Challenges at the APOE locus: a robust quality control approach for accurate APOE genotyping

Background Genetic variants within the APOE locus may modulate Alzheimer’s disease (AD) risk independently or in conjunction with APOE*2/3/4 genotypes. Identifying such variants and mechanisms would importantly advance our understanding of APOE pathophysiology and provide critical guidance for AD therapies aimed at APOE. The APOE locus however remains relatively poorly understood in AD, owing to multiple challenges that include its complex linkage structure and uncertainty in APOE*2/3/4 genotype quality. Here, we present a novel APOE*2/3/4 filtering approach and showcase its relevance on AD risk association analyses for the rs439401 variant, which is located 1801 base pairs downstream of APOE and has been associated with a potential regulatory effect on APOE. Methods We used thirty-two AD-related cohorts, with genetic data from various high-density single-nucleotide polymorphism microarrays, whole-genome sequencing, and whole-exome sequencing. Study participants were filtered to be ages 60 and older, non-Hispanic, of European ancestry, and diagnosed as cognitively normal or AD (n = 65,701). Primary analyses investigated AD risk in APOE*4/4 carriers. Additional supporting analyses were performed in APOE*3/4 and 3/3 strata. Outcomes were compared under two different APOE*2/3/4 filtering approaches. Results Using more conventional APOE*2/3/4 filtering criteria (approach 1), we showed that, when in-phase with APOE*4, rs439401 was variably associated with protective effects on AD case-control status. However, when applying a novel filter that increases the certainty of the APOE*2/3/4 genotypes by applying more stringent criteria for concordance between the provided APOE genotype and imputed APOE genotype (approach 2), we observed that all significant effects were lost. Conclusions We showed that careful consideration of APOE genotype and appropriate sample filtering were crucial to robustly interrogate the role of the APOE locus on AD risk. Our study presents a novel APOE filtering approach and provides important guidelines for research into the APOE locus, as well as for elucidating genetic interaction effects with APOE*2/3/4. Supplementary Information The online version contains supplementary material available at 10.1186/s13195-022-00962-4.


Introduction
APOLIPOPROTEIN E*4 (APOE*4) is the strongest genetic risk factor for late-onset Alzheimer's disease (AD) [1]. In subjects of European ancestry, one copy of APOE*4 increases the risk of a clinical diagnosis of AD by about 3-fold and two copies increase the risk by about 12-fold [2,3]. APOE*2 on the other hand decreases the risk of AD by about half [3], while APOE*3 is the reference allele. Beyond the two common missense variants that compose APOE*2/3/4 (rs429358 and rs7412), there may be other coding variants on APOE or non-coding regulatory variants in the APOE locus that further impact AD risk, either independently or in conjunction with APOE*2/3/4 [4][5][6][7][8][9][10][11][12][13][14][15]. This pertains, by example, to a crucial question in the field: why do some APOE*4 carriers remain asymptomatic even into advanced old age? One possibility is that there may be genetic variants in the APOE locus that affect APOE*4 availability and in turn mitigate APOE*4-related risk for AD. Identifying such variants would importantly advance our understanding of APOE*4 pathophysiology and provide critical guidance for AD therapies aimed at APOE*4 [16,17].
Despite its therapeutic promise and three active decades of research, the APOE locus remains relatively poorly understood in AD. While there are multiple reasons contributing to this, one prominent one is that the APOE locus harbors multiple nearby genes and shows a complex linkage disequilibrium (LD) structure with APOE*2/3/4, making it difficult to identify causal variants and interaction effects [18,19]. Other important reasons are that relevant risk variants may be rare, thus requiring large sample sizes, and that the quality of the APOE*2/3/4 genotype can bear heavily on correctly identifying interaction effects and causal haplotypes. The latter may be of particular relevance given the plethora of available protein-based (e.g., two-dimensional gel electrophoresis and MALDI-TOF mass spectrometry) and DNA-based methods (e.g., TaqMan assays, high-resolution melting analysis, PCR sequencing, etc.) for APOE*2/3/4 genotyping [20][21][22][23][24]. Importantly, these methods have variable quality and limitations related to the haplotypic nature of APOE*2/3/4. For instance, protein-based assays may suffer from biases in detecting different APOE isoforms, while DNA-based assays can be affected by rare variants in the genomic region near APOE*2/3/4 (cf. Huang et al. for a detailed review) [25]. In turn, cohorts that are commonly included in genetic association studies of AD have used variable APOE genotyping methods [26][27][28][29][30], which has thus led to variable APOE*2/3/4 genotype quality across cohorts used in meta-analyses. The approach used to quality control the APOE*2/3/4 genotype is therefore critical to ensure robust association analyses. While the need for stringent APOE quality control is not necessarily novel, to our knowledge, there is currently no specific study that clearly addresses this issue, nor are there are any consensus guidelines.
In this study, we present analysis approaches and related findings to guide future research in the APOE locus. Specifically, we show findings for a large-scale analysis of rs439401 and its association with AD risk. This variant, located 1801 base pairs downstream of APOE, was recently identified as a brain APOE splice quantitative trait locus (sQTL) in GTEx [31,32], spurring our interest to investigate it. We hypothesized it may affect APOE*4-related risk for AD and observed that it is most often seen on the same chromosome copy as APOE*3 (i.e., is in-phase with APOE*3), but in rare instances was seen together with APOE*4. We thus stratified analyses according to APOE*3 and APOE*4 genotypes to evaluate whether effects depended on the variant being inphase with APOE*4. We use analyses on this variant to illustrate how critical it is to have accurate APOE*2/3/4 genotype data. Based on initial analyses using a conventional APOE filtering approach and a subsequent robustness assessment, we designed and present a novel APOE filtering approach that we believe will be highly relevant to help guide further reproducible research in this area.
whole-genome sequencing (WGS) ( Table S1). The discovery samples comprised publicly available case-control (majority), family-based, population-based, and longitudinal cohorts. Independent replication samples, genotyped on SNP microarrays, were available from three large cohorts: the Rotterdam study, a population-based prospective study, the European Alzheimer Disease Initiative (EADI), roughly two-thirds of which is from a prospective population-based study and one third from case-control samples, and the European Alzheimer & Dementia BioBank (EADB), which collated AD case-control samples from 15 European countries. Ascertainment of genotype/phenotype data for each cohort/project are described in detail elsewhere [33, 40-44, 46, 47, 54]. Cross-sample genotype/phenotype harmonization for the discovery samples is summarized in Supplementary Methods. Phenotypes from respective cohorts were updated as of March 2021. Data were analyzed between December 2019 and June 2021.

