Human whole genome genotype and transcriptome data for Alzheimer’s and other neurodegenerative diseases

Previous genome-wide association studies (GWAS), conducted by our group and others, have identified loci that harbor risk variants for neurodegenerative diseases, including Alzheimer's disease (AD). Human disease variants are enriched for polymorphisms that affect gene expression, including some that are known to associate with expression changes in the brain. Postulating that many variants confer risk to neurodegenerative disease via transcriptional regulatory mechanisms, we have analyzed gene expression levels in the brain tissue of subjects with AD and related diseases. Herein, we describe our collective datasets comprised of GWAS data from 2,099 subjects; microarray gene expression data from 773 brain samples, 186 of which also have RNAseq; and an independent cohort of 556 brain samples with RNAseq. We expect that these datasets, which are available to all qualified researchers, will enable investigators to explore and identify transcriptional mechanisms contributing to neurodegenerative diseases.


Methods
The repository of human whole genome genotype and transcriptome data described herein (Table 1, Fig. 1) consist of the following resources some of which have previously been published: Previously published datasets include whole genome genotype data from the Mayo LOAD GWAS 1 (Data Citation 2) and microarray-based whole transcriptome data from the Mayo eGWAS 10 (Data Citations 3,4). Next-generation RNA-sequencing (RNAseq) data from a subset of the patients from the Mayo Clinic eGWAS, referred to as the 'Mayo Pilot RNAseq' (Data Citation 5), was published in part 18 . A non-overlapping cohort with RNAseq-based transcriptome data named 'Mayo RNAseq' (Data Citations 6,7) has also been published in part 26 . For a comprehensive description of the overall repository, the data from the published studies are also described herein, albeit in an abbreviated fashion. These four study cohorts will be referred to by their names as mentioned above, preceded by letters A-D (Table 1) henceforth.

Study Populations
All of this work was approved by the Mayo Clinic Institutional Review Board. All human subjects or their next of kin provided informed consent. The characteristics of the four study populations are as follows: Mayo LOAD GWAS. The characteristics of the cohort for this study (Data Citation 2) were previously described in detail 1 . Briefly, this is a LOAD case versus control study composed in total of 2,099 subjects sourced from three different series, namely: Mayo Clinic Jacksonville, Mayo Clinic Rochester and Mayo Clinic Brain Bank series. These series are respectively termed as JS, RS and AUT in the GWAS publication 1 (Table 1). Subjects in the Mayo Clinic Jacksonville and Mayo Clinic Rochester series were diagnosed clinically. These series consisted of 353 LOAD cases versus 331 controls; and 245 LOAD cases versus 701 controls. The Mayo Clinic Brain Bank series is a post-mortem cohort that consists of 246 LOAD cases versus 223 controls. All subjects were North American Caucasians. All clinical LOAD subjects were diagnosed as probable or possible AD, according to NINCDS-ADRDA criteria 27 . All clinical controls had a clinical dementia rating score of 0. LOAD subjects in the Mayo Clinic Brain Bank series met neuropathologic criteria for definite AD and had a Braak score of ≥4.0 (ref. 28), while controls did not meet neuropathologic criteria for AD, and each had Braak score of ≤2.5, which is an intermediary level of neurofibrillary tangle pathology between Braak score of 2 and 3; but most controls had neuropathologies unrelated to AD, including vascular dementia, frontotemporal dementia, dementia with Lewy bodies, multi-system atrophy, amyotrophic lateral sclerosis, and progressive supranuclear palsy. Ages, APOE ε4 genotype and sex distribution for the Mayo LOAD GWAS cohort are shown in Table 2. This study only included subjects with ages between 60 and 80 years, based on the assumption that much of the genetic risk for LOAD will be concentrated in this age group, especially given the age-dependent effects of the strongest AD risk variant apolipoprotein E ε4 (APOE4) 28 . Age for the clinically diagnosed LOAD cases is defined as age at first diagnosis of AD, since age at onset is not always available. Age at entry into the study is used for the clinically diagnosed controls. Age at death is utilized for the cases and controls in the postmortem Mayo Clinic Brain Bank series, given that for this cohort, age at clinical diagnosis/ evaluation is not always available. Illumina Hap300 microarray genotypes from the subjects in these three case-control series were utilized to conduct a GWAS of LOAD risk 1 .
