The TMEM106B FTLD-protective variant, rs1990621, is also associated with increased neuronal proportion

Abstract

Apart from amyloid β deposition and tau neurofibrillary tangles, Alzheimer disease (AD) is a neurodegenerative disorder characterized by neuronal loss and astrocytosis in the cerebral cortex. The goal of this study is to investigate genetic factors associated with the neuronal proportion in health and disease. To identify cell-autonomous genetic variants associated with neuronal proportion in cortical tissues, we inferred cellular population structure from bulk RNA-seq derived from 1,536 individuals. We identified the variant rs1990621 located in the TMEM106B gene region as significantly associated with neuronal proportion (p-value = 6.40×10⁻⁰⁷) and replicated this finding in an independent dataset (p-value= 7.41×10⁻⁰⁴) surpassing the genome-wide threshold in the meta-analyses (p-value = 9.42×10⁻⁰⁹). This variant is in high LD with the TMEM106B non-synonymous variant p.T185S (rs3173615; r² = 0.98) which was previously identified as a protective variant for Frontotemporal lobar degeneration (FTLD). We stratified the samples by disease status, and discovered that this variant modulates neuronal proportion not only in AD cases but also several neurovegetative diseases and in elderly cognitively healthy controls. Furthermore, we did not find a significant association in younger controls or schizophrenia patients, suggesting that this variant might increase neuronal survival or confer resilience to the neurodegenerative process. The single variant and gene-based analyses also identified an overall genetic association between neuronal proportion, AD and FTLD risk. These results suggest that common pathways are implicated in these neurodegenerative diseases, that implicate neuronal survival. In summary, we identified a protective variant in the TMEM106B gene that may have a neuronal protection effect against general aging, independent of disease status, which could help elucidate the relationship between aging and neuronal survival in the presence or absence of neurodegenerative disorders. Our findings suggest that TMEM106B could be a potential target for neuronal protection therapies to ameliorate cognitive and functional deficits.

INTRODUCTION

Alzheimer disease (AD) is characterized by the presence of extracellular amyloid plaques and intracellular deposits of tau protein that form tangles [36]. AD is characterized by brain volume loss, and neuronal death [37]. Multiple studies have shown that neuronal loss and synapse dysfunction precede cognitive deficits in AD, occurring even 15–20 years before clinical onset [5].

Previous studies from our lab have shown that there are genetic variants associated with neuronal proportion. We developed, optimized and validated an in-silico digital deconvolution approach to infer brain cell proportion from RNA-Seq data [56]. As expected, we found that AD cases present with lower neuronal proportion than cognitively healthy controls. We also demonstrated that individuals that carry the rare APP, PSEN1 or PSEN2 pathogenic mutations and the common APOE ε4 variants present significantly lower neuronal proportion. The goal of this study is to identify common genetic variants associated with neuronal proportion in very large cohorts. Using neuronal proportion as a quantitative trait for genetic studies provides several advantages over classical genetic studies, as quantitative traits provide more power than regular case-control studies [17, 20–22, 42, 65, 75]. This phenotype (neuronal proportion), similar to other AD-related endophenotypes such as cerebrospinal fluid (CSF) tau and Aβ levels is closer to the underlying biology than case-control status, and provides a biological model of the diseases [17, 21, 22, 26, 35]. We have successfully used quantitative endophenotypes (CSF protein levels [20, 21, 42, 65, 75]) to identify novel genetic variants and genes implicated in AD.

Previous studies have already leveraged neuronal proportion derived from digital deconvolved RNA-seq data to identify genes implicated in disease. In a recently published study by the PsychENCODE consortium [83], a very similar approach to the deconvolution method we previously reported [56] was used to infer cell type composition from RNA-Seq obtained from brain tissue in psychiatric disorder cohorts. This study includes of participants from all age groups, but the median age was 51 years old. This study found that common variants in the FZD9 gene were associated with neuronal proportion [83]. Interestingly, variants in FZD9 are associated with Williams syndrome, a developmental disorder associated with mild to moderate intellectual disabilities with learning deficits and cardiovascular problems [12].

These results suggest that using neuronal proportion as a quantitate trait for genetic studies can identify variants and genes implicated in disease. However, the PsychENCODE study was focused on psychiatric disorders and included relatively young individuals. No similar studies have been performed in neurodegenerative disorders. In this study, we utilized neuronal proportions inferred from our deconvolution method [56] to perform cell type quantitative trait loci (cQTL) analysis in 2,008 brain samples derived from 1,536 individuals enriched for AD cases to search for potential new loci that are associated with neurodegenerative disorders.

METHODS

Study participants

Our analysis includes participants from seven studies with a total sample size of 1,536 (Table 1). Five of these studies are mainly focused on neurodegenerative disorders (mean age at death = 82.76 ± 9.86) that include brains affected by Alzheimer’s disease (AD; n = 640), Frontotemporal lobar degeneration (FTLD; n = 11), Progressive Supranuclear Palsy (PSP; n = 75), and pathological aging (PA; n = 28). PA is defined as a form of cerebral amyloidosis in older people, who have widespread extracellular amyloid-beta (Aβ) senile plaque deposits in the setting of limited neurofibrillary tau pathology [58].

Table 1.

Demographic information for cohorts included in this study.

	N	Region	Age	% Male	RIN	TIN	Control	AD	FTLD	PSP	PA	SCZ	BP	Unknown
Discovery
ROSMAP	484	1	86.7 ± 4.48	35.3	7.17 ±0.93	73.7 ±4.82	100	317	0	0	0	0	0	67
Replication
Mayo	235	1	80.4 ± 8.24	49.8	8.23 ± 0.82	77.7± 5.99	54	78	0	75	28	0	0	0
MSSM	206	4	84.1 ± 7.30	36.9	6.62 ± 1.65	76.4 ±2.51	45	161	0	0	0	0	0	0
Knight-ADRC	96	1	83.0 ± 12.50	44.8	6.49 ± 1.19	79.4 ±2.00	12	68	11	0	0	0	0	5
DIAN	15	1	50.9 ± 7.08	73.3	5.55 ± 1.09	78.9 ±0.99	0	15	0	0	0	0	0	0
GTEx	118	3	57.9 ± 10.10	67.8	7.00 ± 0.85	74.0 ±2.86	114	1	0	0	0	0	0	3
CommonMind	382	1	65.6 ± 17.50	61.8	7.68 ± 0.89	50.1 ±7.08	163	0	0	0	0	189	30	0
Replication Total	1,052						388	323	11	75	28	189	30	8
Merged Total	1,536						488	640	11	75	28	189	30	75

N: sample size; Region: number of cortical brain regions involved in the analysis; % Male: percentage of male participants in the cohorts; RIN: RNA integrity number; TIN: transcript integrity number; AD: Alzheimer’s Disease; FTLD: Frontotemporal lobar degeneration; PSP: progressive supranuclear palsy; PA: pathological aging; SCZ: schizophrenia; BP: bipolar disease; Unknown: other unknown dementia or no diagnosis information. Age, RIN and TIN are shown as mean ± standard deviation.