Genetic data quality control and processing
Genetic data in the discovery samples underwent standard quality control (QC; Plink v1.9) and ancestry determination (SNPweights v2.1; Fig. S1) [57]. Only non-Hispanic subjects of European ancestry (representing the vast majority of samples) were selected for processing. Data were restricted to those providing coverage of the rs439401 variant. Principal component analysis of genotyped SNPs provided principal components (PCs) capturing population substructure (PC-AiR, Fig. S2) [58]. Identity-by-descent (IBD) analyses reliably identified kinship down to 3rd degree relatedness (PC-Relate, Fig. S3) [58]. Sparse genetic relationship matrices (GRM) were constructed to enable analyses including related individuals [59]. SNP array data were used to perform genotype imputation with regard to the TOPMed imputation reference panel [60,61]. Genetic processing of Rotterdam, EADI, and EADB replication samples is described elsewhere [33,54]. Detailed descriptions of all processing steps are in Supplementary Methods and Table S2.

Ascertainment of rs439401
The rs439401 variant was originally included in our analyses as it had a cross-cohort genotyping rate >80% in the discovery samples. Genotypes were considered from either the direct call on the SNP array data (i.e., called from probe intensity data) or the call from WGS data. We specifically relied solely on directly genotyped data rather than using imputed data in order to obtain unbiased results. This choice was additionally motivated reasoning that putative rare haplotypes may not be accurately imputed, particularly when using the commonly younger (non-AD) individuals in imputation reference panels [60,62,63]. Genotype reliability for the variant was verified by cross-correspondence across 3804 duplicate samples in the discovery and by assessing genotype intensity data on the SNP microarray in EADB.

Ascertainment of APOE genotypes
Throughout, we will refer to APOE*2/3/4 genotypes as APOE genotypes. APOE genotypes were available from (1) cohort demographics (i.e., "provided" APOE), which generally had APOE genotype status determined through various direct genotyping methods (detailed elsewhere [33,54]), (2) directly from WES/WGS calls, or (3) through imputation of rs429358 (which captures the APOE*4 allele) and rs7412 (which captures the APOE*2 allele). It is relevant to note that rs429358 was never directly available on the SNP microarrays. It is further relevant to note that for the current WES data from ADSP, rs7412 was not available, with only rs429358 being reliably called in most subjects. The WES data could thus be used only to verify subjects with a provided APOE*3/3, 3/4, or 4/4 status (cf. Supplementary Methods).

