Figure 1

Performance of random signatures levels off at around 300 probesets. The dotted lines are the accuracies obtained (for increasing number of nearest neighbours from bottom to top) for randomly shuffled compound-target associations. This shows that although the signature probesets were selected randomly, they nonetheless yield a better target prediction accuracy than chance alone. One nearest neighbour: cross; two nearest neighbours: triangle; three nearest neighbours: square. Displayed data are the average accuracy values (n = 50) and the total length of the error bars is the corresponding standard deviation.