These samples come from the Mayo Clinic (Mayo), Mount Sinai School of Medicine (MSSM), the Charles F. and Joanne Knight Alzheimer’s Disease Research Center (Knight ADRC), the Dominantly Inherited Alzheimer Network (DIAN), and the Religious Orders Study and Memory and Aging Project (ROSMAP) studies. In addition, we analyzed brains from the CommonMind (CMC; mean age at death = 65.6 ± 17.5) and GTEx (mean age at death = 57.9 ± 10.1) studies, whose age at death is significantly lower than the neurodegenerative aging cohorts (p-value < 4.41 × 10⁻⁶⁷). Apart from neurodegenerative disorders, this study also includes samples from individuals with schizophrenia (SCZ; n = 189), bipolar disorders (BP; n = 30), and cognitively healthy controls (n = 488) (Table 1, Online Resource Table 1).

Additionally, two studies, MSSM and GTEx, contain data from more than one tissue per participant. In our analyses of the MSSM samples, we used RNA-Seq data generated from the anterior prefrontal cortex (BA10), superior temporal gyrus (BA22), ventral anterior cingulate cortex (BA24), parahippocampal gyrus (BA36), and inferior frontal gyrus (BA44). For the GTEx Study we used RNA-Seq data derived from the Brodmann area (BA) BA24 (anterior cingulate cortex), BA9 (frontal cortex), and general cortical regions (CTX). See Online Resource Table 2 and each study reference(s) for more data collection and generation specifications.

Standard protocol approvals, registrations and patient consents

The protocols for the DIAN and Knight-ADRC studies have been approved by the institutional review board of Washington University in St. Louis. The protocol for the Mayo dataset was approved by the Mayo Clinic Institutional Review Board (IRB). All neuropsychological, diagnostic and autopsy protocols for the MSSM dataset were approved by the Mount Sinai and JJ Peters VA Medical Center Institutional Review Boards. The religious orders study and the memory and aging project of ROSMAP were approved by the IRB of Rush University Medical Center. The NIH Common Fund’s GTEx program protocol was reviewed by Chesapeake Research Review Inc., Roswell Park Cancer Institute’s Office of Research Subject Protection, and the institutional review board of the University of Pennsylvania. Within the CommonMind consortium, the MSSM sample protocol was approved by the Icahn School of Medicine at Mount Sinai IRB; the Pitt sample protocol was approved by the University of Pittsburgh’s Committee for the Oversight of Research involving the Dead and the IRB for Biomedical Research; the Penn sample protocol was approved by the Committee on Studies Involving Human Beings at the University of Pennsylvania. All participants were recruited with informed consent for research use.

Data collection and generation

Cortical tissues from different regions of post-mortem brains were collected (Online Resource Table 2). RNA was extracted from lysed tissues and libraries prepared for next-generation sequencing. Ribosomal RNA constitutes 80%−90% of total RNA, but it is not the target of this study. To focus on mRNA quantification, rRNA was either depleted or the mRNA was captured during RNA-Seq library preparation. The DIAN [56], Knight ADRC [56], MSSM [85], and CommonMind [30] studies used an rRNA-depletion approach to remove ribosomal RNA. The Mayo [1], ROSMAP [7, 8], and GTEx [6, 15] studies used a poly-A enrichment approach to enrich for mRNA. DNA was also collected and sequenced to obtain genotype information. RNA-Seq paired with genotype data for each participant was either sequenced at Washington University for the DIAN and Knight-ADRC studies or downloaded from public databases for all the other studies.

Data QC and preprocessing

Genetic Data

Stringent quality control (QC) steps were applied to all genotyping array or sequence data. The minimum call rate for single nucleotide polymorphisms (SNPs) and individuals was 98%, and autosomal SNPs not in Hardy-Weinberg equilibrium (p-value < 1×10⁻⁰⁶) were excluded. X-chromosome SNPs were analyzed to verify gender identification. Unanticipated duplicates and cryptic relatedness (PiHat ≥ 0.25) among samples were tested by genome-wide estimates of pairwise identity-by-descent. PLINK1.9 [13] was used to calculate principal components. The 1000 Genomes Project Phase 3 data (October 2014), SHAPEIT v2.r837 [19], and IMPUTE2 v2.3.2 [38] were used for phasing and imputation. Individual genotypes imputed with a probability < 0.90 were set to missing and imputed genotypes with probability ≥ 0.90 were analyzed as fully observed. Genotyped and imputed variants with MAF < 0.02 or IMPUTE2 information score < 0.30 were excluded. WGS QC was performed by filtering out reads with sequencing depth DP < 6 and quality GQ < 20, followed by similar QC approaches as described above for genotyping data. After the QC, all studies including imputed genotypes and WGS data were merged into a binary file using PLINK for downstream analysis. In order to identify related individuals, principal component analyses (PCA) for population stratification and identity by descent (IBD) analyses, were performed on the merged binary files using PLINK to retain unrelated participants of European ancestry and to ensure sample identity (Online Resource Figure 1 and 2).

Expression Data

FastQC was applied to RNA-Seq data to examine various aspects of sequencing quality [2]. Outlier samples with high rRNA content (> 10.9%) were removed from the study. The remaining samples were aligned to the human GRCh37 primary assembly using STAR with 2-Pass Basic mode (ver 2.5.4b) [23]. Alignment metrics were ascertained by applying Picard CollectRnaSeqMetrics [10] including reads bias, coverage, ribosomal contents, coding bases, etc. Transcript integrity number (TIN) was calculated directly from post-sequencing results to model the RNA degradation by measuring mRNA integrity directly [84]. We calculated the TIN for each library using RSeQC tin.py [84] (ver 2.6.5). RNA-Seq coding genes and transcript expression was quantified using Salmon transcript expression quantification (ver 0.7.2) with the GENCODE Homo sapiens GRCh37.75 reference genome [64].