APOE genotype filtering criteria
To our understanding, common criteria across prior studies regarding APOE genotypes can be summarized as giving priority to provided APOE genotypes when available (as direct genotyping methods are generally considered the gold standard), followed by using APOE genotypes derived from rs429358 and rs7412 when directly called on a SNP microarray, followed by inference of APOE genotypes through (high quality) imputation of rs429358 and rs7412. There is no clear consensus on whether or how any discrepancies across available APOE genotypes for a given subject should be adjudicated. Furthermore, with the recent increasing availability of WGS/ WES data in the AD field [42,46,51], these data can now also be used to verify APOE genotypes. When high-quality WGS/WES calls are available for rs429358 and rs7412 (i.e., good read depth/quality with a clear reference/alternate allele distribution) [64], the derived APOE genotype may be considered the ground truth. Recent work indeed suggests that a higher APOE genotype accuracy can be achieved using next-generation sequencing compared to conventional gold standard methods [65].

APOE filtering approach 1
Based on the above considerations, we designed criteria to use APOE genotypes according to the highest available quality. Specifically, when multiple APOE genotypes were available for a given subject, the APOE genotype we selected followed the priority of WGS/WES over provided/demographic sources (for details regarding "provided/demographic" APOE sources, please cf. above in the section "Ascertainment of APOE genotypes"). If APOE genotype was only available from provided/demographic sources and was discordant across duplicate samples, then those samples were flagged for exclusion (N = 73 out of 1501 (4.86%) unique subjects). Similarly, the correspondence between APOE genotypes derived from WES and WGS across duplicate samples was checked and only showed discordance in five subjects differing for APOE*2/3 and APOE*3/3 genotypes across the WES and WGS data (these subjects were excluded). The final set of samples used for association analyses thus did not display any mismatches in prioritized APOE genotypes across duplicates, but in some instances, the APOE genotype from provided/demographic versus WES/WGS sources differed. APOE status as inferred from imputation was entirely ignored, reasoning this was less reliable and that rare haplotypes of potential interest in the APOE locus may lead to false imputation of APOE*2/3/4 genotype.

APOE filtering approach 2
After further assessment of the initial results, we had concerns about the reliability of APOE genotype status in some APOE*4 subjects carrying rs439401 (cf. Results). We therefore expanded the first approach to exclude any subjects who had their prioritized APOE genotype determined from provided/demographic APOE but were still discordant with their imputed APOE genotype (N = 632 out of 12,753 (4.96%) in the discovery sample after passing all other filtering steps). Note that imputation scores (R 2 ) for rs429358 and rs7412 were never lower than 0.8. Information regarding APOE imputation, as well as several correspondence checks across different sources of APOE genotypes, are provided in the supplementary and referenced in the "Results" section. An additional check for APOE genotype consistency was also performed using newly released sequencing data from the ADSP (NG00067.v5) [66], processed in May 2021 (cf. Supplementary Methods and the "Results" section).
An overview of the study design and APOE filtering approaches is presented in Fig. 1.

Simulations of concordance rates between observed and true APOE*4/4 genotypes
In order to understand potential uncertainty in APOE*4/4 genotypes, we simulated different type I and II error rates for APOE*4/4 status. Type I error rate was defined as the probability, p1, to mis-classify non-APOE*4/4 carriers as APOE*4/4. Type II error rate was defined as the probability, p2, to mis-classify APOE*4/4 carriers as non-APOE*4/4. We considered a range of true frequencies, f true , for APOE*4/4 cases and controls respectively with regard to all cases and controls (that is, all APOE strata). This range for f true was centered on observations in the current discovery samples, which should represent a reasonable approximation of expected frequencies in casecontrol samples. The observed frequency, f obs , was then defined as f true *(1-p2)+(1-f true )*p1. The concordance rate between observed and true APOE*4/4 was finally defined as f true *(1-p2)/f obs .

