Fig. 6. Predictive value of an identical set of blood gene expression markers identified by network analysis in the classification of response to recombinant human growth hormone (r-hGH) in patients with growth hormone deficiency (GHD) and Turner syndrome (TS).

First year growth response is used as an example. Similarities in the interactome models of the response of GHD and TS to r-hGH were identified by overlap at each year of therapy. Genes were selected that were significantly related to growth response in either or both GHD and TS, generating an identical set of gene probesets used for prediction of both high and low response in both GHD and TS. BORUTA, an all relevant feature selection wrapper random forest-based algorithm, was used to confirm the importance of gene expression probe-sets used for classification of response to r-hGH. The BORUTA algorithm uses a 100-fold permutation to define the noise present in the data; the noise is modelled as shadow variables and used as a basis to assess confidence in the data. Green = confirmed gene probeset, yellow = tentative gene probeset, red = rejected gene probeset, blue = shadow variables (high, medium and low shadow variables are derived to define the noise within the dataset). Low quartile (left column—LoQ) and high quartile (right column—HiQ) are shown for first year growth response to r-hGH in GHD and TS. The same group of gene probesets are used in each case (colour figure online).