We inferred neuronal proportions from RNA-Seq gene expression quantification using the same digital deconvolution method and panel reported by Li et al. [56]. Briefly, we first assembled and curated a cell-type specific reference panel to model the transcriptomic signature of neurons, astrocytes, oligodendrocytes and microglia, and selected a set of gene markers to capture the distinctive expression profile of these cell types. Gene markers for the four major cell types include STMN2, SYN1, SYT1, GAD1, CCK for neuron; GFAP, ALDH1L1, AQP4, GJA1, SOX9 for astrocyte; MOG, MOBP, SOX10, GPR37 for oligodendrocyte; TLR2, CX3CR1, IL1A for microglia. Using the reference panel and the method population-specific expression analysis (PSEA, also named meanProfile in the CellMix implementation [33]) and semi-supervised non-negative matrix factorization (ssNMF, also named ssFrobenius in the CellMix implementation), we calculated the cellular composition from the RNA-Seq data. This method estimates the proportion of neurons, astrocytes, oligodendrocytes and microglia in a specific sample. The proportions of these four cell types for a specific sample always add up to 1. Based on our previous studies, a decrease of neuronal proportion is associated with an increase of astrocytes although this is not perfectly correlated, indicating that the information of each cell proportion is unique and captures cell-type specific profile.

Outlier values for each cell-type proportion were removed from each study or from each brain region if the study included more than one region. Outliers are defined as cell-type proportion values that fall below the first quantile minus 1.5 × IQR (interquartile range) or above the third quantile plus 1.5 × IQR. If a subject contains an outlier proportion in any of the four major cell types from any of the two independent deconvolution methods, this subject would be removed from the analysis. For example, 39 subjects were removed from the ROSMAP study as they were outliers in cell deconvolution analyses. These 39 subjects were removed from 523 ROSMAP samples during QC (Online Resource Table 1 before QC), for demographics before QC), and 484 samples were included in the neuronal proportion QTL analysis (Table 1 for demographics after QC). This way we can ensure that the cell-type proportion phenotypes are validated by two independent deconvolution methods in the context of all four major CNS cell types, and any spurious deconvolution results will be excluded from downstream analysis. After removing outliers, neuronal distributions were zero centered for each cortical region and study by subtracting the mean value (Online Resource Figure 3).

Data analysis

We performed a three-stage genome-wide association study (GWAS) analysis: discovery, replication and meta-analysis. For the discovery phase, we used the ROSMAP dataset (n = 484). For the replication phase, we combined the MSSM, GTEx, CMC, WU (including Knight ADRC and DIAN), and Mayo studies (n = 1,052).

For all analyses, we applied a linear regression model using normalized neuronal proportions including age, sex, PC1 and PC2 derived from GWAS data to account for population stratification, and median TIN as covariates. TIN is a measure of RNA integrity. Association analyses were performed using the Meta-Tissue QTL software [34, 76] (http://genetics.cs.ucla.edu/metatissue/install.html). Meta-Tissue is the software developed by GTEx, designed to take into account the heterogeneity of multiple tissues. Meta-Tissue is designed to handle a wide array of tissues, from skin, to liver, to brain tissue donated from a diverse set of individuals. In this study we are using different brain regions, but all of them are cortical regions, therefore the heterogeneity of this study is lower than that in GTEx study in general. This method was used because MSSM and GTEx contain subjects with multiple cortical tissues. The Meta-Tissue QTL allows for the performance of association analyses when multiple samples from the same individual are included. The Meta-Tissue [76] processing pipeline calls two main functions, Meta TissueMM [76] followed by Metasoft [34]. Meta TissueMM applies a mixed-model to account for the heterogeneity of multiple tissue QTL effects. Metasoft performs a meta-analysis while providing a more accurate random effect p-value for multiple tissue analyses and an m-value based on Bayesian inference to indicate how likely a locus is to be a QTL in each tissue. Genome-wide association results were also visualized as Manhattan plots. For the signal of interest, we generated a PM-Plot that integrates evidence from both frequentist (p-value) and Bayesian (m-value) sides to interpret the heterogeneity of multi-tissue QTL effects. Variants with m-values greater than 0.9 are predicted to have an effect and are colored red. Variants with m-values less than 0.1 are predicted not to have an effect and are colored blue. All the other studies with m-values between 0.1 and 0.9 are predicted to have an ambiguous effect and are colored green [34].

For the meta-analysis, both discovery and replication studies were combined using the MetaTissue [76] method described above. Forest plots for each dataset were generated using the visualization function provided by Metasoft [34].

Because our study includes brain samples with different neurodegenerative diseases, we also analyzed each disease separately, including AD (n = 639), other neurodegenerative diseases (n = 103; PSP and PA), schizophrenia (n = 189), young subjects (Age at death < 65; n = 277) and cognitively normal elderly individuals (n = 199). Disease subcategories with less than 20 subjects were removed from the analysis to avoid false results due to a small sample size. Similar data preparation and analysis pipelines were employed as documented above. Results were depicted as Manhattan plots using R’s (ver 3.4.3) qqman package [78] (ver 0.1.4).

Genetic variance estimation

The Genome-wide Complex Trait Analysis (GCTA) v1.25.2 tool was used to estimate the proportion of phenotypic variance explained by the common (MAF > 0.02) imputed and genotyped autosomal variants. The restricted maximum likelihood (REML) analysis was performed on the normalized neuronal proportion values adjusted for age, sex, median TIN, and the first two principal components as covariates. Key variants rs1990621 (TMEM106B) and rs429358 (APOE) among others that were significantly associated with neuronal proportion were analyzed individually to determine the amount of variance contributed by these variants to the observed neuronal proportion phenotype using R’s linear regression function.

Functional annotation

All variants below the threshold for suggestive significance (p-value < 1×10⁻⁰⁵) were uploaded to FUMA (v1.3.3d) [86] to annotate significant SNPs with GWAS catalog (e91_r2018-02-06) and ANNOVAR (updated 2017-07-17). Gene-based analysis was also performed by MAGMA (v1.06) [18] implemented in FUMA. MAGMA is an algorithm to perform gene and gene-set analysis, based on GWAS summary statistics. This gene-set analysis uses a regression structure to allow generalization analysis of continuous properties of genes and simultaneous analysis of multiple gene-sets and other gene properties, and provides a p-value for each gene in the genome, taking into account all the independent SNPs in the gene region. Tissue specific gene expression of related genes was performed with data from publicly available databases, Genotype-Tissue Expression (GTEx) Analysis V7 [6] and the Brain eQTL Almanac (Braineac) [67] to determine the tissue(s) in which the genes were most highly expressed.

