktail (Flt3-ligand, SCF, TPO, IL-3), and transduced by 3 cycle-exposure to the GIADA1 or the γ_c receptor RV vector supernatant as previously described.^2,3 

Integration sites were cloned by linker-mediated polymerase chain reaction (LM-PCR) or linear amplification–mediated PCR (LAM-PCR), as described.^18,19  Briefly, genomic DNA was extracted from 0.5 to 5 × 10⁶ infected cells and digested with MseI and a second enzyme to prevent amplification of internal 5′ LTR fragments (PstI for RV vectors and SacI/NarI for LV vectors). An MseI double-stranded linker was then ligated, and LM-PCR was performed with nested primers specific for the linker and the 3′ LTR (MLV: 5′-GACTTGTGGTCTCGCTGTTCCTTGG-3′ and 5′- GGTCTCCTCTGAGTGATTGACTACC-3′; and HIV: 5′- AGTGCTTCAAGTAGTGTGTGCC-3′ and 5′- GTCTGTTGTGTGACTCTGGTAAC-3′). PCR products were shotgun-cloned (TOPO TA cloning kit; Invitrogen, Carlsbad, CA) into libraries of integration junctions, which were then sequenced to saturation. A valid integration contained the MLV or HIV nested primer, the entire MLV or HIV genome up to a CA dinucleotide, and the linker nested primer. Sequences between the 3′ LTR and the linker primers were mapped onto the human genome by the University of California Santa Cruz (UCSC) BLAT alignment tool (http://genome.ucsc.edu), accessed May 2004. Random genomic sequences originated by LM-PCR (genomic MseI-MseI, PstI-MseI, NarI-MseI, or SacI-MseI fragments) were used as controls. Sequences featuring a unique best hit with 95% or greater identity to the human genome were considered genuine integration sites, and classified as intergenic when occurring at an arbitrarily chosen distance of more than 30 kb from any “known gene” (UCSC definition), perigenic when 30 kb or less upstream or downstream of a known gene, and intragenic when within the transcribed portion of at least 1 known gene. In case of multiple transcript variants, the most represented and/or the longest isoform was chosen. Gene density analysis was performed using the UCSC Table Browser tool (http://genome.ucsc.edu). For each integration, the number of known genes (a single isoform in case of multiple variants) contained in a range of 1 Mb around the insertion site was calculated. For all pairwise comparisons, we applied a 2-sample test for equality of proportions with continuity correction using the Rweb 1.03 statistical analysis package (http://math.montana.edu/Rweb/).

A genomic region was defined as a “hot spot” for retroviral integration according to criteria developed for defining cancer-related CISs, with minor modifications.^11,20  Cutoff values were set at 36 kb for 2 insertions, 56 kb for 3 insertions, and 104 kb for 4 or more insertions.

The expression profile of CD34⁺ cells was determined by microarray analysis. RNA was isolated from 1 to 2 × 10⁶ CB- and BM-derived CD34⁺ cells stimulated with cytokines according to the same protocols used for RV (CB- and BM-derived cells) or LV (CB-derived cells) vector transduction, transcribed into biotinylated cRNA, hybridized to Affymetrix HG-U133A Gene Chip arrays (Santa Clara, CA) and analyzed as previously described.⁹  To correlate retroviral integration and gene activity, expression values from the CD34⁺ cell microarrays were divided into 4 classes (ie, absent, low below the 25th percentile in a normalized distribution, intermediate between the 25th and the 75th percentiles, and high above the 75th percentile).

Functional cluster analysis of genes targeted by retroviral integrations and from control sequences was performed using the DAVID 2.1 Functional Annotation Tool^21,22  (http://david.abcc.ncifcrf.gov). In the DAVID annotation system, a Fisher exact test corrected for multiple comparisons (DAVID's EASE score²¹ ) is adopted to measure the level of gene enrichment in Gene Ontology (GO) annotation terms with respect to a background population, and GO categories considered overrepresented when yielding an EASE score lower than 0.05. A list of 417 cancer-associated CISs was obtained from the Mouse Retrovirus Tagged Cancer Gene Database,²³  where murine genes were replaced with human homologs. Genes were analyzed also by the network-based Ingenuity pathways analysis tool (Ingenuity Systems, www.ingenuity.com). Gene identifiers were uploaded into the application, and mapped to their corresponding Focus Gene in the Ingenuity Pathways Knowledge Base. Networks were algorithmically generated based on the direct or indirect interaction between Focus Genes. The functional analysis of each network identified the biological functions and/or diseases that were most significant to the genes in the network. (Fisher exact test). A list of 596 human proto-oncogenes was compiled from the University of New South Wales (UNSW) Embryology DNA-Tumor Suppressor and Oncogene Database²⁴  and the Tumor Gene Database.²⁵ 

