An intronic VNTR affects splicing of ABCA7 and increases risk of Alzheimer’s disease

Participants were part of a large prospective cohort of Belgian AD patients and healthy elderly control individuals. Patients were ascertained at the Memory Clinic of Middelheim and Hoge Beuken (Hospital Network Antwerp, Belgium), and at the Memory Clinic of the University Hospitals of Leuven, Belgium. Possible, probable, and definite AD diagnosis was based on NINCDS-ADRA [28] and/or NIA-AA [18, 29] criteria. Control individuals were recruited from partners of patients, or were volunteers from the Belgian community. All control individuals scored > 25 on the Montreal Cognitive Assessment (MoCA) test [32], and were negative for subjective memory complaints, neurological or psychiatric antecedents, and family history of neurodegeneration. All participants and/or their legal guardian provided written informed consent before inclusion. The study protocols were approved by the ethics committees of the Antwerp University Hospital and the participating neurological centers at the different hospitals of the BELNEU consortium and by the University of Antwerp. Table S1 provides an overview of individuals for whom DNA, Epstein-Barr virus transformed lymphoblastoid cell lines (LCL), or CSF were included.

Genomic start and end coordinates of the ABCA7 VNTR were delineated with the Tandem Repeat Finder algorithm [6], as implemented in the Tandem Repeats Database (TRDB) [14]. To further assess whether the VNTR length could contribute to AD risk, we used the genotyping methods described below.

We analyzed the depth of NGS sequencing reads aligning to the ABCA7 VNTR as a proxy for VNTR length. BAM files from 772 AD patients, and 757 matched healthy elderly control individuals—generated as previously described [10]—were used for this purpose. The total number of reads in the BAM file and the number of reads mapping to the VNTR core sequence (chr19:1049514–1049953 [hg19]), which excludes sequencing reads aligning to the VNTR breakpoints, were calculated with Rsamtools [30]. Individuals with very low coverage (≤ 5 VNTR sequencing reads) were not included. VNTR coverage was then normalized by division by the total number of reads for that sample. Association between normalized coverage and SNP genotypes was calculated with a genotypic Kruskal–Wallis test. D′ and r² LD values were established previously [10].

To analyze the association between VNTR size and GWAS SNP genotypes in an extended population, we used NGS data of the 1000 Genomes Project [1]. Aligned low coverage whole genome sequencing data from European individuals (i.e. British in England and Scotland, Finnish in Finland, Iberian populations in Spain, Utah residents with Northern and Western European ancestry, and Tuscany in Italy) were processed in an analogous manner as the Belgian NGS dataset described above.

Southern blotting was used to genotype VNTR lengths in 275 patients and 177 control individuals (Table S1, Fig. S1), including individuals without pathogenic mutations [8] in APP, PSEN1, and PSEN2, or ABCA7 PTC mutations (incl. rs200538373) [10]. Furthermore, we excluded individuals for whom only a single band on Southern blotting was observed, since this could reflect failed detection of the second VNTR allele. The Southern blotting protocol was adapted from the DIG Application Manual For Filter Hybridization (Roche, Basel, Switzerland) (details provided in Supplementary Methods).

Statistical analyses on Southern blotted VNTR alleles were performed in R3.3.2 [35], unless mentioned otherwise. VNTR allele distribution was compared to genotypes of rs3764650 and rs78117248, which were determined previously [10, 45]. Association between SNP genotypes and phenotype was assessed with allelic Fisher exact tests. VNTR length association testing per SNP was done with a Kruskal–Wallis Rank Sum test. Optimal VNTR size cutoff to distinguish expanded and wild-type VNTR lengths was determined as the largest VNTR allele length observed in control individuals without an rs3764650 risk allele. Association testing and calculation of odds ratio (OR) with 95% confidence interval (95% CI) between expanded and wild-type allele carriers in patients and controls was calculated with Fisher exact testing. Calculations of heritability attributable to an ABCA7 VNTR expansion were performed with INDI-V [52], assuming AD heritability of liability of 79% [13], a 13% risk of AD after age 65 years [42], and a multiplicative model. Association of VNTR length with age at onset (AAO) within patients was calculated with linear regression, using the sum of VNTR alleles or longest VNTR allele, APOE ε4 dosage, and gender as independent variables. AAO difference between patients with and without an expanded VNTR allele was assessed using a Mann–Whitney U test.