Mayo eGWAS. This cohort was previously described in detail 10  WG-DASL gene expression measures for a subset of Mayo Brain Bank subjects that were included in the Mayo LOAD GWAS: RNA was isolated from two brain regions: TCX and CER. This data is utilized to identify loci associated with brain gene expression in subjects with AD, subjects with Other brain pathologies that do not meet criteria for AD (Non-AD), and the combined cohort. (Data Citation 2) (Fig. 1). The Mayo eGWAS is a whole transcriptome expression study in which brain samples from two different regions were analyzed, namely cerebellum (CER), which is relatively spared in AD, and temporal cortex (TCX), which is typically one of the first regions to be affected with AD neuropathology 29 . Transcriptome measurements were obtained from TCX of 202 AD subjects and from CER of 197 AD (Table 1). This study also included subjects without AD neuropathology, which are referred to as non-AD, given that many of these subjects had other neuropathologies. There were 197 non-AD subjects with TCX transcriptome measurements with the following neuropathologic diagnoses: progressive supranuclear palsy (PSP, n = 107); Lewy body disease (LBD, n = 25); corticobasal degeneration (CBD, n = 22); frontotemporal lobar degeneration (FTLD, n = 16); multiple system atrophy (MSA, n = 11), vascular dementia (VaD, n = 6); other (n = 10). There were 177 non-AD subjects with CER transcriptome measurements that had the following neuropathologies: PSP (n = 98); LBD (n = 23); CBD (n = 22); FTLD (n = 15); MSA (n = 7); VaD (n = 4); other (n = 8). Eighty-five percent of the subjects in the TCX cohort overlapped with those in the CER cohort. Demographics for the Mayo eGWAS subjects and samples, including RNA quality as assessed by RNA Integrity Numbers (RIN) are shown in Table 2.
Mayo Pilot RNAseq. All subjects in the Mayo Pilot RNAseq study (Data Citation 5) are a subset of the Mayo eGWAS (Data Citations 3,4), and are therefore also participants of the Mayo Clinic Brain Bank series that was included in the Mayo LOAD GWAS (Data Citation 2) (Fig. 1). The diagnostic categories in the Mayo Pilot RNAseq consist of 94 subjects with AD neuropathology and 92 PSP subjects, previously described 18,26 . PSP is a primary tauopathy characterized neuropathologically by neurofibrillary tangles (NFT) and tau-positive glial lesions 29,30 ; and often presents clinically as a parkinsonian disorder. All PSP   28 NFT stage of 3.0 or less, CERAD 31 neuritic and cortical plaque densities of 0 (none) or 1 (sparse) and lacked any of the following pathologic diagnoses: AD, Parkinson's disease (PD), DLB, VaD, PSP, motor neuron disease (MND), CBD, Pick's disease (PiD), Huntington's disease (HD), FTLD, hippocampal sclerosis (HipScl) or dementia lacking distinctive histology (DLDH). Subjects with pathologic aging also lacked the above diagnoses and had Braak NFT stage of 3.0 or less, but had CERAD neuritic and cortical plaque densities of 2 or more. None of the pathologic aging subjects had a clinical diagnosis of dementia or mild cognitive impairment. Given the presence of amyloid plaques, but not tangles and the absence of dementia, pathologic aging is considered to be either a prodrome of AD or a condition, in which there is resistance to the development of NFT and/or dementia 32 . Within the Mayo RNAseq cohort (Data Citations 6,7), all AD and PSP subjects were from the Mayo Clinic Brain Bank, and all pathologic aging subjects were obtained from the Banner Sun Health Institute. Thirty-four control CER