Human CD34⁺ HSCs were purified from umbilical CB pools, BM from patients with ADA-SCID and X-SCID, or PB from a healthy donor. CB CD34⁺ cells were transduced with MLV-derived gamma-RV or HIV-derived LV vectors carrying a GFP reporter gene and either a wild-type or a U3-deleted (SIN) LTR. BM CD34⁺ cells were transduced with MLV-derived RV vectors expressing either ADA²  or γ_c receptor³  from a wild-type LTR. PB CD34⁺ cells were transduced with the vector expressing γ_c receptor.³  Transduction efficiency ranged from 15% (SIN-RV) to more than 90% (SIN-LV) depending on the vector and target cell type, and remained stable throughout the culture period. DNA was obtained 1 to 12 days after infection, from cells that underwent 1 (all BM and PB samples) to 5 to 6 (all CB samples) cell doublings in culture. Vector-genome junctions were cloned and sequenced by a LM-PCR or LAM-PCR approach adapted to the different vector types, and mapped onto the human genome. Cumulatively, we mapped 1030 RV and 849 LV integrations in CB- or BM-derived CD34⁺ cells. A total of 595 RV integrations were obtained from CB cells transduced with wild-type (395) or SIN (200) LTR vectors expressing ΔLNGFR from an internal promoter, and 435 RV integrations were obtained from BM cells transduced with wild-type LTR vectors expressing ADA (190) or γ_c receptor (245). All LV integrations were obtained from CB cells transduced with wild-type (404) or SIN (445) LTR vectors expressing GFP or ΔLNGFR from an internal promoter.

Among RV integrations, 172 (16.7%) were in an intergenic position, 566 (55.0%) were within the transcribed portion of at least 1 gene, and 292 (28.3%) were at a distance of 30 kb or less upstream or downstream of 1 or more genes (Table 1; the complete list of sequences is available at GenBank with the accession number ER916114–ER918350). Among LV integrations, 148 (17.4%) were in an intergenic position, 609 (71.7%) were in an intragenic position, and 92 (10.9%) were in a perigenic position. Conversely, a collection of 798 control sequences randomly cloned by LM-PCR contained 369 (46.2%) intergenic, 308 (38.6%) intragenic, and 121 (15.2%) perigenic sequences. Compared with controls, RV vectors showed a preference for intragenic (2-sample test for equality of proportions with continuity correction; P < .001) and perigenic (P < .001) integration, while LV vectors showed a much higher preference for intragenic positions (P < .001).

The position of the integrated proviruses with respect to known genes is shown in Figure 1, which considers the total number of vector-gene interactions in an interval of 30 kb around each insertion site (1517 and 1241 for RV and LV vectors, respectively). Compared with randomly cloned or computer-generated²⁶  control sequences, a significant clustering around transcription start sites was observed for RV but not LV vectors. Overall, 29.3% of the total RV vector-gene interactions were within 10 kb from the + 1 position of known genes, compared with 16.1% for LV vectors (P < .001; Table 1; Figure 1). The RV general integration preferences were similar in CD34⁺ and HeLa cells, as indicated by the analysis of 869 insertions from a previously published collection²⁷  (Table 1).

In CD34⁺ cells, RV integrations showed a significant preference for gene-dense regions: more than 60% of proviruses were found in genomic regions containing 6 to 20 genes/Mb, with a peak of 35% at a density of 6 to 10 genes/Mb, while more than 60% of control sequences mapped to regions with a gene density of less than 5 genes/Mb (P < .001; Figure 2A). On the contrary, LV integrations followed a distribution within regions of different gene density more similar to that of the control sequences and of the human genome, and different from that of RV (P < .001; Figure 2B).

