Reply: Evidence that APP gene copy number changes reflect recombinant vector contamination

Abstract

In the accompanying comment¹, Kim et al. concluded that somatic gene recombination (SGR) and amyloid precursor protein (APP) genomic complementary DNAs (gencDNAs) in brain are contamination artifacts and do not naturally exist. We disagree. Here we address the three types of analyses used by Kim et al. to reach their conclusions: informatic contaminant identification, plasmid PCR, and single-cell sequencing. Additionally, Kim et al. requested “reads supporting novel APP insertion breakpoints,” and we now provide 10 different examples that support APP gencDNA insertion within eight chromosomes beyond wildtype APP on chromosome 21 from Alzheimer’s disease (AD) samples. If SGR exists as experimentally supported here and previously^2,3, contamination scenarios become moot. Our informatic analyses of data generated by an independent laboratory (Park et al.)⁴, complement and are entirely consistent with what Lee et al.² presented via nine distinct lines of evidence, in addition to three from a prior publication³. Plasmid contamination was identified in a single pull-down dataset after publication of Lee et al.²; however subsequent analyses did not alter any of our conclusions including those of our prior publications^3,5 and plasmid contamination-free replication of this approach by ourselves and others supported the original conclusions. Novel retro-insertion sites, alterations of APP gencDNA number and form with cell type from the same brain and pathogenic SNVs occurring only in AD, all support the existence of APP gencDNAs produced by SGR.

Identification of novel APP gencDNA insertion sites

One predicted outcome of SGR is the generation of novel retro-insertion sites distinct from the wildtype locus, as we demonstrated using DNA in situ hybridization (DISH; Lee et al. Figure 2n). Analyses of independently published datasets (Park et al.)⁴ produced by whole-exome pull-down of DNA from laser-captured hippocampus or blood revealed 10 different APP insertion sites within eight different chromosomes (Figure 1, Supplementary Table 1). We identified clipped reads spanning APP UTRs and novel genomic insertion sites on chromosomes 1, 3, 9, 10, and 12 (Figure 1a; wildtype APP is located on chromosome 21). The corresponding paired-end reads mapped to the same inserted chromosome. We also identified reads spanning APP exon::exon junctions of gencDNAs that had mate-reads mapping to other genomic sites on chromosomes 1, 3, 5, 6, and 13 (Figure 1b). We are unaware of contamination sources capable of producing these results that are entirely consistent with our DISH data showing APP gencDNA locations distinct from wildtype APP. These novel APP gencDNA insertion sites strongly support the natural occurrence of APP gencDNAs.

Figure 1. Identification of novel APP insertion sites in the human genome.

a) Clipped reads spanning APP UTRs and novel chromosomal insertion sites were identified. The paired matereads of the clipped reads (black stripes) uniquely mapped to the same chromosomes. b) Discordant read-pairs were identified where one read spanned an APP exon::exon junction and the corresponding mate-read mapped to a novel chromosome. Each chromosome has a unique color. Arrowhead direction represents the read orientation after mapping to the human reference genome. Arrows oriented in the same direction support sequence inversions. See detailed sequence and alignment information in Supplementary Table 1.

Figure 2. Identification of APP gencDNA sequences in 10 new whole-exome pull-down datasets from two independent laboratories.

a) Method schematic depicting the standard protocol for whole-exome pull-downs and highlighted methodological differences between the independent laboratories are presented. b) APP-751 sequence with non-duplicate gencDNA reads from the 10 new datasets; color key indicates the source reads for all panels. c) Reads mapping to junctions between APP exons 7, 8, and 9 that are absent from APP-751. d,e) Paired reads that represent a DNA fragment containing both an exon::exon junction and an APP 3’ or 5’ UTR.

