Independent and epistatic effects of variants in VPS10-d receptors on Alzheimer disease risk and processing of the amyloid precursor protein (APP).

Genetic variants in the sortilin-related receptor (SORL1) and the sortilin-related vacuolar protein sorting 10 (VPS10) domain-containing receptor 1 (SORCS1) are associated with increased risk of Alzheimer's disease (AD), declining cognitive function and altered amyloid precursor protein (APP) processing. We explored whether other members of the (VPS10) domain-containing receptor protein family (the sortilin-related VPS10 domain-containing receptors 2 and 3 (SORCS2 and SORCS3) and sortilin (SORT1)) would have similar effects either independently or together. We conducted the analyses in a large Caucasian case control data set (n=11,840 cases, 10,931 controls) to determine the associations between single nucleotide polymorphisms (SNPs) in all the five homologous genes and AD risk. Evidence for interactions between SNPs in the five VPS10 domain receptor family genes was determined in epistatic statistical models. We also compared expression levels of SORCS2, SORCS3 and SORT1 in AD and control brains using microarray gene expression analyses and assessed the effects of these genes on γ-secretase processing of APP. Several SNPs in SORL1, SORCS1, SORCS2 and SORCS3 were associated with AD. In addition, four specific linkage disequilibrium blocks in SORCS1, SORCS2 and SORCS3 showed additive epistatic effects on the risk of AD (P≤0.0006). SORCS3, but not SORCS2 or SORT1, showed reduced expression in AD compared with control brains, but knockdown of all the three genes using short hairpin RNAs in HEK293 cells caused a significant threefold increase in APP processing (from P<0.001 to P<0.05). These findings indicate that in addition to SORL1 and SORCS1, variants in other members of the VPS10 domain receptor family (that is, SORCS1, SORCS2, SORCS3) are associated with AD risk and alter APP processing. More importantly, the results indicate that variants within these genes have epistatic effects on AD risk.

Introduction

A central event in the pathogenesis of Alzheimer's disease (AD) is the deposition of amyloid β (Aβ) 1–40 and Aβ1–42 peptides generated by proteolytic cleavage by β- and γ-secretase from a larger membrane-bound protein, the amyloid precursor protein (APP).[1] APP and the secretases are integral transmembrane proteins dynamically sorted through the plasma membrane. Modulation of APP sorting through the membrane or altering APP cleavage by secretase enzymes could affect the regulation of Aβ production or processing.

Variants in two members of the vacuolar protein sorting 10 (VPS10) domain-containing receptor protein family, sortilin-related receptor (SORL1) and sortilin-related VPS10 domain-containing receptor 1 (SORCS1), are associated with late-onset AD presumably through effects on APP sorting and cleavage.[2, 3, 4] The VPS10 domain-containing receptor protein family contains five type I membrane homologs (SORL1, sortilin (SORT1), SorCS1, SorCS2 and SorCS3),[5, 6, 7, 8, 9] that are expressed in the central nervous system. All contain a single Vps10p-D situated at the N-terminus of their luminal/extracellular moiety. The VPS10 motif functions as a sorting receptor in the Golgi compartment required for the intracellular sorting and delivery of proteins, including APP. In SORT1, also known as neurotensin receptor-3, the Vps10p-D makes up the entire luminal extracellular part of the receptor, but the other four receptors have additional modules. In SORL1, the Vps10p-D is followed by five low-density lipoprotein receptor class B repeats flanked by an epidermal growth factor precursor-type repeat, a cluster of 11 low-density lipoprotein receptor class A repeats and 6 fibronectin type-III repeats. The mutually highly homologous SorCS1, SorCS2 and SorCS3 contain a leucine-rich segment between the Vps10p-D and the transmembrane domain. Structure prediction of the leucine-rich segment suggests a beta-sandwich fold and relates the domain to the immunoglobulin-like fold (E-set) superfamily. Following the extracellular and transmembrane segment, each receptor carries a short (40–80 amino acids) cytoplasmic domain comprising typical motifs for interaction with cytosolic adaptor molecules. In genomic DNA, members of this family are large with many exons but the coding sequence lengths are usually <3700 nucleotides. Very large introns (introns 1–2) typically separate the exons encoding the VPS10 domain; the remaining exons are separated by much smaller introns. Exons 1–3 encode the VPS10 domain.