Statistical analyses
Primary analyses evaluated associations of rs439401 with relative risk for AD in APOE*4/4 carriers using additive genetic models. In additional supporting analyses, associations were evaluated in APOE*3/4 carriers, comparing wild-type (WT) to homozygote (HOM) genotypes, ensuring rs439401 was in-phase with APOE*4. The expectation here was to observe similar but attenuated effects compared to associations in APOE*4/4 carriers. Additional associations were evaluated in APOE*3/4 and 3/3 carriers using additive genetic models, with the expectation of observing little or no effect if associations were conditional on being in-phase with APOE*4. APOE*2/4 carriers were not considered given sample paucity. Analyses were restricted to subjects aged 60 and above, consistent with age cutoffs in prior genetic studies of AD [54]. Replication analyses focused only on evaluating variants in-phase with APOE*4. Lastly, to provide additional insight into the putative role of rs439401 in AD, we evaluated the association of rs439401 with relative risk for AD in the full discovery sample, while adjusting for APOE*2 and APOE*4 dosage.
Cohorts in the discovery were combined into a single mega-analysis, included related subjects, and outcome measures were adjusted for age, sex, the first five genetic PCs, and the GRM. In full sample analyses, models further included APOE*2 and APOE*4 dosage as covariates. In replications, models included only unrelated subjects and were not adjusted for the GRM. EADI and Rotterdam further adjusted for the first three genetic PCs, while EADB adjusted for the first 20 genetic PCs and genotyping center. Notably, models in the discovery mega-analyses did not adjust for cohort, reasoning that this may inadvertently diminish power given variable cohort sizes and carrier distributions. This is especially relevant in case of lower frequency variants in the APOE*4/4 stratum, where cohort bins and the number of allele observations become very small. Still, to address potential concerns regarding cohort biases, in sensitivity analyses, the effect of cohort adjustment in the discovery was evaluated.
Associations with AD risk were evaluated under a case-control design using linear mixed-model regression in analyses of related subjects and logistic regression in analyses of unrelated subjects. Additional details for model/inclusion criteria are in Supplementary Methods. Association analyses were considered significant below a threshold P-value of 0.05. All analyses were performed in R v3.6.0.

Participant demographics and rs439401 linkage structure
Across all 142,075 genotyped samples considered in this study (Table S1), 65,701 unique participants passed filtering and inclusion criteria. Participant demographics for APOE*4/4 and 3/4 carriers are in Table 1, while detailed full sample demographics are in Table S3-4. In the discovery, rs439401 displayed high LD (D'>0.9) with APOE*3, but in rarer instances was observed in-phase with APOE*4, thereby deviating from the expected LD structure (Table S5).

APOE filtering approach 1: Rs439401 shows variable association with Alzheimer's disease risk
Primary case-control findings in APOE*4/4 carriers in the discovery showed that rs439401 displayed a strong, Fig. 1 Schematic overview of the study design and two APOE*2/3/4 filtering approaches protective, and significant effect on case-control status (Table 2). It displayed similar protective effect sizes in EADI and Rotterdam replication samples, but was risk increasing in EADB, and did not reach significance in any replication sample. When in-phase with APOE*4 in APOE*3/4 (WT-HOM) stratified analyses, rs439401 showed a protective significant effect in the discovery, but variable non-significant results in the replication samples ( Table 2). In contrast, in the discovery, rs439401 did not associate with AD risk in APOE*3/4 (additive model) or 3/3 stratified analyses (Table S6), nor in the full sample analysis (odds ratio = 0.99; 95% confidence interval = [0.95, 1.03], P-value = 0.61).
Because of the use of a mega-analysis design that does not adjust for cohort, there may still be concern for potential cohort biases. Therefore, as a sensitivity Table 1 Sample demographics for association analyses with Alzheimer's disease case-control status   analysis, we re-evaluated the case-control discovery findings, now adjusting for cohort or cohort/array/center (Fig. S5). These analyses indicated diminished significances, but effect sizes remained comparable and rs439401 remained strongly significant in APOE*4/4 carriers.

