Evaluation of a Genetic Risk Score to Improve Risk Prediction for Alzheimer’s Disease

INTRODUCTION

With increasing prevalence and no effective treatment, Alzheimer’s disease (AD) has become a global public health priority and a threat to healthcare systems [1]. Failure of recent clinical trials to improve cognitive functions in AD patients [2, 3] and studies suggesting a long preclinical phase of the disease [4] underline the need for earlier interventions, moving from early clinical dementia with mild cognitive complaints to a stage of preclinical AD with pathological changes of AD but no discernible clinical signs.

In this context, building accessible, easy to use prediction tools is of key importance to predict who, or which groups of persons, will develop the pathology of AD. Tools based on several genetic or imaging markers are already being used to improve enrollment in clinical trials by including asymptomatic, at-risk participants [5]. In the Dominantly Inherited Alzheimer Network Trial Units (DIAN-TU) [6] and Alzheimer Prevention Initiative (API) trials [7], risk of developing AD is assessed based on carriage of known autosomal dominant AD mutations, which represents a perfect predictor of early-onset AD, or APOE ɛ4 status, an important but imperfect predictor of late-onset AD [8–10]. In the Anti-Amyloid Treatment in Asymptomatic Alzheimer’s Disease (A4) trials, risk is assessed using positron emission tomography (PET) imaging of brain amyloid load but such imaging is expensive and exposes the screened population to radiation [11]. Other biomarkers can be measured in the cerebrospinal fluid (CSF) but lumbar puncture has relatively low acceptance rates in asymptomatic adults.

Recent genome-wide association studies (GWAS) and follow-up genotyping studies have broadened the spectrum of genetic variation known to underlie late-onset AD and provided a promising ground on which to build a prediction tool [8, 12–16]. Since the effect sizes associated with individual genetic variants are small, ranging from a 23% reduction to a 22% increase in risk, an aggregated genetic risk score (GRS) could perform better than any individual variant. Such an approach has been successfully used to replicate association of genetic variants with AD, [17–23] but also to show associations of AD-related genetic variants with mild cognitive impairment [21, 22] and endophenotypes such as brain volumes measured by MRI, [24–27] cognitive tests, [18, 28] age at onset [23, 29], or CSF biomarkers of AD [23, 30]. These results hinted at potential benefits of using genetic information to improve the selection of asymptomatic, at-risk participants in preclinical biomarker or drug trials.

So far, the genetic variations associated with AD outside the APOE locus have shown little improvement in risk prediction over known predictors [8, 26]. Moreover, only two of them have used the most recent set of single nucleotide polymorphisms (SNPs) available [21, 23] and only one used a prospective design [21] which might be better suited in a preclinical context.

Our objectives were to extend those results by constructing a GRS based on the most up-to-date results and assess its association with risk of developing incident AD and its utility for risk prediction of AD over age, sex, education, and APOE ɛ4 in eight prospective cohorts.

MATERIALS AND METHODS

RESULTS

Overall, 19,687 participants at risk were included in this project, of whom 2,782 developed AD after times of follow-up ranging from a mean of 5.9 years in AGES to 10.9 years in Rotterdam. Other demographic characteristics of participants in the individual studies are presented in Table 1. After a fixed period of 7 years of follow-up, observed risk of AD ranged from 0.068 [0.047–0.097] in WHICAP to 0.214 [0.190–0.240] in ROSMAP (Supplementary Table 1). These varying risks reflect the varying ages and baseline cognitive status of the cohorts. As expected, observed risk of AD increased with time in all cohorts.

We first considered associations between individual SNPs and risk of incident AD. Most of the associations were not statistically significant (Table 2). Effect sizes were globally small, ranging from 0.92 [0.85–1.01] for CR1 to 1.07 [0.93–1.22] for CASS4 and were different from the estimations drawn from the IGAP meta-analysis (Table 2). As expected, having at least one copy of APOE ɛ4 was strongly associated with risk of incident AD (HR = 2.08; 95% CI =   [1.92–2.26]; p-value =  2.09×10^–72; I² = 0%). As the samples used in this project were partially overlapping with the ones used in the original IGAP study, we ran an additional IGAP meta-analysis after excluding those and did not find significant changes in the estimations of HRs for the SNPs considered (Supplementary Table 2).

Distributions of the GRS across studies are provided in Supplementary Figure 1. In nearly all studies, association of the GRS reached or was close to statistical significance, using a p-value threshold of 0.05 (Fig. 1). Effect sizes were remarkably consistent across studies, with HRs ranging from 1.12 in CHS to 1.24 in ROSMAP. When meta-analyzed, those results yielded a pooled HR of 1.17 (95% CI =   [1.13–1.21] per standard deviation (SD) increase in GRS; p-value =  2.86×10^–16) with no evidence of important heterogeneity (I² = 0%). When considering tertiles of the GRS, being in the third tertile was associated with a HR of 1.42 (95% CI =   [1.29–1.55]; p-value =  1.09×10^–13) when compared to the first tertile (Supplementary Figure 2).