Previously, we demonstrated that SORL1 modulates the translocation and retention of APP in subcellular compartments, which are less favorable for secretase processing, thereby reducing the extent of proteolytic breakdown into both amyloidogenic and non-amyloidogenic products.[3] Furthermore, we showed that under-expression of SORL1 leads to overexpression of Aβ and an increased risk of AD. Subsequently, we demonstrated that genetic variation in SORCS1 also influences AD risk, cognitive performance, APP processing and Aβ40 and Aβ42 levels through an effect on γ-secretase processing of APP.[2, 10] Overexpression of SorCS1 reduced Aβ40 and Aβ42 levels, whereas suppression of SorCS1 increased γ-secretase processing of APP. The association of SORL1 with AD has been supported by a meta-analysis of Caucasian and Asian data sets that included a total of 12 464 cases and 17 929 controls11 and has been further validated in various ethnic groups, including African Americans, Israeli Arabs and Caribbean Hispanics, although with some degree of allelic heterogeneity.[3, 11, 12, 13, 14, 15, 16, 17, 18] In addition, these data are supported by a study in which overexpression of SorCS1cβ-myc in cultured cells caused a significant reduction in Aß generation, whereas, conversely, endogenous murine Aß40 andAß42 levels were increased in the brains of Sorcs1 hypomorphic mice.[19]

We hypothesized that variants in other members of the sortilin-related VPS10 domain containing receptor family, namely SORCS2, SORCS3 and SORT1, would also be associated with AD risk either independently or through epistatic effects. These homologous genes are expressed in different brain regions with different subcellular localisations,[20, 21, 22] but there are many brain regions, such as the hippocampus, in which these genes are co-expressed albeit at low levels.[20]

We conducted single-marker association and epistasis analyses of all the five homologous genes in a large Caucasian case-control data set, with sufficient power to detect modest effect sizes and interactive effects. In addition, we conducted microarray gene expression analyses and γ-secretase assays for SORCS1, SORCS2, SORCS3 and SORT1.

Participants and methods

Ethics statement

Informed consent was obtained from all the participants using procedures approved by institutional review boards at each of the clinical research centers collecting human subjects for the ADGC project.

Participants

The data set included 11 840 cases and 10 931controls from the ADGC data set.[23] The clinical characteristics are summarized in Table 1. The diagnoses of ‘probable' or ‘possible' AD were defined based on the National Institute of Neurological and Communication Disorders and Stroke–Alzheimer's Disease and Related Disorders Association (NINCDS-ADRDA) diagnosis criteria at clinics specializing in memory disorders or in clinical investigations. Persons were classified as ‘controls' when they were without cognitive impairment or dementia at last visit.

Table 1

Characteristics of the study sample

Characteristics
Number of cases with AD	11 840
Number of controls	10 931
Age at Onset for AD cases (s.d.)	74.55 (6.8)
Age at last exam for controls (s.d.)	76.26 (7.2)
Proportion of females	59.66%
Frequency of APOE e4+
Cases	0.38
Controls	0.14
Frequency of APOE e4−
Cases	0.62
Controls	0.86

Abbreviation: AD, Alzheimer's disease; APOE, apolipoprotein E.

Genotyping

HapMap2-imputed genotypic data for single nucleotide polymorphisms (SNPs) in SORCS1 (n=648), SORCS2 (n=740), SORCS3 (n=742), SORL1 (n=160) and SORT1 (n=40) was obtained from the previously published genome-wide association study.[23] The SNPs assessed included both intronic and exonic SNPs. The SORCS1 SNPs were not identical to the SNPs assessed in our previous study, which had been selected based on previous reports.[2] Details regarding apolipoprotein E (APOE) genotyping are described in the Supplementary methods.

Cell culture and transfection

Using HEK293 cell lines, reverse transcriptase–PCR (RT–PCR) and western analysis were used to detect all five VPS10 proteins and to verify the knockdown and specificity of each short hairpin RNA (shRNA) as previously described.[2] The corresponding shRNA DNA sequences are shown in Supplementary Table 3.

APP-GV Assay

The γ-secretase activity and nuclear translocation of the APP/Fe65/TIP60 protein complex was monitored with the APP-GV assay.[24] The APP-GV assay is a luciferase-based assay[24] consisting of the APP gene's C-terminus (AICD) fused to a transcription factor composed of the GAL4 DNA-binding domain with VP16 transcriptional activator (GV). In addition, the AICD fragment is fused to the GV domains as a positive control of AICD generation and allows for the evaluation of the AICD-specific contribution to the observed modulation in the APP-GV assay. Briefly, SorCS2 cDNA or SorCS2 shRNAs transiently transfected were evaluated in either the APP-GV or the AICD-GV assay, as previously described[24] in the HEK293 cell line. SorCS3 cDNA or SorCS3 shRNAs, and SORT1 cDNA or SORT1 shRNAs, were evaluated in a similar fashion.