To correlate vector integration with gene activity, we determined the expression profile of more than 16 000 genes by microarray analysis in CB- and BM-derived CD34⁺ cells activated in culture in the same conditions used for RV and LV transduction. As shown in Figure 3, approximately 60% of 1571 probesets representing 866 genes hit by a RV vector detected a transcript in activated CD34⁺ cells; among them, 13% were classified as lowly abundant, 30% were classified as intermediately abundant, and 17% were classified as highly abundant, compared with a 45% to 47% “present” call on the whole microarrays (percentages were slightly different between CB- and BM-derived cells) and a 11% to 12%, 23%, and 11% to 12% breakdown in the 3 abundance classes. With the exception of the lowest expression class, all differences were statistically significant (P < .001), indicating that RV vectors integrate preferentially into genes active in CD34⁺ cells at the time of transduction, and particularly in the fraction of genes expressed at a higher level. A similar correlation with gene activity was observed, as already reported in T cells,²⁸  for genes hit by LV vectors: approximately 56% of 1346 probesets representing 757 hit genes detected a transcript in activated CD34⁺ cells, with a 13%, 31%, and 12% breakdown in the 3 abundance classes, respectively. Compared with the whole microarray, the fraction of probesets with a present call was significantly higher (56% vs 46%; P < .001), but the difference was accounted for essentially by the intermediately abundant transcripts (31% vs 23%; P < .001; Figure 3), indicating that LV vectors tend to integrate into active genes in CD34⁺ cells but have no specific preference for genes expressed at high levels when compared with RV vectors (P < .001).

A functional classification by the GO criteria²⁹  of genes hit by RV and LV vectors in CD34⁺ cells (Tables S1, S2, available on the Blood website; see the Supplemental Materials link at the top of the online article) showed statistically significant biases toward several gene categories (Figure 4A). In particular, genes involved in the establishment and/or maintenance of chromatin architecture, signal transduction, and cell cycle were significantly more represented in the collection of genes hit by RV integrations compared with their expected frequency in the human genome (EASE score < .005). Genes involved in chromatin remodeling and phosphorylation also were hit at a higher-than-expected frequency by LV vectors (EASE score < .0005 and < 0.005, respectively), particularly those with serine/threonine kinase and GTPase activity (EASE score < .0005). Two additional categories (ie, transcription and apoptosis) were overrepresented in genes hit by RV and/or LV vectors, although at less significant levels (EASE score < .05). A different analysis, carried out by the Ingenuity network-based pathway analysis software, indicated that genes involved in cell signaling, cell growth/proliferation, cell death, cancer, and hematopoietic system development were significantly overrepresented in the collection of RV and/or LV integrations with respect to genes annotated in the Ingenuity Pathways Knowledge Base software (.005 < P < .05). These categories were therefore chosen to carry out a direct frequency comparison between RV and LV target genes and our control gene list (Tables S1Table S2. List of all known genes (UCSC definition) hit by lentiviral (LV) proviruses, inside or at a distance of ≤30 kb upstream or downstream, in human CD34+ HSCs (PDF, 86.4 KB)–S3). Genes involved in cell signaling, growth/proliferation, and death were overrepresented in both RV and LV integrations with respect to control sequences (P < .001; Figure 4B), while genes involved in hematopoietic and immune system development, immune response and cancer were significantly overrepresented only in RV integrations (P < .001; Figure 4B). The comparison was then extended to genes specifically annotated in cancer-related databases (see “Materials and methods, Functional clustering analysis” for definitions and data source). RV integrations hit 77 proto-oncogenes and 64 cancer-associated CISs, corresponding to 7.5% and 6.2%, respectively, of the 1030 integrations (Figure 5). Both categories were significantly overrepresented (P < .001) compared with control sequences (27 proto-oncogenes and 17 CISs out of 798 sequences). On the contrary, LV integrations hit 49 proto-oncogenes and 32 CISs out of 849 integrations (Figure 5), a borderline significant difference compared with controls (P = .03 and .07, respectively). Interestingly, HeLa cell integrations show overrepresentation of proto-oncogenes but not CISs (data not shown). This finding is not surprising considering that CIS have been mostly defined in hematopoietic malignancies.

