E coli genome 2.0 array

Prior to building a prediction model, we transformed the adjusted gene expression data into categorical values (under-expressed, UE; wild-type, WT; over-expressed, OE) in order to deal with biases arising from combining different platforms and improve the classification accuracy [63 (link)]. We first measured the log₂Fold Change (FC) of gene expression with respect to the WT expression for each gene. WT samples were identified from experiments that didn’t undergo genetic and environmental perturbations from the three platforms (7 for Affymetrix E. coli Antisense Genome Array, 6 for Affymetrix E. coli Genome 2.0 Array, and 6 for RNA-Seq). log₂Fold Change (FC) was separately measured for each platform by comparing the mean of WT data. Using transformed data, we estimated a normal distribution N(μ, σ²) for each gene and finally converted each log₂FC gene value into one of the 3 categorical values by measuring deviation from the mean (UE when g_ij < μ_i – σ_i; WT when μi – σ_i ≤ g_ij ≤ μ_i + σ_i; OE when μ_i + σ_i< gij; g_ij is the log₂FC for gene i in sample j, μ_i is mean of gene i and σ_i is standard deviation of gene i). The platform-specific categorization of gene expression effectively removes platform biases (Fig. 1B).