We developed a probability-based machine-learning program, Colander, to identify tandem mass spectra that are highly likely to represent phosphopeptides prior to database search. We identified statistically significant diagnostic features of phosphopeptide tandem mass spectra based on ion trap CID MS/MS experiments. Statistics for the features are calculated from 376 validated phosphopeptide spectra and 376 nonphosphopeptide spectra. A probability-based support vector machine (SVM) program, Colander, was then trained on five selected features. Data sets were assembled both from LC/LC-MS/MS analyses of large-scale phosphopeptide enrichments from proteolyzed cells, tissues and synthetic phosphopeptides. These data sets were used to evaluate the capability of Colander to select pS/pT-containing phosphopeptide tandem mass spectra. When applied to unknown tandem mass spectra, Colander can routinely remove 80% of tandem mass spectra while retaining 95% of phosphopeptide tandem mass spectra. The program significantly reduced computational time spent on database search by 60-90%. Furthermore, prefiltering tandem mass spectra representing phosphopeptides can increase the number of phosphopeptide identifications under a predefined false positive rate.