Hypothesis-based investigation of known AD risk variants reveals the genetic underpinnings of neuropathological lesions observed in Alzheimer’s-type dementia

Abstract

Alzheimer’s disease (AD) is the leading cause of dementia worldwide. Besides neurofibrillary tangles and amyloid beta (Aβ) plaques, a wide range of co-morbid neuropathological features can be observed in AD brains. Since AD has a very strong genetic background and displays a wide phenotypic heterogeneity, this study aims at investigating the genetic underpinnings of co-morbid and hallmark neuropathological lesions. This was realized by obtaining the genotypes for 75 AD risk variants from low-coverage whole-genome sequencing data for 325 individuals from the Leuven Brain Collection. Association testing with deeply characterized neuropathological lesions revealed a strong and likely direct effect of rs117618017, a SNP in exon 1 of APH1B, with tau-related pathology. Second, a relation between APOE and granulovacuolar degeneration, a proxy for necroptosis, was also discovered in addition to replication of the well-known association of APOE with AD hallmark neuropathological lesions. Additionally, several nominal associations with AD risk genes were detected for pTDP pathology, α-synuclein lesions and pTau-related pathology. These findings were confirmed in a meta-analysis with three independent cohorts. For example, we replicated a prior association between TPCN1 (rs6489896) and LATE-NC risk. Furthermore, we identified new putative LATE-NC-linked SNPs, including rs7068231, located upstream of ANK3. We found association between BIN1 (rs6733839) and α-synuclein pathology, and replicated a prior association between USP6NL (rs7912495) and Lewy body pathology. Additionally, we also found that UMAD1 (rs6943429) was nominally associated with Lewy body pathology. Overall, these results contribute to a broader general understanding of how AD risk variants discovered in large-scale clinical genome-wide association studies are involved in the pathological mechanisms of AD and indicate the importance of downstream elimination of phenotypic heterogeneity introduced in these studies.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00401-024-02815-w.

Introduction

In the broad spectrum of modern-day diseases, few are so complex and elude a complete understanding as Alzheimer’s disease (AD). This insidious form of dementia imposes a grave burden upon the daily lives of patients and family. The first disease-modifying treatments are now emerging [59, 63]. However, they only focus on one aspect of the pathological profile of AD and have limited efficacy in slowing cognitive decline. This indicates that the pathogenesis underlying the disease remains to be unraveled further. Besides the development of novel treatments, the extensive preclinical phase of AD offers significant opportunity for early detection and intervention [56]. Therefore, the strong genetic background of AD, with an estimated heritability of 60–80% [7], offers added value for genetic studies to identify early therapeutic targets and risk genes underlying disease mechanisms to elucidate the biological pathways involved in the development of the disease and aid in the discovery of biomarkers.

A definite diagnosis of AD can only be made after a post-mortem examination in which the presence of the neuropathological hallmarks of AD is confirmed. These hallmarks comprise extracellular plaques of aggregated Aβ_1-42 protein and intracellular, neurofibrillary tangles composed of hyperphosphorylated tau protein (pTau) [33]. These hallmark lesions are accompanied by a general state of neuroinflammation, ultimately resulting in neuronal loss. Tangles and plaques are rarely the sole neuropathological phenotypes observed in the AD brain. Additional co-morbid lesions are present in the majority of cases which could have a synergistic effect on the clinical presentation and disease progression.

Cerebral amyloid angiopathy (CAA) is a frequently observed co-morbidity in AD and is characterized by the accumulation of Aβ in blood vessel walls, occurring in over 90% of AD cases [32]. There are two types of CAA, Type I and Type II. CAA type I distinguishes itself from type II by the presence of Aβ deposits in the cortical capillaries, additional to Aβ inclusions in the leptomeningeal and cortical vessels [61]. Genetically, a distinction can be made since the frequency of APOE ε4 alleles is up to four times higher in CAA type I cases compared to type II and controls. CAA in the brain of AD cases can contribute to the clinical gravity of the disease by increasing risk for weakening of the vessel wall, eventually resulting in ruptures. Granulovacuolar degeneration (GVD) bodies are membrane-bound vacuolar structures containing dense protein granules. GVD bodies are most frequently observed in hippocampal pyramidal neurons of patients with AD and other tauopathies. It was hypothesized that GVD can be induced by tau accumulation [66] and the two pathologies often co-occur in the brain. Evidence has also shown a direct relation between GVD and a cell-death mechanism referred to as necroptosis, posing that GVD is in fact a proxy for necroptosis [38]. Furthermore, GVD is associated with limbic-predominant age-related TDP-43 encephalopathy neuropathological changes (LATE-NC), with LATE-NC being suspected to aggravate GVD pathology [37]. LATE-NC is a pathology where phosphorylated transactive-response DNA binding protein 43 (pTDP-43) accumulates in the neuronal cytoplasm. TDP-43 inclusions in the brain are hallmark lesions of frontotemporal lobe degeneration with TDP-43 brain pathology (FTLD-TDP) and amyotrophic lateral sclerosis (ALS). However, inclusions are also observed in up to 70% of symptomatic AD cases [43]. Genetically, LATE-NC has been associated with 5 risk genes: TMEM106B, GRN, KCNMB2, ABCC9 and APOE [48]. TDP-43 inclusions have been associated with a more severe clinical profile of AD [34], indicating that it is an important co-morbidity in AD pathology. Another pathology coinciding with AD is the presence of α-synuclein inclusions observed as Lewy bodies and Lewy neurites [57]. Lewy body dementia (LBD) accounts for an average 5% of reported dementia cases [35] and can occur alongside AD, resulting in a form of mixed dementia. Reports show that up to 50% of AD cases have co-morbid Lewy bodies in the neocortical and limbic brain regions [23]. Several genes have been linked to Lewy body dementia with some overlap with the genetic risk profile of AD and PD as well as some unrelated genes. Some examples are APOE, SNCA, GBA and LRRK2 [50].

Genome-wide association studies (GWAS) are rapidly discovering common (MAF > 0.01) genetic variation associated with disease. Some examples of the earliest associated loci are BIN1, CLU, CR1 and PICALM [26, 40, 71]. Now, over 75 genetic loci have been associated with sporadic AD [5]. Pathway enrichment analysis of genes located in or prioritized to these loci has highlighted a variety of biological pathways, pointing to intuitive disease mechanisms, like APP and tau processing, as well as pathways like endocytosis, lipid metabolism and innate immunity [31, 39, 65, 69]. GWASs for AD have mainly been using clinical AD diagnosis as the predicted variable in regression models due to the necessity of ever-increasing sample sizes to gain sufficient statistical power for detecting more significant relations. Another valuable approach would be to use specific endophenotypes observed in AD pathogenesis. Endophenotypes have proven to add value for risk variant discovery using them as predicting variables in regression models in GWASs. Investigating association with endophenotypes, like CSF or serum biomarker levels and brain imaging has allowed the discovery of novel risk loci and confirmation of various loci discovered with case–control designed studies [27, 30, 41]. In this study, a hypothesis-based investigation of several known AD risk genes and their contribution to individual neuropathological lesions was performed in a European cohort of 325 individuals aged 50 or older at death. Our analyses highlight the importance of APOE in AD pathogenesis and uncover additional strong associations between a known AD risk variant situated in APH1B and several distinct neuropathological phenotypes.