Microarray gene expression and quantitative RT–PCR

Expression profiling was performed separately for the cerebellum, parietal-occipital neocortex and amygdala regions from 19 AD and 10 control brains from the New York Brain Bank (www.nybb.hs.columbia.edu). This three-region approach allowed us to enhance the signal-to-noise ratio[25] and to determine those changes in expression that are specific for late-onset AD and consistent with the distribution of AD pathology. For the expression profiling of AD and control brains, the Affymetrix GeneChip Human Exon 1.0 ST Arrays (Affymetrix, Santa Clara, CA, USA) were used. Frozen brain tissue was ground over liquid nitrogen and stored at −80 °C until use. Total RNA was extracted and purified using the TRIzol Plus RNA purification kit (Invitrogen, Life Technologies, Grand Island, NY, USA). All RNA preparations were analyzed using an Agilent Bioanalyzer (Agilent Technologies, Santa Clara, CA, USA; RNA 6000 nano-kit) to determine RNAquantity/quality and only samples with RNA integrity number >8 were used in the subsequent RNA amplification and hybridization steps. The Genechip expression two-cycle target labeling kit (Affymetrix) was used for all samples according to Affymetrix's protocols. Briefly, the procedure consists of an initial ribosomal RNA reduction step and two cycles of reverse transcription followed by in vitro transcription. For each sample, 1 μg of total RNA is initially subjected to removal of ribosomal RNA using the RiboMinus Transcriptome Isolation Kit (Invitrogen) and spiked with Eukaryotic PolyA RNA controls (Affymetrix). The ribosomal RNA-depleted fraction was used for cDNA synthesis by reverse transcription primed with T7-random hexamer primers, followed by second strand synthesis. This cDNA served as the template for in vitro transcription to obtain amplified antisense cRNA. Subsequently, cRNA from the first round was reverse transcribed using random primers to obtain single-stranded sense DNA. In this second reverse transcription, dUTP (2′-deoxyuridine, 5′-triphosphate) is incorporated into the DNA to allow for subsequent enzymatic fragmentation using a combination of UDG (uracil-DNA glycosylase) and APE1 (apyrimidinic endonuclease 1). All reverse and in vitro transcription steps were performed using the GeneChip WT cDNA synthesis and amplification Kit (Affymetrix). The resulting fragmented DNA was labeled with Affymetrix DNA Labeling Reagent. Labeled fragmented DNA was hybridized to Affymetrix Human Exon 1.0 ST arrays, washed and stained using the GeneChip Hybridization, Wash and Stain Kit (Affymetrix). Fluorescent images were recorded on a Genechip scanner 3000 and analyzed with the GeneChip operating software.

Significant results obtained from the microarray study were validated by quantitative RT–PCR using the same set of AD and control samples. Total RNA (1 μg) from of the amygdala region was used to generate cDNAs using the AffinityScript first-strand synthesis kit (Agilent Stratagene, CA, USA). RT–PCR primers were designed for three randomly selected exons of SorCS3 (10, 17 and 21). The housekeeping gene, TBP (TATA-binding protein) , was used as the endogenous control; and samples were analyzed in triplicate. The primers used in the quantitative RT-PCR are available from Supplementary Table 1. Real-time RT-PCR was done using SYBR Green reagent (TaKaRa Mirus Bio, Madison, WI, USA) on an ABI7500 system (Applied Biosystems, Foster City, CA, USA).

Statistical methods

Extensive quality review of SNPs and samples were previously completed.[23] Then multivariate logistic regression analyses in PLINK (http://pngu.mgh.harvard.edu/~purcell/plink/) were used to assess additive genotypic and allelic associations with AD risk in the case-control data sets, and generalized estimating equation models were used for family-based data sets. All models were first adjusted for age at examination, sex and population stratification and subsequently for APOE-ɛ4 (additive effect) as well. For adjustment for population stratification, the first two, three or four estimated principal components were used, as described previously.[23] Logistic generalized estimating equation models[26, 27] were used to evaluate association in the family-based data sets, using the same adjustments. Then, a meta-analysis of the individual study results was performed using inverse variance weights for the effect estimates as implemented in METAL (http://www.sph.umich.edu/csg/abecasis/Metal/). In order to take linkage disequilibrium (LD) between the markers into account, the P-value threshold for multiple testing correction was, in both single-marker and epistasis analyses, determined by applying the algorithm by Li and Ji.[28] As this was a candidate gene study with the a priori hypothesis of an association between each of the explored genes and AD, the calculation was done separately for each gene.

