Polygenic Risk Score Analysis of Pathologically Confirmed Alzheimer's Disease

Introduction

Polygenic risk score (PRS) analysis enhances the predictability of the diagnosis of Alzheimer's disease (AD) over the use of just the apolipoprotein E locus (1). In a recent PRS analysis, we showed that the area under the curve (AUC) in the recent genome wide association study (GWAS), was 0.79 (1). However, the study samples in these cohorts were largely comprised of clinical cases of AD, and the diagnostic accuracy of these is not perfect as recent clinical trial failures have highlighted (2). In addition, the majority of controls which are used in GWAS, are sampled from a general population and are often underaged to develop AD. This diagnostic uncertainty has also been demonstrated by the observation of c9orf72 expansions (a locus causing frontotemporal dementia) within some of the clinical AD cohorts used in the generation of the GWAS and AD sequencing data (3).

Having a better understanding of the diagnostic utility of PRS is of importance for two reasons: first, because it enables the accurate assessment of how much risk for disease there is still left to be found and this is important in setting research goals, and second, because this type of analysis could be used in the refinement of inclusion criteria for clinical trials and eventually, in clinical health care recommendations.

We have previously reported a GWAS in clinically characterized and neuropathologically confirmed samples of AD and matched controls (4): in this analysis, we apply PRS to these pathological data to determine whether some of the “missing heritability” of AD is due to clinical misdiagnosis.

Methods

The sample characteristics of the dataset used in this study were the same as in our original analysis. This project was declared IRB exempt (Medstar Project #.2003-118) under the Code of Federal Regulations, 45 CFR, 46. Eight cases and eight controls had corrupted data files and were omitted (4). This left 1011 cases and 583 controls. The total number of imputed single nucleotide polymorphisms (SNPs) was 36,481,940. The number of SNPs with Info score above 0.8 was 11,016,052. From these, the number of SNPs with MAF>=0.01 was 7,868,100 and these were used in the analysis. Association analysis was performed for each SNP using logistic regression analysis as implemented in snptest (5).

We performed predictive modelling using polygenic score based upon SNPs with p-value cut-off p=10^-4, 10^-3, 0.01, 0,05, … 0.5 as in (7) as predictor variables. These sets of SNPs are capturing APOE and index GWAS SNPs (7) either directly or via their proxies. For prediction modelling we converted imputed “dosage” genotypes in our data to “most probable genotype” with probability over 90%. The individual polygenic risk scores were generated as sum of the risk alleles weighted by effect sizes as in the International Genomics of Alzheimer's Project (IGAP) study (7), then were further adjusted for first 10 principal components and standardized. The models were fitted using the above mentioned individual polygenic risk scores and predicting AD/control status in our study. This is the most powerful way of testing the prediction ability of the strongest genetic predictors to date: however our study was part of the IGAP study (7) (∼3% overlap) and therefore the results will be marginally overfitted. We accounted for this overfitting in our analysis as below.

Since summary statistics for the IGAP (7) data excluding our sample was not available to us, we estimated the effect of possible bias using simulations. For that we first simulated a sample of 17008 cases and 37154 controls, matching the IGAP stage-I study, for a typical SNP with minor allele frequencies=0.2 and effect size of odds ratio (OR)=1.05. This OR matches the average effect size for IGAP pruned SNPs with association p-value≤0.5, mean(B_IGAP)=0.05, mean(SE_IGAP)=0.035; OR_IGAP=exponential(0.05)=1.05. Then we randomly removed 1101 cases and 583 controls (matching our study size) and recalculated the association effect size 1000 times de novo. The “removal-based-simulated” effect sizes for a single typical SNP were found to be normally distributed with mean B_SIM=0.05 (SD_SIM=0.004) (not shown). Assuming that the removed sample is a random subset of cases and controls, the expected distribution of the IGAP pruned SNPs effect sizes should have the same mean but slightly increased standard error:

where N is the IGAP sample size, N_o is the overlap sample size. In particular, we can roughly expect the mean(SE_IGAP)_ADJ=0.035/sqrt(0.97)=0.0355, where 0.035 is the mean(SE_IGAP) of the effect size for IGAP pruned SNPs with association p-value≤0.5.

To adjust prediction modelling for overlapping samples, we ran further simulations where the effect sizes for each SNP in the IGAP study were simulated as b∼N(B_IGAP, sd=0.12*SE_IGAP), where B_IGAP is the beta-coefficient and SE_IGAP is the standard error for that SNP in the IGAP study. The sd=0.12*SE_IGAP was chosen empirically to allow for both the variability due to IGAP B-coefficient estimate and due to random subsample removal. As a rough example, multiplying the mean(SE_IGAP)_ADJ by 0.12 results in a standard deviation, which is approximately matching the “removal-based-simulated” SD_SIM: 0.12*mean(SE_IGAP)_ADJ =0.12*0.0355 ≈ 0.004= SD_SIM. Thus, in each simulation step, each SNP in IGAP had a simulated effect size and p-value corresponding to this effect size. Then the SNPs were reselected, repruned and the polygenic scores recalculated. The prediction accuracy of the simulated PRS was calculated at each simulation (N simulations =1000) and mean of simulated AUC for SNPs with p≤0.5 was reported and is discussed below.

Results

The primary results (QQ-plot and Manhattan plot) were consistent with our previous analysis of these data (4). There were no genome-wide significant hits apart from APOE locus.

We compared the results of our analysis with 21 index genome-wide significant SNPs identified in the IGAP (7) study (see Table 1). Sixty three percent of IGAP GWAS index SNPs (14 out of 22) show larger effects in our dataset compared to the original report (7), of which 5 have significantly larger effect sizes (see the last column of Table 1), including the two SNPs tagging the APOE status.

The results of predictive modelling are presented in Table 2. Training on the whole IGAP the prediction accuracy AUC reaches 86% (Figure 1) when all SNPs with p≤0.5 are included in the model, However, as discussed above there is an element of overfitting in this analysis as our data was part of the IGAP analysis. Accounting for this possible inflation using simulation (see Methods), the prediction accuracy is 84% (95%CI 82-86%).

Discussion

These data systematically confirm, in the context of genome wide data, our results examining the APOE locus (4): genetic prediction is better in the context of autopsy confirmed cases and controls. This has implications for our view of how much genetic variability remains to be found: in an earlier analysis, we estimated that the theoretical maximal genetic variance to be found would generate an AUC of 82% (95% confidence interval 78%-85%) (8). The figure now identified, based on the genome wide analysis of a pathological cohort is 84% (95% confidence interval 82-86%). Thus the theoretical and assessed the figures for risk prediction accuracy overlap and both are larger than the AUC of 0.75 assessed using clinical cohorts (1). There is thus further evidence that polygenic risk profiling captures the SNP-heritability very well with regards to common variation in AD, although of course, heritability estimates (8) were constructed on clinical diagnoses of AD so strict comparisons are hazardous. This does not imply that there are no genetic findings of very rare variants (f<0.1%) still to be made, although the increasing predictability of genetic findings (9-12) and the fact that most new findings relate to already identified pathways implies that research may be better focused on the targeted sequencing of established pathways, bioinformatic analyses of multi-omics data sets, and cell biology rather than on large scale genome wide sequencing projects in unrelated sporadic AD individuals. These data also illustrate that there is a degree of misdiagnosis in the clinical AD series (3), and even more so in population based controls.

A final implication of these data is that genome wide genotyping and PRS based analytic strategies are reasonably effective at predicting those who will develop disease. They also suggest that this predictive utility is unlikely to improve much more. This strategy may therefore now be useful for designing clinical trials and eventually in clinical practice.