Methods

Cohort

The analyses were performed on the Leuven Brain Collection, which includes human brain tissue received from university or municipal hospitals in Leuven (Belgium), Bonn, Offenbach am Main and Ulm (Germany). Collection was in accordance with the local ethical committee guidelines and laws regarding the use of human tissue for research. Brains of 325 individuals were included in the present study, aged 50 years or older at death. Individuals were labeled as AD when they were clinically diagnosed with dementia (based on the CDR score) and had at least an intermediate degree of AD neuropathological change (ADNC) according to NIA-AA criteria (ABC score) [28]. Individuals without dementia but with ADNC were labeled as asymptomatic AD. Individuals with no ADNC were labeled as ‘noAD’, meaning they had no AD pathology, but not excluding the presence of other pathologies, thus comprising individuals with healthy brains as well as individuals suffering from neurodegenerative brain diseases without ADNC. For each individual, 500 mg fresh frozen cerebellar tissue was stored in Biobank Antwerp, Antwerp, Belgium; ID: BE 71030031000 for DNA extraction.

Staging and staining of neuropathological lesions

Staging and staining of the lesions was performed at the Laboratory of Neuropathology at the KU Leuven. Neurofibrillary tangles were stained using an AT8 phospho-tau monoclonal antibody (Thermo Fisher Scientific, United States) in a 1:1000 dilution and/or with the Gallyas silver staining method. Braak stage was determined as described elsewhere, with stages I and II indicating the presence of tangles in the entorhinal cortex, stage III and IV the spread to hippocampus and the limbic system, stage V spread to the neocortex and finally the most advanced stage VI where tangles are numerously present throughout the neocortex [10]. CERAD score was also assessed on a staining with an AT8 antibody, scored 0-III ranging from no neuritic plaque pathology to severe and widespread neuritic plaques [44]. The use of the AT8 phospho-tau immunohistochemistry for detecting neuritic plaques has been chosen since immunohistochemical stainings are more reproducible than silver methods such as the Bielschowsky staining [2] and the consensus guidelines allow Gallyas or anti-tau immunohistochemistry for neuritic plaque detection [29]. Aβ staining was performed using an 4G8 antibody against Aβ_17-24 (Biolegend, USA) in a 1:5000 dilution. Aβ score in the medial–temporal lobe (MTL) reflects the spread of Aβ in the MTL in four phases [62]. The severity score of CAA was determined according to Vonsattel et al. [64]. TDP43 lesions were stained using a rabbit antibody against pTDP-43 S409/10 (Cosmobio Co., Japan) in a 1:5000 dilution after heat pretreatment at pH 6. pTDP staining was used to determine LATE-NC stages as well as GVD stages. LATE-NC was staged according to consensus guidelines [48] with three distinct stages: in stage I, TDP lesions are restricted to the amygdala or are present only in the hippocampus, stage II indicates lesions in both amygdala and hippocampus and in stage III the middle frontal gyrus is involved. GVD stages were determined based on earlier described guidelines [49]. In stage I, GVD lesions are restricted to the CA1/2-subiculum region, in stage II, the entorhinal cortex and/or CA3-4 are affected together with the basal nucleus of Meynert. Stage III indicates additional involvement of the temporal neocortex and in stage IV, the amygdala and/or hypothalamus are affected as well. In stage V, the pathology is widespread and affects the oral raphe nuclei, pedunculopontine tegmental nucleus, cingulate gyrus, thalamus, and the frontal and parietal neocortex. Inclusions of α-synuclein were stained using an anti-α-synuclein antibody, clone 5G4 (Merck Life Science, Belgium) in a 1:2000 dilution. Staging was performed following the Braak scoring system for Parkinson’s disease (PD) [11]. Stage I describes Lewy bodies restricted to the olfactory bulb and dorsal motor nucleus of the vagal nerve. In stage II, the brainstem and locus coeruleus become involved, whereas in stage III, there is a spread of the pathology to the substantia nigra and other brainstem regions. Stage IV indicates spread of Lewy body pathology to the limbic system, affecting the hippocampus and the amygdala. In stage V, the neocortex is affected and in the most advanced stage, stage VI, the pathology is widespread throughout the neocortex. Finally, general AD pathology was described by the NIA-AA classification system (ABC score), which is a score compiled by Braak NFT score, Aβ score and CERAD score, indicating the level of AD pathology (absent, low, intermediate or high) [28].

DNA extraction and quality control

DNA was extracted from fresh frozen cerebellar tissue using a QIAGEN DNeasy 96 Blood & Tissue kit (QIAGEN, Benelux). After extraction, an AMPure XP magnetic bead (Beckman Coulter life sciences, Analis, Belgium) purification was done to remove leftover waste products from the DNA extraction process. Concentration of the extracted DNA was measured by UV/Vis (Trinean NV, Belgium) and normalized to 100 ng/µl based on initial dsDNA concentrations. Normalized plates were subjected to an additional concentration measurement using a Qubit Quant-it kit (Thermo Fisher Scientific, Belgium) for validation. Quality of the extracted DNA was investigated using an Agilent Fragment Analyzer Capillary array (Agilent Technologies, Belgium). The DNA was checked for fragment length and contamination peaks. In case of insufficient quality (concentration < 1 ng/μl), DNA extraction was repeated. APOE genotyping was performed on all individuals and for participants with pre-existing information on APOE genotype, cross-validation was performed.

AVITI low-coverage whole-genome sequencing

Genotype information was acquired using a low-coverage whole-genome sequencing approach. Library preparation was performed using the xGen cfDNA & FFPE DNA Library Preparation Kit (IDT, Belgium) with complementary xGen UDI 10nt primers (IDT, Belgium) for adding a specific barcode to each of the samples. The xGen cfDNA & FFPE kit utilizes a double adaptor ligation step, minimizing chimera or dimer formation, and is tailored for low-input materials. Afterward, library concentration was measured, and samples were pooled equimolarly. Sequencing was performed at the VIB Nucleomics Core (Ghent, Belgium) on an AVITI instrument (Element Bioscience, San Diego) which employs avidity chemistry for sequencing [3]. In short, this technique employs rolling circle amplification, creating DNA polonies. Subsequently, nucleotide identification is achieved by fixing multivalent nucleotide ligands on dye-labeled cores which form polymerase–polymer–nucleotide complexes, then binding the polonies. The resulting fluorescent signal can be detected, and base-calling can be performed from the raw fluorescent signals. AVITI sequencing has improved data quality and provides flexibility in the workflow, allowing the adjustment of the library pools in between sequencing runs. For our purpose, 325 samples were sequenced at an average depth of 1 × coverage, followed by imputation.