Robustness assessment: limitations to APOE filtering approach 1
After the initial analyses, we assessed the robustness of the primary discovery findings. This appeared particularly relevant considering the very low frequency of rs439401 carriers in APOE*4/4 controls in EADB versus other cohorts, suggesting potential biases in the controls across the cohorts. The concordance rate of rs439401 from duplicate samples across microarrays and WGS (99.97%) supported genotype reliability (Table S7). Similarly, the variant appeared confidently called from the EADB microarray intensity data (Fig. S4). Overall, we concluded there were no specific genotyping issues for rs439401.
Another important consideration is that some error rate is expected for the different direct APOE genotyping methods used across cohorts. Overall, the reliability of the APOE*4 genotype may thus be of concern especially when considering the rare APOE*4-rs439401 haplotype. After assessing all APOE*4/4-rs439401 carriers, it was apparent that one cohort, MIRAGE, contributed a large amount of APOE*4/4-rs439401 controls for which APOE status was available only from provided/demographic sources ( Fig. 2A, Table S8). We then assessed the concordance rate between provided and imputed APOE genotypes across all respective cohorts and observed that MIRAGE displayed the lowest concordance rate of all cohorts included in the discovery analyses (Fig. 2B), despite comparably high imputation scores for rs429358 and rs7412 to other cohorts (Table S9). Overall, this supported concern for the APOE*4/4-rs439401 controls from MIRAGE.
Extending on the above considerations, we assessed discordance rates between imputed and provided APOE for different strata (Fig. 2C, Table S10). Importantly, while the discordance rate was only 4.3% in the full sample, it increased to 7.2% in APOE*4/4 cases, further increased to 16.1% in APOE*4/4 controls, and then drastically increased to 47.4% in APOE*4/4-rs439401 carrier cases and 85.7% in APOE*4/4-rs439401 carrier controls. While our a priori assumption for approach 1 reasoned that imputed APOE may be discordant with provided APOE in case of subjects with rare haplotypes (e.g., APOE*4/4-rs439401 carriers), the observation that this discordance was 2-fold higher in controls compared to cases would not be expected. Rather, it more likely indicates that a miscall of the APOE genotype was true in at least some of these individuals. To better understand these observations, we performed simulation studies using different type I and type II error rates (0-5%) for APOE*4/4 genotyping and observed that APOE*4/4 controls were more likely than APOE*4/4 cases to not actually be APOE*4/4 carriers (Fig. S6-7). This was the result of the low frequency of APOE*4/4 controls and the strong case-control imbalance in APOE*4/4 carriers. Overall, this supported concern for the validity of approach 1.
We then used the recently released new ADSP WGS and WES data, which now cover additional subjects that are duplicated on SNP array samples included in our discovery analyses (N = 3644 as determined by identity-by-descent). We assessed the APOE genotype calls from the novel WES/WGS data and observed that three APOE*4/4-rs439401 control subjects (not from the MIRAGE cohort) in the prior discovery samples were in fact APOE*3/3 or APOE*3/4 carriers, which was also the imputed APOE genotype (Table S8). Overall, this again raised concern about the validity of approach 1.
In sum, these additional checks for robustness of the findings suggested problems with APOE genotype reliability in subjects with APOE*4-rs439401 haplotypes and APOE*4/4 carriers overall, indicating a limitation to the first (conventional) APOE filtering approach. In a final check, we observed that despite good concordance between provided and WGS APOE (99.1%), imputed and WGS APOE was more concordant (97.2%) than imputed and provided APOE (95.7%), indicating that at least in some subjects imputed APOE was likely more correct than provided APOE (Table S10).

APOE filtering approach 2: Rs439401 shows no association with Alzheimer's disease risk
In light of the identified APOE reliability limitations, we extended approach 1 to filter out any subjects that did not have WGS/WES APOE and at the same time were discordant for provided and (high-quality) imputed APOE. We also filtered out any discordant APOE calls with regard to the new ADSP WES/WGS data since this information was available (in case of APOE*4-rs439401 carriers, this overlapped with samples where provided and imputed APOE were discordant). We then applied this to the discovery samples and reran analyses. Exclusion of subjects with discordant APOE status with the newly released ASDP WES/WGS data removed 61 (out of 12,367) subjects from the SNP-array samples. Further applying the new APOE filter excluded 632 (out of 12,753 considered) subjects from the discovery SNP-array samples. APOE*4-rs439401 carrier frequencies dropped substantially, particularly in controls, and became more consistent with those observed in the haplotype reference consortium (Fig. 3A). Case-control association analyses now indicated no effects for APOE*4-rs439401 carriers (Table 3 and Fig. 3B, C) and still no effect in full sample analyses (odds ratio = 1.00; 95% confidence interval = [0.96, 1.05], P-value = 0.93). In sum, approach 2 produced results that were more realistic in terms of expected linkage structure and more consistent with the lack of significant replication findings.

