I see the point in applying such algorithms in genome-wide association studies (GWAS), but I'm not sure about it when few SNPs are available. In this case, would it be better to keep applying univariate approaches to relate genotype and phenotype?
15 SNPs are very few for a GWAS. Unless you already selected them around a specific region suspected to harbour some effect. Any technique that will reduce the number of SNPs to be tested is anyway useful (lassp, PLS, PCA and so on)