Abstract

Introduction

Two recent genome-wide association studies independently identified CLU as a risk gene for Alzheimer's disease (AD).^{1, 2} Follow-up studies and meta analyses have replicated these results, although the strongest associated variant sometimes differed.^{3, 4, 5, 6, 7} Efforts to identify functional variations through exon sequencing and examining effects of single-nucleotide polymorphisms (SNPs) on CLU expression have not yet provided a functional link between the associated polymorphisms and AD, but they have excluded the involvement of common coding variation.⁸ The same study examined the effect of SNPs on the gene's expression with negative results; however, the microarray platform used did not examine individual splice variants.

Clusterin, also known as apolipoprotein J, is a glycoprotein first identified in 1988⁹ and discussed as a candidate gene for AD for more than 15 years.^{10, 11} Its multiple functions include roles in apoptosis, complement regulation, lipid transport, sperm maturation, endocrine secretion, membrane protection, promotion of cell interactions and as a chaperone.^{12, 13, 14, 15, 16} Secreted soluble and nuclear forms of clusterin have been described and their production is likely regulated by use of alternative transcription start sites¹⁷ or alternative splicing.¹³ This is achieved through use of discrete translation initiation sites, alternatively introducing an endoplasmic reticulum-targeting signal upstream of a nuclear localization signal. The nuclear form of clusterin is specifically induced in epithelial cells by tumor growth factor-β,¹⁷ whereas in prostate cells, different CLU isoforms have been shown to have different responses to androgens and opposing functions with regard to apoptosis.^{18, 19}

The importance of CLU alternative splicing on its function led us to the hypothesis that the reported association with AD, although it is shown not to have a significant impact on the overall transcript levels as measured by microarrays,⁸ might reflect a disruption of the balance between transcripts. We tested our hypothesis on a set of 190 temporal lobe samples without brain pathology (controls) and followed up in another set of 115 temporal lobe samples from AD cases and controls.

Materials and methods

Luciferase reporter assays

Constructs: The upstream (U), short (S) and downstream (D) inserts (shown in Figure 1c were PCR amplified from genomic DNA from individuals homozygous for the reference or risk alleles for rs9331888 (primers in Supplementary Table 2). Sequencing revealed no other variation in the amplified sequence. Fragments were amplified using PfuTurbo Polymerase (Stratagene, Santa Clara, CA, USA) high-fidelity taq, and A' overhangs were added by incubating the products for 5min with Taq DNA polymerase (Invitrogen) and excess deoxyadenosine triphosphate. The amplicons were TA cloned into pCR 8/GW/TOPO (lnvitrogen) entry vector containing attL1 and attL2 recombination sites. Inserts were subcloned via recombination, using the Gateway LR Clonase Enzyme Mix (Invitrogen) to a pDSma_promoter vector.²² This plasmid is a pGL3 firefly luciferase reporter vector containing an SV40 promoter (Promega, Madison, WI, USA), modified to contain the Gateway cassette containing the attA and attB recombination sites as described in Grice et al.²² Inserts were verified by Sanger sequencing using the commercial primer RVprimer3 (Promega) and pGLR 5′ (Supplementary Table 2). Neuroblastoma (SK-N-SH, ATCC no. HTB-11) cells were grown in an ATCC-suggested medium (http://www.atcc.org), without antibiotics. Approximately 0.8 × 10⁵ cells were plated 24h before transfection for cells to reach 90–95% confluency. Primary cortical neuron cultures were prepared from day 16–18 embryonic C57BL/6 mice, following established protocols.²³ Animal protocols were approved by Johns Hopkins University Animal Care and Use Committee. Primary cortical neurons were plated in 24-well plates at a density of 2 × 10⁵ in 1ml of growth medium per well. The neurons were switched to 500μl antibiotic-free medium on the third day in vitro and transfected on day 4. SK-N-SH and primary cortical neurons were co-transfected (Lipofectamine 2000, Invitrogen) using 0.8μg of plasmid DNA and 0.08μg phRL-SV40 control renilla plasmid, using standard 24-well protocol. Medium was replaced 5h post transfection. Dual luciferase assays (Promega) were performed in quadruplicate 24h post transfection, following the manufacturers standard protocol (Tecan Genios Microplate Reader, Männedorf, Switzerland). Relative luciferase units were calculated by dividing the firefly luciferase values by the renilla control values for each transfection reaction.

Open in a separate window

Figure 1

(a) CLU alternative transcripts from reference sequence (RefSeq) and UCSC examined in this study. Single-nucleotide polymorphisms (SNPs) genotyped within the gene are shown and SNPs reported in recent genome-wide association studies (GWAS) are marked with an asterisk. (b) Enlargement of the region around rs9331888. The SNP location is marked by a vertical red line. Selected tracks from the UCSC genome browser ENCODE data, as well as phylogenetic conservation data are shown. For clarifications, see the UCSC genome database website. (c) The sequences inserted into reporter constructs are shown named U (upstream) S (short) and D (downstream). Each sequence was made to carry either a reference or a risk allele at rs9331888, shown here by a white asterisk and found to influence Alzheimer's disease (AD) risk in the Lambert et al.² GWAS.