RESULTS

Study Design

We performed a three-stage study design: discovery, replication and meta-analysis. The discovery phase included RNA-seq data from 484 subjects (ROSMAP). The other six studies were collapsed into replication dataset with 1,052 subjects (Table 1). To our knowledge, this is the largest study (n = 1,536) that analyzes genetic architecture underlying neuronal proportion in aging brains (Figure 1, and Table 1). The GTEx and MSSM studies include brain RNA-seq data from multiple cortical regions. For this reason we utilized the Meta-Tissue software [76], which was specifically designed for multi-tissue QTL analysis and performed a mixed model analysis with random effects that account for correlated measurements from multi-tissue individuals. To attain the largest available sample size for this study, the discovery and replication sets were merged to perform a multi-tissue QTL meta-analysis in a search for additional signals hidden in the previously separated discovery or replication analyses due to lack of power. After merged analysis, the cohorts were stratified into four major disease status groups (AD, control, schizophrenia, and other non-AD neurodegenerative disorders) to explore how different disease strata could impact the results.

Figure 1. Study Design.

RNA-Seq and paired genotype or whole genome sequencing (WGS) data were accessed and preprocessed for downstream analysis. Genotype data quality was ensured based on our quality control criteria and imputed as needed. WGS and imputed genotypes were merged, followed by principal component analysis (PCA) and identity by descent (IBD) to select unrelated subjects of European ancestry. RNA-Seq data quality was checked with FastQC and aligned to human GRCh37 primary assembly with STAR, from which transcript integrity number (TIN) was inferred with RSeQC to account for RNA integrity variation that we later incorporated into the analysis. Gene expression was quantified from unaligned RNA-Seq with the psuedo-aligner Salmon for deconvolution. Cell type composition comprised of the four major CNS cell types were inferred by performing deconvolution on gene expression quantification results. Using cell type proportions as quantitative traits, we identified loci in the TMEM106B gene region associated with neuronal proportion in our assembled dataset. CMC: CommonMind Consorsium; GTEx: The Genotype-Tissue Expression; Mayo: the Mayo Clinic; MSSM: Mount Sinai School of Medicine; Knight ADRC: the Charles F. and Joanne Knight Alzheimer’s Disease Research Center; DIAN: the Dominantly Inherited Alzheimer Network; ROSMAP: the Religious Orders Study and Memory and Aging Project; TCX: temporal cortex; PAR: parietal cortex; CTX: cortex; FCX: frontal cortex; DLPFC: dorsal lateral prefrontal cortex. BA9: dorsal lateral prefrontal cortex; BA10: Anterior prefrontal cortex; BA22: superior temporal gyrus; BA24: ventral anterior cingulate cortex; BA36: parahippocampal gyrus; BA44: inferior frontal gyrus.

We performed functional annotation to identify the variants and genes that drive the association with neuronal proportion. Single variants were annotated by ANNOVAR and other gene mapping tools integrated by FUMA to identify functional genes or pathways. In addition, neuronal proportion heritability contributed by the target variants was analyzed using genome-wide complex trait analysis (GCTA). To determine the overlap between the genetic architecture and other complex traits, including AD, FTLD, and schizophrenia, we analyzed if the SNPs and genes associated with these diseases were also associated with neuronal proportion.

TMEM106B variants associated with neuronal proportion

Using the normalized neuronal proportion as a quantitative trait in the discovery dataset (n = 484), we identified 1,340 loci that have p-values < 1.0×10⁻⁰³; and 22 loci with p-values < 1.0×10⁻⁰⁵, including: LIMCH1, GLRA1/ATOX1, DOCK2, FOXO3, TMEM106B, DLC1, PRKCA, RNF152, and CCDC102B (Online Resource Table 3, Online Resource Figure 4ab). One of the most significant SNPs, rs1990621 (β = 0.3; p-value = 6.40×10⁻⁰⁷) is located in the TMEM106B gene region and was one of the few loci with p-value around 1.0×10⁻⁰⁷ (Table 2, Figure 2e, and Online Resource Table 3).

Table 2.

SNPs passing genome-wide suggestive level in final meta-analysis merging discovery and replication data.

SNP	Chr.	BP	A1	A2	MAF	Discovery		Replication		Meta-analyses
						Effect	P-value	Effect	P-value	Effect	P-value	% Variance
rs6750387	2	76,967,335	G	A	0.381	−0.068	2.91×10−01	−0.049	4.80×10−07	−0.056	9.23×10−07	0.17%
rs665133	3	6,676,306	C	T	0.029	0.425	2.33×10−02	0.364	7.26×10−06	0.376	9.50×10−07	0.83%
rs6908749	6	111,962,944	A	G	0.073	0.180	1.73×10−01	0.042	8.08×10−06	0.084	7.43×10−06	0.02%
rs6925190	6	153,744,859	A	T	0.120	−0.305	3.10×10−03	−0.260	4.99×10−05	−0.270	8.03×10−07	1.03%
rs1990621	7	12,283,873	G	C	0.466	0.300	6.40×10−07	0.130	7.41×10−04	0.160	9.42×10−09	2.28%
rs11239562	10	46,063,198	G	A	0.438	−0.121	5.02×10−02	−0.174	1.74×10−05	−0.158	3.97×10−06	1.38%
rs112142762	12	14,384,935	C	T	0.035	−0.173	2.97×10−01	0.183	3.53×10−06	0.069	8.11×10−06	0.01%
rs78937948	12	81,558,976	G	C	0.055	0.504	2.52×10−03	0.391	1.01×10−04	0.424	1.25×10−06	1.24%
rs441407	13	86,379,248	T	C	0.275	−0.168	2.18×10−02	−0.195	2.88×10−05	−0.187	2.27×10−06	1.42%
rs283815	19	45,390,333	G	A	0.250	−0.151	5.86×10−02	−0.229	3.36×10−06	−0.214	1.94×10−06	0.91%
rs429358	19	45,411,941	C	T	0.151	−0.126	1.79×10−01	−0.258	8.15×10−06	−0.235	1.22×10−05	1.27%
rs73487196	22	48,793,169	T	C	0.071	0.453	7.42×10−03	0.374	1.75×10−04	0.398	6.62×10−06	0.62%

Statistical analyses were performed with both PLINK and Meta-Tissue for the Discovery dataset and with Meta-Tissue for the Replication and Meta-analyses Nominally, significant p-values are in bold for the discovery and replication datasets. Only genome-wide significant p-values are in bold for the meta-analyses

Figure 2. The single variant and gene-based multi-tissue meta-analyses identified TMEM106B and APOE as genes associated with neuronal proportion.

