Fig. 3

Composition of sub-nuclear compartments in a LocNuclei's development set, b the human, c chimp, d mouse and e yeast proteome. Proteomes show in B-E are corrected for prediction bias. Most machine learning tools have a prediction bias leading to wrong estimates of distributions. Using the confusion matrix of the development set, this bias can be corrected leading to more realistic estimates of the distribution [14]. When applying this correction to the sub-nuclear predictions for human, chimp, mouse and yeast, the distributions for human, chimp and mouse look more similar to the development set. For all four organisms, the fraction of proteins annotated to the nucleoplasm decreases after the correction