Overall, these analyses show that both RV and LV vectors have a general tendency to integrate into genes involved in the regulation of cell growth and proliferation, and that RV integrations have a specific bias for genes associated with oncogenic transformation. An Ingenuity network analysis confirmed these biases and showed, in addition, that a significant number of genes hit by RV integrations are functionally linked in gene networks involved in apoptosis (Table S4; Figure 6A), signal transduction, transcriptional regulation, and cancer (Table S4; Figure 6B).

The RV and LV insertion site collections were analyzed for the presence of integrations at recurrent sites (hot spots), using essentially the same criteria previously applied to the definition of cancer-associated CISs (at least 2 independent insertions in less than 30 kb, 3 in less than 50 kb, and 4 in less than 100 kb^11,20 ). Overall, 219 (21.3%) of 1030 RV insertion sites met these criteria, identifying 97 hot spots in the genome of CD34⁺ cells (Table S5). A total of 109 (12.5%) of 869 integrations met the same criteria in HeLa cells, defining 52 hot spots (data not shown). LV vectors showed a significantly lower propensity to integrate at recurrent sites, with only 70 (8.2%) of 849 integrations meeting the definition criteria, and identifying 33 hot spots (Table S5). Comparing the 3 collections, 1 hot spot appeared to be a recurrent site for both RV (4 hits) and LV (3 hits) integration (Chromosome 17q23.2: 55188652–55285672), while 3 hot spots were found in common between CD34⁺ and HeLa cells (data not shown). It is worth noting that 22 (2.8%) of 798 control sequences also met the hot spot definition criteria (Table S5), defining a background level of false positivity in the LM-PCR analysis. The different subgroups of RV integrations contributed to the hot spot list proportionally to their size, with no apparent bias related to the type of transduced cell (CB, BM, or PB), the vector used for transduction (wt-LTR or SIN-LTR), or the number of cell doublings undergone in culture before harvesting (Table S6). In particular, nonexpanded cell populations (BM- and PB-derived), which collectively contributed less than half of the 1030 total RV integrations, contributed with at least 1 integration to 56 (58%) of the 97 RV hot spots (Table S5).

The position of RV hot spot integrations with respects to known genes reflected the RV general integration preferences, with intergenic, perigenic, and gene-dense regions overrepresented to the same extent observed in the entire collection of RV integrations, and clustering around transcription start sites (TSSs) only slightly decreased (P = .015; Table 1; Figures 1A, 2A). On the contrary, LV hot spots showed a higher frequency of integration in intragenic (81.4% vs 71.7%) and gene-dense (65.7% vs 35.6% in the more than 11 genes/Mb range) regions (Table 1; Figure 2B). Similarly, RV hot spots occurred in the same proportion of expressed genes observed for all RV integrations (Figure 3A), while LV hot spots contained a significantly higher proportion of expressed genes (73.2% vs 55.9%; P = .003; Figure 3B).

Interestingly, the maximum distance between independent integrations defining a hot spot was significantly lower for RV vectors compared with LV vectors and control sequences with hot spot characteristics. Overall, 52% and 67% of the RV hot spots in CD34⁺ and HeLa cells span less than 10 kb, including those containing 3 or 4 independent integrations, compared with 36% and 27% for LV and control sequences, respectively (Figure 7). One-fourth (26.0%) of the RV hot spots in CD34⁺ cells and almost one-half (40.4%) of those in HeLa cells contained 2 independent integrations in less than 2 kb, compared with only 3% of the LV hot spots.

The list of RV integration hot spots in CD34⁺ cells includes proto-oncogenes (eg, LYL1, MYB), cancer-associated CISs (eg, FLI1, EVI2A, EVI2B, NF1), and genes involved in chromosomal translocations in hematopoietic malignancies (eg, LMO2, MKL1, ETV6) (Table 2), all of them occurring at frequencies significantly higher than expected (P < .001) and higher than in the overall list of RV integrations (Figure 5). Interestingly, nonexpanded cell populations contributed with at least 1 integration to 9 (53%) of the 17 hot spots containing a proto-oncogene or a cancer-associated CIS (Table 2), again indicating the absence of biases related to the number of cell doublings in culture. On the contrary, LV hot spots showed little enrichment for proto-oncogenes or CISs, although in this case, low numbers make comparisons poorly significant (Figure 5). Furthermore, RV but not LV hot spots included a very high proportion of genes belonging to the intracellular signaling cascade category (25.3%), which were significantly overrepresented using either the human genome or the total RV integrations as a background population in a GO analysis (EASE score, 1.2 × 10⁻⁶ and 2.2 × 10⁻⁴, respectively), despite their relatively small number (ie, 22). Interestingly, genes involved in hematopoietic and immune system development and in immune response by Ingenuity pathway analysis were further and significantly enriched in RV hot spots with respect to the entire list of RV integrations (P < .1; Figure 4B).

