The gammaretroviral vector, SFG.iCasp9.2A.ΔCD19, has been previously described.17 It encodes a safety switch, inducible caspase 9 (iCasp9), and truncated CD19 (ΔCD19), which enables the detection of transduced cells by flow cytometry. The vector was pseudotyped with Gibbon ape leukemia virus envelope.17 Patient samples and generation of transduced Jurkat cell clones are described in the online supplementary information.

Genomic DNA was extracted using Purelink genomic DNA mini kit (Invitrogen, Waltham, MA) according to the manufacturer’s instructions. Genomic DNA was digested with two 6-cutter restriction enzymes which generate compatible cohesive ends (underlined): NcoI, which cuts at C/CATGG and BspHI, which cuts at T/CATGA (both from New England BioLabs, Ipswich, MA). NcoI has three cut sites and BspHI has one cut site within the transgene, with the most distal cut site generated by NcoI at 1185 base pair (bp) from the junction between the vector insert and the flanking genomic DNA (figure 1A). Note that these restriction enzymes do not have any cut sites within the 5′ and 3′ long terminal repeats (LTR), which flank the transgene and are identical in sequence and orientation. This assay design avoids restriction enzyme digestion within the 5′LTR, which can result in the downstream amplification of the internal transgene sequence, producing non-informative reads. The resulting DNA fragments were circularized for inverse PCR or ligated to linker cassettes for downstream PCR amplification.

Inverse PCR was performed as previously described.15 16 In brief, DNA fragments were circularized in a dilute mixture of 1 ng/µL DNA in T4 DNA ligase buffer with one cohesive end unit/μL T4 DNA ligase (New England Biolabs) for 16 hours at 16°C. The ligation product was purified by ethanol precipitation. The first PCR reaction was 35 cycles, using a forward primer that was complementary to the distal 3′LTR and directed towards the flanking genomic DNA; and a reverse primer complementary to the junction between the distal transgene and 3′LTR (figure 1B). The PCR mixture was as follows: 50–400 ng circularized DNA template, forward and reverse primers (100 nM each), dNTP (250 nM), 1.25 Unit HotStart Taq DNA polymerase (Qiagen, Hilden, Germany), 1X PCR buffer, and 1X Q-Solution (Qiagen), in a final reaction volume of 50 µL. A 0.5 µL aliquot of the first PCR product was amplified in a nested PCR reaction for 35 cycles, and tailing sequences were added for barcoding and nanopore sequencing. All PCR reactions were performed on a Bio-Rad T100 thermal cycler (Bio-Rad Laboratories, Hercules, CA). Primer sequences and thermocycling conditions are listed in online supplementary table S1.

Cassette ligation PCR was performed according to the schema in figure 1C. The restriction enzymes generated 5′ overhangs. A single dideoxy-CTP (ddCTP) was filled in to prevent elongation of the recessed 3′ ends using the following reaction mixture: NcoI/BspHI digested DNA 16.7 ng/µL, ddCTP 33 µM (GE Healthcare, Chicago, IL), DNA polymerase I, large (Klenow) fragment 0.167 U/µL (New England BioLabs) in CutSmart buffer (New England BioLabs) in a final reaction volume of 15 µL. The mixture was incubated at 25°C for 30 min, heat inactivated at 75°C for 20 min, and purified by ethanol precipitation.