Results

Three different CLU splice forms with different transcription start sites (Figure 1a) are reported in RefSeq²⁴ (NM_001831, NM_203339 and NM_001171138) and can be differentiated using sequence-specific primers for quantitative real-time PCR. All RefSeq transcripts include coding sequence for both the endoplasmic reticulum-targeting signal and the nuclear localization signal, but whether they all use the upstream translation start site and include both signals in the protein is unknown. In our experiments using brain RNA, we did not observe the transcript lacking exon 2 described in a breast cancer cell line,¹³ but we found a transcript reported in the UCSC genome browser (GenBank acc# CR617497) lacking exons 1, 3 and 4, thus missing both the endoplasmic reticulum-targeting signal and nuclear localization signal. We designed specific primers (Supplementary Table 2) and successfully amplified NM_203339, NM_001171138 and CR617497. We confirmed that single and specific amplicons were amplified from each transcript by Sanger sequencing, agarose gel electrophoresis and DNA melting curves. Using real-time PCR, we quantified the transcripts in tissue from the superior temporal lobe of 190 individuals without gross brain pathogy (‘controls') obtained from the HBTRC (see Supplementary Table 2 for sample details). We extracted brain DNA from the same individuals and successfully genotyped 42 HapMap SNPs in and around CLU, chosen to be inter-correlated at r²0.8 and including the associated SNPs reported in the two first genome-wide association studies reporting CLU.^{1, 2} Further, we genotyped 27 ancestry informative SNP markers,²⁵ as the HBTRC did not retain ancestry information (Supplementary Table 3 describes all study SNPs).

We first examined the ancestry informative SNP markers in a PC analysis and identified two outlier individuals for PC-1, likely of African ancestry, who were removed from further analysis. No additional PCs showed outliers. Using log-transformed transcript quantity as an outcome, we applied a generalized linear model including age, sex, PMI, PCR plate ID (identity) (to account for plate effects as we used two 384-well real-time plates) and the first PC of the ancestry informative SNP markers to account for possible residual admixture. We found a highly significant increase for all transcripts with increasing age (about 1% increase per year, all P-values <0.0002, see Table 1), consistent with our recent report of significant overlap between genes changing expression in AD and with normal aging.²⁶ Transcript CR617497 was significantly lower in males (~17%, P=0.001) and NM_203339 approached significance for an approximate 13% decrease in males (P=0.058). A significant plate effect was identified only for transcript NM_001171138 and no effects of PMI were identified. Plate ID and PMI were both included in all analyses to account for possible small effects, regardless of significance. Transcript expression levels showed strong positive inter-correlations persisting after correcting for all the effects in our model, possibly suggesting significant common regulation (all P<10⁻⁴). We then included each SNP genotype sequentially as a predictor in the model. The results for the three transcripts for SNPs showing at least nominally significant effects on transcription are shown on the left side of Table 2. The SNP variants had little or no effect on NM_001171138 and CR617497. In contrast, NM_203339 showed multiple, nominally significant correlations with genotypes, the strongest with rs9331888. This SNP, located in exon 1 of NM_203339 (Figure 1b) ranked third in the recent genome-wide association studies on Caucasians by Lambert et al.² and was the only one of the three SNPs that replicated in an independent study of a Chinese population,⁶ although contradicting results have been reported,²⁷ providing strongest support for a different SNP, rs11136000. The risk alleles of rs11136000 and rs9331888 are in strong linkage disequilibrium (D'=0.96, r²=0.224 based on our data). We then examined the effect of SNP genotypes on each transcript, adjusting for the amount of the other two transcripts, which were added as predictors in the model. This analysis reflects regulation of alternative splicing by removing variability common to all transcripts and highlighting their differences. The results of this analysis (Table 2, right side) showed a highly significant effect of rs9331888 on the relative levels of NM_203339 (P=4.3 × 10⁻¹²). To test whether other less significant effects on NM_203339 observed for SNPs rs11136000, rs9331908, rs9331931, rs3087554 and rs17466684 were due to linkage disequilibrium, we analyzed each in a model that also included rs9331888. In all cases, only rs9331888 remained significant, suggesting that other SNP effects on splicing likely reflect their linkage disequilibrium with rs9331888.

To exclude artifacts, we performed multiple quality-control steps. We excluded nucleotide variation under a PCR primer by sequencing the regions under the NM_203339-specific primers in all individuals. Additionally, we designed a new set of transcript-specific primers for NM_203339 and CR617497, and performed new quantitative PCR experiments starting from RNA, obtaining similar highly significant results. We only used CR617497 in this experiment, because the original data showed that correcting NM_203339 for just one other transcript revealed the effect almost equally well as correcting for both. We further tested for genotyping errors and confirmed by nucleotide sequencing all rs9331888 genotypes.