RV integration preferences have significant consequences on the potential genotoxicity of different families of vectors used to transfer genes into HSCs. The probability of dominant activation of potentially cancer-causing genes (eg, those involved in the control of stem-cell self-renewal, growth, and differentiation) may in fact differ significantly between RV and LV vectors simply based on the different frequency by which they may target those genes. Here we report a detailed analysis of the RV and LV integration preferences in human CB-, PB- and BM-derived CD34⁺ HSCs transduced in the same conditions used in clinical applications and analyzed without selection. The general integration preferences of the 2 vector families were similar to those previously described for other mammalian hematopoietic or nonhematopoietic cells (reviewed in Bushman et al⁸ ), and showed on average a 2-fold higher probability for RV vectors to target gene-dense regions, highly active genes, and promoter-proximal regions. However, RV but not LV integration occurs at high frequency (> 20%) at genomic locations (hot spots) that are significantly enriched in proto-oncogenes and genes involved in the control of cell proliferation.

A high frequency of hot spots, defined by a statistical criterion previously applied to define cancer-associated CISs,^11,20  appears to be a hallmark of RV integration in human CD34⁺ HSCs. We found that more than one-fifth of the RV integrations meet the definition criteria, a frequency more than 7-fold higher than expected from the analysis of a randomly cloned collection of human DNA sequences, and almost 3-fold higher than that found in a collection of LV integrations of comparable size. The average extension of RV hot spots (ie, the maximum distance between all insertions within each spot) was well within the definition criteria, and significantly smaller than that of LV hot spots, spanning less than 10 kb in half of the cases and less than 2 kb in one-fourth of the cases. RV integration appears therefore to have high preference for restricted genomic locations, which may exhibit specific chromatin conformations or features that favor tethering of the preintegration complexes (PICs) with higher probability. These features do not include gene density, proximity to promoters, or gene expression per se, since hot spots integrations show exactly the same preferences observed in the entire collection of RV integrations. Interestingly, we observed that the frequency of hot spots increased progressively during the study, following the increase of the sample size in an almost linear fashion. This may suggest that by analyzing a much higher number of sequences, all RV integrations could be clustered in a defined subset of genomic regions, all having the appropriate features recognized by the PICs. Unfortunately, the molecular bases of the interactions between RV PICs and the mammalian chromatin are poorly understood, and it is difficult to correlate our finding with any specific mechanism. The situation was completely different in the case of LV hot spots, the frequency of which increased only slightly with the increase in the sample size and appeared to plateau. More importantly, insertions in LV hot spots showed strikingly different characteristics with respect to the general LV integration preferences, and were greatly enriched in gene-dense regions and expressed genes. These data suggest that LV integration may happen in a much wider portion of the HSC genome, and that hot spots are generated at low frequency by locations that are more favorable than others to PIC interaction, and are apparently those with a high density of expressed genes. This explanation is consistent with the available evidence that LV PICs are tethered to the human genome by a widely distributed chromatin component loosely associated with gene activity, such as chromatin-remodeling³⁰  or DNA-repair³¹  complexes, high mobility group (HMG)³²  and polycomb group proteins,³³  and LEDGF.^34,35 