and 31 control TCX samples were from the Mayo Clinic Brain Bank, and the remaining control tissue was from the Banner Sun Health Institute. All subjects were North American Caucasians. All but control subjects, had ages at death ≥60, and a more relaxed lower age cutoff of ≥50 was applied for normal controls to achieve sample sizes similar to that of AD and PSP subjects. No upper age limit was imposed on this cohort, however when subjects had ages at death of ≥90, their ages were recorded as '90_or_above' and shown as '90' in Table 2 to protect patient confidentiality. Table 2 details the demographic characteristics of the Mayo RNAseq cohort (Data Citations 6,7). PSP subjects tended to be younger than the other diagnostic groups. As expected, there was a greater frequency of APOE4 positive subjects in the AD group, followed by pathologic aging, then PSP and control subjects. AD and pathologic aging subjects had greater female sex frequency (57%), followed by controls (49%), then PSP subjects (39%). RIN for all samples were selected to be ≥5.0. Pathologic aging and control samples had slightly lower RINs than AD and PSP samples, due to limitations in availability of samples in these former diagnostic categories.

Molecular Data
Sample collection and processing. For the Mayo LOAD GWAS (A) (Data Citation 2), DNA samples were collected and processed as previously described 1 . For the antemortem Mayo Clinic Jacksonville and Mayo Clinic Rochester series, whole blood samples were collected in 10 ml EDTA tubes followed by DNA    extraction using AutoGenFlex STAR instrument (AutoGen), whereas cerebellar tissue was used for DNA extraction from the postmortem Mayo Clinic Brain Bank series using the Wizard Genomic DNA purification kit (Promega). Given limited amounts of DNA from samples in the Mayo Clinic Rochester series and Mayo Clinic Brain Bank series, whole genome amplification (WGA) was applied using the Illustra GenomiPhi V2 DNA Amplification Kit (GE Healthcare Bio-Sciences), in four 5 ml reactions that utilized 5-15 ng genomic DNA as a template. Subsequent to the pooling of these reaction products, WGA DNA was subjected to quality control (QC) using SNP genotyping as previously described. RNA extraction methods for the Mayo eGWAS 10 (B) (Data Citations 3,4) and Mayo Pilot RNAseq 18 (C) (Data Citation 5) were previously described. Total RNA was extracted from frozen brain samples using the Ambion RNAqueous kit (Life Technologies, Grand Island, NY) according to the manufacturer's instructions. Brain samples for the Mayo RNAseq (D) (Data Citations 6,7) study underwent RNA extractions via the Trizol/chloroform/ethanol method, followed by DNase and Cleanup of RNA using Qiagen RNeasy Mini Kit and Qiagen RNase -Free DNase Set. The quantity and quality of all RNA samples were determined by the Agilent 2100 Bioanalyzer using the Agilent RNA 6000 Nano Chip (Agilent Technologies, Santa Clara, CA). Samples had to have an RNA Integrity Number (RIN) ≥5.0 for inclusion in either study ( Table 2).
Data generation. The genotype data for the Mayo LOAD GWAS (A) (Data Citation 2) was generated using HumanHap300-Duo Genotyping BeadChips 1 , which were processed with an Illumina BeadLab station at the Mayo Clinic Genotyping Shared Resource (currently Mayo Clinic Medical Genome Facility = MGF, Rochester, Minnesota) according to the manufacturer's protocols. Two samples were genotyped per chip for 318,237 SNPs across the genome. Genotype calls were made using the auto-calling algorithm in Illumina's BeadStudio 2.0 software.