All panels in this figure were produced using meta-analysis results from Meta-Tissue analytic pipeline. a) Manhattan plot showing the SNP-based genome-wide significant hit located in chromosome 7 with other suggestive SNP hits labeled. rs1990621, located in the chromosome 7 TMEM106B gene region, was significantly associated with neuronal proportion in meta-analyses. b) QQ plot of the SNP-based analysis. c) Manhattan plot showing the gene- based genome-wide significant hit located in chromosome 7 with other suggestive genes. d) QQ plot of the gene-based analysis. e) Local plot showing the zoomed-in view of the hit in chromosome 7 with the target SNP rs1990622 labeled with dark purple. The top leading SNP is rs1990621. Nearby SNPs were also mainly located in the TMEM106B gene region and color coded with their linkage disequilibrium (LD) r2 thresholds. f) Local plot showing the zoomed-in view of the hit in chromosome 19 with the target SNP rs2075650 labeled with dark purple. The top three leading SNPs are rs283815, rs769449, and rs429358. Nearby SNPs were also mainly located in the TOMM40/APOE gene region and color coded with their LD r² thresholds. One gene omitted in this region is SNRPD2.

We used the replication cohorts (n = 1,052) to determine if any of these loci showed consistent association with neuronal proportion (Online Resource Table 4). The association of rs1990621 with neuronal proportion was in the same direction in all cohorts and also in all regions in the studies with multiple regions, except in two regions: BA24 and CTX, from the GTEx samples. This result indicates that there is some tissue heterogeneity in the effect of the association across the studies we analyzed (Figure 3a). For this reason, we decided to use a statistical framework that can analyze all of these cohorts and regions together while accounting for tissue heterogeneity. Meta-Tissue calculates a m-value [34] to predict if a variant has a tissue-specific significant effect. The m-value is similar to the posterior probability of an association based on the Bayes factor [34], but allows for the performance of meta-analyses on different datasets and regions accounting for heterogeneity. It is able to detect whether a specific variant has a significant effect in a study overall. Using this approach, rs1990621 in the TMEM106B gene region showed a significant association in the same direction in the replication cohort (rs1990621; β = 0.13; p-value = 7.41×10⁻⁰⁴; Figure 3c, Table 2, and Online Resource Figure 4cd). This analysis included only individuals not related to those from the discovery dataset, therefore this represents an independent replication of this signal. None of the other 21 loci with suggestive p-value showed a nominal association in the replication dataset (Online Resource Table 4).

Figure 3. Meta-Tissue analysis results of rs1990621.

a) Forest plot showing the p-value and confidence interval for rs1990621 for each tissue site of each dataset that is included in the Meta-Tissue meta-analysis. Summary random effects were depicted at the bottom as RE Summary. b) Based on Meta-Tissue meta-analysis, PM-Plot of rs1990621 is plotted while combining both p-value (y axis) and m-value (x axis). Red dots indicate that the variant is predicted to have an effect in that particular dataset, blue dots mean that the variant is predicted to not have an effect, and green dots represent ambiguous predictions. c) Forest plot p-value and confidence interval for rs1990621 for discovery, replication, and merged datasets. d) Forest plot p-value and confidence interval for rs1990621 when splitting the merged dataset into four main disease categories. CMC: CommonMind Consorsium; GTEx: The Genotype-Tissue Expression; Mayo: the Mayo Clinic; MSSM: Mount Sinai School of Medicine; Knight ADRC: the Charles F. and Joanne Knight Alzheimer’s Disease Research Center; DIAN: the Dominantly Inherited Alzheimer Network; ROSMAP: the Religious Orders Study and Memory and Aging Project; TCX: temporal cortex; PAR: parietal cortex; CTX: cortex; FCX: frontal cortex; DLPFC: dorsal lateral prefrontal cortex. BA9: dorsal lateral prefrontal cortex; BA10: Anterior prefrontal cortex; BA22: superior temporal gyrus; BA24: ventral anterior cingulate cortex; BA36: parahippocampal gyrus; BA44: inferior frontal gyrus; AD: Alzheimer’s Disease; SCZ: schizophrenia; Other: other non-AD neurodegenerative disorders.

We used the same approach to perform a meta-analysis of the discovery and replication datasets, including multiple brain regions. In the multi-tissue meta-analyses including all samples (n = 1,536), rs1990621 passed the genome-wide threshold for significance (p-value = 9.42×10⁻⁰⁹; Online Resource Table 5 and 6). The minor allele of rs1990621 was associated with increased neuronal proportion (β = 0.16; Figure 3ab, Online Resource Table 6). No other loci passed the GWAS threshold but there were eleven other loci that passed the suggestive threshold (Table 2 and Online Resource Table 5).

To determine the joint effects of genetic markers, genome-wide gene-based analyses were performed with MAGMA, based on the multi-tissue meta-analyses results. MAGMA provides a p-value for each gene by taking into account all the independent SNPs in a specific gene region. TMEM106B (p-value = 2.96×10⁻⁰⁸) was the only gene that passed the genome-wide significance threshold, followed by APOE (p-value = 3.2×10⁻⁰⁵), and APOE which is the most important genetic risk factor for sporadic AD (Figure 2cd, and Online Resource Table 7). However, a total of 27 genes showed a p-value < 1×10⁻⁰³ (Online Resource Table 7). Among them, PDE11A [32, 44, 45] and PRND [39, 71], which have been associated with AD; and KIAA1586 [24, 49], which has been previously associated with schizophrenia.

Disease-specific analyses: TMEM106B and neuronal proportion

TMEM106B is a known risk locus for FTLD [16, 29, 60, 80]. To determine if the association of TMEM106B with neuronal proportion was driven by FTLD or other diseases involved in this study, we stratified the cohorts by disease type to determine if the association of rs1990621 was driven by a single disease or multiple diseases. Samples were stratified by the following categories: AD (n = 639), non-AD neurodegenerative disorders (n = 103), schizophrenia (n = 189), young (n = 277) and old (n = 199) cognitively healthy control individuals. The non-AD neurodegenerative disorder category also includes other neurodegenerative diseases, mainly PSP and PA.

Rs1990621 in TMEM106B showed a strong association with neuronal proportion in AD (β = 0.26; p-value = 1.95×10⁻⁰⁷), and cognitively normal elderly individuals (β = 0.27; p-value = 1.34×10⁻⁰²), but not in the young controls (Figure 3d, and Online Resource Table 8). Rs1990621 was also associated with neuronal proportion in the non-AD dementia category, with a higher effect size (β = 0.45; p-value = 8.19×10⁻⁰⁴) than any other categories. These analyses indicate that the association of rs1990621 in TMEM106B with neuronal proportion is not unique to AD, but can be also found in other neurodegenerative diseases. The fact that this association is found in the normal elderly individuals with an effect size similar to that of the AD group also indicates that the association is not driven by pathological neuronal death.

