Polygenic resilience scores capture protective genetic effects for Alzheimer's disease.

Polygenic risk scores (PRSs) can boost risk prediction in late-onset Alzheimer's disease (LOAD) beyond apolipoprotein E (APOE) but have not been leveraged to identify genetic resilience factors. Here, we sought to identify resilience-conferring common genetic variants in (1) unaffected individuals having high PRSs for LOAD, and (2) unaffected APOE-ε4 carriers also having high PRSs for LOAD. We used genome-wide association study (GWAS) to contrast "resilient" unaffected individuals at the highest genetic risk for LOAD with LOAD cases at comparable risk. From GWAS results, we constructed polygenic resilience scores to aggregate the addictive contributions of risk-orthogonal common variants that promote resilience to LOAD. Replication of resilience scores was undertaken in eight independent studies. We successfully replicated two polygenic resilience scores that reduce genetic risk penetrance for LOAD. We also showed that polygenic resilience scores positively correlate with polygenic risk scores in unaffected individuals, perhaps aiding in staving off disease. Our findings align with the hypothesis that a combination of risk-independent common variants mediates resilience to LOAD by moderating genetic disease risk.

Introduction

Alzheimer’s disease (AD) is the leading cause of dementia [1]. AD exists as two genetically distinct forms: early-onset AD, which is caused by autosomal dominant mutations in one of several genes (PSEN1, PSEN2, APP, SORL1) and typically has an onset of symptoms between the ages of 40 and 60 years [2], and the more common late-onset AD (LOAD), which is sporadic, polygenic, and typically has an onset of symptoms in the mid-60s [3]. Elevated risk of LOAD is associated with a host of lifestyle factors and medical conditions, such as a high-fat diet, heavy drinking and smoking, cardiovascular disease, type-2 diabetes, and traumatic brain injury [4]. More importantly, the heritability of LOAD from twin studies was estimated at 58–79% [5], and its estimates from single-nucleotide polymorphisms (SNPs) range from 13 to 33% [6-9]. The goal of this study is to determine whether genes also play a role in resilience to LOAD. We used an innovative approach first introduced and applied in schizophrenia as a general framework for resilience research [10], focusing on individuals at the highest levels of genetic risk.

To date, genome-wide association studies (GWASs) have discovered close to 50 genome-wide significant loci (P < 5e-08) associated with LOAD risk [9, 11–20]. The ε4 allele of apolipoprotein E (APOE) is the polymorphism with the strongest effect on LOAD susceptibility [21]. Beyond APOE-ε4, there may be thousands of additional genetic polymorphisms that make small individual contributions to the overall risk for LOAD [22-25]. A polygenic risk score (PRS) [26] can be derived by summing the weighted effect of SNPs to identify a single genetic risk variable that reflects one’s relative susceptibility to LOAD. Recent LOAD PRSs capture most of the SNP heritability for LOAD [9, 24, 27]. Extensive research shows that PRSs boost the accuracy of LOAD diagnosis beyond the performance of APOE [22-25], and capture LOAD phenotypic variability not explained by APOE status [28, 29].

Revealing the genetic architecture of LOAD is vital for understanding its etiology and identifying molecular targets for innovative therapeutic interventions. Yet, knowledge of risk factors might be fruitfully complemented by an understanding of resilience-associated or -promoting mechanisms as well. As such, some AD research has shifted focus from symptomatic cases to healthy aging individuals or asymptomatic individuals at elevated risk [30]. This was motivated by the premise that high-risk asymptomatic individuals, yet unaffected, may provide clues that protect them against AD. Here, we employ the term “resilience” to indicate individuals who show better than expected outcomes in the face of high genetic risk for disease [30-35].