For the Mayo eGWAS study (B) (Data Citations 3,4), transcript levels were measured using the Whole Genome DASL assay (Illumina, San Diego, CA) as previously described 10 . Probe annotations were done based on NCBI RefSeq, Build 36.2. The RNA samples were randomized across the chips and plates using a stratified approach to ensure balance with respect to diagnosis, age, gender, RIN and APOE genotype. Raw probe mRNA expression data were exported from GenomeStudio software (Illumina Inc.) and preprocessed for background correction, variance stabilizing transformation, quantile normalization and probe filtering using the lumi package of BioConductor 33 .
Samples for both Mayo Pilot RNAseq (C) (Data Citation 5) and Mayo RNAseq (D) (Data Citations 6,7) studies were randomized prior to transfer to the Mayo Clinic MGF Gene Expression Core for library preparation and then the Sequencing Core for RNA sequencing. Mayo Pilot RNAseq (C) (Data Citation 5) AD and PSP samples were randomized across flowcells, taking into account age at death, sex and RIN. These samples underwent library preparation and sequencing at different times and therefore should be considered as separate datasets. Likewise, Mayo RNAseq (D) of TCX 26  Data Processing. Mayo LOAD GWAS (A) (Data Citation 2) genotypes from Illumina BeadStudio 2.0 software were utilized to generate lgen, map and fam files that were imported into PLINK 34 and converted to binary ped (.bed) and map (.bim) files, which are deposited together with PLINK format fam and covariate files (DOI and descriptions for each these files are provided in Table 3 (available online only)).
The Mayo eGWAS WG-DASL microarray expression dataset from TCX and CER (B) includes covariates and probe expression levels (Data Citation 3), which are preprocessed as published 10 and described above. The Mayo eGWAS 'eSNP Results' (Data Citation 4) are the eQTL results from the test of association between the Mayo LOAD GWAS (Data Citation 2) genotypes and the WG-DASL gene expression measures analyzed by multivariable linear regression using an additive model in PLINK 34 , as published previously 10 (DOI and descriptions for each these files are provided in Table 3 (available online only)). These analysis used preprocessed probe transcript levels as traits, SNP minor allele dosage as the independent variable, and adjusted for the following covariates: APOE ε4 dosage (0, 1, 2), age at death, sex, PCR plate, RIN and adjusted RIN squared (RIN-RINmean) 2 . Analyses were limited to SNP-probe pairs that were in-cis, defined as +/ − 100 kb of the targeted gene according to NCBI Build 36. The ADs and non-ADs were analyzed both separately and jointly. The joint analyses included diagnosis as an additional covariate (AD = 1, non-AD = 0). Results of analyses for both the genotyped SNPs as well as genotypes imputed to HapMap2 reference are provided. HapMap2 imputations were done as described 10 . The eGWAS results were previously made available through the NIAGADS repository (https://www. niagads.org/datasets/ng00025).
The Mayo Pilot RNAseq 18 (Data Citation 5), Mayo RNAseq TCX 26 and CER data (Data Citations 6,7, respectively) were processed using the same analytic pipeline. Read alignments were done using the www.nature.com/sdata/ SCIENTIFIC DATA | 3:160089 | DOI: 10.1038/sdata.2016.89 SNAPR software 35 , an RNA sequence aligner based on SNAP, using GRCh38 reference and Ensembl v77 gene models. Outputs include per-sample gene and transcript counts, which are merged into a single file per data type (gene or transcript) that contains data for all samples across all genes/transcripts (DOI and descriptions for each these files are provided in Table 3 (available online only)). Alignment with SNAPR starts with the creation of hash indices built from both a reference genome GRCh38 and transcriptome GRCh38.77. SNAPR filters fastq reads by Phred score (>80% of the read must have a Phred score > = 20) and simultaneously aligns each read (or read pair) to both the genome and transcriptome. The best alignment is written to a sorted BAM file with read counts simultaneously tabulated and written for each sample. Read counts are given by gene ID and transcript ID (two separate files). We have previously tested the read counts generated by SNAPR to the read counts generated by HT-Seq and found them to be very comparable.