In addition, no association was found in the in schizophrenic cohort (β = 0.01; p-value = 9.32×10⁻⁰¹) or cognitively normal young individuals (β = 0.05; p-value = 5.61×10⁻⁰¹; Figure 3d, and Online Resource Table 8). The samples included in the schizophrenic cohort and GTEx were significantly younger than the AD or non-AD neurodegeneration cohorts (p-value < 7.82 ×10⁻²⁴). This suggests that rs1990621 is associated with a protective mechanism mainly in neurons of aging brains and neurodegenerative diseases.

Functional annotation

The variant rs1990621 is located in the TMEM106B gene region. Although the CADD score and RegulomeDB score for this variant are not high enough to suggest any functional consequences (Online Resource Figure 5bc), this variant is in high LD with rs1990622 (r² = 0.98), a TMEM106B variant previously identified to be associated with FTLD risk [80], particularly in granulin (GRN) mutation carriers [16, 29]. Rs1990621 is also in high LD with rs1990620 (r² = 0.99), which is a cis-eQTL for TMEM106B [31]. Another variant in LD with rs1990621 is rs3173615 (r² = 0.98), which is located in the exon 6 of TMEM106B (dark blue dot in Online Resource Figure 5b), and produces a nonsynonymous amino acid change in TMEM106B: p.T185S. Previous studies indicate that the rs3173615 coding variant or rs1990620, is the functional variant driving the association in the FTLD risk GWAS [31, 60]. Our study finds the same signal related to this nonsynonymous change in TMEM106B. TMEM106B is expressed in neurons and microglia, with the highest protein expression detected in the late endosome/lysosomal compartments of neurons [9, 52, 72, 74]. The nonsynonymous SNP rs3173615 p.T185S affects TMEM106B protein levels by increasing TMEM106B protein degradation [9, 14, 60].

Estimation of neuronal proportion variance explained by associated genetic loci

To determine the proportion of variance in neuronal proportion explained by the common genetic loci, we analyzed all of the tested genotyped and imputed autosomal common variants (MAF > 0.02) using one-stage joint analysis. The approach we took is GCTA [90], which fits the effects of all the common SNPs as random effects in a mixed linear model. Then SNP variances are estimated based on the genetic relationship matrix between individuals. After correcting for age, sex, median TIN and the first two principal components, approximately 50.87% (p-value = 3.73×10⁻⁰³) of the variability in neuronal proportion was explained by common variants. The top SNP, rs1990621, accounts for 2.28% of the variability in neuronal proportion (Table 2); the SNP rs429358 in APOE gene region accounts for 1.27% of neuronal proportion variability. The top 12 loci in the meta-analyses, that include APOE and TMEM106B explain 11.18% of the total variation in neuronal proportion. This result indicates that a large proportion of the genetic hereditability is not explained by the top 12 loci and that that many genetic loci are yet to be discovered.

The impact of other neurodegenerative risk loci on neuronal proportion

To investigate what other AD or FTLD variants might have an effect on neuronal proportion in QTL analysis, we analyzed the 38 genome-wide significant loci for AD [51], FTLD [27], and FTLD with TDP pathology [80]. Among those, only common variants (MAF > 0.02) located in the TMEM106B and APOE gene were associated with neuronal proportion. Both rs1990622 (TMEM106B; Figure 2e) and rs2075650 (TOMM40/APOE; Figure 2f) were found to be associated with FTLD (Table 3). The top signals in the APOE region are rs283815, rs769449, and rs429358 with p-values < 1.22×10⁻⁰⁵. Note that rs429358 is one of the two SNPs that determines APOE isoforms. APOE ε4 and ε 2 alleles, coded by rs429358(C) and rs7412(C), confer the largest effect for AD risk. We observed that the C allele of rs429358 was associated with decreased neuronal proportion significantly, but no significant association was observed between rs7412 and neuronal proportion. Our gene-based analysis also indicates that there is an overlap in the genetic architecture of neuronal proportion between AD and FTLD, as the top two genes were TMEM106B (p-value = 2.96×10⁻⁰⁸) and APOE (p-value = 3.2×10⁻⁰⁵; Figure 2cd).

Table 3.

Neuronal proportion cQTL p-values were reported for variants previously identified in AD risk (by Lambert et al.), FTLD risk (by Ferrari et al.), and FTLD-TDP risk (by Van Deerlin et al.) studies.