Data processing and imputation

Raw FASTQ files were aligned to GRCh38 genome build using the Burrows–Wheeler Alignment (BWA) tool. Alignment was followed by imputation using GLIMPSE (Genotype Likelihoods IMputation and PHasing mEthod) [53], a set of tools used for phasing and imputation of low-coverage whole-genome sequencing data. GLIMPSE utilizes a matrix of genotype likelihoods at all variable positions of the reference panel as input. Genotype likelihoods are then refined by iterative genotype imputation and haplotype phasing utilizing the Gibbs sampling procedure [53]. For this study, the 1000 Genomes haplotype reference panel [4] was employed. Data quality control was carried out with PLINK version 2.0, filtering out positions with a low imputation quality score (R² < 0.3) and rare variants (MAF < 0.01). Relatedness in the population was investigated, but no sample pairs exceeded a PI-HAT threshold of 0.125. From the 325 individuals, allele dosages were extracted for 85 AD risk variants [83 index variants described by Bellenguez et al. and two APOE SNPs, rs429358 and rs74121 (Supplementary Table 1)]). The imputation pipeline could not perform imputation on indels, excluding four variants (rs139643391, rs1160871, rs35048651 and rs149080927) from downstream analysis. Additionally, five variants did not pass the MAF threshold (rs143080277, rs72824905, rs141749679, rs75932628 and rs60755019) as expected based on earlier reported MAF. Finally, one variant in the IGH gene cluster locus was not imputed and thus also not included in further analysis (rs7157106).

Genotype validation with ONT multiplex assay

Imputation following low-coverage whole-genome sequencing (lcWGS) data results in genotype likelihoods rather than hard-called genotypes, entailing the use of allele dosages in downstream regression modeling. In order to assess the reliability of the genotypes assigned by GLIMPSE based on genotype likelihoods, an alternative genotyping strategy was applied to investigate the concordance for SNPs with significant associations. The selected SNPs were genotyped utilizing an in-house designed Oxford Nanopore Technology (ONT) multiplex assay. Primers for all SNPs were developed using primerBLAST. Primer pairs were pooled together in one plex. In short, a multiplex PCR (mPCR) was carried out per 96 samples with previously tested optimized conditions. PlatinumTaq polymerase was used as a DNA polymerase. After mPCR, a barcoding PCR (bPCR) was performed resulting in the attachment of a unique barcode to the amplified DNA fragments. Tapestation (Agilent Technologies) was utilized for assessment of sample length and concentration. Consecutively, samples were pooled equimolarly. Library preparation was performed using the ligation sequencing kit LSK110 (Oxford Nanopore Technology) according to the accompanying protocol. Sequencing was performed on a Flongle flow cell (Oxford Nanopore Technology). Sequencing reads were aligned to GRCh38 genome assembly and variant calling was performed using Longshot [18], generating a phased VCF file used for downstream QC and data analysis.

Statistical analysis

Data QC and association analyses were performed with PLINK 2.0 (version v2.00a2.3LM). For the 12 investigated phenotypes, either linear or logistic regression was applied to assess the relation between known AD risk variants and neuropathological lesions. Linear regression was applied for semi-quantitative phenotypes (i.e., stages of pathology or severity scores), whereas logistic regression was utilized for binary phenotypes (presence or absence of a certain lesion). Regression models were built, correcting for age and sex, using allele dosages as an independent variable. Bonferroni correction was applied for multiple testing with a significance cutoff of p = 0.0006, correcting for the number of SNPs that was independently investigated (n = 75). For the three SNPs where true significant associations were discovered, a concordance check was performed comparing the hard-called genotypes of the SNPs between the distinct sequencing approaches. To explore the relationship between neuropathological lesions and AD risk SNPs, construction of a graphical lasso model [21] and network visualization was accomplished using glasso and qgraph packages in R studio respectively. The model included significantly associated genotypes and phenotypes. Variables in the model were standardized followed by computation of the Spearman partial correlation matrix. Graphical lasso was then used to generate the sparse inverse covariance matrix applying an L1 penalty to promote sparsity. The regularization parameter (ρ) was chosen based on the Extended Bayesian Information Criterion (EBIC) [20] using the recommended EBIC hyperparameter (γ = 0.5). The robustness of the model was investigated by testing stability with the corStability function from the bootnet R package [19].

Meta-analysis

Validation of significant genotype–phenotype associations was realized by performing a meta-analysis of the discovery cohort summary statistics and published summary statistics [36] of the Religious Orders Study/Memory and Aging Project (ROSMAP) cohort [6], Alzheimer’s Disease Sequencing Project (ADSP) and Alzheimer’s Disease genetic Consortium (ADGC) [14, 47]. In these cohorts, association of AD risk variants was performed with dichotomized neuropathological phenotypes. Therefore, we harmonized the neuropathological phenotypes from the Leuven Brain Collection prior to meta-analysis. More specifically, for pTau pathology, Braak NFT stages 0–IV were set to ‘0’ and Braak NFT stages V–VI were referred to as ‘1’. For the neuritic plaque score, CERAD scores 0–II were converted to ‘0’ and a CERAD score of III was set to ‘1’. Neuropathological data regarding TDP43 pathology from LATE-NC stage, pTDP in CA1 and pTDP in dentate gyrus were combined into one score reflecting the total absence (0) or presence (1) in any of these variables of pTDP lesions. For α-synuclein, the discovery cohort phenotype reflecting the general presence (1) and absence (0) of α-synuclein pathology was used. Employing these four dichotomized variables, association testing was performed in PLINK 2.0 (version v2.00a2.3LM) correcting for age and gender. Resulting summary statistics were used for meta-analysis using METAL (version 2010-03-25) employing the standard scheme ‘SAMPLESIZE’, implementing the studies' respective sample sizes as weight for the meta-analysis. Respective sample sizes varied for each phenotype based on the availability of phenotypic data. Variants selected for meta-analysis had at least one association with p < 0.05 for a specific phenotype in the discovery association analysis.

Results

Table 1 presents the demographic and neuropathological characteristics of the samples of the Leuven Brain Collection included in this study. Of 325 individuals, 98 had ADNC and a clinical diagnosis of AD, 113 had ADNC without clinical AD, and 114 had no ADNC. This latter group included true controls but also patients with other neurodegenerative brain diseases, such as ALS, FTLD, PD, vascular dementia, and mixed neuropathological profiles. The samples were not systematically screened for known causal mutations in genes implicated in neurodegenerative brain diseases.

Table 1.

Demographic and neuropathological characteristics of the study cohort of 325 individuals