Discussion
Our results demonstrate that the filtering criteria for APOE*2/3/4 genotypes can heavily impact association finding for variants that exert their effect in conjunction with APOE*2/3/4. Specifically, we used the APOE sQTL variant rs439401 to illustrate this point. Using more conventional filtering criteria regarding APOE genotypes (approach 1), we showed that, when in-phase with APOE*4, rs439401 was variably associated with protective effects on AD case-control status. However, when assessing the reliability of APOE*2/3/4 genotypes with more scrutiny and applying a novel filter to increase certainty of the APOE genotypes (approach 2), we observed that all significant effects were lost. The findings and methodology presented here are thus of high relevance to guide future research into the APOE locus. Specifically, we propose that our approach 2 can serve as a consensus APOE genotyping approach for future studies, namely, to prioritize first WGS/WES APOE*2/3/4 genotypes if  Table S8). The red arrow indicates that a large fraction of control rs439401 carriers was contributed by MIRAGE. B Concordance rates between provided and imputed APOE per cohort (additional data in Table S9). The red arrow indicates that MIRAGE had the lowest concordance rate, suggesting potential limitations with its provided APOE data that could explain observations in A. C Concordance rates between provided and imputed APOE for the discovery sample, considering multiple strata (additional data in Table S10). APOE*4/4 strata considered provided APOE*4/4 genotypes after applying APOE filtering approach 1. Note decreased concordance in APOE*4/4 controls compared to cases. Note strongly decreased concordance for rs439401 carriers, specifically controls. Simulations confirmed that APOE*4/4 controls are more likely than cases to not actually be APOE*4/4 carriers (cf. Fig. S6-7). Abbreviations: CN, cognitively normal; AD, Alzheimer's disease; OR available (and if only either rs429358 or rs7412 is available from WGS/WES, to use those genotype data to verify the provided/demographic APOE*2/3/4 genotypes); second to use provided/demographic APOE*2/3/4 genotypes; and third, in subjects without WGS/WES information, to exclude those for whom the provided/ demographic and imputed (R2>0.8) APOE*2/3/4 genotypes are discordant. Another important step to ensure the highest quality of APOE*2/3/4 genotypes is to verify and harmonize this information across available duplicate samples.
The rs439401 variant considered in the current study has previously been investigated with regard to AD risk in different contexts and using variable strategies and study designs [8][9][10][11]13]. Our analyses however considered a substantially larger sample size, essentially incorporating most European ancestry AD cohorts included in prior studies, specifically focused on evaluating effects Fig. 3 Overview of rs439401 frequencies and case-control association findings, comparing APOE filtering approach 1 to approach 2. A Carrier frequencies across both approaches for APOE*4/4 and APOE*3/4 WT vs HOM groups, as well as in the Haplotype reference consortium v1.1 (HRC). Note decreased frequencies for rs439401 in approach 2 that appear concordant with the HRC. B, C Overview of association findings for all evaluated strata, comparing B approach 1 to C approach 2. Significant effects are denoted by an asterisk (*). Error bars show 95% confidence intervals. Note loss of significant effects in approach 2 stratified to respective APOE genotypes, and tested only directly genotyped variants. Further, up-to-date genotype and phenotype data for a large set of AD cohorts was jointly harmonized to compose a parsimonious discovery sample. Non-European ancestries were not investigated here owing to the paucity of publicly available data. When compared to similar prior studies [6, 13-15], our discovery group was larger and we incorporated three large replication cohorts. Furthermore, through the implementation of linear mixed modeling and cross-sample harmonization, we were able to increase the power and specificity for variant discovery, while additionally verifying genotype reliability across nearly 4000 duplicate samples. In sum, our analyses should provide a robust assessment of the presented APOE filtering approaches and rs439401's association with AD risk. A recent study, using samples largely overlapping with the current discovery (but smaller in size) and an APOE filtering approach similar to our approach 1, evaluated the association of variants on the larger APOE locus with AD risk in APOE*4/4 carriers and did not identify the strong association of rs439401 that we observed in approach 1 [13]. Beyond differences in sample size and harmonization, the latter study adjusted models by study/cohort and made use of imputed genotypes. We specifically decided in primary analyses not to adjust for cohort, as we reasoned that this may inadvertently diminish power given variable cohort sizes and carrier distributions, especially in APOE*4/4 carriers. We further reasoned that through our extensive phenotype/ genotype harmonization and the use of a mixed model mega-analysis design, which may capture some latent cohort effects, there was less concern for potential cohort bias. Additionally, given the complex LD structure of the APOE locus, we were concerned about the reliability of imputation and focused only on directly called genotypes. A similar limitation regarding imputation was recognized by the authors of the prior study [13]. These differences likely explain why rs439401 was not observed in their study. Regardless of our considerations and of cohort adjustment, we determined that the APOE filtering criteria were the most relevant factor for variable rs439401 association findings.
One important insight from our study was that subjects, particularly controls, with a provided APOE*4/4 genotype had a higher probability of discordance between their imputed and provided genotype than did subjects in the full sample. Such biases are, however, not limited to APOE*4/4 carriers. The six APOE genotypes (*2/2, 2/3, 3/3, 2/4, 3/4, 4/4) show large differences in numbers of carriers and case-control ratios, owing to the allele frequencies of rs429358/rs7412 and their effect on AD risk. As a result, the different APOE genotypes will be expected to have different concordance rates between true and observed APOE genotypes. We observed varying concordance between imputed and provided APOE across the six APOE genotypes, with particularly lower concordance rates in APOE*2 carriers (Fig. S8). Just as the APOE*4/4 provided genotype was most likely to be incorrect here in controls (a phenotype for which APOE*4/4 is a particularly rare genotype), the APOE*2/2 genotype is more likely to be incorrect in cases (a phenotype for which APOE*2/2 is a particularly rare genotype). The proposed APOE genotype filter will therefore also be specifically relevant for studies focusing on APOE*2.
Our study highlights several important considerations for further work on the APOE locus. Most notably, we illustrate how APOE genotype filtering criteria can strongly impact association findings for variants in the APOE locus, especially when studying haplotypes or interaction effects with APOE*2/3/4. The same will hold true when considering non-local variants in, for instance, a genome-wide association study of AD in APOE*4/4 subjects, or when aiming to disentangle genetic interaction effects with APOE*2/3/4. Based on our observations, we suggest that future studies consider implementing the methodology that we proposed in approach 2 and subject their assessment of APOE genotypes to extensive scrutiny. The limitations observed for APOE*2/3/4 genotype reliability also emphasize that next-generation sequencing data will be crucial to interrogate the APOE locus with higher confidence and to ensure that putative rare haplotypes are not missed because of the need for sample filtering in SNP array data. Lastly, in order to have higher confidence in local haplotypes, long read sequencing approaches will additionally be crucial to help disentangle the local haplotype structure on APOE with regard to AD.

