Purchase this article with an account.
S. K. Iyengar, Y. Wang, R. P. Igo, Jr., J. P. SanGiovanni, A. K. Henning, T. Clemons, E. Y. Chew; Comparison of Multiple Predictive Models for AMD Susceptibility. Invest. Ophthalmol. Vis. Sci. 2008;49(13):2654.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To develop predictive models for age-related macular degeneration (AMD) based on SNP combinations
The ca.106,000 (100K) Illumina SNP set and the associated phenotypes for any AMD (N= 391), advanced AMD (AAMD) (N=377), choroidal neovascularization (CNV) (N=194) and geographic atrophy (GA) (N=134) were obtained from the dbGaP AREDS project. Data cleaning led to removal of SNPs and samples that consistently showed low call rates (< 90% calls with GenCall score < 0.5) or poor quality (median GenCall score < 0.6) leaving 98,470 SNPs for analysis. Initial testing under additive (trend) models was used to find disease-associated SNPs for covariate-adjusted outcome variables. SNPs with p-values < 0.001 were examined using forward stepwise logistic regression to identify the best set of predictive SNPs for AMD, AAMD, CNV and GA after adjusting for age. A variety of methods including neural networks (NN) and classification trees (CT) were used to cross-validate the results of the predictive models.
Association testing showed that 98 - 278 SNPs were associated with AMD, AAMD, GA or CNV. Of these, SNPs in CFH and PLEKHA1-HTRA1 showed genome-wide significance. Using forward selection it was possible to reduce the set of greatest importance to 43-69 SNPs. Our final predictive models suggested that at least 10 SNPs, at independent genomic locations, were necessary to explain 75-80% of the variance for AMD, GA and CNV, but the AAMD model saturated at 4 SNPs and did not include CFH or PLEKHA1-HTRA1. Cross-validation using NN and CT showed that the predictive model generated via regression recapitulated poorly using alternative approaches. We are in the process of improving the initial SNP generation mechanism to derive more robust models.
This evaluation shows that AMD has a pleiotropic basis and many loci are necessary to cause disease. We are currently evaluating the validity of the models for risk prediction.
This PDF is available to Subscribers Only