Cognitive Decline in Alzheimer’s Disease: Limited Clinical Utility for GWAS or Polygenic Risk Scores in a Clinical Trial Setting

Introduction: Alzheimer’s disease (AD) is a progressive and irreversible neurological disease. The genetics and molecular mechanisms underpinning differential cognitive decline in AD are not well understood; the genetics of AD risk have been studied far more assiduously. Materials and Methods: Two phase III clinical trials measuring cognitive decline over 48 weeks using Alzheimer’s Disease Assessment Scale-cognitive subscale (ADAS-cog, n = 2060) and Clinical Dementia Rating-Sum of Boxes (CDR-SB, n = 1996) were retrospectively genotyped. A Genome-Wide Association Study (GWAS) was performed to identify and replicate genetic variants associated with cognitive decline. The relationship between polygenic risk score (PRS) and cognitive decline was tested to investigate the predictive power of aggregating many variants of individually small effect. Results: No loci met candidate gene or genome-wide significance. PRS explained a very small percentage of variance in rates of cognitive decline (ADAS-cog: 0.54%). Conclusions: These results suggest that incorporating genetic information in the prediction of cognitive decline in AD currently appears to have limited utility in clinical trials, consistent with small effect sizes estimated elsewhere. If AD progression is more heritable soon after disease onset, genetics may have more clinical utility.

Supplementary 1: Trial Design AVA102670 and AVA102672 were international, phase III, multicentre, randomised, doubleblind, placebo-controlled, parallel-group studies designed to assess efficacy and safety of rosiglitazone XR (RSG XR) as adjunctive therapy in mild-to-moderate Alzheimer's Disease patients already being treated with an approved acetylcholinesterase inhibitor (AChEI, AVA102670) or donezepil (AVA102672) and stratified by APOE e4 allele status.
Full details of the clinical trial design and treatment arms is available on the GSK study register: https://www.gsk-studyregister.com/en/. A schematic representation of the study design is shown in Figure S1. Figure S1. Study Schematic.
Following a 28-day (±7 days) screening period in which patients received standard of care (SoC, AChEI or donezepil), patients were randomised by APOE 4 allele status into one of three arms. Double-blind treatment ended at week 48 with all patients receiving SoC and placebo until study end at week 54.

Supplementary 2: GWAS Results
There were no significant associations with any of the 39 candidate variants, or genomewide variants with either primary endpoints (Table below). Here a positive effect size/change score represented faster cognitive decline. Significance thresholds were p<6.41x10 -4 for candidate variants and p<2.5x10 -8 for genome-wide variants.
Genome-wide differences in allele frequency that correlate with phenotype can be an indicator of various artefacts and are routinely corrected for using genomic control (GC, Dadd et al 2009). This process calculates an inflation factor, lambda, which summarises evidence for genome-wide deviation from expected test statistics. A lambda value of 1 indicates no evidence of inflation, and typically values below 1.045 are of no concern. Inflation can be tested by adjusting test statistics and recalculating lambda. Lambda for ADAS-cog before GC adjustment is 0.987308 and 0.987308 after. Lambda for CDR-SB before GC adjustment is 1.000035 and 1 after. This indicates that there is little concern over population structure not captured by Principal Components in this analysis.
The effect estimate is with respect to the minor allele. 24 candidate variants have a MAF>20% (see supplementary Table S2a and S2b) Candidate variants are sorted by p-value   Under a genetic additive model, candidate variants have at least 80% power for ADAS-Cog and CDR-SB change if the absolute effect size is greater than 1.36 and 0.49 phenotypic units respectively and minor allele frequency is >20%. Twenty-four of the 39 candidate variants had minor allele frequencies >20% (Table S2a).
GWAS variants have at least 80% power for ADAS-Cog and CDR-SB change if the absolute effect size is greater than 3.39 and 1.23 phenotypic units respectively and minor allele frequency is >20%.  On the figures below a horizontal dotted line indicates 80% power. In all cases, the calculations indicated that there was good power (>80%) to observe an association between the AD PRS and disease progression (ADAS-Cog and CDR-SB) provided the genetic correlation between the two phenotypes was greater than ~5%. (http://cnsgenomics.com/shiny/gctaPower/ ), power for GCTA was derived. This calculation required an assumption on the variance of SNP-derived genetic relationshipsi.e. the variance in relatedness as measured by DNA, between pairs of individuals. As the current samples were unrelated individuals measured genome-wide, the default value of 2x10 -5 was used. Heritability (h 2 ) is the proportion of variance in phenotype in a population that can be explained by common SNPs genome-wide. Here, power to detect a range of heritability values was estimated. These results indicated that power to detect even large heritability estimates (20%) was poor given the sample sizes used here.

Supplementary 8: Mixed Model Analysis
Congruency between the present data and two recent publications was evaluated using mixed effect models. These publications suggest associations between genetic variants in IL1RAP (Ramanan et al 2018) and TREM2 (Del Aguila et al 2018) and AD progression. The IL1RAP variant was associated with higher rates of amyloid accumulation (independent from APOE) and rs12053868-G carriers were more likely to progress from mild cognitive impairment to AD and exhibited greater longitudinal temporal cortex atrophy on magnetic resonance imaging. For TREM2 Del Aguila report, rs143332484 reached "nominal" significance with progression as measured by CDR-SB (p=0.02) and the Free and Cued Selective Reminding Test (FCSRT)-Free Recall measure of episodic memory (p=0.02). However, the TREM2 analysis did not include correction for multiple testing. Another TREM2 variant (rs75932628), not in LD with rs143332484, was previously associated with AD risk (Guerreiro et al, 2013), not AD progression. While these publications have limitations to understanding genetic predictors of cognitive decline, we undertook a similar analysis approach using the data herenamely, a linear mixed-model repeated measure (MMRM) framework that incorporates additional cognitive assessment data than simply the change from Baseline to Week 48 approach employed in our main analysis.
There were no significant results with any of the 3 candidate variants. A liberal significance threshold was set at p<0.017 for 3 candidate variants after adjusting for Type I error, to evaluate whether the results of the previous study replicated here. The impact of PRS on phenotype can be interpreted through splitting individuals into quantiles based on their PRS and comparing the average phenotype amongst individuals in each quantile with individuals in a reference quantile. Here, individuals were split into 5 quintiles, based on the AD PRS at the most predictive threshold for each phenotype, and compared with reference to the 3 rd quantile, i.e. the middle quantile. This is equivalent to asking how an individual in the top 20% for genetic risk compares to an 'average' individual, and so on. These results are presented below graphically and as tables.

Discussion
The table below shows the results for the interaction term for PRS by candidate genotype regressed on ADAS-Cog endpoint in a linear model with clinical covariates used earlier and 6 PCs.
This analysis applied novel approaches to assess the heritability of cognitive declineevaluating PRS by candidate SNP interaction and genome-wide data to assess the heritability of cognitive decline.
The interaction analysis for ADAS-Cog showed no significant interaction effectsthe two variants with the lowest p-values were CNTNAP2 (P=0.024) and MS4A4E-MS4A4A (P=0.027), which did not meet our statistical significance threshold. While APOE is an important contributor to AD risk, there was no evidence for an interaction effect between PRS and APOE (p=0.88) in ADAS-cog change score.