					cQTL		AD risk		FTLD risk		FTLD-TDP risk
SNP	Gene	Minor	MAF	SNP proxy	Effect	p-value	OR	p-value	OR	p-value	OR	p-value
rs6656401	CR1	A	0.157	-	−0.09	5.59×10−02	1.18	5.7×10−24	-	-	-	-
rs730482	BIN1	T	0.327	rs6733839	−0.03	3.73×10−01	1.22	6.9×10−44	-	-	-	-
rs35349669	INPP5D	T	0.426	-	−2.33×10−03	9.80×10−01	1.08	3.2×10−08	-	-	-	-
rs190982	MEF2C	G	0.353	-	3.08×10−04	5.88×10−01	0.93	3.2×10−08	-	-	-	-
rs9271192	HLA-DRB5–HLA-DRB1	C	0.268	-	−0.04	4.31×10−01	1.11	2.9×10−12	-	-	-	-
rs10948363	CD2AP	G	0.249	-	−0.01	8.62×10−01	1.1	5.2×10−11	-	-	-	-
rs2718058	NME8	G	0.386	-	3.76×10−04	9.99×10−01	0.93	4.8×10−09	-	-	-	-
rs1476679	ZCWPW1	G	0.266	-	−0.02	7.71×10−01	0.91	5.6×10−10	-	-	-	-
rs11771145	EPHA1	A	0.371	-	0	9.67×10−01	0.9	1.1×10−13	-	-	-	-
rs28834970	PTK2B	C	0.332	-	−0.07	6.38×10−02	1.1	7.4×10−14	-	-	-	-
rs1532277	CLU	T	0.362	rs9331896	1.42×10−03	9.92×10−01	0.86	2.8×10−25	-	-	-	-
rs10838725	CELF1	C	0.275	-	−0.07	8.60×10−02	1.08	1.1×10−08	-	-	-	-
rs7124974	MS4A6A	T	0.365	rs983392	−0.05	2.05×10−01	0.9	6.1×10−16	-	-	-	-
rs10792832	PICALM	A	0.340	-	0.01	4.50×10−01	0.87	9.3×10−26	-	-	-	-
rs11218343	SORL1	C	0.045	-	0.12	1.35×10−01	0.77	9.7×10−15	-	-	-	-
rs17125944	FERMT2	G	0.084	-	−0.1	1.71×10−01	1.14	7.9×10−09	-	-	-	-
rs10498633	SLC24A4-RIN3	A	0.213		0.01	8.32×10−01	0.91	5.5×10−09	-	-	-	-
rs1460595	DSG2	A	0.038	rs8093731	−0.05	3.81×10−01	0.73	1.0×10−04	-	-	-	-
rs4147929	ABCA7	A	0.159		0.04	4.78×10−01	1.15	1.1×10−15	-	-	-	-
rs429358	APOE	C	0.151	-	−0.24	1.22×10−05	1.35	6.7×10−536	-	-	-	-
rs3865444	CD33	A	0.287	-	0.05	2.02×10−01	0.94	3.0×10−06	-	-	-	-
rs7274581	CASS4	G	0.105	-	0.02	3.01×10−01	0.88	2.5×10−08	-	-	-	-
rs1980493	BTNL2	G	0.144	-	−0.02	7.83×10−01	-	-	0.77	1.57×10−08	-	-
rs9268856	HLA-DRB5	A	0.270	-	−0.02	8.12×10−01	-	-	0.81	5.51×10−09	-	-
rs9268877	HLA-DRB5	A	0.427	-	−0.05	1.91×10−01	-	-	1.21	1.05×10−08	-	-
rs1020004	TMEM106B	G	0.315	-	0.1	1.35×10−03	-	-	1.03	4.59×10−01	0.60	5.00×10−11
rs6966915	TMEM106B	A	0.461		0.16	1.24×10−08	-	-	1.07	1.21×10−01	0.61	1.63×10−11
rs1990622	TMEM106B	G	0.467	-	0.16	1.44×10−08	-		1.08	7.88×10−02	0.61	1.08×10−11
rs10128715	RAB38/CTSC	A	0.122	rs74977128	0.03	7.29×10−01	-	-	1.81	3.06×10−08	-	-
rs302668	RAB38	C	0.308	-	−0.08	4.57×10−02	-	-	0.81	2.44×10−07	-	-
rs302665	RAB38	G	0.246	rs302652	−0.07	3.65×10−02			0.73	2.02×1008	-	-
rs2075650	APOE	G	0.141	-	−0.2	2.11×10−04			1.30	8.81×1007	-	-

Minor: minor allele of the SNP; MAF: minor allele frequency in our genomic data; SNP proxy: the original SNP reported in previous studies but not present in our genomic data, and is replaced by the closest SNP in LD found in our data; OR: odds ratio reported in previous studies.

DISCUSSION

We have assembled the largest dataset to date to analyze the genetic architecture underlying aging brains. In this study, we analyzed RNA-seq data from more than 2,008 brain samples (1,536 unique individuals). Using neuronal proportion as a quantitative trait, we found genetic variants associated with neuronal proportion. Our study identified a protective variant, rs1990621, in TMEM106B, which was associated with increased neuronal proportion (p-value = 6.40×10⁻⁰⁷). This association was replicated in an independent dataset (p-value = 7.41×10⁻⁰⁴) showing a significant genome-wide p-value in the meta-analyses (p-value = 9.42×10⁻⁰⁹). In this study we used multiple datasets that include different cortical regions, and different overall characteristics (for example, RNA library preparation, TIN, post-mortem interval, and age of death, between other variables (see Table 1)). This will lead to heterogeneity in the study and lower the statistical power of the analyses. It is possible that additional variants that influence neuronal proportion would require larger studies or studies with less heterogeneity in order to identify them. However, the fact that the rs1990621 replicated in a completely independent dataset indicates that this association is very robust and that this is in fact a real signal.

In addition, previous studies have already linked TMEM106B and rs1990621 with aging [69]. Rhinn et al, identified rs1990621 as the most significant signal for differential aging. Differential aging was calculated as the difference between the chronological and biological age. Although there is some overlap between the datasets used in Rhinn et al, mainly ROSMAP, most of the datasets used in Rihnn et al and the present study are different. Therefore, the fact that the same gene was identified and validated using different datasets and different phenotypes further supports the role of TMEM106B in neurodegeneration. Rhinn et al., also found that the effect of the TMEM106B variant started around 65 years old, which support our findings in which we observed an association of this variant in elderly, but not in young controls. At the same time, this study significantly extends the results from Rihnn et al. Firstly, the Rihnn et al study is focused on aging while this study is focused on neurodegenerative disease, especially Alzheimer Disease. We demonstrate that the TMEM106B is associated with neuronal proportion in AD, which is also supported by the Jun et al study, that reported that TMEM106B is associated with AD risk when taking into account APOE [41]. At the same time, this study goes beyond AD, and also indicates the TMEM106B is involved in other neurodegenerative diseases, such as PSP and FTLD. These findings point to a common underlying mechanism in these neurodegenerative diseases, which could have large implications in the field.

Because the top signal, rs1990621, is in LD with rs1990622, which is known to be associated with FTLD with TDP-43 inclusions [80], and our dataset includes FTLD individuals, we performed additional analyses to determine if this association was driven by individuals with FTLD or other neurodegenerative diseases such as PSP or PA. As expected, the strength of the association was larger in the group that included the FTLD samples (β = 0.45; p-value = 8.19×10⁻⁰⁴), but the association of TMEM106B with neuronal proportion was also found in AD (β = 0.26; p-value = 1.95×10⁻⁰⁷). More importantly, the effect size of TMEM106B with neuronal proportion is similar in AD and in elderly cognitively normal individuals (β = 0.27; p-value = 1.34×10⁻⁰²). This effect was not observed in a younger schizophrenia cohort with a mean age of death less than 65 years old (β = 0.01; p-value = 9.32×10⁻⁰¹), or in the younger cognitively normal individuals (β = 0.05; p-value = 5.61×10⁻⁰¹). These results indicate that rs1990621 in TMEM106B can be considered as a protective variant for neurodegeneration in general, not only for FTLD, as it also leads to a higher neuronal proportion in aging brains in general. The protective effect of TMEM106B is also found in healthy brains and could explain the lower risk of developing neurodegenerative diseases in individuals carrying these variants. As we did not see any association with the younger controls, it would be more likely that TMEM106B would be associated with better neuronal survival under challenging conditions, like aging, although additional studies are needed to test this hypothesis. On a similar note, it is important to recall that TMEM106B, was initially reported to be associated with FTLD [80], which is characterized by TDP-43 pathology. TDP-43 pathology can be also found in around 40% of the pathologically diagnosed AD cases [77], and increases with healthy aging [43, 79, 88]. Therefore, it is possible that individuals with the TMEM106B protective allele would develop less TDP-43 pathology and would be protected from neurodegenerative diseases. These results suggest a common pathway involving TMEM106B on aging brains in the presence or absence of neurodegenerative pathology that may contribute to cognitive resilience and neuronal protection. In addition, the second most significant signal in our gene-based analyses was APOE (p-value = 3.2×10⁻⁰⁵). The rs429358 variant that codes for the APOE ε4 isoform also showed a suggestive p-value for neuronal proportion (p-value = 1.22×10⁻⁰⁵). This observation suggests that there is an overlap in the genetic architecture of neuronal proportion with AD, neurodegeneration and aging.