Epistasis

Using only the SNPs that were associated with AD in the single-marker analyses (P⩽0.05), we tested for an interaction between SNPs in the five homologous genes. The analysis was carried out using PLINK (http://pngu.mgh.harvard.edu/~purcell/plink/) adjusting for population stratification. The model based on generalized estimating equations yields a list of SNP-by-SNP comparisons with beta coefficients and P-values. Based on the number of independent interactions tested, we accepted a P-value of ⩽0.0001 as statistically significant. As described above, LD between the markers was taken into account applying the algorithm by Li and Ji.[28]

Statistical analysis for the gene expression and quantitative RT–PCR data

To determine in which genes expression levels differ between affected and unaffected brain regions, as well as between AD and control brains, we performed both within- and between-group factors' analysis of variance using PARTEK GENOMICS SUITE 6.4 (http://www.partek.com/partekgs). Before the expression analysis, we log10-transformed the Rank invariant normalized expression data. False discovery rate was used to account for the error in multiple comparisons. The real-time RT–PCR data were analyzed by the comparative CT method integrated in the DataAssist Software (Life Technologies).[29]

Statistical analysis for the cell biology assays

Mean expression levels were compared by analysis of variance with post hoc correction using Graphpad Statistical software (Graphpad, Inc., San Diego, CA, USA). All data were normalized to transfection efficiency (for example, green fluorescent protein) and then to the control values on each plate for every assay to allow for comparisons across experiments.

Results

Single-marker analyses

Table 1 shows the characteristics of the study populations. In all, 15 SNPs in SORL1, 23 SNPs in SORCS1, 18 SNPs in SORCS2 and 12 SNPs in SORCS3 were associated with AD (Table 2). These SNPs belonged to distinct LD blocks in these genes (Supplementary Figure S1). All SNPs in SORL1 reached statistical significance after correction for multiple testing and taking LD between the markers into account. One of these SNPs, rs1784933, corresponds to SNP26 in the original study by Rogaeva et al.[3] and is located 6 kb from SNP 25, which is part of one of the two SORL1 clusters that have been repeatedly associated with AD in different ethnic groups.[3, 18] The SNPs in SORCS1, SORCS2 and SORCS3 were close, but not statistically significant. Interestingly, in line with previous reports,[2] most of the significant SNPs in SORCS1 and SORCS2 are located in intron 1, which is adjacent to the exons encoding the VPS10 domain. In addition, in all the four homologs (SORL1, SORCS1, SORCS2, SORCS3) some of the disease-associated SNPs were close to splice sites (Table 2). None of the genotyped SNPs in SORT1 were significantly associated with AD (Supplementary Table 2).

Table 2

SNPs in SORL1, SORCS1 SORCS2 and SORCS3 associated with AD: (a) SNPs in SORL1, (b) SNPs in SORCS1, (c) SNPs in SORCS2, (d) SNPs in SORCS3

(a)
SNP	CHR	BP	Allele 1	Allele 2	Freq1	β	s.e.	Pa	NUMOBS	Locationb	Splice site distance
rs7946599	11	121 423 640	a	g	0.0155	−0.4545	0.1027	9.66E-06	20131	Intron 16	1006
rs2298814	11	121 424 882	a	g	0.0157	−0.4234	0.0916	3.76E-06	23821	Intron 17	64
rs6589885	11	121 426 042	a	g	0.0162	−0.4004	0.0851	2.57E-06	23821	Intron 18	15
rs7131432	11	121 426 870	a	t	0.0165	−0.3804	0.0811	2.75E-06	23821	Intron 18	843
rs720099	11	121 433 793	t	c	0.9829	0.3607	0.0766	2.51E-06	23821	Intron 21	3427
rs11218342	11	121 434 428	t	c	0.9829	0.3575	0.0761	2.60E-06	23821	Intron 21	3221
rs11218343	11	121 435 587	t	c	0.9828	0.3539	0.0755	2.72E-06	23821	Intron 21	2062
rs1784919	11	121 439 665	t	c	0.0172	−0.3586	0.0752	1.88E-06	23821	Intron 22	1201
rs1792124	11	121 441 520	a	g	0.0172	−0.359	0.0752	1.83E-06	23821	Intron 23	541
rs3781835	11	121 448 254	a	g	0.0173	−0.3501	0.0751	3.14E-06	23821	Intron 23	145
rs3781838	11	121 453 517	t	g	0.9825	0.3506	0.0754	3.29E-06	23821	Intron 25	650
rs12272618	11	121 460 324	t	c	0.9821	0.3536	0.0754	2.74E-06	23821	Intron 29	225
rs2276412	11	121 460 846	t	c	0.018	−0.3531	0.0752	2.64E-06	23821	CDS-synon (exon 30)
rs7939826	11	121 468 256	t	c	0.0198	−0.3463	0.0738	2.69E-06	23821	Intron 32	1775
rs1784933	11	121 489 416	a	g	0.9532	0.216	0.0606	0.000363	21455	Intron 41	67

Abbreviations: BP, base pair position; β, beta coefficient; CDS, coding sequence; CHR, chromosome; location, single nucleotide polymorphism (SNP) location; NUMOBS, number of subjects; s.e., standard error of beta coefficient; SORCS1, sortilin-related VPS10 domain-containing receptor 1; SORL1, sortilin-related receptor 1; UTR, untranslated rgion.