Linker cassettes were made by annealing two single-stranded DNAs: 5′-ATGTCCCATGGTCA-3′ and 5′-CATAGTTGTTCCACTCCGACCATGGGA-3′ (both at 20 µM) in Tris-HCl (50 mM) with MgCl₂ (5 mM). The mixture was heated to 95°C for 5 min, then gradually cooled to room temperature by turning off the heating block. The linker cassettes were aliquoted, stored at –20°C, and used without repeated freeze/thaw. The ddCTP-filled in DNA fragments were ligated to the linker cassettes using the following reaction mixture: DNA 5 ng/µL, linker cassette 500 nM, and T7 DNA ligase 30 U/µL in T7 ligase buffer (New England BioLabs). The mixture was incubated at 25°C for 30 min, heat inactivated at 65°C for 10 min, and purified by ethanol precipitation. Because the 3′ recessed end of the DNA fragments had a ddCTP, this reaction resulted in ligation of the longer linker cassette strand and a nick in the shorter linker cassette strand. The first PCR reaction was 30 cycles in 50 µL: the forward primer was complementary to the junction between the transgene and the 3′LTR to minimize non-informative reads from 5′LTR priming; and the reverse primer was identical to the longer linker cassette strand. The unligated linker cassette strand was short (14 nucleotides) relative to the ligated cassette strand (27 nucleotides), which reduced unwanted PCR priming of non-flanking genomic fragments. A 0.5 µL aliquot of the first PCR product was amplified 30 cycles in a nested PCR reaction (50 µL) which included tailing sequences for barcoding and nanopore sequencing.

Fragment size estimation was performed by agarose gel electrophoresis and imaged on Vilber Lourmat Ebox CX5 (Vilber Lourmat). The nested PCR products (8 µL) were run on 1% agarose gel. To check for specificity of the amplicons, the PCR products were digested with SmaI, which has cut sites within the 3′LTR and distal transgene, generating DNA fragments of 244 bp and 391 bp from inverse PCR amplicons and fragments of 244 bp from cassette ligation PCR amplicons (figure 2).

Sequencing libraries were prepared according to 1D PCR barcoding amplicons protocol from Oxford Nanopore Technologies (ONT, Oxford UK). First, PCR amplicons from inverse PCR and cassette ligation PCR were barcoded using PCR barcodes provided by the supplier. Barcoding PCR mixture was as follows: 24 µL of nested PCR product (0.5 nM), 25 µL of LongAmp Taq master mix (New England Biolabs), and 1 µL of PCR barcode (ONT). Reactions were amplified in a thermal cycler with the following conditions: 95°C for 3 min, 15 cycles of amplification at 95°C for 15 s, 62°C for 15 s, and 65°C for 4 min, followed by final extension of 65°C for 1 min. Barcoded PCR products were purified using 0.8 X Agencourt AMPure XP beads (Beckman Coulter, Brea, CA). Concentrations of purified PCR products were measured using Qubit High Sensitivity Kit (Thermo Fisher, Waltham MA). Based on the concentrations, PCR products were pooled using 10 times more polyclonal clinical samples as compared with oligoclonal samples to increase sequencing coverage. Sequencing libraries were prepared from pooled barcoded PCR products according to manufacturer’s instruction. Briefly, end repair was performed in the following mixture: 48 µL of pooled PCR product, 3.5 µL of NEBNext FFPE DNA repair buffer, 2 µL of NEBNext FFPE DNA repair mix, 3.5 µL of Ultra II End-prep reaction buffer, and 3 µL of Ultra II End-prep enzyme mix (all from New England Biolabs); the mixture was incubated at 20°C for 15 min and 65°C for 15 min. End repaired products were purified with 1X Ampure XP beads (Agencourt) and eluted in 60 µL of nuclease-free water. Purified end-repaired products were ligated with 25 µL Ligation Buffer (ONT), 10 µL NEBNext Quick T4 DNA Ligase (New England Biolabs), and 5 µL Adapter Mix (ONT) and incubated at room temperature for 30 min. Ligated products were purified with 0.4X Ampure XP beads (Agencourt) and Short Fragment Buffer (ONT) and final sequencing library was eluted in Elution Buffer (ONT). Sequencing library was loaded into a single ONT PromethION flowcell and sequenced for 40 hours.