Limitations
One limitation of our proposed approach is that it relies on the availability of high-quality imputed genotypes for rs429358 and rs7412, as well as careful phenotype/genotype harmonization across multiple data sources, which may not always be feasible for different research groups. Nonetheless, our findings show that efforts to increase APOE*2/3/4 genotype reliability should be pursued and that collaborative large-scale AD harmonization initiatives should consider this as an important focus. Furthermore, our approach may be considered to be highly conservative when excluding subjects for which the imputed and provided APOE*2/3/4 genotypes are discordant, since some of the imputed APOE*2/3/4 genotypes may in fact be the correct ones. Future studies may thus also consider retaining those subjects, using their imputed APOE*2/3/4 genotypes. Lastly, we propose to prioritize WES/WGS APOE*2/3/4 genotypes given the high quality and reliability of these sequencing technologies. However, as detailed in the supplement, careful consideration of genotyping quality and depth, integrated with provided APOE*2/3/4 genotype information, were crucial to maximize APOE*2/3/4 genotype reliability. It will therefore be critical that such information is made readily available and evaluated in future studies.

Conclusion
We showed that careful consideration of APOE genotype and appropriate sample filtering was crucial to robustly interrogate the role of the APOE locus on AD risk. Our study presents a novel APOE filtering approach and provides important guidelines for research in this area, as well as for elucidating genetic interaction effects with APOE*2/3/4.