Based on the number of tests performed, a P-value of 0.0009 can be considered statistically significant.

Exons 1–16 encode the VPS10 domain.

Based on the number of tests performed, a P-value of 0.0002 can be considered statistically significant.

Exons 1–18 encode the VPS10 domain.

Based on the number of tests performed, a P-value of 0.0006 can be considered statistically significant.

Epistasis analysis

Upon testing for epistatic effects between the SNPs that were associated in the single-marker analyses (Table 3a), 34 pairs of SNPs showed epistatic effects at a P-value of <0.01. The vast majority (n=26 pairs) included a specific LD block in SORCS3 with two specific LD blocks in SORCS2 (Table 3, Figures 1a and b). Consistent with the single-marker analyses, the interacting SNPs were located in introns 1 and 2 (Table 3, Figure 1), adjacent to the exons coding for the VPS10 domain. The epistasis β for these SORCS2/SORCS3 interactions ranged from −0.94 to 0.94, reflecting larger effects (additive) than the single-marker effects (−0.15<β<0.20).

Table 3

Epistasis between two specific LD blocks in (a) SORCS2 and SORCS3, and (b) SORCS1/SORCS2 and SORCS1/SORCS3

(a)
								SORCS3				SORCS2
SORCS3	SORCS2	A1	A2	BETA	s.e.	Pa	DIR	CHR1	BP1	GENE1	Locationb	CHR2	BP2	GENE2	Locationb
rs749304	rs7665496	ac	gt	−0.2347	0.0731	0.00133	−−−−−−?−−−−??−−	10	106 990 399	SORCS3	Intron 20	4	7 328 734	SORCS2	Intron 1
rs12263804	rs7665496	cc	tt	−0.2345	0.0731	0.00134	−−−−−−?−−−−??−−	10	106 993 770	SORCS3	Intron 20	4	7 328 734	SORCS2	Intron 1
rs3750261	rs7665496	cc	tt	−0.2163	0.0723	0.00279	−−−−−−?−−−−??−−	10	107 023 390	SORCS3	UTR-3	4	7 328 734	SORCS2	Intron 1
rs10884126	rs7665496	ac	gt	0.2153	0.0724	0.00292	++++++?++++??++	10	107 025 028	SORCS3	Near gene-3	4	7 328 734	SORCS2	Intron 1
rs749304	rs6840423	ag	gt	0.1592	0.061	0.00903	+−−+++?++++++−+	10	106 990 399	SORCS3	Intron 20	4	7 329 324	SORCS2	Intron 1
rs12263804	rs6840423	cg	tt	0.1584	0.061	0.00936	+−−+++?++++++−+	10	106 993 770	SORCS3	Intron 20	4	7 329 324	SORCS2	Intron 1
rs749304	rs7684383	ac	gt	−0.4259	0.125	0.00066	−−−−−+?−−−−??−−	10	106 990 399	SORCS3	Intron 20	4	7 417 241	SORCS2	Intron 2
rs12263804	rs7684383	cc	tt	−0.4249	0.125	0.00067	−−−−−+?−−−−??−−	10	106 993 770	SORCS3	Intron 20	4	7 417 241	SORCS2	Intron 2
rs10884126	rs7684383	ac	gt	0.4128	0.1224	0.00074	+++++−?++++??++	10	107 025 028	SORCS3	NearGene-3	4	7 417 241	SORCS2	Intron 2
rs3750261	rs7684383	cc	tt	−0.4121	0.1223	0.00076	−−−−−+?−−−−??−−	10	107 023 390	SORCS3	UTR-3	4	7 417 241	SORCS2	Intron 2
rs12262245	rs7684383	cc	gt	−0.7045	0.2318	0.00237	−−−−−+?−−−−??+−	10	106 621 704	SORCS3	Intron 2	4	7 417 241	SORCS2	Intron 2
rs7086583	rs7684383	ac	ct	−0.7032	0.2317	0.00241	−−−−−+?−−−−??+−	10	106 622 009	SORCS3	Intron 2	4	7 417 241	SORCS2	Intron 2
rs3976793	rs7684383	ac	gt	0.6954	0.2332	0.00287	++++−−?++++??−+	10	106 616 736	SORCS3	Intron 2	4	7 417 241	SORCS2	Intron 2