We observed a significant interaction between APOE ɛ4 status and the GRS (p-value =  3.45×10^–2), with a higher effect size in APOE ɛ4 carriers (HR = 1.24; 95% CI =   [1.15–1.34] per SD increase in GRS; p-value =  7.63×10^–8; I² = 31%) than in non-carriers (HR = 1.13; 95% CI =   [1.08–1.18] per SD increase in GRS; p-value =  2.40×10^–7; I² = 0%) (Fig. 1). APOE ɛ4 carriers that were in the third tertile of GRS had an increased risk of incident AD of 62% (HR = 1.62; 95% CI =   [1.26–2.07]; p-value =  1.33×10^–4; I² = 55%) compared to APOE ɛ4 carriers in the first tertile of GRS. In APOE ɛ4 non-carriers, this risk was increased by 33% (HR = 1.33; 95% CI =   [1.19–1.49]; p-value =  8.18×10^–7; I² = 0%) in the third tertile of GRS compared to the first tertile (Supplementary Figure 2).

Measures of improvement in risk prediction after 7 years of follow-up are presented in Fig. 2 and Supplementary Table 3. The “base” model, comprised of age, sex, education, and APOE ɛ4 status had good discrimination abilities, as illustrated by C-index ranging from 0.688 [0.641–0.735] in AGES to 0.803 [0.776–0.830] in FHS (Supplementary Table 3). Calibration of the models at 7 years of follow-up was good across studies (Supplementary Figure 3).

When adding the GRS in this model, the difference in C-index was small, ranging from 0.002 in FHS to 0.009 in WHICAP, and was not statistically different from zero, except in the Rotterdam study (Fig. 2). Meta-analysis of the difference in C-index yielded a pooled estimation of 0.0043 [0.0019–0.0067], homogeneous across studies (I² = 0%) and statistically different from zero (p-value =  4.92×  10^–4). As for hazard ratio, difference in C-index was higher in APOE ɛ4 carriers (Δ-C =  0.0112 [0.0015–0.0208]; p-value = 2.32×10^–2; I² = 42%) than in non-carriers (Δ-C =  0.0018 [–0.0003–0.0039]; p-value =  9.83×10^–2; I² = 0%) (Fig. 2). Similar patterns were observed for IDI and NRI>0 across studies (Supplementary Table 3). Using fixed follow-up periods of 5, 6, and 8 years did not modify those figures, except in APOE ɛ4 carriers where a slight improvement was observed with increasing follow-up time (Supplementary Table 3 and Supplementary Figure 4).

DISCUSSION

In this large set of prospective cohorts from IGAP, we studied the association of an aggregate measure of the genetic variants identified in a previous work [16] with incident AD. We observed that, compared to SNPs considered separately, the aggregate measure showed a significant association with incident AD, even without inclusion of APOE ɛ4. Persons in the top tertile of this genetic risk score had a 42% higher risk of AD compared to persons in the bottom tertile, with higher risk among APOE ɛ4 carriers (62%) than in non-carriers (33%). When assessing the utility of this measure for individual risk prediction at 7 years, we observed small improvement over age, sex, education, and APOE ɛ4 status, with stronger effect in APOE ɛ4 carriers.

In this study, an increase in one standard deviation of the GRS, corresponding approximately to 2.5 risk alleles, was associated with a 17% increase in risk of developing AD. GRSs have been previously used to increase statistical power when studying associations with either AD status or AD-related endophenotypes. Results have often shown significant associations, depending greatly on whether APOE ɛ4 was included or not in the GRS [17–30]. Previous work from the Rotterdam Study, which was included in this study, reported similar increases of 14% (p = 0.010) [18] and 15% (p = 0.058) [21] in AD risk for an increase of one standard deviation of the GRS. These results confirm the small increase in risk carried on average by common genetic variants identified by GWAS and legitimate the use of GRSs to increase statistical power to identify such associations.

Persons in the top tertile of our GRS had a 42% higher risk of incident AD compared to persons in the bottom tertile. A recent study by Sleegers et al. reported an odds ratio of AD of 5.69 [4.17–7.75] when comparing the fifth quintile of a GRS to the third quintile [23]. Although their GRS included APOE ɛ4 and had different units which makes the comparison to our GRS difficult, these results highlight the benefit of using a sufficient number of variants outside APOE ɛ4 to improve stratification of participants at higher risk of AD in the context of preclinical biomarkers or drug trials.

