Fig. 4.

A machine learning algorithm separates direct from indirect Cdc28 targets. (a) Correction for the 1NM-PP1 off-target effect yielded a list of dephosphorylated sites that correspond to both direct and indirect Cdc28 targets. A mixture of direct and indirect Cdc28 targets (dephosphorylated phosphosites) along with part of the negative control set of up-regulated phosphosites across all experiments was used to build a support vector machine (SVM) model. Accuracy was estimated by cross-validation on the remaining part of the negative control group. Known CDK substrates as well as CDK motifs were enriched among direct targets predicted by the model. (b) The KID was used to extract known Cdc28 partners from literature-based studies, which were compared with direct and indirect Cdc28 sites predicted by the SVM model.