	Total (n = 325)	No AD (n = 114)	Asymptomatic AD (n = 113)	AD (n = 98)
Demographics
Age (Mean ± stdev)	71 ± 10	66 ± 10	72 ± 9	76 ± 9
Sex (% female)	58.7	51	66.3	58
APOE ε3/3 (%)	57.9	68.4	55.8	48
APOE ε3/4 (%)	24.9	12.3	29.2	34.7
APOE ε4/4 (%)	4.3	0	0.88	13.3
APOE ε2/3 (%)	11.7	17.5	12.4	4.1
APOE ε2/4 (%)	0.62	0	1.8	0
APOE ε2/2 (%)	0.62	1.8	0	0
Braak NFT stage
Braak 0 (%)	8.3	20.2	2.7	1
Braak 1 (%)	36	48.3	42.5	14.3
Braak 2 (%)	25.2	22	33.6	19.4
Braak 3 (%)	13.5	8.8	17.7	14.3
Braak 4 (%)	6.2	0.9	2.7	16.3
Braak 5 (%)	5.5	0	0.9	17.4
Braak 6 (%)	5.2	0	0	17.4
Absent (%)	0	0	0	0
Aβ score
Aβ score 0 (%)	38.5	100	8	2
Aβ score 1 (%)	15.1	0	31.9	13.3
Aβ score 2 (%)	14.2	0	30.1	12.2
Aβ score 3 (%)	12.6	0	17.7	21.4
Aβ score 4 (%)	19.4	0	12.4	50
Absent (%)	0.31	0	0	1
Granulovacuolar degeneration
Stage 0 (%)	53.5	76.3	59.3	20.4
Stage 1 (%)	20	12.3	27.4	20.4
Stage 2 (%)	8.6	7	4.4	15.3
Stage 3 (%)	4.6	0	5.3	9.2
Stage 4 (%)	7.4	0.9	3.5	19.4
Stage 5 (%)	4	0	0	13.3
Absent (%)	1.9	3.5	0	2
TDP-LATE
Stage 0 (%)	61.2	57.9	69.9	55.1
Stage 1 (%)	4.6	3.5	4.4	6.1
Stage 2 (%)	3.4	1.8	0.9	8.2
Stage 3 (%)	2.5	0	0.9	7.1
Absent (%)*	28.3	36.8	23.9	23.5
TDP presence
TDP in CA1 (%)	14.5	11.4	8.9	24.5
Absent (%)	4	8.8	1.8	1
TDP in dentate gyrus (%)	10.8	10.5	5.3	17.4
Absent (%)	3.7	8.8	0.9	1
Braak α-synuclein stage
Braak 0 (%)	74.5	80.7	77.9	63.3
Braak 1 (%)	7.4	8.8	7.1	6.1
Braak 2 (%)	1.5	0	1.8	3.1
Braak 3 (%)	2.5	3.5	1.8	2
Braak 4 (%)	5.2	2.6	4.4	9.2
Braak 5 (%)	2.8	0	0.9	8.2
Braak 6 (%)	4.9	2.6	4.4	8.2
Absent (%)	1.2	1.8	1.8	0
α-synuclein presence
α-synuclein positive (%)	24.6	17.5	20.4	37.8
Absent (%)	1.2	1.8	1.8	0
α-synuclein outside brainstem (%)	13.2	5.3	9.7	26.5
Absent (%)	1.2	1.8	1.8	0
CAA severity score
CAA severity 0 (%)	59.4	93.9	55.8	23.5
CAA severity 1 (%)	15.1	3.5	23.9	18.4
CAA severity 2 (%)	22.8	1.8	18.6	52
CAA severity 3 (%)	2.8	0.9	1.8	6.1
Absent (%)	0	0	0	0
CERAD neuritic plaque score
CERAD 0 (%)	74.5	100	84.1	33.7
CERAD 1 (%)	9.9	0	12.4	18.4
CERAD 2 (%)	8.9	0	1.8	27.6
CERAD 3 (%)	6.8	0	1.8	20.4
Absent (%)	0	0	0	0
ABC score
ABC 0 (%)	35.1	100	0	0
ABC 1 (%)	42.5	0	87.6	39.8
ABC 2 (%)	12	0	11.5	26.5
ABC 3 (%)	10.5	0	0.9	33.7
Absent (%)	0	0	0	0

Percentages in each column show distribution of the lesions over the different stages for the respective phenotype and disease group (no AD, asymptomatic AD or AD). Absent shows the percentage of individuals for whom phenotype data are missing per neuropathological lesion

*TDP-LATE staging was not performed in cases with FTLD or ALS

Hypothesis-based analysis on 75 AD risk variants

Regression models were generated for 12 neuropathological phenotypes, summarized in Table 1. The distinct phenotypes can be mapped to different parts of the AD neuropathological profile. These distinct parts being the hallmark pathological features of AD, such as the severity of AD pathology (ABC score), neuritic plaques (CERAD score), Aβ (Aβ score and CAA score) and pTau (Braak NFT stage and CERAD score) pathology, as well as co-morbid granulovacuolar degeneration (GVD stage), α-synuclein (Braak PD stage), TDP43 (presence in posterior middle temporal lobe, presence in dentate gyrus and LATE-NC stages).

Regression modeling was done for 75 AD risk variants which passed quality control. A complete overview of all (nominally) significant associations is depicted in Fig. 1. An overview of all SNPs that were investigated is provided in Table 2 and summary statistics are listed in Supplementary Table 1.

Fig. 1.

Genotype–phenotype associations. 31 SNPs for which at least one nominal association with a neuropathological phenotype was discovered, are shown on the y-axis. The 12 investigated neuropathological phenotypes are listed on the x-axis. Each data point on the graph represents information on an individual association test. The color is indicative of the direction and magnitude of the effect, while the shape itself indicates the p-value bin

Table 2.