Post-processing was also performed using the same pipeline for these three RNAseq datasets as follows: The individual read count files produced by SNAPR are merged into a single file using two scripts: merge_count_files.R and a dataset-specific read-count merge script. These scripts generate the corresponding _counts.txt.gz files. The merged count files are normalized with the normalize_ readcounts.R script, which uses the edgeR implementation of the trimmed mean of M-values (TMM) normalization method to calculate counts per million (CPM). These normalized counts are saved for both gene and transcript levels (DOI and descriptions for each these files are provided in Table 3 (available online only)).
Code Availability. The R script called merge_count_files.R 36 was used to merge the RNAseq read count files produced by SNAPR into a single file, and can be found at https://github.com/CoryFunk/ AMP-AD-scripts/blob/master/combine_count_files.pl. Also, the R script used to normalize the merged RNAseq read counts, called normalize_readcounts.R 36 , can be found at https://github.com/CoryFunk/ AMP-AD-scripts/blob/master/tmm_normalization.R.

Data Records
Data available for studies A-D (Data Citations 2-7; Table 3 (available online only)) consists of a set of files that contain genomic, genetic or covariate data for a defined set of samples; analytic results are also provided when available. Data files can be found in the Sage Bionetworks AMP-AD Knowledge Portal (Data Citation 1) in study specific folders (and subfolders). Users can identify and search for data files and data descriptions using the unique Synapse ID and corresponding DOI provided in Table 3 (available online only). Each sample within a study has a unique sample ID, this sample ID is consistent across all files within the study, and files in other studies where applicable. The relationship between studies and sample overlaps is illustrated in Fig. 1

Technical Validation Data QC
Mayo LOAD GWAS (A) (Data Citation 2) QC methods were previously published 1 . Briefly, using PLINK 34 , subjects with genotyping call rates of o90%, duplicate genotyping and/or sex-mismatches between recorded and deduced sex were eliminated from the dataset. All SNPs with genotyping call rates o90%, minor allele frequencies o0.01, and/or Hardy-Weinberg p values o0.001 were also eliminated. Prior to QC, 318,237 SNPs were genotyped in 2,465 subjects. The available data includes the 313,504 SNP genotypes from 2,099 subjects that passed these QC parameters.
The Mayo eGWAS 10 (B) (Data Citations 3,4) data was generated as follows: We annotated probes for presence of genetic variants by comparing their positions according to NCBI RefSeq, Build 36.3 to those of all variants within dbSNP131 and identified the list of probes that have ≥1 variants within their sequence. We depict this information in the files for the Mayo eGWAS, 'eSNP Results' (Data Citation 4) ( Table 3 (available online only)), by including 'SNP-In-Probe' column, which has 'TRUE' if the probe sequence harbors ≥1 SNP, and 'FALSE', otherwise. We also calculated for each probe within each analytic group, percent detection rate above background. Probes that are detected in >12.5%, >25%, >50% and >75% of the subjects in each analytic group are annotated by four separate columns within the 'eSNP Results' (Data Citation 4) from the eGWAS that included HapMap2 imputed genotypes, described below. The purpose of these annotation columns is to enable others the flexibility to impose cutoffs based on presence/absence of variants within probe sequence and/or probe detection rates while providing the full dataset for completeness. The Mayo eGWAS (Data Citation 3,Data Citation 4) also included replicate samples as described for QC and to estimate intraclass coefficients (ICC), which is the between-subject variance, as a percentage of the total variance in probe expression 10 . There were 4 AD and 4 non-AD temporal cortex samples that were measured in 5 replicates; and 10 AD and 5 non-AD cerebellar sample replicates across five plates. Universal human RNA (UHR) samples were also run on each PCR plate as part of QC. The expression phenotypes include results from only one of the replicate subjects selected randomly and exclude UHR results. It should be noted that 3 AD and 9 non-AD subjects for TCX, and 4 AD subjects for CER, do not have associated GWAS genotypes as they did not pass ≥ 1 GWAS QC parameter described above.