Increasing evidence suggests that several factors—including education, literacy, physical activity, and mental activity—can moderate the risk for LOAD [31, 32, 36], and it is estimated that one-third to 40% of dementia cases might be preventable [36, 37]. These moderation effects may be explained by reverse causation [38], but genetic influences—which are not subject to reverse causation—also underlie these factors. Educational attainment [39, 40] and, particularly, general cognitive ability [40, 41] are heritable. Thus, some of these factors may also confer resilience-enhancing genetic effects. Notably, some genetic variants, such as APOE-ε2 [42] and the APP A673T variant [43], have been identified as protective for LOAD. However, the biological mechanisms that drive the protective effects remain largely unknown. Importantly, we consider such protective effects to be fundamentally different from the “resilience” effects we sought in our study, in that protective factors are generally operative across the full range of risk, whereas resilience factors are only operative in those at the highest risk for disease. Very little work has been aimed at identifying additional genetic resilience factors that potentially moderate the genetic risk established by the cumulative effects of risk-associated alleles and their corresponding protective alleles. Genetic resilience against risk for LOAD has been investigated through diverse approaches based on varying conceptualizations and measurements used to identify individuals at high risk. As aggregation of beta-amyloid plaques and tau tangles in the brain are two of the neuropathological hallmarks of LOAD [44], a principal focus of resilience has been on asymptomatic individuals who have cognition levels that are better than predicted based on these pathologies [45-47]. Other studies have leveraged known genetic risk factors to study resilience. For example, in APOE-ε4 carriers, over a dozen SNPs have been reported to potentially facilitate resilience, such as rs10553596 in CASP7 [48] and the rs4934 nonsynonymous variant in SERPINA3 [49, 50]. However, a substantial part of the genetic risk for LOAD is neglected without incorporating the effects of genes other than APOE. Thus, although composite genetic risk indices (such as the PRS) are growing in popularity and utility, they have not been employed in the service of identifying genetic resilience for LOAD. Now, with very large numbers of LOAD samples and a more comprehensive profile of the genetic factors that confer LOAD risk, we are entering a period in which it is possible to study the interplay of genetic risk factors and genetic modifiers that reduce their penetrance.

Here, we posit the existence of common genetic variants, which have not been identified by GWAS as associated with AD as either risk or protective factors, that can help older adults remain LOAD-free despite a high genetic risk burden. We hypothesize that there exist resilience-associated variants that lower LOAD susceptibility in a manner that is statistically independent of the effects of risk-associated alleles (or their alternative protective alleles). We tested this hypothesis by capitalizing on the most comprehensive known PRS for LOAD [18] and APOE allelic status to develop two designs identifying unaffected individuals with the highest genetic likelihood of developing LOAD. Design 1 defined “resilient” individuals as normal controls with the highest PRSs for LOAD. Design 2 defined “resilient” individuals as normal controls with at least one APOE-ε4 allele and the highest LOAD PRSs (excluding the APOE region). We aimed to discover residual common genetic variants that confer resilience to unaffected individuals in the highest genetic risk tiers for LOAD. We then leveraged this profile of resilience-promoting genetic variants to build a polygenic resilience score for LOAD. We hypothesized that polygenic resilience scores would account for significant variation in affection status for LOAD among individuals with high genetic risk, and would show a significant positive correlation with PRSs in unaffected controls.

Methods

Research design

Our workflow is shown in Fig. 1. In stage 1, a recent GWAS meta-analysis for LOAD [18] was leveraged for identifying risk variants and polygenic risk scoring. In stage 2, we compared two analytic designs to identify high-risk “resilient” normal controls and “risk-matched” LOAD cases. In stage 3, a resilience GWAS was conducted for each design using the identified high-risk individuals. Then the polygenic resilience score weights were derived from resilience GWAS meta-analysis summary statistics. Finally, polygenic resilience scores were replicated in independent external studies for evaluating the performance in distinguishing high-risk “resilient” normal controls from “risk-matched” LOAD cases. The parameters of each analysis step are summarized in Supplementary Table 3.

Fig. 1

An illustration of the workflow of deriving polygenic resilience scores for late-onset Alzheimer’s disease (LOAD) for design 1 and design 2.

Stage 1: Using prior LOAD genome-wide association study (GWAS) results to calculate polygenic risk scores (PRSs). Stage 2: Identifying resilient individuals. In stage 2, we deployed two analysis designs differing in the definition of “resilient” individuals. In design 1, normal controls with LOAD PRSs ≥90th percentile were defined as “resilient” participants. In design 2, within the subset of normal controls who had at least one apolipoprotein E (APOE)-ε4 allele, a threshold of ≥80th percentile of PRSs (excluding SNPs in the APOE region) was used to define high-risk controls as “resilient”. Stage 3: Resilience GWAS and replication of polygenic resilience scores. GWAS was performed using “resilient” individuals and risk-matched affected cases from each of the two designs. For each design, polygenic resilience scores were derived and evaluated in external replication datasets. LD linkage disequilibrium, OR odds ratio, SNPs single-nucleotide polymorphisms.