An APP plasmid contaminant (pGEM-T Easy APP) was found in our single pull-down dataset, however we could not definitively determine which APP exon::exon reads were due to gencDNAs vs. plasmid contamination, especially in view of the 11 other distinct and uncontaminated approaches that had independently supported and/or identified APP gencDNAs. Three other pull-down datasets from our laboratory were informatically analyzed and found to contain APP gencDNA reads while being free from APP plasmid contamination by both VecScreen⁶ and subsequent use of Kim et al.’s Vecuum script⁷ (Figure 2a,b). Possible external source contamination noted by Kim et al. in two of three datasets could not definitively account for all APP exon::exon junctions.

The recent availability of independently generated AD datasets⁴ provided a test for the reproducibility of APP gencDNA identification. Five different sporadic AD (SAD) brains and two AD blood samples contained APP gencDNA sequences and were plasmid-free by Vecuum⁷ screening (Figure 2a-e). In addition to exon::exon junction reads and novel insertion sites, we also identified APP UTR sequences paired with reads containing APP gencDNA exon::exon junctions (Figure 2d,e). This may be explained by a key experimental design factor: Park et al.’s pull-down probes contain sequences corresponding to APP 5’ and 3’ UTRs.

In addition to APP plasmid and amplicon contaminants, Kim et al. invoked genome-wide mouse and human mRNA contamination in the Park et al. dataset. We cannot address conditions in the Park et al. laboratory but note that it is completely independent of our own. Kim et al.’s explanation implicates the generation of DNA from mRNA: a process that requires reverse transcriptase activity. The Agilent SureSelect pull-down employed by Park et al. and in our experiments do not use reverse transcriptase (Figure 2a and Supplementary Methods), and we are unaware of any mechanism that would generate DNA from RNA in the absence of reverse transcriptase activity under the employed conditions. An alternative explanation is the existence of gencDNAs affecting other genes as we previously detected in non-APP intra-exonic junctions (IEJs) found in commercial cDNA Iso-Seq datasets (Extended Data Figure 1). Additional validation would be required for new genes, however we note that an average of 450 megabase-pairs of extra DNA exist within AD neurons³ that could accommodate new gencDNA sequences. Kim et al. further invoked genome-wide mouse and human mRNA contamination in the Park et al. dataset to account for APP gencDNAs, an explanation conflicting with available data. Mouse-specific single nucleotide polymorphisms (SNPs) in the Park et al. dataset cannot account for all APP gencDNA-supporting reads: five of seven APP exon::exon junction sequences do not contain putative mouse-specific SNPs at the specific region reported by Kim et al. (Figure 3; Kim et al. Figure 2d). Most critically, novel APP gencDNA insertion sites identified here cannot be explained by genome-wide mRNA contamination.

Extended Data Figure 1. IEJs identified from commercially available long-read transcriptome datasets in two genes other than APP.

Sequences containing IEJs were identified and shown for a) Gene 1 and b) Gene 2. Gene 2 is shown in two parts. Grey dashed lines show ends of RefSeq exons; solid purple lines denote IEJs. All splice isoforms were examined. The Alzheimer brain Iso-Seq dataset was generated by Pacific Biosciences, Menlo Park, California, and additional information about the sequencing and analysis is provided at https://downloads.pacbcloud.com/public/dataset/Alzheimer_IsoSeq_2016/.

Figure 3. Five APP gencDNA-supporting reads spanning exon::exon junctions that do not contain mouse-specific SNPs.

APP gencDNA reads were identified that span the APP exon10::exon11 junction from the Park et al. datasets. The reference sequences of human and mouse exons are indicated and the positions where the nucleotides differ are highlighted. Five of the seven exon::exon junction-spanning reads do not contain mouse-specific SNPs.