To test whether VNTR allele lengths share a haplotype with ABCA7 PTC mutations, we used Southern blot to genotype VNTR length in an additional seventeen individuals carrying a previously identified PTC mutation [10]. These individuals were not included in any statistical analyses.

CSF concentrations of amyloid β_1–42 (Aβ_1–42), phosphorylated tau_181P (P-tau_181P), and total tau (T-tau) were determined for a subset of 168 patients from the Southern blotting cohort (Table S1) as part of their diagnostic work-up with commercially available single parameter ELISA kits (Fujirebio Europe, Ghent, Belgium). Linear regression with the sum of VNTR alleles or longest VNTR allele as the independent variable and log₂ transformed biomarker levels as the dependent variable was calculated in R [35]. Results are reported as β-regression coefficients with standard error (SE) and p value.

Procedures for RNA extraction, cDNA generation and characteristics of the LCL cohort are given in Supplementary Methods.

Quantitative real time PCR (qRT-PCR) was performed with the TaqMan Fast Advanced Master Mix (Thermo Fisher Scientific), according to the manufacturer’s protocol in a 384-well plate setting on LCL RNA. Each sample-gene pair was conducted in triplicate and on two different cDNA batches from separate RNA extractions. ABCA7 transcripts were targeted using the Hs01105117_m1 Taqman assay, with primers located in exon 45 and 46 of ABCA7 (Fig. 1), and three reference genes were used: HPRT1 (Hs02800695_m1), GAPDH (Hs99999905_m1), and YWHAZ (Hs00237047_m1). The PCR protocol (2 min at 50 °C, then 20 s at 95 °C, followed by 40 cycles of 1 s at 95 °C, and 20 s at 60 °C), and fluorescence detection was performed on a ViiA 7 Real-Time PCR System (Thermo Fisher Scientific). Ct values were exported to Qbase+ (Biogazelle, Gent, Belgium) to calculate relative expression.

Splicing events in proximity to the VNTR were identified using cDNA PCR amplification followed by Sanger sequencing in both LCL and brain cDNA. Forward primers were positioned in exon 17 (5′-TTTCGGAGGAGCTACTGGTG-3′), or on the exon18-cryptic-acceptor-splice-site-breakpoint (5′-GACCCAAAGGCCTGGAGA-3′), and paired with a reverse primer in exon 20 (5′-GTGCTCGTCCACGGTCAG-3′). Amplification was performed with Titanium Taq (Clontech Laboratories, Mountain View, CA, USA) in a 1 M betain solution using the following thermal cycler protocol: 2 min at 95 °C, followed by 35 cycles of 30 s at 95 °C and 1.5 min at 68 °C, and finished by 3 min at 68 °C. The Sanger sequencing reaction was then carried out with BigDye Terminator v3.1 Cycle Sequencing Kit (Thermo Fisher Scientific), and subsequently processed on an ABI 3730 DNA Analyzer (Thermo Fisher Scientific). Analysis was performed with Seqman software (DNASTAR, Madison, WI, USA). All splicing events were further confirmed by analyzing in-house RNAseq data (described in Verheijen et al. [50]) with the use of Integrative Genomics Viewer (IGV) software [37].

Capillary fragment analysis was used to quantify eight observed isoform combinations. PCR amplification of cDNA was analogous to the isoform identification method described above, using a 5′ FAM-labeled forward primer in exon 17 (5′-TTTCGGAGGAGCTACTGGTG-3′), and an exon 20 reverse primer (5′-GTGCTCGTCCACGGTCAG-3′). The PCR amplicons were then supplemented with formamide and GeneScan 600 LIZ Size Standard (Thermo Fisher Scientific) and size separated on an ABI 3730 DNA Analyzer (Thermo Fisher Scientific). Analysis was performed with MAQ-S v1.5.0 software (Agilent). Manual inspection was performed to confirm that fluorescence signals corresponded to the expected size, and signal heights were within quantifiable range. Ratios for canonical, exon 18 cryptic splice acceptor, and exon 19 skipping isoforms were calculated as the signal area of the corresponding isoform peaks divided by the total area of these peaks.