Acknowledgements
NCRAD. Biological samples used in this study were stored at study investigators' institutions and at the National Cell Repository for Alzheimer's Disease (NCRAD) at Indiana University, which receives government support under a cooperative agreement grant (U24 AG21886) awarded by the National Institute on Aging (NIA). We thank contributors who collected samples used in this study, as well as patients and their families, whose help and participation made this work possible. EADB. We thank the many study participants, researchers, and staff for collecting and contributing to the data, the high-performance computing service at the University of Lille, and the staff at CEA-CNRGH for their help with sample preparation and genotyping, and excellent technical assistance. We thank Antonio Pardinas for his help. This research was conducted using the UK Biobank resource (application number: 61054).  Genotyping of the Dutch case-control samples was performed in the context of EADB (European Alzheimer & Dementia biobank) funded by the JPco-fuND FP-829-029 (ZonMW project number #733051061). This research is performed by using data from the Parelsnoer Institute an initiative of the Dutch Federation of University Medical Centres (www. parel snoer. org). 100-Plus study: We are grateful for the collaborative efforts of all participating centenarians and their family members and/or relations. We thank the Netherlands Brain Bank for supplying DNA for genotyping. This work was supported by Stichting AlzheimerNederland (WE09.2014-03), Stichting Diorapthe, Horstingstuit foundation, Memorabel (ZonMW project number #733050814, #733050512) and Stichting VUmcFonds. Additional support for EADB cohorts: WF, SL, and HH are recipients of ABOARD, a public-private partnership receiving funding from ZonMW (#73305095007) and Health~Holland, Topsector Life Sciences & Health (PPP-allowance; #LSHM20106). The DELCODE study was funded by the German Center for Neurodegenerative Diseases (Deutsches Zentrum für Neurodegenerative Erkrankungen (DZNE)), reference number BN012. Gra@ce. The Genome Research @ Fundació ACE project (GR@ACE) is supported by Grifols SA, Fundación bancaria 'La Caixa' , Fundació ACE, and CIBERNED (Centro de Investigación Biomédica en Red Enfermedades Neurodegenerativas (Program 1, Alzheimer Disease to MB and AR)). A.R. and M.B. receive support from the European Union/EFPIA Innovative Medicines Initiative Joint undertaking ADAPTED and MOPEAD projects (grant numbers 115975 and 115985, respectively). M.B. and A.R. are also supported by national grants PI13/02434, PI16/01861, PI17/01474 and PI19/01240. Acción Estratégica en Salud is integrated into the Spanish National R + D + I Plan and funded by ISCIII (Instituto de Salud Carlos III)-Subdirección General de Evaluación and the Fondo Europeo de Desarrollo Regional (FEDER-'Una manera de hacer Europa'). Some control samples and data from patients included in this study were provided in part by the National DNA Bank Carlos III (www. banco adn. org, University of Salamanca, Spain) and Hospital Universitario Virgen de Valme (Sevilla, Spain); they were processed following standard operating procedures with the appropriate approval of the Ethical and Scientific Committee. The present work has been performed as part of the doctoral program of I. de Rojas at the Universitat de Barcelona (Barcelona, Spain). EADI. This work has been developed and supported by the LABEX (laboratory of excellence program investment for the future) DISTALZ grant (Development of Innovative Strategies for a Transdisciplinary approach to ALZheimer's disease) including funding from MEL (Metropole européenne de Lille), ERDF (European Regional Development Fund), and Conseil Régional Nord Pas de Calais. This work was supported by INSERM, the National Foundation for Alzheimer's disease and related disorders, the Institut Pasteur de Lille and the Centre National de Recherche en Génomique Humaine, CEA, the JPND PERADES, the Laboratory of Excellence GENMED (Medical Genomics) grant no. ANR-10-LABX-0013 is managed by the National Research Agency (ANR) part of the Investment for the Future program, and the FP7 AgedBrainSysBio. The Three-City Study was performed as part of collaboration between the Institut National de la Santé et de la Recherche Médicale (Inserm), the Victor Segalen Bordeaux II University and Sanofi-Synthélabo. The Fondation pour la Recherche Médicale funded the preparation and initiation of the study. The 3C Study was also funded by the Caisse Nationale Maladie des Travailleurs Salariés, Direction Générale de la Santé, MGEN, Institut de la Longévité, Agence Française de Sécurité Sanitaire des Produits de Santé, the Aquitaine and Bourgogne Regional Councils, Agence Nationale de la Recherche, ANR supported the COGINUT and COVADIS projects. Fondation de France and the joint French Ministry of Research/INSERM "Cohortes et collections de données biologiques" programme. Lille Génopôle received an unconditional grant from Eisai. The Three-city biological bank was developed and maintained by the laboratory for genomic analysis LAG-BRC -Institut Pasteur de Lille. Rotterdam Study. The Rotterdam Study is funded by Erasmus Medical Center and Erasmus University, Rotterdam, Netherlands Organization for the Health Research and Development (ZonMw), the Research Institute for Diseases in the Elderly (RIDE), the Ministry of Education, Culture and Science, the Ministry for Health, Welfare and Sports, the European Commission (DG XII), and the Municipality of Rotterdam. The authors are grateful to the study participants, the staff from the Rotterdam Study, and the participating general practitioners and pharmacists. The generation and management of GWAS genotype data for the Rotterdam Study (RS-I, RS-II, RSIII) were executed by the Human