rs12259189	rs7684383	cc	tt	−0.6935	0.2333	0.00295	−−−−++?−−−−??+−	10	106 615 387	SORCS3	Intron 2	4	7 417 241	SORCS2	Intron 2
rs6584629	rs7684383	ac	gt	0.6881	0.2349	0.00341	++++−−?++++??−+	10	106 608 435	SORCS3	Intron 2	4	7 417 241	SORCS2	Intron 2
rs12249460	rs7684383	ac	gt	0.9499	0.3247	0.00345	+++++−?+++−??−+	10	106 605 440	SORCS3	Intron 2	4	7 417 241	SORCS2	Intron 2
rs749304	rs4234804	aa	gg	0.4236	0.125	0.00071	+++++−?++++??++	10	106 990 399	SORCS3	Intron 20	4	7 417 632	SORCS2	Intron 2
rs10884126	rs4234804	aa	gg	−0.4137	0.1223	0.00072	−−−−−−?−−−−??−−	10	107 025 028	SORCS3	NearGene-3	4	7 417 632	SORCS2	Intron 2
rs12263804	rs4234804	ca	tg	0.4225	0.125	0.00072	+++++−?++++??++	10	106 993 770	SORCS3	Intron 20	4	7 417 632	SORCS2	Intron 2
rs3750261	rs4234804	ca	tg	0.413	0.1223	0.00073	++++++?++++??++	10	107 023 390	SORCS3	UTR-3	4	7 417 632	SORCS2	Intron 2
rs12262245	rs4234804	ca	gg	0.6971	0.2313	0.00258	+++++−?++++??−+	10	106 621 704	SORCS3	Intron 2	4	7 417 632	SORCS2	Intron 2
rs7086583	rs4234804	aa	cg	0.6958	0.2312	0.00262	+++++−?++++??−+	10	106 622 009	SORCS3	Intron 2	4	7 417 632	SORCS2	Intron 2
rs3976793	rs4234804	aa	gg	−0.6885	0.2327	0.00309	−−−−++?−−−−??+−	10	106 616 736	SORCS3	Intron 2	4	7 417 632	SORCS2	Intron 2
rs12259189	rs4234804	ca	tg	0.6867	0.2328	0.00318	++++−−?++++??−+	10	106 615 387	SORCS3	Intron 2	4	7 417 632	SORCS2	Intron 2
rs6584629	rs4234804	aa	gg	−0.6823	0.2344	0.0036	−−−−++?−−−−??+−	10	106 608 435	SORCS3	Intron 2	4	7 417 632	SORCS2	Intron 2
rs12249460	rs4234804	aa	gg	−0.9405	0.3241	0.00371	−−−−−+?−−−+??+−	10	106 605 440	SORCS3	Intron 2	4	7 417 632	SORCS2	Intron 2

Abbreviations: A1, Allele 1; A2, Allele 2; Beta, beta coefficient; CHR, chromosome; DIR, direction of effect; location, single nucleotide polymorphism (SNP) LD, linkage disequilibrium; location; s.e., standard error of beta coefficient; SORCS1, sortilin-related VPS10 domain-containing receptor 1.

P-value cutoff for significance after correction for multiple testing: P=0.0001.

Exons 1–18 encode the VPS10 domain.

Figure 1

Black arrows: (a) The linkage disequilibrium (LD) block in sortilin-related VPS10 domain-containing receptor 3 (SORCS3) showing significant interaction with (b) two specific LD blocks in SORCS2 and (c) one specific LD block in SORCS1. Red arrows: Additional single nucleotide polymorphisms (SNPs) showing epistasis between SORCS2 and SORCS3. (a) The single LD block in SORCS3 showing epistasis with two specific LD blocks in SORCS2 (Figure 3b) and one specific LD block in SORCS1 (Figure 3c). (b) The two specific two LD blocks in SORCS2 (black arrows) showing epistasis with SORCS3, and SNPs showing in addition epistasis with SORCS1 (red arrows). (c) The specific LD block in SORCS1 (black arrows) showing epistasis with the specific LD block in SORCS3, and SNPs showing epistasis with SNPs in SORCS2 (red arrows).