Samples and genotypes

We acquired the largest available collection of genome-wide SNP data for clinically diagnosed or autopsy-confirmed LOAD to ensure adequate power. Table 1 shows the number of normal controls, LOAD cases, high-risk “resilient” normal controls, and “risk-matched” LOAD cases in each study. Summary statistics of age-at-onset (AAO) for LOAD cases and age-at-last-examination (AAE) for normal controls are presented in Supplementary Table 1. In design 1 and design 2, the mean AAE of high-risk “resilient” normal controls and the mean AAO of “risk-matched” LOAD cases ranged from 70.3 to 80.9, and there were no significant age differences between groups. A common lower bound for AAO of LOAD is 65; however, the age cutoff has no specific biological significance [3], and many genetic studies of LOAD have included cases with AAO as low as 60 (and the same AAE for unaffected comparison subjects). Therefore, we included participants in our analysis having AAO/AAE ≥ 60 years old. The full name and accessibility of each study can be found in Supplementary Table 2. All 26 studies in the discovery stage came from the stage-1 AD GWAS meta-analysis of Kunkle et al. [18]. The eight studies in the replication stage are fully independent of the discovery studies. Full descriptions of the discovery and replication samples were published previously [9, 17, 18, 51, 52]. Genotypes for all studies were imputed using the Haplotype Reference Consortium (HRC) r1.1 2016 reference panel [53]. Detailed quality control (QC) steps for samples and genotypes are described in Supplementary Methods.

Table 1

The number of LOAD cases and normal controls, high-risk normal controls (“resilient” individuals), and risk-matched LOAD cases identified in each of the discovery and replication studies.

			Design 1						Design 2
			Normal controls			LOAD cases			Normal controls			LOAD cases
	Study	Sub-study	N High-risk	N Total	% Retained	N Risk-matched	N Total	% Retained	N High-risk	N Total	% Retained	N Risk-matched	N Total	% Retained
Discovery	ADGC	ADC1	52	512	10.2	842	1524	55.2	30	150	20.0	773	1021	75.7
		ADC2	16	155	10.3	386	620	62.3	9	41	22.0	212	353	60.1
		ADC3	57	566	10.1	379	666	56.9	27	132	20.5	357	391	91.3
		ADC4	38	376	10.1	222	304	73.0	20	100	20.0	115	160	71.9
		ADC5	51	503	10.1	187	285	65.6	24	119	20.2	147	178	82.6
		ADC6	34	337	10.1	68	213	31.9	19	93	20.4	94	121	77.7
		MAYO	112	1117	10.0	626	754	83.0	63	313	20.1	436	492	88.6
		UMVUMSSM	113	1108	10.2	753	1123	67.1	51	251	20.3	430	661	65.1
		ACT + GenDiff	157	1556	10.1	425	524	81.1	63	314	20.1	198	234	84.6
		MIRAGE	51	509	10.0	101	137	73.7	38	190	20.0	68	78	87.2
		WASHU1	18	176	10.2	137	253	54.2	10	48	20.8	65	141	46.1
		ROSMAP	72	717	10.0	192	237	81.0	23	114	20.2	18	87	20.7
		UPITT	82	811	10.1	794	1152	68.9	32	158	20.3	545	664	82.1
		OHSU	18	180	10.0	119	174	68.4	5	25	20.0	49	73	67.1
		Tgen II	36	360	10.0	364	613	59.4	16	76	21.1	251	396	63.4
		NIA-LOAD	102	1007	10.1	582	760	76.6	58	289	20.1	463	571	81.1
		ROSMAP2	22	214	10.3	50	59	84.7	3	15	20.0	0	18	0.0
		WASHU2	8	71	11.3	29	36	80.6	4	16	25.0	16	19	84.2
		MTC	19	188	10.1	220	252	87.3	5	22	22.7	48	143	33.6
		TARCC	18	176	10.2	241	306	78.8	10	47	21.3	122	184	66.3
		WHICAP	56	553	10.1	17	72	23.6	23	113	20.4	8	16	50.0
	ADNI	ADNI-1	13	127	10.2	174	350	49.7	6	28	21.4	97	230	42.2
	CHARGE	CHS	168	1661	10.1	274	451	60.8	66	330	20.0	99	148	66.9
		FHS	191	1901	10.0	137	288	47.6	74	370	20.0	54	94	57.4
	EADI		620	6172	10.0	1636	2167	75.5	247	1235	20.0	870	1072	81.2
	GERAD		139	1388	10.0	2354	2992	78.7	62	310	20.0	1006	1623	62.0
N Total			2263	22441	10.1	11309	16312	69.3	988	4899	20.2	6541	9168	71.3
Replication	AddNeuroMed		19	183	10.4	38	223	17.0	9	44	20.5	19	121	15.7
	ADGC	ADC7	79	784	10.1	45	513	8.8	51	252	20.2	57	320	17.8
	ADNI	ADNI-GO/2/3	38	374	10.2	37	211	17.5	23	113	20.4	38	142	26.8
	PGC-ALZ	Norwegian DemGene Network	75	741	10.1	163	1085	15.0	34	168	20.2	130	671	19.4
		GENDER/SATSA/HARMONY	77	764	10.1	39	306	12.7	40	196	20.4	34	148	23.0
		TwinGene	608	6080	10.0	30	287	10.5	353	1765	20.0	31	156	19.9
	AIBL		77	762	10.1	13	111	11.7	36	180	20.0	12	71	16.9
	Sydney MAS		83	830	10.0	16	95	16.8	37	182	20.3	10	31	32.3
N Total			1056	10518	10.0	381	2831	13.5	583	2900	20.1	331	1660	19.9

