Figure 2

Computerized selection of the most predictive peaks for discrimination between high and low responders to GH treatment. The figure shows the rotated partial least squares regression using the 2 (A) and 5 (B) most predictive protein peaks. In all analyses, changes in the peak intensity during the first year of GH treatment are used. The left panels demonstrate which peaks are selected and the degree to which they correlate with the first dimension (low/high response on the x-axis). The first dimension (Predicted response) is a weighted sum of the selected peaks and gives the best prediction of the binary low/high-response variable. A positive correlation with the first dimension indicates that the maximum peak intensity has increased more for the high responders than low responders during 1 year of treatment, whereas a negative correlation indicates that the maximum peak intensity has increased more for the low responders. The second dimension is a residual component. Residual, in the sense that it has no correlation with the response variable and, component, in the sense that it of all the selected predictor variables it explains the maximum variance. The use of this dimension facilitates the discovery of clusters of related peaks, or peaks conveying comparable information. In the middle panels, correlations of dimensions 1 and 2 with phenotype are shown. The right panels illustrate the discrimination of the individual low and high responders using the selected peaks. Low responders are shown as triangles; high responders as circles.