Eight pairs resembled additional epistatic effects between SORCS1/SORCS3 and between SORCS1/SORCS2 (Table 3b). Of note, the single SORCS3 SNP (rs1670036) interacting with SORCS1 is located in the specific LD block also showing interaction with SORCS2 (Figure 1a), and again all SNPs constituting the LD block in SORCS1 are located in introns 1 and 2 (Figure 1c) adjacent to the exons encoding the VPS10 domain. The epistasis β for these SORCS1/SORCS3 and SORCS1/SORCS2 interactions ranged from −0.39 to 0.40, again reflecting larger effects (additive) than the corresponding single-marker analyses (−0.15<β<0.20). Although the P-values for epistatic effects just missed the multiple testing threshold of 0.0001, the number of significant interactions was clearly higher than expected by chance (expected SORCS2/SORCS3 interactions: 10.8). Forest plots for the SNPs with epistatic effects or strongest individual associations with AD (rs7665496, rs6840423, rs7684383, rs4234804, rs1670036, rs1792124, rs12248379, rs13110208) are shown in Supplementary Figure S2.

Microarray gene expression and quantitative RT-PCR analyses

Microarray expression analyses showed lower expression of SorCS3 in AD brains compared with control brains (mean gene expression intensity: 4.17±0.43 vs 5.03±0.49 (P=5.1E−5; Figure 2a), in line with what we had previously observed in SORL1 and SORCS1.[2, 3] To validate the significant results of the Affymetrix array, we conducted a quantitative RT-PCR for the SORCS3 gene, using brain tissue from the amygdala region. Calculation of the fold change rate with the Relative Quantitation method of DataAssist software confirmed the results of the expression array for all the three investigated SORCS3 exons. Compared with the control samples, the AD samples showed significantly reduced expression of exons 10, 17 and 21 of SORCS3; 87% (P=0.012), 74% (P=0.003) and 83% (P=0.003), respectively (Supplementary Figure S3). Notably, these findings were also validated by comparison with publicly available gene expression results (188 cases, 176 controls, P<0.0001, http://labs.med.miami.edu/myers/).[30] We did not find a significant difference between AD and controls in the expression levels of SORCS2 or SORT1 (Figures 2b and c). There was no significant difference in the expression levels between AD and control brains in brain tissue from regions unaffected by the disease process (occipital lobe, cerebellum) for any of the homologs.

Figure 2

(a) View of sortilin-related VPS10 domain-containing receptor 3 (SORCS3) exon expression profile in 19 Alzheimer's disease (AD; red triangles) and 10 control (CTRL; blue squares) amygdala tissue. Each triangle dot represents least squares mean expression of an exon in AD tissue; each square dot represents least squares mean expression of an exon in control tissue. The mean gene expression intensity of AD vs controls was 4.17±0.43 vs 5.03±0.49 (P=5.1E-5) across all exons. (b) View of SORCS2 exon expression profile in 19 AD (red triangles) and 10 control (blue squares) amygdala tissue. Each triangle dot represents least squares mean expression of an exon in AD tissue; each square dot represents least squares mean expression of an exon in control tissue. The mean gene expression intensity of AD vs controls was 5.29±0.33 vs 5.08±0.32 (P=0.12) across all exons. (c) View of sortilin 1 (SORT1) exon expression profile in 19 AD (red triangles) and 10 control (blue squares) amygdala tissue. Each triangle dot represents least squares mean expression of an exon in AD tissue; each square dot represents least squares mean expression of an exon in control tissue. The mean gene expression intensity of AD vs controls was 9.43±0.38 vs 9.21±0.39 (P=0.17) across all exons.

cDNA transfection of Vps10 family members in Hek 293 cells demonstrated a significant (0.01⩽P⩽0.05) decrease of γ-secretase (APP-GV), whereas there was no effect on the AICD-GV translocation assay (Figure 3).

Figure 3

cDNA overexpression in transfected HEK293 cells. AICD, APP gene's C-terminus; APP, amyloid precursor protein; GV, GAL4 DNA-binding domain with VP16 transcriptional activator; SORCS, sortilin-related VPS10 domain-containing receptor; SORLA, sortilin-related receptor.

γ-Secretase processing

In HEK293 cells (Figure 4), 4/4 SorCS2-shRNAs, 3/3 SorC3-shRNAs and 2/3 SORT1-shRNAs caused a significant increase greater than threefold in APP processing (from P<0.001 to P<0.05) as compared with the result with the scrambled shRNA (analysis of variance with Bonferroni correction) while not affecting the nuclear translocation of the control AICD-GV only-fragment.