Association between 75 AD risk SNPs and 12 neuropathological phenotypes

Phenotype	Chr	Variant ID	Gene	REF	ALT	Effect	Beta/OR	L95	U95	P value	R2
Braak NFT (n = 325)	19	rs429358	APOE	T	C	C	0.75	0.47	1.02	1.58 × 10–07	0.100
	15	rs117618017	APH1B	C	T	T	0.87	0.53	1.20	6.90 × 10–07	0.047
	8	rs73223431	PTK2B	C	T	T	0.31	0.09	0.53	0.005	0.024
	3	rs16824536	MME	G	A	G	0.67	− 1.19	− 0.14	0.013	0.026
	19	rs7412	APOE	C	T	T	− 0.48	− 0.91	− 0.04	0.032	0.005
	4	rs3822030	IDUA	G	T	T	0.25	− 0.47	− 0.02	0.033	0.025
	19	rs9304690	SIGLEC11	C	T	T	0.29	0.02	0.56	0.037	0.010
Aβ score (n = 324)	19	rs429358	APOE	T	C	C	1.01	0.74	1.28	1.71 × 10–12	0.155
	19	rs7412	APOE	C	T	T	− 0.88	− 1.31	− 0.45	7.22 × 10–05	0.031
	11	rs3851179	EED	T	C	T	− 0.34	− 0.58	− 0.10	0.005	0.014
	1	rs679515	CR1	T	C	T	0.37	0.09	0.65	0.009	0.037
	15	rs117618017	APH1B	C	T	T	0.44	0.09	0.79	0.015	0.011
	21	rs2830489	ADAMTS1	C	T	T	0.28	0.03	0.52	0.026	0.018
CERAD (n = 325)	19	rs429358	APOE	T	C	C	0.48	0.31	0.65	8.04 × 10–08	0.105
	15	rs117618017	APH1B	C	T	T	0.38	0.17	0.59	0.0005	0.028
	8	rs73223431	PTK2B	C	T	T	0.23	0.09	0.36	0.001	0.033
	19	rs7412	APOE	C	T	T	− 0.35	− 0.62	− 0.08	0.011	0.013
	20	rs6742	SLC2A4RG	T	C	C	− 0.20	0.03	0.36	0.02	0.013
	21	rs2830489	ADAMTS1	C	T	C	− 0.17	0.02	0.31	0.032	0.018
	7	rs13237518	TMEM106B	C	A	C	0.15	− 0.28	− 0.01	0.039	0.015
Braak PD (n = 321)	7	rs6943429	UMAD1	T	C	T	0.40	0.14	0.65	0.003	0.028
	2	rs6733839	BIN1	C	T	T	0.41	0.12	0.70	0.005	0.022
	10	rs7912495	USP6NL	A	G	G	0.36	0.10	0.63	0.008	0.023
	19	rs429358	APOE	T	C	C	0.38	0.04	0.73	0.030	0.023
	20	rs6742	SLC2A4RG	T	C	T	0.36	0.03	0.68	0.031	0.009
	15	rs3848143	SNX1	G	A	G	0.37	0.03	0.70	0.034	0.015
α-synuclein in CA1 (n = 321)	19	rs429358	APOE	T	C	C	1.77	1.13	2.75	0.012	0.026
	17	rs7225151	SCIMP	G	A	A	1.90	1.13	3.18	0.015	0.021
	7	rs6943429	UMAD1	T	C	T	1.54	1.08	2.19	0.018	0.017
	15	rs3848143	SNX1	G	A	G	1.57	1.01	2.43	0.045	0.139
α-synuclein beyond brainstem (n = 321)	10	rs7912495	USP6NL	A	G	G	2.03	1.24	3.32	0.005	0.026
	2	rs6733839	BIN1	C	T	T	1.97	1.17	3.31	0.011	0.017
	19	rs429358	APOE	T	C	C	1.89	1.11	3.22	0.020	0.028
	7	rs6943429	UMAD1	T	C	T	1.71	1.09	2.70	0.021	0.017
	2	rs17020490	PRKD3	T	C	C	2.00	1.02	3.91	0.044	0.008
	20	rs6742	SLC2A4RG	T	C	T	1.75	1.01	3.02	0.044	0.007
LATE-NC (n = 233)	10	rs7068231	ANK3	T	G	T	− 0.18	− 0.31	− 0.06	0.004	0.036
	16	rs450674	MAF	T	C	C	0.19	0.05	0.32	0.006	0.023
	19	rs429358	APOE	T	C	C	0.19	0.04	0.35	0.017	0.035
	12	rs6489896	TPCN1	T	C	C	0.32	0.06	0.57	0.017	0.036
	10	rs6586028	TSPAN14	C	T	C	− 0.19	− 0.36	− 0.03	0.021	0.024
	2	rs6733839	BIN1	C	T	T	0.16	0.02	0.30	0.022	0.021
	15	rs117618017	APH1B	C	T	T	0.20	0.02	0.39	0.034	0.010
	19	rs7412	APOE	C	T	T	− 0.26	− 0.51	− 0.01	0.045	0.011
pTDP-43 in CA1 (n = 321)	19	rs429358	APOE	T	C	C	2.05	1.23	3.41	0.006	0.035
	10	rs7068231	ANK3	T	G	T	0.54	0.33	0.89	0.015	0.019
	20	rs6742	SLC2A4RG	T	C	T	0.43	0.22	0.86	0.016	0.016
	2	rs6733839	BIN1	C	T	T	1.78	1.09	2.90	0.021	0.015
	9	rs1800978	ABCA1	C	G	G	2.00	1.04	3.87	0.038	0.013
	2	rs72777026	ADAM17	A	G	G	1.87	1.02	3.42	0.042	0.012
pTDP-43 in dentate gyrus (n = 313)	20	rs6742	SLC2A4RG	T	C	T	0.21	0.08	0.58	0.002	0.032
	9	rs1800978	ABCA1	C	G	G	2.44	1.18	5.02	0.016	0.018
	2	rs6733839	BIN1	C	T	T	1.94	1.12	3.36	0.019	0.018
	10	rs7068231	ANK3	T	G	T	0.55	0.32	0.96	0.035	0.014
	6	rs143332484	TREM2	C	T	T	5.17	1.09	24.60	0.039	0.012
GVD stage (n = 319)	19	rs429358	APOE	T	C	C	0.49	0.22	0.75	0.0003	0.057
	15	rs117618017	APH1B	C	T	T	0.57	0.25	0.90	0.0007	0.022
	19	rs7412	APOE	C	T	T	− 0.56	− 0.97	− 0.15	0.008	0.010
	21	rs2830489	ADAMTS1	C	T	T	0.27	0.04	0.50	0.021	0.023
	8	rs73223431	PTK2B	C	T	T	0.22	0.02	0.43	0.035	0.014
CAA (n = 325)	19	rs429358	APOE	T	C	C	0.55	0.38	0.71	3.84 × 10–10	0.133
	15	rs117618017	APH1B	C	T	T	0.28	0.07	0.49	0.010	0.017
	6	rs7767350	CD2AP	C	T	T	− 0.17	− 0.32	− 0.02	0.024	0.022
ABC score (n = 325)	19	rs429358	APOE	T	C	C	0.63	0.46	0.79	1.12 × 10–12	0.165
	19	rs7412	APOE	C	T	T	− 0.60	− 0.86	− 0.33	1.30 × 10–05	0.037
	15	rs117618017	APH1B	C	T	T	0.42	0.21	0.63	0.0001	0.029
	21	rs2830489	ADAMTS1	C	T	C	− 0.19	0.04	0.34	0.014	0.021
	17	rs4277405	ACE	C	T	T	0.16	− 0.30	− 0.01	0.034	0.016
	11	rs3851179	EED	T	C	C	0.16	− 0.30	− 0.01	0.035	0.008
	20	rs6742	SLC2A4RG	T	C	C	− 0.17	0.01	0.34	0.044	0.008
	8	rs73223431	PTK2B	C	T	T	0.14	0.002	0.28	0.047	0.014

Results of the regression models are summarized here. For each phenotype that was investigated, all hits reaching nominal significance of p < 0.05 are shown, as well as information regarding the chromosome, variant ID, the prioritized gene in the variant region following Bellenguez et al., the reference allele, alternative allele, and the effect allele, as well as total sample size per phenotype and the variance explained for each phenotype-genotype relation (R²). Of note, LATE-NC staging was not performed for ALS and FTLD cases. Effect sizes/ORs and standard errors are described as well. The p-values marked in bold are passing Bonferroni multiple testing correction (p < 0.0006) and were considered as the significant association discovered in this study

In total, 27 SNPs were at least nominally associated with one neuropathological phenotype (p < 0.05). Three of the SNPs, rs429358, rs7412 and rs117618017, retained significant associations with neuropathological markers after Bonferroni correction. rs429358 and rs7412 are both SNPs determining the APOE risk haplotype in exon four. Significant associations with rs429358 were found with six neuropathological phenotypes: Aβ score (p = 1.17 × 10^–12), CAA (p = 3.84 × 10^–10), Braak NFT stage p = 1.88 × 10^–07), CERAD score (p = 8.04 × 10^–08), ABC score (p = 1.12 × 10^–12) and GVD stage (p = 0.0003). For rs7412, a significant association was found with ABC score (p = 1.30 × 10^–05) and with Aβ score (p = 7.22 × 10^–05). As expected, APOE SNPs contributed most to the variance of each trait they were associated with.