To examine the effect of individual VNTR alleles on ABCA7 isoform abundancy we performed allele-specific isoform expression on an individual with a small (837 bp) and large (9711 bp) VNTR allele, and heterozygosis (A/G) for nearby exonic polymorphism rs3752240. VNTR alleles were phased with rs3752240 on genomic DNA according to the “short VNTR allele PCR protocol” (Supplementary Methods) using primers 5′-CAAGACCACCACCCTGTGA-3′ and 5′-AGAGATGGGGAAGGGACCTC-3′. Since the largest allele is too large for PCR amplification, only the shortest allele will be observed, allowing phasing. RNA extraction and cDNA synthesis was performed as described above. Subsequently the “Isoform quantification PCR protocol” was used for which the forward and reverse primers were 5′ supplemented with Illumina Nextera Adapters (5′-TCGTCGGCAGCGTCAGATGTGTATAAGAGACAG-3′ and 5′-GTCTCGTGGGCTCGGAGATGTGTATAAGAGACAG-3′ respectively). The amplicon was then PCR barcoded with i5 and i7 adapters and 300 bp paired end sequenced on MiSeq with v3 chemistry (Illumina, San Diego, CA, USA). GMAP was used to align and output splicing events of sequencing reads [53]. Genotypes for rs3752240 per read were determined with samtools [25].

The qRT-PCR and isoform quantifications were compared to the sum of the VNTR alleles. Since the difference between two VNTR alleles within an LCL sample was kept to a minimum (Supplementary Methods), the VNTR sum value corresponds well to the underlying individual alleles. Association between expression and VNTR sum was calculated with linear mixed-effects models, using lme4 in R [5, 35], correcting for gender and batch in which LCL were grown. The full model was compared to the null model (without VNTR sum) using a Likelihood Ratio test. Linear regression was used to calculate the expression fold change between different VNTR sizes. The analyses were additionally performed using the longest VNTR allele as independent variable.

We examined repetitive regions in ABCA7, which were poorly assessed by conventional NGS data analysis, in our previously generated NGS dataset [10]. This revealed many sequencing reads with low mapping quality in intron 18, close to exon 18, aligning to a low complexity C-rich region. Further inspection indicated a 25 bp tandem repeating pattern. The VNTR comprises 592 bp of the human reference genome (chr19:1049437–1050028 [hg19]), corresponding to 23.7 repetitions of the 25 bp VNTR core repeat motif (Fig. 1). The first two nucleotides (GT) of the VNTR correspond to the canonical splice donor site of exon 18. The sequence between VNTR repeat units can vary slightly (Fig. S2).

First, we investigated the depth of NGS sequencing reads aligning to the VNTR (chr19:1049514–1049953 [hg19]) as a proxy for VNTR length, in relation to the genotype of ABCA7 GWAS index SNP rs3764650. Carriers of the risk increasing rs3764650[G]-allele had increased read depth in the Belgian cohort (n = 1529; Fig. S3, p = 2.2 × 10⁻¹⁶) [10], as well as in low coverage whole genome sequencing data in European subjects derived from the 1000 Genomes Project (n = 403; Fig. S4, p = 0.01) [1].

To directly genotype VNTR length, we performed Southern blotting on 452 individuals. For each individual, two size-separated bands were imaged corresponding to the VNTR alleles; no signs of mosaicism were present (Fig. S1). We observed high variability in VNTR lengths between individuals. The smallest allele was 298 bp (approx. 12 repeat units), which could be confirmed with Sanger sequencing. The largest observed allele size was 10678 bp (approx. 427 tandem repeats). Relative to the ABCA7 VNTR in the human reference genome (chr19:1049437–1050028 [hg19], 592 bp), the longest VNTR length causes an eightfold size increase of intron 18 and a 39% extension of ABCA7 pre-mRNA. Increasing VNTR lengths were associated with risk alleles of rs3764650 (p = 6.7 × 10⁻⁸) and rs78117248 (p = 6.3 × 10⁻⁴) (Fig. 2a and Fig. S5), confirming our observations based on NGS-based approximation.