Figure 4

γ-Secretase activity and nuclear translocation of amyloid precursor protein (APP) assays with sortilin-related VPS10 domain-containing receptor 2 (SORCS2), SORCS3 or SORT1 short hairpin RNAs (shRNAs). Both the APP-GV (GAL4 DNA-binding domain with VP16 transcriptional activator) and APP gene's C-terminus (AICD)-GV assay were performed in HEK293 cells. The data from SorCS2, SorCS3 or sortilin 1 (SORT1) shRNA was normalized to either APP-GV only or AICD-GV with the scrambled sequence shRNA (shRNA-scrambled), which was included as a negative control. The data are representative for the assays, were performed in ⩾3 experiments in replicates of eight samples per condition (96-well format, s.d. bars are shown, *P<0.05, **P<0.01, ***P<0.001 as compared with APP-GV only (analysis of variance, Bonferroni's correction).

Discussion

Taken together with previous studies,[2, 3] the findings here indicate that variants in SORL1, SORCS1, SORCS2 and SORCS3 of the VPS10-d receptor family are associated with AD risk. The results are consistent with previous studies showing associations between the SNPs in SORL1 and SORCS1 with AD[2, 3, 11, 13, 14, 15, 16, 18] and with cognitive performance.[10] Similar to previous reports, the associated SNPs in SORCS1, SORCS2 and SORCS3 were mostly located in introns 1–3, implicating the VPS10 domain.

The effect sizes of the associated SNPs were small (β: −0.45 to 0.36), but this is consistent with previous observations for the homologous genes SORL1 and SORCS1[2, 3, 18] as well as all recently detected novel AD susceptibility loci identified by large genome-wide association studies.[23, 31, 32, 33, 34]

The epistasis models of SNPs significant in single-marker analyses further revealed pairwise SNP associations between specific LD blocks in the highly homologous SORCS1, SORCS2 and SORCS3. One single LD block in SORCS3 showed epistasis with both a single LD block in SORCS1 and two specific LD blocks in SORCS2. In addition, the same two regions of SORCS2 and SORCS3 interacted. Of note, the epistasis β ranging from −0.94 to 0.94 reflected larger effects (additive) than the corresponding single-marker analyses (−0.15<β<0.20), and the interacting SNPs are almost exclusively located in introns 1 and 2, adjacent to the exons encoding the VPS10 domain. This region has also been demonstrated to include the majority of disease-associated SNPs for both SORCS1 and SORL1.[2, 3, 10, 18, 20] Our findings indicate that there are sequences within these specific LD blocks that are biologically important and that are interacting. The mechanism underlying these interactions is presently unclear. It could arise from direct interaction between the homologs, interactions with a mutual binding partner or interactions with a common substrate, such as APP or APP-CTF. However, it could also result from quite remote interactions that do not require a first- or second-order interaction between these proteins.

Suppression of SORCS2, SORCS3 or SORT1 increased γ-secretase processing of APP, findings consistent with reported effects by SORL1 and SORCS1 on γ-secretase processing of APP and changes in Aβ40 and Aβ42 levels.[2, 3] Although SorCS3 and SorCS1 do not convey trans-Golgi network to late endosome sorting,[35, 36] SORT1 is—similar to SORL1—also capable of mediating sorting of ligands from the trans-Golgi network to late endosomes or lysosomes.[22] Interestingly, sortilin-mediated endocytosis has been shown to determine levels of progranulin involved in frontotemporal dementia.[37] Results from the current genome-wide association studies suggest that other genes PICALM, BIN1 and CD2AP, modulate intracellular trafficking of cell surface proteins. Thus, it is appears that in addition to their effect on γ-secretase processing of APP, some members of the VPS10-d receptor family exert their effect on AD through modulation of APP trafficking.

Reduced expression of SORCS3 may be a secondary effect of the disease, but it is consistent with the γ-secretase assays, which indicated that suppression of SORCS3 activates Aβ production. If correct, this would provide a potential explanation for how downregulation of SORCS3 might increase risk for AD. We are unable to see any difference in the expression of SORCS2 and SORT1. However, we cannot yet exclude the possibility that this was the result of the small sample size or chosen phenotype. In a previous study by Mufson et al.,[38] SORT1 levels were—consistent with our findings—not associated with clinical diagnosis or antemortem cognitive test scores. However, there was an association with severity of neuropathology by Braak and NIA-Reagan diagnoses.

The significant strengths of this study are the large sample size, allowing us to detect small effects and explore epistasis. Limitations include that the SNPs assessed were derived from available genome-wide arrays. Thus, they do not cover the complete genetic variation in these genes, and it is possible that there are additional disease-associated markers that have not been genotyped. It is also possible that we lacked the power to detect additional disease-associated markers or interactions of SNPs with lower allele frequencies or effect sizes.

Taken together, our results indicate that in addition to SORL1 and SORCS1, the variants in other members of the VPS 10-D receptor family (SORCS2, SORCS3 and SORT1) are associated with AD either independently or through epistatic mechanisms.