LOAD late-onset Alzheimer’s disease.

Note: “Retained” column indicates the percentage of high-risk normal controls of all normal controls retained for resilience genome-wide association analysis per study, or the percentage of risk-matched LOAD cases of all LOAD cases retained in analysis per study. ROSMAP2 study had no LOAD cases (in italic) whose risk matched with high-risk normal controls in design 2, and was not included in the analysis for design 2.

A list of study full names is in Supplementary Table 2.

Identifying individuals at high genetic risk

In design 1, a PRS was used to select individuals with high genetic risk. At the time of deploying our analyses, the Kunkle et al., 2019 study [18] was the largest publicly available GWAS using clinically diagnosed or autopsy-confirmed AD cases and CN controls, as opposed to proxy AD cases and controls that might lead to inaccurate risk estimation [27]. Therefore, we consider that this study will give the most accurate measure of AD risk and derived the PRS weights from its stage-1 AD GWAS meta-analysis summary statistics [18]. See Supplementary Methods for further details. The variance in AD explained by PRS maximizes at a P-value threshold of 0.5 in participants from GERAD (Genetic and Environmental Risk for Alzheimer’s disease) [23] and 22 locally available ADGC (Alzheimer’s Disease Genetics Consortium) studies (Supplementary Fig. 4). We, therefore, adopted this threshold to ensure that our risk measure captures as much of the genetic risk for AD as possible. This very conservative threshold will thus ensure that potential risk SNPs with even very small effect sizes will not be advanced for consideration as resilience SNPs. However, if studies other than Kunkle et al. 2019 were used to estimate genetic risk for AD, it may be the case that smaller P-value thresholds may be optimal (e.g., 5e−08, 1e−05, 0.1) [9, 22, 54, 55]. Within each study, LOAD cases and normal controls were ranked based on their PRSs. Note that as the true prevalence of resilience to AD in the population is unknown, we adopted the same high-risk percentile cutoff that proved effective in our original workflow [10], and classified the 10% of controls with the highest PRSs as “resilient”. The LOAD cases whose risk scores were between the 90th percentile and the maximum PRS in controls were retained as risk-matched LOAD cases for comparison.

In design 2, we restricted the analysis to APOE-ε4 carriers. APOE and its flanking region (chr19: 44,400 kb–46,500 kb) [23] were removed from the PRS. As this analysis was restricted to fewer individuals due to the APOE-ε4 stratification, we chose a more lenient high-PRS cutoff (80th percentile) for identifying “resilient” individuals to retain more participants and preserve power. In this design, “resilient” normal controls were identified as those with at least one APOE-ε4 allele, and a risk score ranked at ≥80th percentile. Risk-matched LOAD cases were defined as APOE-ε4 carriers whose PRSs fell within the high-PRS range of “resilient” normal controls.

Derivation, replication, and statistical analysis of polygenic resilience scores

GWASs of resilience were performed using logistic regression with Plink (version 1.9) [56]. Selected principal components, AAO/AAE, and sex were used as covariates. A GWAS meta-analysis was conducted in METAL [57] software using an inverse-variance random-effect model with genomic control. In accord with the pipeline described by Hess et al. [10], SNPs known to be associated with LOAD risk were excluded from the resilience-scoring algorithm; these were defined as those SNPs that showed an association with AD risk (P < 0.5) from the GWAS meta-analysis summary statistics [18], and variants that were in linkage disequilibrium (LD) (r ≥ 0.2 in a 1-Mb window) with those risk variants with associations of P < 0.5. This pruning step of excluding risk variants from consideration as resilience loci serves as a conservative measure to avoid re-discovering risk variants for resilience scoring. For both resilience designs, the polygenic resilience score weights were generated from the marginal SNPs of resilience GWAS meta-analysis summary statistics following the same series of QC steps (see Supplementary Methods).