Non-biological data are generated by PCR of APP plasmids in Kim et al

Kim et al. used PCR of APP splice variant plasmids which generated sequences containing IEJs. However, multiple discrepancies in this approach and results differ from our biological IEJs and gencDNAs: 1) experimental conditions beyond our primer sequences were different: Kim et al. employed twice the concentration of primers and >1 million times more template (250 picograms of APP plasmid is 4.6 x 10⁷ copies vs. ~40 gencDNA copies in our PCR of 20 nuclei (based on Lee et al. Figure 5²: DISH 16/17 averaged ~1.8 copies/SAD nucleus)); 2) both gencDNA and IEJ sequences can be detected with as few as 30 cycles of PCR as we used in single molecule real-time (SMRT) sequencing (Lee et al. Figure 3)² vs. 40 cycles used by Kim et al.; 3) agarose gels in Kim et al. are uniformly and unambiguously dominated by a vastly over-amplified ~2 kb band (Kim et al. Figure 1c and Extended Data Figure 3a) that is never seen in human neurons despite our routine identification of myriad smaller bands (c.f., Lee et al. Figure 2b)². We did observe an over-amplified ~2 kb band in our purposeful plasmid transfection experiments that also utilized PCR; however, gencDNA and IEJ formation was comparatively limited, and critically, required both reverse transcriptase activity and DNA strand breakage (Lee et al., Figure 4²); and 4) only 45 unique IEJs from AD and 20 from non-diseased brains were identified (Lee et al. Figure 3 with some overlap, fewer than 65 total)² compared to the 12,426 identified by Kim et al. (~200-fold increase over biological IEJs; Kim et al. Supplementary Table 1). We wish to note that microhomology regions within APP exons are intrinsic to APP’s DNA sequence and that microhomology mediated repair mechanisms involve DNA polymerases^8,9. Kim et al.’s PCR results differ from our biological data yet may inadvertently support endogenous formation of at least some IEJs within DNA rather than requiring RNA.

Detection of IEJs without use of APP PCR

Despite these differences between the non-biological plasmid PCR data generated by Kim et al. and our data, Kim et al. concludes that IEJs from our original study² might have originated from contaminants. To eliminate this possibility, Lee et al.² presented four lines of evidence for APP gencDNAs containing IEJs that are independent of APP PCR: two different commercially produced cDNA SMRT sequencing libraries, DISH, and RNA in situ hybridization (RISH). The SMRT sequencing libraries revealed IEJs within APP (Lee et al. Extended Data Figure 1E)² as well as other genes (Extended Data Figure 1), which cannot be attributed to plasmid contamination or PCR amplification. DISH and RISH results support the existence of APP gencDNAs and IEJs (see Supplementary Discussion and Lee et al., Figure 2, Extended Data Figures 1 and 2)² by using custom-designed and validated commercial probe technology (Advanced Cell Diagnostics, ACD), which was independently shown to detect exon::exon junctions¹⁰ and single nucleotide mutations¹¹. Thus, gencDNAs and IEJs are detectable in the absence of targeted PCR. Importantly, the contamination proposed by Kim et al. cannot account for the dramatic change in the number and forms of APP gencDNAs occurring with disease state. The change is also apparent when comparing cell types, where signals are vastly more prevalent in SAD neurons compared to non-neurons from the same brain and processed at the same time by DISH (Lee et al. Figures 5)². Independent PNA-FISH and dual-point-paint experiments from our previous work further support APP gencDNAs³ (Table 1). Critically, SMRT sequencing identified 11 single nucleotide variations that are considered pathogenic in familial AD, which were only present in our SAD samples, none of which exist as plasmids in our laboratory.

Table 1.

Summary of targeted and non-targeted APP PCR methods and lines of evidence that support APP gencDNAs and IEJs.

Kim et al. compared APP gencDNA copy number estimates from pull-down sequencing and DISH. However, a direct comparison is not possible since the two methodologies are fundamentally different. For example, pull-downs employ solution hybridization on isolated DNA, while DISH uses solid-phase hybridization on fixed and sorted single nuclei. Moreover, the sequences targeted between the two are not the same. Pull-down probes target wildtype sequences for endogenous and gencDNA loci, resulting in pull-down competition. By contrast, DISH probes target only gencDNA sequences to provide greater sensitivity. Competition by wildtype loci reduced the efficiency of capture, which is underscored by 32% to 40% of nuclei that do not contain gencDNAs and would contribute only wildtype sequences (Lee et al., Figure 5c,f). Moreover, a majority of gencDNA positive nuclei (62% to 73%) showed two or fewer signals (Lee et al., Figure 5c,f) which reduced the relative representation of gencDNA loci. Since IEJs do not contain the full exon sequence, there is inefficient hybridization and a lack of sequence capture and detection. This limitation is overcome by SMRT sequencing (Extended Data Figure 1 and Lee et al., Extended Data Figure 1e). Lastly, multiple other protocol variations exist which explain the hypothesized discrepancies including tissue preparation, fixation, and hybridization conditions.

