Urinary nucleosides based potential biomarker selection by support vector machine for bladder cancer recognition

Urinary nucleosides are potential biomarkers for many kinds of cancers. But up to now, it has been little focused in bladder cancer recognition. The aim of present study is try to validate the potential of urinary nucleoside as biomarker for bladder cancer diagnosis by finding out some urinary nucleosides with good discriminative performance for bladder cancer recognition in urinary nucleoside profile.

20 urinary samples for cancer and the same number for control are collected and treated by capillary electrophoresis-mass spectrometry experiments to achieve urinary nucleoside profile, in which 44 peaks were integrated and the ratios of the relative peak area to the concentration of urinary creatinine were used as features to describe all samples. Support vector machine based recursive feature elimination (SVM-RFE) and a new feature selection method called support vector machine based partial exhaustive search algorithm (SVM-PESA) were used for biomarker identification and seeking optimal feature subsets for bladder cancer recognition.

Based on the urinary nucleoside profile, 22 optimal feature subsets consist of 3-4 features were found with 95% 5-fold cross validation accuracy, 100% sensitivity and 90% specificity by SVM-PESA, whose performance were much better than that of optimal feature subset selected by SVM-RFE. By analyzing the statistical histogram of features' appearance frequency in several best feature subsets, urinary nucleosides with m/z 317, 290 and 304 were thought as potential biomarkers for bladder cancer recognition.

These results indicated urinary nucleosides may be useful as tumor biomarkers for bladder cancer, and the new method for biomarker selection is effective.