C-indexes for the base model were relatively high in all the studies indicating good predictive value. Our decision to include or exclude variables in this model was based on literature, expected availability and expected heterogeneity in the measure across studies. In terms of clinical utility for individual risk prediction, the estimations of differences in C-index, NRI>0 and IDI were small across all of the studies, indicating limited clinical utility over age, sex, education, and APOE ɛ4 status. Those results are in agreement with previous studies that reported little clinical utility of genetic data outside APOE ɛ in improving risk prediction of AD over other predictors such as age and education level [8, 21–23]. Given our results, the GRS seems currently more useful for risk stratification in a population than for individual risk prediction.

The significant interaction of the GRS with APOE ɛ4 status hinted at potentially useful strategies. Such results could be the basis for a genetic risk prediction decision tree. One could assess the APOE genotype, and according to it, follow-up with more complex genetic analysis. The GRS could help refine risk and age-at-onset predictions in the APOE ɛ4 positive subgroup that is currently the focus of several prevention efforts. Further, this measure could be a useful adjunct for risk stratification in clinical trials among the 75% of the population who are APOE ɛ4 negative, as 50% of the population burden of AD will manifest in this subgroup.

All the samples we used came from population studies and it would be interesting to see if the predictive performances of this GRS are improved in selected population, either persons consulting for memory complaints, mild cognitive impairment, or persons with other risk factors such as diabetes or hypertension. The timing is also of importance in assessing the utility of such a genetic risk prediction tool. We restricted our samples to persons over 65 years of age but the GRS could work best in young to middle-aged populations where gene-environment interactions leading to AD have not had the chance to express themselves. Such questions will require an even larger sample size to answer reliably.

The main limitation of this study is the use of study populations that were already part of the initial IGAP study where the variants used in the GRS were discovered [16]. As such, it is currently not feasible to obtain a large sample outside IGAP for independent validation. Nevertheless, the eight cohorts studied here were a small fraction of the larger sample used in the discovery of the SNPs constituting the GRS. We could have used the ORs derived from the independent replication stage of IGAP but as these were very close to the ones we used to construct our GRS, this should not have dramatically affected our results. The encouraging results we obtained should therefore be replicated in an independent cohort study.

There are other decisions we made in the construction of the GRS that could be argued but in this initial attempt to construct and validate a GRS, we chose a conservative approach. Thus our selection of the SNPs used in the GRS was based purely on statistical grounds. By selecting SNPs based on GWAS p-value threshold, we limited the number of false positive signals in our GRS, but we also excluded signals that could be replicated in future studies. For example, a follow-up study has already reported replication of the TRIP4 locus that was suggestive in the IGAP meta-analysis [33]. We also excluded some signals that failed to reach genome-wide association threshold in IGAP but have persuasive biological evidence in favor of a real implication in the pathophysiology of AD, such as CD33. Future work on alternative versions of a GRS could expand the list of SNPs included using a less stringent threshold of p-values and other criteria, e.g. predicted functionality or biological evidence.

Moreover, inclusion of rare variants could also improve the GRS. It is unlikely that a unique rare variant could dramatically improve risk stratification due to its low prevalence in any population, but they could improve risk prediction in selected individuals and inclusion of genetic information in genes known to harbor multiple rare variants that alter risk for the disease could represent an interesting addition.

The SNPs used to create the GRS were based on studies using a case-control design. Thus, it is possible that the genetic risk factors identified so far only act very late in the pathophysiology of AD. Nevertheless, designing a GWAS of incident AD with a number of incident cases similar to IGAP would require prospective follow-up of hundreds of thousands of participants, which is currently not feasible.

Finally our GRS did not consider any gene-gene or gene-environment interactions which could be of importance [19].

Our study has several strengths. First it presents data from the largest set of prospective studies of AD to date. The large sample sizes ensure sufficient statistical power and fairly precise estimations of risk while the prospective design ensure more accurate assessment of AD onset date. Second, we accounted for the differences in follow-up time by choosing a fixed time for estimating measures of improvement in risk prediction, ensuring comparability across studies. Finally, all studies used homogeneous definitions for the variables used as confounders and the statistical analyses were performed using the same statistical script.

In conclusion, we have shown that evaluation of the clinical utility of a GRS requires more than study of its association with incidence of a disease. In the context of AD, a risk prediction tool build using newly identified genetic information outside APOE ɛ4 showed a statistically significant, small improvement in individual AD risk prediction after 7 years of follow-up over age, sex, education, and APOE ɛ4 status. Nevertheless, this GRS may be a useful tool for improving risk stratification in clinical trials aimed at preventing the onset of clinical AD.