Rs117618017 is a C > T polymorphism in exon 1 of APH1B, located on chromosome 15. The variant encodes a missense mutation on a protein level (p.T27I). We observed a significant association between rs117618017 and three neuropathological phenotypes: ABC score (p = 0.0001), CERAD score (p = 0.0005) and Braak NFT stage (p = 6.90 × 10^–07). Additionally, there was an indication for a relation between rs117618017 and Aβ score (p = 0.015), CAA (p = 0.010), LATE-NC stages (p = 0.034) and GVD stage (p = 0.0007), albeit not passing multiple testing correction. For all these observations, an increase in T allele dosage is associated with an increase in pathology levels.

Investigating top hit in APH1B on directly sequenced data

For the top hits, a concordance check comparing hard-called genotypes from lcWGS data with directly sequenced genotypes was performed. A concordance of 97.5% was reported for APH1B genotypes and for both APOE SNPs, a concordance of 99.7% was recorded (Table 3). For the top associated SNP, rs117618017, regression models were reconstructed with the associated phenotypes. However, for these models, hard-called genotypes acquired with the ONT approach were used to check whether these results are comparable to the results obtained with regression models constructed with allele dosages from imputed lcWGS data.

Table 3.

Top associated SNPs for which concordance was checked with Oxford Nanopore sequencing

rsID SNP	Chromosome	Genomic position	Minor allele	Major allele	MAF	Concordance	Gene
rs117618017	15	63,277,703	T	C	0.12	97.50	APH1B
rs429358	19	44,908,684	C	T	0.16	99.70	APOE
rs7412	19	44,908,822	T	C	0.07	99.70	APOE

Minor allele frequencies (MAF) reported here are calculated based on allele frequencies observed in the low-coverage whole-genome sequencing dataset

rsID reference SNP identifier

Association tests were performed for six neuropathological phenotypes, all rendering significant signals as expected based on the previous results (Fig. 2). P-values remained in the same order of magnitude for most phenotypes except for CAA severity (p = 0.003, β = 0.31, SE = 0.1) and Aβ score (p = 0.0007, β = 0.47, SE = 2.7) where the signal was stronger using hard-called genotypes compared to dosages. For Braak NFT stage (p = 5 × 10^–07, β = 0.83, SE = 0.16), CERAD score (p = 0.0003, β = 0.38, SE = 0.1), GVD stage (p = 0.0003, β = 0.07, SE = 0.02), ABC neuropathological score (p = 1.8 × 10^–05, β = 0.45, SE = 0.1) and Aβ score, there was a clear increase in the level of pathology with each T allele for rs117618017 where homozygous SNP carriers generally had the highest levels of pathology. Only for the CAA severity score, the group of individuals which were heterozygous for the SNP seemed to have higher levels of pathology compared to either homozygous genotype groups.

Fig. 2.

APH1B association with neuropathological phenotypes. Panel A–F; Panels show boxplots for Braak NFT stage, CERAD neuritic plaque score, GVD stage, ABC score, Aβ score and CAA severity respectively. The y-axis represents the stage or score of the phenotype and the x-axis represents the genotype for rs117618017

Additional statistical models were used portraying the relation between pTau, Aβ, APH1B genotype and age at death to model the directionality of the effects observed in the initial analysis. First, regression models were constructed implementing Braak NFT and Aβ score as covariates. When constructing the model investigating association between pTau and rs117618017 (APH1B), using Aβ score as a covariate, there is a persisting significant effect (p = 1.24 × 10–05, β = 0.58, SE = 0.13). On the other hand, for the regression model investigating the relation with Aβ, including Braak NFT stage as a covariate omits the significant signal (p = 0.64, β = − 0.07, SE = 0.14). The disappearance of the effect of Aβ when correcting for pTau is an indication that the relation with rs117618017 is likely not direct but rather modulated by pTau.

To further explore the connection between neuropathological lesions and AD risk SNPs, a graphical lasso model was constructed for n = 267 individuals who had phenotyping data available for age at death, Braak NFT, Aβ score and GVD stage (Fig. 3). Additionally, the two SNPs for which significant associations with these neuropathological phenotypes were discovered, were also included in the model. This network approach revealed nine edges, each indicating a positive dependency between nodes (i.e., variables). The strongest dependency in the network was observed between GVD stage and Braak NFT stage (r = 0.48). Braak NFT stage also had an edge with the age at death (r = 0.1). There were also edges connecting GVD stage with Aβ score (r = 0.17) and age at death (r = 0.13). The second strongest dependency was observed between Aβ score and Braak NFT stage (r = 0.33). Finally, the genetic factors included in this model, rs117618017 (APH1B) and rs429358 (APOE) showed a positive dependency with Braak NFT stage (r = 0.11) and Aβ score (r = 0.17) respectively. Between rs429358 and Braak NFT stage, there was a small dependency as well (r = 0.06). From this graphical network, it became apparent that the relation that was discovered between APOE SNP rs429358 and GVD stage is likely not a direct association but is modulated by Aβ pathology and pTau since there is no direct edge connecting APOE to GVD. On the other hand, the association discovered between APH1B SNP rs117618017 and pTau seems to be a direct effect, where the association between rs117618017 and Aβ seems to be modulated through pTau. These results complement our hypothesis that the effect of rs117618017 on pTau is a direct effect while the effect on Aβ is mainly mediated through pTau. Model stability was validated using bootstrapping and the CS-coefficient. Results showed that the model is robust with CS coefficients for both node strength and edge weights being 0.752, indicating the model metrics remaining stable when up to 75.2% of the data was dropped, implicating reliability of the model.

Fig. 3.

Network analysis. A graphical network constructed employing a graphical lasso model. Every edge (i.e., line) that connects two nodes (i.e., variables) represents a positive conditional dependency between those variables. The strength of the dependency is represented by the width of the edge and the r value, which is displayed at the center of every edge. From this graphical representation, the interaction between different risk factors can be deduced. Relevant to this study is the direct dependency between rs117618017 and pTau (Braak NFT), indirect dependency between rs117618017 and Aβ (Aβ score) and the indirect dependency between rs429358 and Aβ

Meta-analysis using three independent cohorts

In order to investigate the robustness of the significant and subthreshold associations discovered in this study, a meta-analysis was performed utilizing published summary statistics on ADSP, ADGC and ROSMAP cohorts [36] for four neuropathological traits for which phenotypes could be harmonized: neuritic plaques (CERAD), pTau (Braak NFT), pTDP and α-synuclein. Of note, scores for Braak NFT, CERAD and the three TDP phenotypes were converted to binary phenotypes in the Leuven Brain Collection to enable this meta-analysis. All results of the meta-analysis are summarized in Table 4. For some variants, summary statistics were only available in some of the cohorts. Therefore, information about the specific cohorts and sample sizes analyzed for every genotype–phenotype combination is indicated. Meta-analysis confirmed the association between APH1B and pTau pathology (p = 0.001) and neuritic plaques (p = 0.0002) with the same directionality as observed in the discovery cohort. Additionally, several of the sub threshold associations of the discovery cohort were confirmed in the meta-analysis: PTK2B was significantly associated with pTau pathology (p = 0.002) and neuritic plaques (p = 0.0006). MME was significantly associated with pTau pathology (p = 0.0006). Regarding pTDP pathology, several variants had nominally significant p-values in the meta-analysis, including TPCN1 (p = 0.001). Finally, significant association was observed between α-synuclein pathology and BIN1 (p = 0.0006).