Polygenic resilience scores were derived for 10 P-value thresholds, in a manner similar to the PRS algorithm, by summing up the weighted effective allele counts of SNPs [26]. Logistic regression was used to assess the likelihood of “resilient” group inclusion based on harboring a higher polygenic resilience score. Selected principal components, AAO/AAE and sex were used as covariates. For each polygenic resilience score, we meta-analyzed the natural logarithm of the odds ratio (OR) of being a high-risk “resilient” normal control versus a risk-matched LOAD case using a random-effects inverse-variance model using the R package metafor, and pooled variance explained in resilience across independent replication studies. All tests were two-tailed unless specified otherwise. See Supplementary Methods for further details.

Results

Resilience GWAS

Design 1 produced 2263 high-risk “resilient” normal controls and 11,309 risk-matched LOAD cases for the resilience GWAS meta-analysis. As expected, the sample size retained in design 2 was smaller, totalling 988 high-risk “resilient” normal controls and 6541 risk-matched LOAD cases (Table 1). Because our analytic approaches used only subsets of all available LOAD case–control GWAS data, we neither had nor anticipated having sufficient power to detect individual SNPs with genome-wide significant association with resilience (Supplementary Fig. 3). Instead, our focus was on deriving and evaluating polygenic resilience scores. As a necessary step to generate SNP-weights for summation in those scores, we performed individual-SNP association tests and briefly reported the results in Supplementary Results.

Replication and evaluation of polygenic resilience scores

After removing risk-associated SNPs (P < 0.5) and SNPs in LD with those risk-associated SNPs (r ≥ 0.2), clumping the remaining marginal SNPs, and applying QC steps, a profile of 18,723 SNPs was included in the resilience score for design 1, and 18,122 SNPs in design 2. Resilience scores for all 10 P-value thresholds were significantly associated with “resilient” group inclusion (“resilient” normal controls versus risk-matched LOAD cases) when tested in locally downloaded discovery datasets. Results of the association between “resilient” group inclusion and polygenic resilience scores from the replication datasets were meta-analyzed, yielding 1056 high-risk “resilient” normal controls and 381 risk-matched LOAD cases in design 1, and 583 high-risk “resilient” normal controls and 331 risk-matched LOAD cases in design 2 (Table 1).

In design 1, the meta-analysis found significant replication of the association between “resilient” group inclusion and polygenic resilience scores at two P-value thresholds (P < 0.1, P < 0.2) (Fig. 2A). The most significant association was found for the polygenic resilience score containing all independent marginal SNPs with resilience GWAS P < 0.1 (OR = 1.24, 95% confidence interval [CI] = 1.05–1.47, P = 0.010). Resilience scores for the 0.1 P-value threshold explained an average of 1.3% (standard deviation [58] = 5.3%) of the variance in “resilient” group inclusion or 1.2% (SD = 4.3%) (Fig. 2B) of the variance on the liability scale, i.e., SNP heritability of resilience. No significant (P < 0.05) replication of the association between “resilient” group inclusion and polygenic resilience scores was observed for any of the 10 polygenic resilience scores in design 2.

Fig. 2

The performance of polygenic resilience scores in capturing resilience variability in independent replication studies.

In design 1, normal controls with late-onset Alzheimer’s disease (LOAD) polygenic risk scores (PRSs) ≥90th percentile were defined as “resilient” participants. In design 2, a threshold of ≥80th percentile of PRSs (excluding SNPs in the apolipoprotein E [APOE] region) was used to define high-risk controls as “resilient” within the normal controls who have at least one APOE-ε4 allele. A, B design 1 (high-risk normal controls, n = 1,056; risk-matched LOAD cases, n = 381). C, D Design 2 (high-risk normal controls, n = 583; risk-matched LOAD cases, n = 331). The odds ratio (OR) and variance explained by polygenic resilience scores reflect meta-analytic results from independent replication samples. Nagelkerke’s pseudo-R2 values on the liability scale are weighted average using the weights from the meta-analysis of ORs. The dot-plots (A, C) show corresponding ORs for resilience scores across 10 P-value thresholds, wherein OR > 1.0 indicates higher resilience scores are associated with a higher likelihood of being a high-risk normal control (“resilient” individual) than being a risk-matched LOAD case. Error bars represent the 95% confidence intervals (CI) around each OR, which are the exponent of the 95% CI of β coefficients. The barplots (B, D) show the amount of variance in resilience (i.e., “resilient” high-risk normal controls versus risk-matched LOAD cases) on the liability scale that is explained by resilience scores. Asterisks (*) indicate P values <0.05 for ORs >1.0.