Single-cell whole-genome sequencing limitations may prevent APP gencDNA detection

Kim et al.’s third type of analysis yielded a negative result via interrogation of their own singlecell whole-genome sequencing (scWGS) data, which cannot disprove the existence of APP gencDNAs. An average of nine neurons from seven SAD brains were examined, raising immediate sampling issues required to detect mosaic APP gencDNAs. Kim et al. identified “uneven genome amplification” (Kim et al. and ^12–14) producing ~20% of the single-cell genome having less than 10X depth of coverage¹⁴ with potential amplification failure at one (~9% allelic dropout rate) or both alleles (~2.3% locus dropout rate)^12,14. These limitations are compounded by potential amplification biases reflected by whole-genome amplification failure rates that may miss neuronal subtypes and/or disease states, which is especially relevant to single copies of APP gencDNAs that are as small as ~0.15 kb (but still detectable by DISH). Kim et al. state that the increased exonic read depth relative to introns reliably detects germline retrogene insertions in single cells from affected individuals (Kim et al., Figure 3b); however, these data also demonstrate that increased exonic read depth is not observed in all cells – or even a majority in some cases – from the same individuals carrying the germline insertions of SKA3 (AD3 and AD4) and ZNF100 (AD2). These results demonstrate inherent technical limitations in Kim et al. that prevent accurate detection of even germline pseudogenes present in all cells, thus explaining an inability to detect the rarer mosaic gencDNAs produced by SGR. Kim et al.’s informatic analysis is also based on the unproven assumption that gencDNA structural features are shared with processed pseudogenes and LINE1 elements (Kim et al. Figure 3a and Extended Data Figure 1a), and possible differences could prevent straightforward detection under even ideal conditions as has been documented for LINE1¹⁵. These issues could explain Kim et al.’s negative results.

Considering these points, we believe that our data and conclusions supporting SGR and APP gencDNAs remain intact and warrant their continued study in the normal and diseased brain.

Author Contributions

MHL, YK, and WR conducted laboratory experiments; CSL and YZ analyzed sequencing data; and JC conceived and oversaw the experiments. All authors wrote and edited the manuscript. This Reply was the work of current laboratory members.

Competing interests

Sanford Burnham Prebys Medical Discovery Institute has filed the following patent applications on the subject matter of this publication: (1) PCT application number PCT/US2018/030520 entitled, “Methods of diagnosing and treating Alzheimer’s disease” filed 1 May 2018, which claims priority to US provisional application 62/500,270 filed 2 May 2017; and (2) US provisional application number 62/687,428 entitled, “Anti-retroviral therapies and reverse transcriptase inhibitors for treatment of Alzheimer’s disease” filed 20 June 2018. JC is a cofounder of Mosaic Pharmaceuticals.

Data availability

Data from Park et al. were deposited in the National Center for Biotechnology Information Sequence Read Archive database, accession number PRJNA532465. Data from the newly reported full exome pull-down datasets will be provided for the APP locus upon request.

Code availability

The source codes of the customized algorithms are available on GitHub https://github.com/christine-liu/exonjunction.

Acknowledgements

We thank Dr. Laura Wolszon and Ms. Danielle Jones for manuscript editing. Research reported in this publication was supported by the NIA of the National Institutes of Health under award number R56AG067489 and P50AG005131 (J.C.); and NINDS R01NS103940 (Y.K.). This work was supported by non-Federal funds from The Shaffer Family Foundation, The Bruce Ford & Anne Smith Bundy Foundation, and Sanford Burnham Prebys Medical Discovery Institute funds (J.C.). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.

Footnotes

• * Co-first authors