Previous studies carried out in patients⁴  as well as in animal models^12,13,36  have indicated that integrations in cancer-associated CISs and growth-controlling genes are enriched in the progeny of RV-transduced, repopulating HSCs. The major conclusion of these studies was that certain viral insertions lead to clonal selection of stem/progenitor cells in vivo. However, the pretransplantation frequency of these insertion events was never accurately measured in the relevant cell population. Our analysis indicates that a bias toward integration into or around certain category of genes (ie, those involved in signal transduction, cell cycle, chromatin remodeling, and transcription) is already present in nontransplanted hematopoietic progenitors, and particularly in integration hot spots. In particular, proto-oncogenes and cancer-associated CISs are enriched at 3- to 5-fold the expected frequency in RV hot spots, indicating a specific preference for genomic locations containing these categories of genes. These include proto-oncogenes expressed in CD34⁺ hematopoietic progenitors and involved in hematopoietic cell neoplasia, such as LMO2 and EVI2-NF1, targeted at a frequency of approximately 1:350; LYL1 and MYB, targeted at a frequency of approximately 1:500; and others (Table 2). Importantly, there was no significant difference in the number of integrations contributing to oncogene-containing hot spots between nonexpanded (BM- and PB-derived) or moderately expanded (all CB-derived) cell populations, arguing against the likelyhood of clonal outgrowth generated in culture by insertional activation of growth-promoting genes.

A network-based pathway analysis indicates that a significant number of genes targeted by RV integration are functionally linked in transcription-, signal transduction–, apoptosis-, and tumorigenesis-related networks. Interestingly, genes involved in hematopoietic and immune system development are targeted at uniquely high frequency by RV integrations, and further enriched in RV hot spots, suggesting that the gene expression program of a cycling hematopoietic cell is at least in part instrumental in directing RV PICs in certain regions of the genome. Consistent with this hypothesis, almost none of the genes present in CD34⁺ cells hot spots are found in hot spots from HeLa cells, which most likely operate different regulatory networks. Kustikova et al³⁶  reached similar conclusions in compiling their “insertional dominance database” from the progeny of serially transplanted HSCs in mice, although they explain the observed overrepresentation of certain gene categories and functional networks with in vivo selection rather than with intrinsic properties of the RV integration machinery. Indeed, 18% to 34% of the genes present in the mouse database, depending on the stringency of the comparison, are present also in our list, arguing against an exclusive role for in vivo selection in determining most of the frequency biases. A notable exception is the EVI1-MDS1 locus, which we found only once in nontransplanted cells, although it was found at exceedingly high frequencies in vivo in mice,^12,36  nonhuman primates,¹³  and, at least in 1 case, humans.⁴  Insertional activation of the EVI1-MDS1 locus should therefore be considered a factor favoring clonal amplification and/or selection in vivo independently from the frequency by which it is targeted by RV integration before transplantation. It should be noted, however, that our data come from a population of hematopoietic progenitors in which the proportion of repopulating stem cells is admittedly low, leaving the possibility that stem cell–specific hot spots went undetected. Unfortunately, an integration analysis in pretransplantation, long-term repopulating stem cells is currently impossible, and it is therefore difficult to come to definitive conclusions as to what proportion of the biases detected in the stem cell progeny in vivo is due to vector preferences, and what proportion is due to in vivo selection. We favor a predominant role of vector-specific factors, also based on our experience with patients with ADA-SCID in whom pretransplantation and posttransplantation integration preferences showed essentially overlapping patterns (A. Aiuti, B.C., A.R., F.M. et al, manuscript submitted).

In conclusion, this study shows previously unrecognized features of RV and LV integration into human HSCs that may have an impact in assessing the prospective genotoxic risk of using either vector system for human gene therapy applications. In particular, the frequency and characteristics of integration hot spots may be substantial factors in determining a differential safety profile for RV and LV vectors of comparable design and content.

This work was supported by grants from Telethon Italy (GGP06101 and TIGET), the European Commission (VI FP, CONSERT), and Fondazione Cariplo.

Contribution: C.C., A.R., and F.M. designed the research and wrote the paper; C.C, G.F., D.S., A. Antonelli, A.M., S.H., and B.C. performed research and analyzed data; M.S., C.v.K., A.T., A. Aiuti, and G.F. contributed vital reagents and data sets.

Conflict-of-interest disclosure: The authors declare no competing financial interests.

Correspondence: Fulvio Mavilio, Department of Biomedical Sciences, University of Modena and Reggio Emilia, Via Campi 287, 41100 Modena, Italy, Phone: +39-059-2055392, Fax: +39-059-2055410; e-mail: fulvio.mavilio@unimore.it.