A flow diagram for sequence data analysis is outlined in figure 3. Analysis parameters and custom scripts are provided in online supplementary information. Sequencing reads were base-called and demultiplexed using Guppy basecaller V.2.3.7 (ONT). Reads were classified based on the read quality score as pass (≥7) or fail (<7) by the basecaller. Adapters and sample barcodes were trimmed using Porechop V.0.2.4 (https://github.com/rrwick/Porechop). Reads that were shorter than the combined lengths of the predicted flanking sequences of the amplicons (885 bp for inverse PCR and 362 bp for cassette ligation PCR) were excluded from analysis. For inverse PCR, filtered reads were aligned to hg38 genome and both flanking sequences (ie, 3′LTR and distal transgene sequence), with masking of the distal transgene sequence using BWA mem (V.0.7.15).18 For cassette ligation PCR, the flanking cassette sequence was short and difficult to align; hence, this was first trimmed with Porechop and the trimmed reads were then aligned to hg38 genome and flanking 3′LTR with BWA mem (V.0.7.15).18 Note that as a result of cassette trimming, the final reads for cassette ligation PCR were 27 bp shorter than the original filtered reads. To determine which of the two ends represented the 3′LTR-genome junction, we developed a tool called FlankDetect (https://github.com/mdcao/japsa) (V.1.9-10b) (online supplementary information), which identifies reads which contain the flanking sequences and reports the integration site that flanks the junctional 3′LTR sequence. This tool also assigns clonality by clustering reads based on the flanking integration sites. Reads which have integration sites within 10 bp of each other are clustered together. During clustering, alignments that were not primary alignment and less than mapping quality of 20 were excluded from further analysis. To eliminate false clusters, Bedtools (V.2.26.0) merge option was used to merge any overlapping clusters which contain reads that align to the same region. The integration site that is most frequently observed within all the reads that are merged together is retained as the integration site of the clone.

Genomic locations of transcription start site (TSS), exonic regions, intronic regions, and intergenic regions were extracted from GENCODE V.28 gene annotation file using custom scripts (online supplementary information). Genomic annotation of integration site was performed using Bedtools (V.2.26.0) intersect option. Genomic distance of the integration site from TSS was calculated using custom scripts (online supplementary information).

The amplicons from inverse PCR and cassette ligation PCR were predicted to have a sandwich structure: the flanking genomic sequence in the middle, which would be contiguous with the 3′LTR on one end, and the sticky-end ligated distal transgene sequence or linker cassette sequence on the other end (figure 2A). The theoretical minimum lengths of productive amplicons, including the tailing sequences, were 929 bp for inverse PCR and 406 bp for cassette ligation PCR. We performed flanking sequence amplification on non-transduced peripheral blood mononuclear cells (PBMC), and SFG.iCasp9.2A.ΔCD19-transduced Jurkat cell clones and polyclonal clinical samples from our previously published phase I clinical trial.15 There was no significant amplification of negative control non-transduced PBMC by inverse PCR (figure 2A., lane 0A) but there was some amplification by cassette ligation PCR (figure 2A, lane 0B). As expected, inverse PCR and cassette ligation PCR on transgenic Jurkat cell clones (clones 1 and 2) resulted in single PCR fragments (figure 2B: lanes 1A, 1B, 2A, 2B); and a 1:1 mixture of the two Jurkat cell clones resulted in a combination of the two bands (lanes 3A and 3B). Amplification of a polyclonal SFG.iCasp9.2A.ΔCD19-transduced clinical cell product yielded a polyclonal smear (lanes 4A and 4B), as did amplification of patient PBMC collected at Day 369 after T cell infusion (lanes 5A and 5B). However, amplification of patient PBMC from Day 1332 after T cell infusion showed a dominant band with inverse PCR and a polyclonal smear with cassette ligation PCR (lanes 6A and 6B), suggesting a possibility of bias in DNA fragment circularization or PCR amplification with the former. The inverse PCR amplicons in lanes 4A and 5A also contained bands that were shorter than the minimum theoretical length of 929 bp. These were later shown on nanopore sequencing to contain 3′LTR and transgene sequences without any intervening genomic DNA, which was consistent with the circularization of degraded DNA fragments near the 3′LTR.

