Figure 1

A BoCaTFBS classifier trained over NF-κB ChIP-chip experimental data. It consists of two cascade stages and 12 features for each stage (partially shown). This cascade predictor was tested by cross-validation and achieved 82% sensitivity (true positive rate) at a 5% false-positive rate. BoCaTFBS classifiers are built on discriminative features, which differentiate positives (the binding sites) from the chosen negative training set (the nonbinding sites). For example, in stage 1, the sequence where position 4 is not C is more likely to have more binding propensity. The consensus sequence of binding sites is GGGRNNYYCC (R is purine, Y is pyrimidine, and N is any nucleotide). The classifier at each stage is built upon a random small subset of the over-represented class at each stage. Moreover, each classifier is dependent on the results of the classifiers in the previous stages. NF-κB, nuclear factor-κB.