For individuals with two alleles within the normal range (< 5720 bp, “wild-type”), distribution was similar between patients and controls (Fig. 2b). However, individuals carrying an allele > 5720 bp (“expanded”), were mostly patients [OR 4.5 (95% CI 1.3–24.2), p = 0.008, Fig. 2b]. Twenty patients (7.3%) carried an expanded VNTR allele in contrast to three controls (1.7%) (Table S2). In this cohort, the heritability of liability attributable to ABCA7 VNTR expansion is estimated at 3.1%. Among carriers of an expanded allele, the second (smaller) allele was longer in patients (2751 ± 1164 bp) than controls (1001 ± 652 bp). In the group with Southern blot data, 25% of patients and 22% of controls carried at least one rs3764650 risk allele. The latter is not significantly associated with AD in this cohort (p = 0.6), precluding conditional regression analysis. For expanded VNTR allele carriers, however, the proportion of rs3764650 risk alleles increased to 63 and 100% for patients and controls, respectively (Table S2). We observed no shared haplotype between expanded VNTR alleles and ABCA7 PTC mutations (Fig. S7). Patients with an expanded VNTR allele had an onset age ranging from 44 to 90 years. The age at inclusion for all three control individuals falls within this range (Table S2). We observed no association between VNTR length and onset age in patients (p = 1, Table S3, Table S4), nor for expanded VNTR allele carriers compared to other patients (p = 0.9). We analyzed the correlation of VNTR length with CSF biomarkers for AD (Fig. 3a–c) in 168 AD patients. A decrease in Aβ_1–42 was observed with increasing VNTR length (β_log(Aβ) = − 4.6 × 10⁻⁵, SE = 2.1 × 10⁻⁵, p = 0.03 for sum of VNTR alleles). For P-tau_181P, a directly proportional trend was observed, albeit not significant (β_log(p-tau) = − 4.0 × 10⁻⁵, SE = 2.3 × 10⁻⁵, p = 0.09 for sum of VNTR alleles). We did not detect an association of VNTR length on T-tau (Table S3).

We observed a decrease in overall ABCA7 expression with increasing VNTR length (β = − 4.7 × 10⁻⁵, SE = 1.9 × 10⁻⁵, p = 0.01 for sum of VNTR alleles) (Fig. 4a, Table S3). This trend was observed for both patients and controls. The average expression for the sample with the largest diploid combination of VNTR alleles (5399 + 4575 bp) was 33% lower than the expression at the smallest diploid VNTR combination (557 + 315 bp).

We observed three separate isoform changes in exon 18 and exon 19 (exons flanking the VNTR), currently unknown to public databases (Fig. 1). These alternative splicing events include usage of a cryptic splice acceptor site in exon 18 (chr19:1049314–1049316 [hg19]), usage of a cryptic splice donor site in intron 18 (chr19:1049461–1049463 [hg19]), and exon 19 skipping. The first two alternative splicing events respectively cause a 52 bp loss of coding sequence, and a 25 bp intron retention (corresponding to one VNTR repeat unit), hence both resulting in an out-of-frame transcript. The third splicing event leads to complete skipping of exon 19, which causes in-frame loss of coding sequence, corresponding to 44 amino acids that are part of the first nucleotide binding domain (NBD) of ABCA7 (Fig. 1). We analyzed the effect of VNTR length on abundance of these splicing events in LCL (Fig. 4b). Exon 18 alternative splicing and intron 18 retention events were not affected by different VNTR sizes (Table S3). Exon 19 skipping, however, strongly increased with expanding VNTR size (p = 3.24 × 10⁻¹³ for sum of VNTR alleles; Table S3). This trend was present in patients and controls (Fig. 4c). Exon 19 skipping was 3.2 times more abundant in the LCL sample with the largest diploid combination of VNTR alleles (5399 + 4575 bp) compared to smallest VNTR combination (557 + 315 bp). To determine the effect of individual alleles on exon 19 skipping, we examined allele-specific expression in LCL of an AD patient with an expanded VNTR allele (9711 bp) and a small wild-type allele (837 bp) (Table S2)—which were respectively phased with the rs3752240[G] and rs3752240[A] allele—and observed a 4.1-fold increase in exon 19 skipping for the expanded allele (Fig. S8).

In addition to quantification in LCL (Fig. 4d), we quantified the three alternative splicing events in the hippocampus, originating from six AD patients and six controls, as well as in Brodmann area 10 (BA10) brain tissue from four patients and six controls (Fig. 4e, f). For each brain tissue, three individuals had VNTR Southern blotted lengths and none carried an expanded VNTR allele. Splicing isoforms causing a frameshift had approximately similar expression level in all tissues: 14 ± 6 and 2 ± 1%, respectively. Exon 19 skipping, however, varied strongly between different brain regions and between brain and LCL: 45 ± 10% in BA10, 29 ± 12% in hippocampus, and 12 ± 6% in LCL. In 90% of BA10 cases and 31% of hippocampal tissue, the cumulative abundance of the alternative splicing transcripts exceeded the abundance of the canonically spliced transcript (Fig. 4e, f).