The difficulties of reproducing conventionally derived results through 500k-chip technology

Based on a "training" sample of 1,042 subjects genotyped for 5,728 single-nucleotide polymorphisms (SNPs) of a conventional 0.4-Mb genome scan and a "test" sample of 746 subjects genotyped for 545,080 SNPs on a 500k-chip, we investigated the extent to which the subjects' immunoglobulin M levels can be reproducibly predicted from a multilocus genotype. We were specifically interested in the reproducibility of predictors across populations (1,042 versus 746 subjects) and across SNP sets (conventional genome scan versus anonymous 500k-chip) because this is a prerequisite for clinical application. For the training sample, neural network (NN) analysis yielded classifiers that predicted immunoglobulin M levels from the subjects' multilocus genotypes at acceptable error rates through a configuration of 15 genomic loci (61 SNPs). With the test sample (746 subjects) we addressed the question of reproducibility across populations and across SNP sets by means of a novel "competitive SNP set" approach. However, the chip data contained several sources of distortion, including greatly elevated noise levels and artifact-prone SNP regions, thus complicating attempts to verify the reproducibility of NN predictors. Though 5 of 15 genomic loci from the training samples appeared to be reproducible, the NN classifiers derived so far from the test samples are insufficiently compatible with the training samples. Nonetheless, our results are promising enough to justify further investigations. Because the underlying algorithm can easily be split into parallel tasks, the proposed "competitive SNP set" approach has turned out to be well suited for computers with today's 64-bit multiprocessor architectures and to offer a valuable extension to genome-wide association analyses.