Based on previous studies [16, 31, 69, 80] and our functional annotation analyses, it is very clear that the functional gene that drives the association with neuronal proportion is TMEM106B. There are two SNPs that implicate TMEM106B that are proposed to be the functional variant underlying the association. The first one is a protein coding variant (rs3173615) in high LD with rs1990621 (r² = 0.98), which leads to protein isoforms (p.T185S). The minor allele of rs1990622, which has a protective effect in FTLD [80], is in-phase with the minor allele of rs1990621, which is associated with increased neuronal proportion in our analysis. The protective role of the rs3173615 variant in TMEM106B is associated with attenuated cognitive deficits or better cognitive performance in ALS [81], hippocampal sclerosis [57], presymptomatic FTLD [66], aging individuals with various neuropathological burdens [89] and in the absence of known brain disease [69]. The S185 allele is protective and is associated with faster protein degradation than the common allele T185 [9, 14, 60]. The second one, rs1990620, is also in LD with rs1990621, and might also be the functional variant as it is a cis-eQTL for TMEM106B [31]. TMEM106B overexpression results in enlarged lysosomes and lysosomal dysfunction [9, 91]. It has also been shown that TMEM106B may interact with GRN in lysosomes [60]. Taken together, these results suggest that genetic variants regulate both neurodegeneration risk and the TMEM106B transcript and protein level in a cell autonomous manner.

Previous GWAS for AD risk performed with the International Genomics of Alzheimer’s Project (IGAP) data, stratified by APOE genotype, showed that AD risk is significantly influenced by the interaction between APOE and TMEM106B [41]. Together with our observation of cellular composition QTL, these results suggest that TMEM106B and APOE may play a role in affecting AD risk/vulnerability by affecting the cellular composition balance between neurons and astrocytes, and modulating endosomal and lysosomal function.

TMEM106B seems to affect risk for neurodegenerative disorders by regulating the endosomal/lysosomal pathway and playing a role in vulnerability to neurodegenerative disorders. Continuous lysosomal turnover of cellular contents through endocytosis and autophagy is crucial for neuronal survival [62]. Impaired lysosomal function reduces lysosomal degradative efficiency, which leads to abnormal build-up of toxic components in the cell. An impaired lysosomal system has been associated with normal aging [11] and a broad range of neurodegenerative disorders [55], including AD [61], PD [3, 70, 87], Huntington disease [28, 82], FTLD [54], ALS [28], Niemann-Pick disease type C [46, 63], Creutzfeldt-Jakob disease [53], Charcot-Marie Tooth disease type 2B [73], Neuronal ceroid lipofuscinoses (Batten disease) [47, 48], autosomal dominant hereditary spastic paraplegia [68], Chediak-Higashi syndrome [50], inclusion body myositis [4], and osteopetrosis [40]. During stress, the autophagylysosome pathway facilitates survival through clearance of damaged molecules and mobilization of intracellular stores of energy and nutrients. Abundant evidence supports the autophagy-lysosome pathway as a critical regulator of lifespan. The autophagy-lysosome activity declines with age and lysosome impairment predisposes one to assorted age-related diseases, including neurodegeneration.

Although each neurodegenerative disorder has its own characteristic proteopathy, the boundaries of protein pathology distribution are never clear-cut across different disorders. In fact, co-pathology or nonspecific proteopathy have been observed in most autopsies of neurodegenerative disorders, such as TDP-43 discussed above, Lewy bodies, α-synuclein and others [25]. Our observation of the association of variants in a lysosomal gene, TMEM106B, with neuronal proportion in aging cohorts supports the involvement of the lysosomal pathway in a common biological mechanism underlying a broad range of neurodegenerative disorders or aging processes in general that contribute to neuronal cell death. It also suggests that modulating the lysosomal function by regulating TMEM106B levels could contribute to the neuroprotective effect observed here. Lysosome-mediated degradation of TMEM106B could be a novel neuroprotective molecular target for aging and neurodegenerative diseases.

Our study has demonstrated the utility of using cell-type composition as a quantitative trait to identify genes associated with changes in cell populations. This approach is particularly powerful for complex disorders that involve considerable changes in cellular composition, for example, neurodegenerative disorders, as well as normal changes during developmental or aging processes. The development of recent single-cell studies will greatly increase the resolution of these methods and will help to advance our knowledge of cellular population changes [59]. With more detailed fine-mapping of cellular composition from single-cell studies and machine learning algorithms, bulk RNA-Seq deconvolution will more accurately capture cellular fraction changes in the samples, such as different types of neurons or different states of astrocytes or microglia. Regarding scalability, this bulk RNA-Seq deconvolution-based QTL is preferable because it is unclear if single-cell sequencing technology can be used on large numbers of subjects due to its current cost. Thus, currently bulk-RNA deconvolution approaches are preferable for carrying out cell-type composition QTL analysis. We predict that with larger sample sizes more hidden signals will be revealed as statistical power increases.

In conclusion, we identified a protective variant, rs1990621, in TMEM106B that is associated with increased neuronal proportion through bulk RNA-Seq deconvolution and neuronal proportion QTL analysis. This observation also confirms previous findings of the protective variant rs1990622 in FTLD risk, which is in high LD with rs19990621 [80]. In addition, we observed a significant association with increased neuronal proportions in the APOE region. This suggests the potential involvement of both APOE and TMEM106B in neuronal protection mechanisms underlying neurodegenerative and normal aging processes, and supports the previous observations of interactions between these two genes [41] in AD cohorts. With larger sample sizes and higher deconvolution resolution, this approach will reveal more biologically relevant and novel loci associated with changes in cellular composition to allow better interpretation of transcriptomic results in the context of both disease etiology and healthy aging. TMEM106B regulates lysosomal function, therefore it is possible that TMEM106B-related lysosomal changes might be involved in a common pathway underlying neuronal death and astrocytosis in neurodegenerative disorders and normal aging cohorts, and may be a potential target for neuronal protection therapies.