Note that the association between “resilient” group inclusion and polygenic resilience scores (P < 0.1, P < 0.2 of design 1) was not significant after multiple-testing correction for 10 P-value thresholds using the false discovery rate (FDR) or Bonferroni method. However, considering polygenic resilience scores were derived by aggregating SNPs within a series of escalating P-value thresholds, they were nested models and not independent. Therefore, a typical FDR or Bonferroni correction under the assumption of independence would be overly conservative.

Interaction of risk and resilience effects

In the full samples from three locally downloaded replication studies (Alzheimer Disease Centers Wave 7 [ADC7], AddNeuroMed, and Alzheimer’s Disease Neuroimaging Initiative stage GO/2/3 [ADNI-GO/2/3]; normal controls, n = 1321; LOAD cases, n = 943) (Table 1 and Supplementary Table 1), we tested for correlations between PRSs and polygenic resilience scores. As hypothesized, the standardized polygenic resilience scores of the optimal P < 0.1 threshold in design 1 exhibited a significant positive correlation with PRSs in normal controls (Pearson’s r = 0.102, 95% CI = 0.048–0.155, degree of freedom [df]=1319, P = 2.1e-04), and no significant correlation was observed in LOAD cases (Pearson’s r = 0.022, 95% CI = −0.042–0.085, df = 941, P = 0.51) (Fig. 3). As expected, the correlation coefficient between polygenic risk scores and polygenic resilience scores in normal controls was significantly larger than the one in LOAD cases (P = 0.03, one-tailed test).

Fig. 3

The correlation of standardized polygenic risk scores (PRSs) and polygenic resilience scores (design 1) in normal controls and late-onset Alzheimer’s disease (LOAD) cases.

The analyses were performed in three independent replication studies not used in the resilience score derivation steps (i.e., ADC7, AddNeuroMed, and ADNI-GO/2/3; normal controls, n = 1321; LOAD cases, n = 943). The optimal P-value threshold for polygenic risk-scoring was 0.5, and the optimal P-value threshold for polygenic resilience scoring was 0.1 (see Fig. 2). The blue round dots indicate normal controls, and the orange circles indicate LOAD cases. The blue and orange lines represent the best fit for correlations between PRSs and resilience scores in normal controls and in LOAD cases, respectively. The blue and orange annotation text shows the Pearson correlation coefficient (r) and the P-value between PRSs and resilience scores in normal controls and LOAD cases, respectively. In this analysis, we excluded ultra-high-risk LOAD cases whose PRSs are higher than the maximum of all normal controls, and ultra-low-risk normal controls whose PRSs are lower than the minimum of all LOAD cases.

Discussion

We applied a validated analytic framework to detect common variants that, when combined into a polygenic resilience score, are associated with lower LOAD risk penetrance among older individuals with relatively high genetic risk of disease. We found reliable evidence to reinforce the notion that unaffected individuals with higher genetic risk loads may be protected from complex diseases, such as LOAD, by the collective effects of risk-independent common variants that reduce the penetrance of one’s overall genetic risk burden. Identifying genetic factors that moderate risk penetrance may prove valuable for explaining the missing heritability and etiologic heterogeneity of LOAD, which in turn could shed light on pathophysiological mechanisms and eventually lead to better interventions and preventive treatments.

Risk-countering effects of polygenic resilience scores

Individuals with higher polygenic resilience scores (P < 0.1 and P < 0.2 thresholds of design 1) had higher odds of being a “resilient” high-risk normal control than a risk-matched LOAD case. Polygenic resilience scores (design 1) significantly increased with higher PRSs in normal controls, but not in LOAD cases. Taken together, these results support the hypothesis that polygenic resilience scores capture risk-countering polygenic effects against the penetrance of high polygenic risk for LOAD, and that normal controls with higher PRSs are protected from LOAD by harboring correspondingly higher polygenic resilience scores. Although no polygenic resilience scores in design 2 demonstrated significant risk-buffering effects, we cannot rule out the possibility that common variants might reduce risk penetrance in normal controls with enriched risk from both APOE and PRSs. In fact, among APOE-ε4 carriers, higher resilience scores in design 1 at the P < 0.1 threshold (OR = 1.64, 95% CI = 1.08–2.50, P = 0.021) and the P < 0.2 threshold (OR = 1.98, 95% CI = 1.24–3.15, P = 3.9e-03) were associated with higher odds of being a “resilient” high-PRS normal control than a risk-matched LOAD case. Among “resilient” high-PRS controls, higher resilience scores in design 1 (P < 0.2 threshold) were significantly associated with increased odds of carrying at least one APOE-ε4 allele (OR = 1.57, 95% CI = 1.07–2.29, P = 0.021). A similar trend was observed when the P < 0.1 threshold was used, although this was not significant (OR = 1.30, 95% CI = 0.90–1.89, P = 0.16) (Supplementary Results). We, therefore, conclude that polygenic resilience scores may moderate the risk effects of the LOAD PRS generally, and the APOE-ε4 allele specifically. However, these analyses were carried out in relatively small studies (ADC7, AddNeuroMed, and ADNI-GO/2/3), and need to be repeated in larger, more powerful, replication samples.