A combination of two 6-cutters (NcoI and BspHI) is expected to cut at approximately every 2048 bp (ie, 4⁶/2), generating an estimated 1.5×10⁶ DNA fragments per diploid human cell, which has a genome of approximately 3×10⁹ bp. Given that a majority of transgenic T cells carry only a small number of transgene inserts per cell, this would mean that only a tiny fraction (<0.005%) of the DNA fragments would include a transgene flanking sequence. Non-specific amplification of non-flanking sequences is therefore a challenge. As shown in figure 2A, specific PCR amplicons following both inverse PCR and cassette ligation PCR should contain a SmaI restriction site within the 3′LTR flanking sequence, with inverse PCR amplicons having an additional SmaI restriction site at the transgene end. Thus, SmaI restriction enzyme digestion of specific PCR amplicons is anticipated to yield a 244 bp DNA fragment from the 3′LTR end and, in the case of inverse PCR, an additional 391 bp DNA fragment from the transgene end, along with flanking genomic DNA fragments of variable lengths. We digested the PCR products with SmaI and confirmed that the amplicons from SFG.iCasp9.2A.ΔCD19-transduced samples were specific, with the generation of 244 bp and 391 bp fragments from inverse PCR and a 244 bp fragment from cassette ligation PCR (figure 2C). In contrast, SmaI digestion of PCR products from negative control non-transduced genomic DNA did not result in any specific 244 bp or 391 bp fragments (figure 2C, lanes 0A and 0B).

All 12 paired PCR products from the six samples were pooled into one nanopore sequencing run. In order to ensure adequate read-depth, the amount of input DNA was limited to the equivalent of less than 10,000 transduced cells per sample (see online supplementary table S2). Because the Jurkat samples were clonal and hence required fewer total reads, they were pooled with the polyclonal patient samples at a ratio of 1:10. The sequencing run produced 23.4 million reads in total and 59% of reads were classified as passing quality score (read quality ≥7). Approximately 6% of the pass reads could not be demultiplexed and were excluded. In most of the samples, >90% of the resulting reads were longer than the minimum theoretical length of 885 bp for inverse PCR and 362 bp for cassette ligation PCR after adapter trimming. A vast majority of the length-filtered reads were specific, with 95% (range 89%–98%) having both expected flanking sequences: the junctional 3′LTR flank on one end and the cassette or transgene sequence on the other end. Table 1 summarizes the number of reads retained after each data analysis step.

Greater than 94% of reads containing the expected flanking sequences from the transduced Jurkat cell lines (samples 1, 2, and 3) could be uniquely aligned to the genome (table 1). The genome alignment rate was lower for the polyclonal clinical samples, with 47%–74% (median 69%) of reads with the expected flanking sequences being successfully aligned. A majority of reads that could not be aligned to genomic DNA were only marginally longer than the minimum theoretical lengths of the PCR amplicons and were therefore too short to be uniquely aligned (figure 4A). Long unaligned reads generally lacked the anticipated flanking 3′LTR/transgene or cassette sequences and therefore represented non-specific amplicons. A significant proportion of inverse PCR amplicons in samples 4 and 5 were below the minimum theoretical lengths. Alignment of three dominant clones (560 bp, 388 bp, and 318 bp) showed that these consisted of the terminal regions of the flanking 3′LTR and transgene sequences, without any intervening genomic DNA (figure 4B), in an orientation that was consistent with the circularization and PCR amplification of degraded DNA fragments within the transgene/3′LTR region (figure 4C).

