Extending Partially Known Networks

Abstract

Besides experimental techniques, computational inference approaches have contributed to the reconstruction of gene regulatory networks in model organisms. Particularly successful are supervised approaches that take the known regulatory interactions and gene expression data into account. However, they have not yet been applied to individuals genotyped by systems genetics data, where genetic polymorphisms are the major source of variation in gene expression profiles. We apply a supervised inference framework to expression datasets, genotype information, and the known gene regulatory interactions that are generated in a standardized setup by the SysGenSim software. We confirmed in this setup that supervised approaches exploiting the known interactions perform better than pure expression-based methods as well as methods exploiting expression data and genotype information. The performance of supervised methods was robust with respect to parameterization and data pre-processing. Furthermore, whether or not the genotype information was explicitly used influenced the performance of supervised approaches only little. We also analyzed differences between real and artificial data and setups to assess the chances of a successful inference in real systems. Due to reasons discussed in this chapter, several extensions of supervised approaches that considerably improve performance on real data were not effective in the SysGenSim case. Our thorough comparison between real and artificial setups suggested that the application of supervised approaches to real systems might be more robust and straightforward in comparison to current unsupervised approaches. In particular, as real genotypes likely are more complex and cause more versatile responses, the finding that supervised approaches are not dependent on the explicit representation of genotype information might prove of advantage.