Interplay of polygenic effects and APOE

In design 2, we hypothesized that a two-stage selection of individuals (with both higher PRSs and one or more APOE-ε4 alleles) would enrich for individuals with the absolute highest genetic risk for LOAD [59, 60]; yet, there was a substantial reduction in the performance of design 2 in contrast to design 1. The lack of significant replication of association with resilience in design 2 simply might be due to lower statistical power in both the resilience score development and replication stages, considering the total sample size of design 2 is approximately half that of design 1. Alternatively, resilience-promoting variants may be found among APOE-ε4 carriers through broader exploration of the model-parameter space (e.g., PRS threshold in particular), separate evaluation of APOE-ε4 homozygotes and various heterozygote combinations, and more accurate modeling of the genetic architecture of resilience (see limitations below). An important question future studies should address is to what extent common variants may influence the penetrance of genetic risk in larger samples of APOE-ε4 carriers, or whether the prevalence of risk-modifying common variants differs between APOE-ε4 carriers and noncarriers.

On the other hand, multiple studies [22–25, 28, 54, 61–63] have revealed that PRSs capture independent risk effects beyond APOE alone, while few studies have explored the risk-predictive performance of PRSs stratified by APOE status. Higher PRSs were found to be associated with increased susceptibility for LOAD in APOE-ε4 noncarriers [25, 29, 59]. Furthermore, the risk effects of PRS deciles across APOE status could be dependent on the ages of participants [29, 59, 62, 64]. Further mining of the complex relationship between the risk effects of PRSs and APOE is outside the scope of the current study; however, further investigations on the penetrance of high PRSs among APOE-ε4 carriers and noncarriers seem warranted.

Strengths and limitations

Our approach has identified candidate resilience loci that may ultimately serve as targets for the promotion of resilience. We examined the performance of two polygenic resilience scores: design 1 selected participants with the highest polygenic risk regardless of APOE-ε4 status, while design 2 restricted analyses to APOE-ε4 carriers. To our knowledge, this is the first study to identify a polygenic resilience score for genetic LOAD risk, comprising thousands of risk-independent common variants that partially offset the genetic risk conferred by a relatively high PRS. An important distinction of the current study relative to prior work on genetic resilience to LOAD is that we accounted not only for the risk from APOE but also the aggregate effect of thousands of additional risk variants throughout the genome via the LOAD PRS.

A conservative variant-filtering strategy was applied, which resulted in the removal of common variants associated with LOAD risk variants (risk association P < 0.5) and those in liberal LD (r > 0.2) with LOAD risk variants. A strength of this approach is that we ensured the polygenic resilience scores derived in the current study are independent of the risk scores so that the SNPs comprising the polygenic resilience score are not sub-threshold risk SNPs. Our design of defining “resilient” groups from the same risk background instead of contrasting high-risk normal controls with low-risk LOAD cases also helped avoid re-discovering variants merely associated with risk. In addition, the resilience alleles of these risk-residual SNPs are not simply protective alleles defined in a risk framework, where each biallelic locus is defined by both a risk allele and a corresponding and opposing protective allele. Thus, this strategy helps identify resilience effects that are conditioned on net risk effects, owing to the combination of risk and protective alleles summed in polygenic risk scores. Yet, although our approach is conservative, it is limited in the identification of a better-performing resilience score because most of the genome has been discarded from the analysis. Biologically, it is plausible that variants nearby risk loci, such as those in the same LD block or in the same gene with risk SNPs, could exert modifying functions [65]. Our conservative strategy, discarding all SNPs with any semblance of risk association, and those in liberal LD threshold with such SNPs, consequently leads to lower power in uncovering variants with potentially higher biological functionality. This notion is borne out in the fact that no significant gene-ontology pathways were enriched by resilience-related common variants identified in this study (results not shown). With larger samples, resilience-conferring SNPs may be investigated using a stricter LD threshold (e.g., r > 0.1) to further restrain the “hitchhiking” of risk variants. More importantly, Mendelian randomization, conditional association testing, or simulation analysis may be better suited to evaluate the hypothesis that resilience signals are more likely to co-localize with risk loci or genes. In addition, filtering variants by LD with risk SNPs results in a low LD structure among the remaining SNPs as demonstrated previously [10], which diminishes our capacity to examine the genetic correlation of resilience to LOAD with other risk- or resilience-related phenotypes (e.g., via LD score regression). A high priority should be placed on the design of new methods that can detect resilience-associated SNPs that may reside in regions of strong LD with risk variants.