For the Mayo Pilot RNAseq 18 (C) (Data Citation 5) data principal components analysis (PCA) identified 2 outliers in the AD and 4 in the PSP cohort. The covariates for these subjects were set to missing ( = NA) in the respective covariate files (DOI and descriptions for these files are provide in Table 3 (available online only)). Hence, although 96 AD and 96 PSP subjects underwent sequencing in the Mayo Pilot RNAseq study, 94 AD and 92 PSP subjects were retained for analyses. It should be noted that of these subjects, 1 AD and 7 PSP subjects lack GWAS data due to either having genotype counts o90% or failing sex checks. PCA identified no outliers in the Mayo RNAseq (D) of TCX 26 samples (Data Citation 6) but 2 such subjects in the CER analyses (Data Citation 7). The covariate data in the relevant CER files for these two subjects were set to missing. We likewise assessed the RNASeq data for sex discrepancies based on Y chromosome gene expression and documented sex and identified 2 subjects with mis-matched sex for both TCX and CER, plus a third subject in the CER cohort. These were also set to missing in the covariate files. At the time of this publication, the Mayo RNAseq subjects did not have GWAS genotypes deposited on Synapse.

Usage Notes
The data described herein is available for use by the research community and has been deposited in the AMP-AD Knowledge Portal (Data Citation 1). Table 3 (available online only) provides a detailed description of the files deposited for the four studies, their specific Synapse identifiers (IDs), DOIs, the types of files and definitions of the column headers. These files (Data Citations 2-7), and their assigned DOIs will be maintained in perpetuity in the AMP-AD Knowledge Portal (Data Citation 1). Access to all of these files is enabled through the Sage Bionetworks, Synapse repository; and a subset of the files for the Mayo LOAD GWAS (Data Citation 2) and the Mayo eGWAS (Data Citations 3,4) are also available via NIAGADS (www.niagads.org).
The AMP-AD Knowledge Portal hosts data derived from multiple cohorts that were generated as part of or used in support of the AMP-AD Target Identification and Preclinical Validation project (Data Citation 1). The portal uses the Synapse software platform 25 for backend support, providing users with both web-based and programmatic access to data files. All data files in the portal are annotated using a standard vocabulary to enable users to search for relevant content across the AMP-AD datasets using programmatic queries. Data is stored in a cloud based manner hosted by Amazon web services (AWS), which enables user to execute cloud-based compute. Detailed descriptions including data processing, QC metrics, and assay and cohort specific variables are provided for each file as applicable.
Access for the data described herein is controlled in a manner set forth by the institutional review board (IRB) at the Mayo Clinic. All data use terms include: (1) maintenance of data in a secure and confidential manner, (2) respect for the privacy of study participants, (3) citation of the data contributors in any publications resulting from data use, and (4) informing data contributors of resultant publications. Specific data use terms are provided for each dataset (Data Citations 3-6) under the header 'Terms of use'; users must register for a Synapse account and provide electronic agreement to these terms prior to accessing the study files. Access to the Mayo LOAD GWAS data (A) (Data Citation 2) requires a data use certificate (doi:10.7303/syn2954402.2). User approvals are managed by the Synapse Access and Compliance Team (ACT).
Data on the AMP-AD Knowledge Portal are annotated with a common dictionary of terms (doi:10.7303/syn5478487.2) to enable querying of the data using the Synapse analytical clients (R client: syn1834618, python client: syn1768504, command line client: syn2375225). Fields, their allowable values specific to the datasets described herein and the dictionary of annotations are shown in Table 3 (available online only). These annotations can be used to identify files of interest within the available datasets and to filter on any of the fields using the allowable values from the dictionary (an example is shown here: doi:10.7303/syn5585666.1).