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

Abstract

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.

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.

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.

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.