Fig. 1. Example of a hypercluster consisting of four clusters for a human gamma-actin peptide (Swissprot P02571). The peptide contains one cysteine which might be labeled with a light or a heavy ICAT tag, and one methionine which may be oxidized or not; i.e. there are four possible modification forms and therefore four possible clusters in total. Each of the four levels in the graph presents one modification form of the peptide. The light/heavy forms of ICAT-labeled cysteine are denoted with C/C^*, respectively. Oxidized methionine is denoted with M#. In this example, all peptide forms have been detected several times, e.g. there were eight spectra measured on the C^*M# precursor ion of this peptide. Indicated masses are average masses as calculated by Sequest.

Fig. 2. Quality control of the classification algorithm. (A) Scan-number distance (corresponding to elution time difference) histograms (bin size 100) of all light/heavy pairs passing the first and those passing the second filter. The triangle indicates the noise model. It is clearly visible that pairs passing the second filter have a strong tendency to co-elute, which is in line with experimental knowledge of the ICAT reagent. The total distribution is close to the expected random distribution plus the second filter ICATcher signal, which is a footprint of a good classification. (B) Histogram of ICATcher scores (bin size 0.05). The two step-histograms are the distribution of all ICATcher scores passing the first filter, and (smaller one) the distribution of scores for all correct pairs identified by Sequest and Peptide Prophet. The continuous curve is a scaled Gaussian model fit for those peptide links which pass the first filter but are false positives. Parameters: σ ≈ 0.13, maximum at μ ≈ 0.54. Only pairs with an ICATcher score below 0.2 are classified as true Δ = 9.03 light/heavy pairs.