Table 4.

Meta-analysis of variants with at least one nominal significant association in the discovery analysis

Gene	rsID	Effect allele	Leuven Brain Collection		ADSP		ADGC		ROSMAP		Meta-analysis
			OR	P value	OR	P value	OR	P value	OR	P value	Z score	P value
pTau pathology			n = 325		n = 1026		n = 2191		n = 1266
SIGLEC11	rs9304690	T	2.01	0.02	0.94	0.62	0.99	0.91	1	0.99	0.90	0.37
PTK2B	rs73223431	T	2.24	0.002	0.98	0.86	1.1	0.17	1.25	0.015	3.07	0.0021
APH1B	rs117618017	T	4.34	0.0001	1.07	0.65	1.08	0.43	1.45	0.0042	3.23	0.0012
MME	rs16824536	A	0.34	0.30	0.54	0.01	0.65	0.0038	1	0.99	3.44	0.0006
Neuritic plaques			n = 325		n = 1028		n = 2195		n = 1266
PTK2B	rs73223431	T	3.17	0.0004	1.14	0.2	1.09	0.23	1.21	0.029	3.45	0.0006
APH1B	rs117618017	T	3.31	0.006	1.05	0.73	1.24	0.026	1.38	0.01	3.70	0.0002
ADAMTS1	rs2830489	T	2.67	0.003	1.08	0.47	1.05	0.49	0.92	0.36	2.03	0.04
SLC2A4RG	rs6742	T	2.26	0.02	1.04	0.72	1.03	0.73	1.06	0.57	1.30	0.20
TMEM106B	rs13237518	A	0.40	0.02	1.07	0.49	1	0.95	0.92	0.35	1.46	0.14
pTDP pathology			n = 313		n = 364		n = 797		n = 1176
TREM2	rs143332484	T	2.33	0.28	0.8	0.77	NA	NA	NA	NA	0.94	0.35
TSPAN14	rs6586028	T	0.63	0.12	NA	NA	0.94	0.66	0.91	0.41	− 1.42	0.16
ABCA1	rs1800978	C	1.56	0.16	NA	NA	NA	NA	1.22	0.13	− 1.99	0.05
TPCN1	rs6489896	T	2.05	0.06	1.72	0.09	1.05	0.84	1.63	0.0069	− 3.17	0.0015
BIN1	rs6733839	T	2.08	0.001	1.05	0.78	1.11	0.33	1.02	0.83	1.87	0.06
SLC2A4RG	rs6742	T	0.44	0.008	1.39	0.13	0.97	0.83	1.02	0.83	2.68	0.0073
ADAM17	rs72777026	A	1.61	0.10	0.88	0.6	1.06	0.71	1.19	0.099	− 1.61	0.11
MAF	rs450674	T	1.25	0.31	NA	NA	0.91	0.4	1.1	0.28	− 1.65	0.10
ANK3	rs7068231	T	0.51	0.003	1.19	0.37	0.92	0.42	0.96	0.6	2.14	0.03
Α-synuclein pathology			n = 321		n = 1023		n = 2185		n = 1225
PRKD3	rs17020490	T	1.67	0.06	1.14	0.29	0.97	0.71	0.9	0.44	− 1.62	0.10
USP6NL	rs7912495	A	1.32	0.14	1.08	0.45	1.12	0.069	1.26	0.021	− 3.14	0.0017
SNX1	rs3848143	A	1.52	0.06	1.12	0.31	0.86	0.06	1.02	0.86	− 2.32	0.02
BIN1	rs6733839	T	1.28	0.22	1.24	0.02	1.13	0.054	1.16	0.13	3.45	0.0006
SLC2A4RG	rs6742	T	1.51	0.06	0.98	0.89	1.03	0.73	1.02	0.89	0.85	0.40
SCIMP	rs7225151	A	1.79	0.03	1.1	0.49	1.26	0.017	0.95	0.75	2.68	0.0074
UMAD1	rs6943429	T	1.54	0.02	0.84	0.08	0.99	0.93	0.95	0.75	1.74	0.08

Meta-analysis performed on harmonized neuropathology traits in Leuven Brain Collection, ADSP, ADGC and ROSMAP cohorts. Z-scores and p-values are reported as well as the effect allele. For this meta-analysis, neuropathology scores for Braak NFT and CERAD were converted to a binary phenotype in the Leuven Brain Collection, and the three TDP-related traits in the Leuven Brain Collection were converted into a single binary trait reflecting pTDP pathology to enable meta-analysis, which results in some SNP-trait associations in the Leuven Brain Collection not fully matching the discovery analysis which was performed on semi-quantitative phenotypes. For every cohort used in the meta-analysis, Odds ratio (OR) and p-value are reported. P-values of 0.002 and lower are passing Bonferroni correction for the number of SNP-phenotype associations tested in the meta-analysis

Discussion

We investigated the relation between known AD risk SNPs and neuropathological lesions in the Leuven Brain Collection, a novel, deeply characterized post-mortem brain cohort. We observed significant association between APH1B and pTau-related phenotypes. Additionally, nominal significant associations were observed between APH1B and GVD, CAA and Aβ pathology. APOE showed associations with hallmark AD lesions, as well as with granulovacuolar degeneration. Numerous subthreshold associations were observed, several of which were substantiated by meta-analysis with three independent cohorts, revealing potentially interesting associations between AD risk genes and non-AD neuropathologies, such as BIN1 and α-synuclein pathology.

The connection between APOE and Aβ pathology is well-studied and is very likely a direct effect, with APOE ε4 contributing to the pathogenesis by for example inducing Aβ oligomerization [68]. The association we observed between GVD and APOE is in contrast with earlier work, investigating human autopsy brains for the relation between LATE-NC, GVD and pTau. This study reported no association between APOE and GVD [37]. A possible explanation for this discrepancy could be the lower power of that study to detect significant genetic associations due to a limited sample size. In our study, accurate genotype and phenotype data for GVD were available for 319 individuals, substantially increasing the power to detect smaller effects compared to earlier work. Network analysis however indicated that the effect of APOE on GVD might not be direct, but rather mediated by Aβ pathology. This warrants further investigation into the relation between the APOE genotype and GVD to better comprehend how they are interacting in AD pathology. To validate these findings, replication of the associations in an independent cohort is needed, but the lack of deeply phenotyped neuropathological cohorts, especially for less-studied lesions such as GVD, is a substantial limitation. Therefore, some caution is warranted in the interpretation of these findings.