Resilience was defined by discrete groups in our analysis, which truncated effective sample sizes to the upper tail of the risk distribution. Choosing a lower percentile cutoff would increase the sample size available for the resilience analysis, while potentially diminishing the signal of resilience genes. In the future, when larger samples are available, higher risk thresholds may be applied and subgroups at more extreme risk could be leveraged to increase power. It is an important task for future studies to investigate which of these factors (i.e., sample size, signal: noise) would have the greater effect on power, and to better understand the prevalence of resilience to AD in the population. Theoretically, resilience may be a continuous measure; thus, our resilience approach might also be improved by leveraging all study samples and modeling the continuity of resilience using either linear or non-linear analysis. Despite the restricted sample sizes in the current study, two resilience scores in design 1 were sufficiently robust to replicate significantly in fully independent studies. Further replication would be key to testing the validity of these resilience scores. It is expected that the strength of our results (in terms of variance explained and the significance of associations) will only increase with the addition of more samples.

Several studies [9, 22, 54, 55] indicated that polygenic risk scores of P-value thresholds less than 0.5 (i.e., 5e−08, 1e−05, 0.1) might show better performance in predicting LOAD risk. Therefore, it may be valuable to compare the performance of resilience scores developed from risk scores at other p-value thresholds. In addition, it is likely that a subgroup of “resilient” normal controls identified in this work will eventually develop LOAD, but with later onset. Thus, all resilient participants demonstrate resilience against high levels of genetic risk for LOAD, but only those who never develop LOAD are additionally resistant against the disease itself. Lastly, the participants in our analysis were of European ancestry, so the degree of generalization of our results to non-European populations is presently unknown.

Future directions

Two analysis designs were deployed in the current study to select individuals with a high genetic risk burden from both PRSs and APOE, and other methods could be devised to expand the capabilities of our resilience approach in LOAD. It has been suggested, for example, that using a PRS with the APOE region removed and adding APOE alleles as a covariate may boost the performance of LOAD risk prediction [66], compared with incorporating APOE alleles as weighted SNPs in PRSs. In addition, it could be important to include the number of APOE-ε4 or ε2 alleles as covariates in resilience analysis models to better reflect the relative risk levels among individuals. In our study, we consider it important to utilize the most comprehensive risk profile of LOAD to identify resilient individuals, i.e., normal controls with the highest genetic risk from all sources. Future studies may be interested in examining the resilience effects that moderate a portion of the LOAD risk. For example, resilience to the risk effects of APOE-ε4 alone can be studied by defining all APOE-ε4 carrying (or homozygotic) normal controls as resilient. To examine whether the resilience scores remain predictive if the APOE region is excluded from the PRS, the resilience to residual polygenic risk effects excluding APOE can be investigated in normal controls without APOE-ε4, who land in the top percentiles of PRSs (excluding the APOE region). Previous studies [42, 67] demonstrated that women carrying APOE-ε4 alleles were at greater risk of developing AD than men with the same APOE-ε4 dosages, especially between the ages of 65 and 75. When larger sample sizes are available, limiting our analysis to females in design 2 may further enrich for high-risk individuals and increase resilience signals.

Potentially, polygenic resilience scores from the current study could be applied to other resilience-related questions. For example, it would be instrumental in discovering the extent to which polygenic resilience score is associated with other phenotypes that have been associated with resilience to LOAD risk (e.g., education, general cognitive ability in early life, and other indices of cognitive reserve, brain reserve, or brain maintenance) [31, 32, 68–72]. In follow-up studies, it might be illuminating to investigate whether these resilience-promoting genetic factors show protective effects for cognitive impairment or LOAD-related pathophysiological changes.

Conclusion

We found evidence to support the hypothesis that thousands of risk-independent common variants underlie resilience among unaffected individuals with higher genetic risk for LOAD. We conclude that common variants not in LD with known LOAD risk variants exert a protective effect on LOAD risk. Our findings provide a significant and novel contribution to the existing understanding of genetic resilience to LOAD risk. This novel approach highlights a window of opportunity for identifying risk-modifying biological mechanisms and potential pathways for intervention in populations at the highest risk for LOAD.

Supplementary information