We proceeded to replicate our result in an independent set of samples, which included AD cases allowing us to test for possible disease-related transcript variation. This replication set included 24 controls from the HBTRC and 29 controls from the Johns Hopkins Brain Resource Center, as well as 22 and 40 AD cases, respectively. We quantified NM_203339, NM_001171138 and CR617497 as above, genotyped rs9331888 by sequencing and by BslI restriction enzyme digestion and applied a similar generalized linear model, with NM_203339 as a dependent variable, age, sex, PMI, sample source, diagnosis and rs9331888 genotype as predictors. The quantitative PCR plate variable no longer applied, as a single plate was used in this experiment. We used existing ancestry information on the Johns Hopkins Brain Resource Center samples and included only Caucasian individuals. Ancestry information was not available for HBTRC samples, but as we identified only two genetic outliers in the previous larger HBTRC sample and observed no significant effects of ethnicity, we included all HBTRC samples. The effect of rs9331888 on relative NM_203339 levels was strongly replicated (P=0.0014, Table 2). As with the initial sample, the significance of the effect on total mRNA levels was more modest, this time only suggestive (P=0.076). Cases had higher levels of all transcripts (Table 1) and suggestively higher levels of relative levels of NM_203339. After correcting for the effect of rs9331888, the difference in relative levels of NM_203339 between AD cases and controls was diminished. Many of the previously observed effects of age and sex on the various transcripts were replicated in this sample as shown in Table 1.

The location of rs9331888 in exon 1 of NM_203339, together with ChIP-seq and DNase hypersensitivity data from ENCODE (encyclopedia of DNA elements; see Figure 1b), suggest that this variant might be directly responsible for regulation of this alternative transcription start. We used the Promega Dual-Luciferase Reporter Assay System to test the six expression constructs shown in Figure 1c (S-ref, S-risk, U-ref, U-risk, D-ref and D-risk; ref and risk indicate the rs9331888 alleles) for enhancer activity in SK-N-SH cells and primary mouse cortical neurons. We observed significantly higher activity of the risk allele for all constructs in primary neurons and the S- and U- constructs in SK-N-SH cells (Figure 2), consistent with our post-mortem brain expression results.

Open in a separate window

Figure 2

Dual luciferase reporter assays comparing constructs carrying reference (ref) or risk alleles in the three different constructs shown in Figure 1c in front of an SV40 promoter. Firefly, relative to renilla luciferase levels, is shown with s.e. bars based on four replicates. All constructs show significant differences between the two rs9331888 alleles in primary neuron culture and two of three also show differences in SK-N-SH cells. As expected, the risk allele shows higher activity. RLU, relative luciferase units.

As shown in Figure 1, rs9331888 lies in a region with significant evidence of regulatory potential. We used the bioinformatics program rVista 2.0 (http://rvista.dcode.org/)²⁸ and found that the risk variant of rs9331888 eliminates binding sites for nuclear factor kappa B and early B-cell factor, whereas it generates a new binding site for heat shock factor protein-1. Pending experimental verifications, this interesting result could provide a guide for further dissection of the relationship between this SNP, this transcript and AD.

Discussion

We have shown that the minor allele of rs9331888, previously associated with increased risk of AD, is associated with increased-relative levels of NM_203339 and is likely the functional variant responsible for this effect. Given the prior genetic association results for this SNP and the distinct roles of CLU transcripts,¹⁸ we hypothesize that alternative splicing is the etiological link between rs9331888 and AD. However, the functional properties of the protein products and whether or how their production varies across the three transcripts are unclear. At least two transcripts, NM_001171138 and NM_203339, potentially contain both the endoplasmic reticulum-targeting signal and nuclear localization signal, but it is unknown whether there is preferential utilization of a specific translation start site, which could dramatically change the functional outcome.^{18, 19} Finally, the function of CR617497 which we were able to reliably detect in the temporal lobe transcriptome is unknown. An alternative hypothesis is that the increased risk is not the result of the different transcript functions, but rather of their specific responses to different signals, responses that might be aberrantly lost or gained for carriers of the rs9331888 risk allele, which abolishes two and introduces one new transcription factor binding site. Regardless of what the true underlying biology will turn out to be, in view of our results and the reported genome-wide association studies, clarifying the properties of these transcripts, the corresponding proteins and their regulation is of great importance for AD research.

As we mentioned in the introduction, although rs9331888 has been independently reported in two populations, it is another CLU SNP, rs11136000 that has shown overall the most consistent associations. As the two risk alleles are in near complete linkage disequilibrium with each other (D'=0.96), it is likely that the functional effect of rs9331888 is responsible for only part of the observed association of the gene with AD. Other rare or common regulatory variants might underlie the remaining risk attributed to CLU and the inconsistencies in association patterns described in the literature. Further clarification of these effects remains a significant task, which will be facilitated by this and future work, testing the new hypotheses and moving translational research forward. Together with the recent many advances in AD through new gene discovery, our understanding of the disease biology will quickly improve, hopefully leading to significant benefits for the patients and those at risk.

Acknowledgments

Notes

The authors declare no conflicts of interest.

Footnotes

References