Aligned reads with 3′LTR-genome junctions that were within 10 bp of each other were clustered together using our customized tool, FlankDetect. The 4,895,820 reads were grouped into 12,186 clusters. Each cluster consisted of 1 to >10⁶ reads: 59% of clusters had single reads, 24% of clusters had 2–10 reads, 7% of clusters had 11–100 reads, and the remaining 10% of clusters had >100 reads (figure 4D). For clusters with two to four reads, the 3′LTR -genome junction was identical in 24% of the clusters. The junction varied by one nucleotide in 28% of clusters, and by two to three nucleotides in 18% of clusters (figure 4E). Overall, for clusters with two to four reads, the spread of 3′LTR -genome junctions was within five nucleotides in 88% of the clusters, but a few clusters had a spread of 10 to 15 nucleotides. For clusters with ≥5 reads, we defined R80 as the span of 3′LTR -genome junction that contained at least 80% of reads within the cluster: 15% of clusters had R80 of 0 nucleotides, meaning 80% of the reads had identical 3′LTR-genome junction; and 80% of clusters had R80 of 1 nucleotide (figure 4F). Overall, for clusters with five or more reads, 97% had an R80 of ≤5 nucleotide, meaning 80% of reads within the clusters were within a span of five nucleotides, although there were isolated clusters with a spread of up to 40 nucleotides.

We observed that a proportion of the clusters were very proximate to each other. In some cases, the read alignments were very closely matched but the clusters were considered separate because the 3′LTR-genome junction was assigned to opposite ends of the read, which would suggest an underlying alignment error. In order to avoid splitting clonal integration sites into multiple artificial clones, clusters with overlapping read alignments were merged and considered as belonging to the same clone. Using this merging algorithm, we identified 6410 unique vector integration clones: 4440 clones (69%) consisted of only one cluster, 697 clones (11%) consisted of two clusters, 383 clones (6%) consisted of four clusters, and the remaining 890 clones (14%) consisted of four to 18 clusters (figure 4G).

Samples 1 and 2 each contained a unique vector integration site, which were detected by both inverse PCR and cassette ligation PCR. Note that both samples had a degree of subclonal contamination, which was more prominent on cassette ligation PCR than inverse PCR. Sample 3 consisted of an equal mixture of samples 1 and 2 but there was a dominance of the integration site from sample 2, with the bias being more pronounced by inverse PCR (figure 5A). The clinical samples were polyclonal: the cell product (sample 4) was anticipated to be the most clonally diverse, with some degree of clonal dominance emerging in the postinfusion patient samples (samples 5 and 6).15 The number of unique vector integration sites identified in samples 4, 5, and 6 were 979, 196, and 602 by inverse PCR and 2258, 1251, and 669 by cassette ligation PCR, with the starting materials containing approximately 7500, 4700, and 2500 iCasp9-transduced cells. In all three samples, clonal skewing was more pronounced with inverse PCR than cassette ligation PCR: the top 10 integration sites in samples 4, 5, and 6 accounted for 49%, 93%, and >99%, respectively, of the total aligned reads by inverse PCR; as compared with 7%, 58%, and 76% of the total aligned reads by cassette ligation PCR (figure 5B). A proportion of the integration sites were detected by both techniques but their percentage representation within the sample differed between the two techniques, and a majority of integration sites were mapped by only one of the techniques (figure 5C). The very large clone representing 97% of reads in sample 6 by inverse PCR, represented only 0.1% of the reads by cassette ligation PCR. This clone also had higher clonal representation by inverse PCR relative to cassette ligation PCR in samples 4 and 5 (0.08% vs 0.02%; and 0.44% vs 0.15%), suggesting amplification bias.

Vector integration sites were identified within all chromosomes, with higher representations within the larger chromosomes, as these represented a larger component of the total genome (figure 6A). A majority (61%–68%) of integration sites were intragenic, with 55%–58% within the introns, and 6%–12% within the exons; with the remaining 32%–39% of integration sites being intergenic (figure 6B). There was a predilection for vector integration near TSS (figure 6C), which was consistent with other reports.3 5 6 There was no predilection for vector integration near LTR73 and LTR76, which are human endogenous retrovirus sequences with some sequence similarity to the SFG vector LTR (online supplementary figure S1).