Besides APOE, the AD risk gene APH1B came forward in this study with significant associations with Braak NFT stage, CERAD score and ABC score in the Leuven Brain Collection. Additionally, trends toward significance were observed for GVD stage, LATE-NC, CAA and Aβ score. The association with Braak NFT stage and CERAD score was observed as a subthreshold association in an independent study [36]. When we meta-analyzed our results with this independent dataset in total meta-analysis sample sizes of 4808 and 4814 respectively, the associations persisted. APH1B encodes a multi-pass transmembrane protein which is a stabilizing subunit of the γ-secretase protein complex. Γ-secretase comprises four subunits: presenilin (PS1 and PS2 homologs), nicastrin, anterior pharynx-defective 1 (APH-1) and presenilin enhancer two [15]. In humans, the two physiological isoforms of APH-1 are APH1A and APH1B. The structural heterogeneity of the isoforms exerts a main effect on the carboxypeptidase activity of γ-secretase, where complexes containing APH1B are less efficient compared to APH1A [1]. Additionally, APH1B containing complexes tend to generate longer, more toxic Aβ species upon cleavage of the most well-known γ-secretase substrate, APP [54]. Due to the well-established function of γ-secretase in APP cleavage, it would be expected that there is a stronger relation between the AD risk SNP in APH1B and Aβ-related phenotypes. However, in this study, the evidence points more toward a relation with tau-related pathologies when directionality of the effect was investigated using Braak NFT and Aβ score as covariates in a regression model and by subsequent construction of a graphical network. Both analyses indicated a direct relation between APH1B rs117618017 and tau pathology rather than Aβ pathology. It can be hypothesized that the effect on tau pathology is exerted through the APP intracellular domain (AICD) rather than through Aβ peptides, as both co-localize in the nucleus of cell where AICD could interact with tau through modulation of the DNA binding capacity and DNA protective function [46]. A study has also shown a relation between tau and AICD in a transgenic AICD mouse model, where ablation of tau seemed protective for AD-like features, indicating that tau is crucial in mediation of the deleterious effects of AICD [22]. Evidence has also been gathered regarding the relation of other γ-secretase subunits with tau pathology, such as presenilin. In mouse models, partial loss of the presenilin subunit resulted in tau aggregation and phosphorylation, increasing overall neurodegeneration [55]. Supporting this relation is the evidence linking presenilin mutations to other tauopathies, such as frontotemporal dementia [8, 52] and Pick’s disease [16]. Additionally, APP metabolism has also been linked to tau proteostasis in cell culture models [45], possibly through regulation of endosome/lysosomal pathways [9]. Interestingly, recent work has also linked an AD risk missense mutation in SORL1, a transmembrane, endosomal protein involved in APP trafficking to tau by showing colocalization in cell lines as well as mediation of tau seeding by mutant SORL1 [13]. This further strengthens the possibility of a relation between γ-secretase, and thus APH1B, to tau pathology.

Rs117618017, encoding a missense mutation in the APH1B protein, has a robust association with clinical AD with evidence compiled from multiple genome-wide association studies and meta-analyses [31, 67]. Using imaging biomarkers, altered expression levels in blood of APH1B RNA have also been linked to amyloid PET and brain atrophy [51], further emphasizing the importance of APH1B in AD pathology. However, this study only focused on markers for Aβ pathology or neurodegeneration and did not investigate the possible relation of APH1B to tau pathology. Functional studies have also investigated the involvement of APH1B in AD. One study expressed the variant in cell lines and reported no significant effects on APP cleavage [70], suggesting that the effect observed here is possibly not directly attributable to the cleave of APP although further investigation is needed. Here, it is proposed that the T allele at rs117618017 increases tau-related pathology, likely through an overall increase of APH1B levels although functional validation experiments are needed to identify the true mechanism of action.

Finally, we observed several subthreshold associations in the discovery cohort, which persisted in the meta-analysis. This includes an association between BIN1 and α-synuclein pathology. Of interest, a whole-genome sequencing study on LBD previously reported a genome-wide association between the BIN1 locus and risk for LBD with the same directionality of effect as observed in AD [12]. In contrast, however, a study investigating the genetic risk of AD cases with and without Lewy body pathology reported that BIN1 risk was associated with AD pathology rather than Lewy body pathology [58]. Further studies are needed to elucidate the potential role of BIN1 in α-synuclein pathology. Additionally, we observed nominal association for UMAD1 and USP6NL with α-synuclein pathology. The latter was also suggested by a prior study [36] and was replicated in our meta-analysis. Meta-analysis did also confirm the association between PTK2B and pTau pathology and neuritic plaques. The PTK2B protein product, Pyk2, has been suggested to be involved in tau toxicity in animal models [17, 24, 42] and the PTK2B risk allele was also associated with longitudinal increase of CSF pTau levels in humans [60]. Finally, meta-analysis confirmed an association between TPCN1 and pTDP pathology, while also showing nominal associations for SLC2A4RG and ANK3. Intriguingly, TDP43 expression seemed to trigger an increase in the expression of ANK3 in a human iPSC model of ALS [25]. We did not detect associations with specific TDP-related risk genes, like GRN and TMEM106B, which could in part be due to the sample size of this study or due to low number of FTLD/ALS cases in our cohort.

Conclusion

Genome-wide association studies rapidly discover genetic variants linked to the clinical diagnosis of AD, but there is a need for studies expanding on these findings to understand which variants are at play in what part of the heterogeneous pathological spectrum of AD. Here, we reported first evidence that genetic variation in the γ-secretase component APH1B is associated with tau pathology. Additionally, a relation between APOE and granulovacuolar degeneration was observed. This effect was proven to be likely indirect through Aβ and pTau, providing accumulating evidence for the relation between APOE and hallmark AD pathology. The results obtained from this investigation offer an opportunity to gain insights into how AD risk variants are exactly involved in AD pathology, more specifically, where certain genetic variants can be mapped to specific neuropathological lesions observed along the AD spectrum.

Supplementary Information

Below is the link to the electronic supplementary material.

Funding

The funding has been received from Fund for Scientific Research Flanders with Grant no. FWO G065721N; FWO with Grant no. FWO G065721N; Onderzoeksraad, KU Leuven with Grant no. C14/22/132; Alzheimer's Association with Grant no. 22-AAIIA-963171; BrightFocus Foundation with Grant no. A2022019F; BOF UAntwerp with Grant no. 49758.

Data availability

All relevant data are provided in the tables of the paper, so an additional data availability is redundant.

Declarations

Conflict of interest

DRT collaborated with Novartis Pharma AG (Switzerland), Probiodrug (Germany), GE-Healthcare (UK), and Janssen Pharmaceutical Companies (Belgium). DRT and KS are members of Acta Neuropathologica editorial board. They were not involved in the assessment or decision-making process for this manuscript. CAFvA received honoraria from serving on the scientific advisory board of Biogen, Roche, Novo Nordisk, Biontech, Lilly, Dr. Willmar Schwabe GmbH &Co. KG, and MindAhead UG and has received funding for travel and speaker honoraria from Biogen, Lilly, Novo Nordisk, Roche diagnostics AG, Novartis, Medical Tribune Verlagsgesellschaft mbH, Landesvereinigung für Gesundheit und Akademie für Sozialmedizin Niedersachsen e. V., FomF GmbH | Forum für medizinische Fortbildung and Dr. Willmar Schwabe GmbH &Co. KG has received research support from Roche diagnostics AG. RVs institution has clinical trial agreements (RV as PI) with Alector, Biogen, Denali, J&J, EliLilly, and UCB and consultancy agreements (RV as DSMB member) with AC Immune. The other authors have nothing to disclose.

Ethical approval

Tissue collection was performed in accordance with the local ethical committee guidelines and laws regarding the use of human tissue for research. The research protocol for the current study was approved by the ethical committees of the UZ/KU Leuven and UZA/University of Antwerp.