Abstract. Phenotypic variation can partly be explained by genetic variation, such as variation in single nucleotide polymorphism (SNP) genotypes. Genomic selection methods seek to predict genetic values (breeding values) based on SNP genotypes. To develop and to optimize these methods, simulated data are often used, which follow a rather simple genotype-phenotype map. Is the conventional approach for data simulation in this field an appropriate basis to optimize such methods in view of experimental data? Here, we present an alternative approach, striving to simulate more realistic data based on a genotype-phenotype map which includes a simulated metabolome level. This level was used to simulate genetic values, implicitly including additive and non-additive genetic effects, whereas in a conventional approach additive and dominance effects were explicitly simulated and assembled to genetic values. For both simulation approaches, different scenarios regarding numbers of quantitative trait loci (QTLs) and SNPs were analysed using fastBayesB as prediction method. We observed that our alternative map showed a smaller prediction precision (at least 3.75 %) compared to the conventional approach in all investigated scenarios. The observed degree of linearity is at least 94.12 % of the conventional approach or less. Additionally, we present results for different simulated data and experimental data to allow a comparison on a purely conceptual level. Concluding, simulating a more complex genotype-phenotype map including a molecular level, allows to study processing of variation from the genetic to the phenotype level in more detail and may prepare the ground for modern methods of genomic selection.