A novel class dependent feature selection method for cancer biomarker discovery

Identifying key biomarkers for different cancer types can improve diagnosis accuracy and treatment. Gene expression data can help differentiate between cancer subtypes. However the limitation of having a small number of samples versus a larger number of genes represented in a dataset leads to the overfitting of classification models. Feature selection methods can help select the most distinguishing feature sets for classifying different cancers. A new class dependent feature selection approach integrates the F-statistic, Maximum Relevance Binary Particle Swarm Optimization (MRBPSO) and Class Dependent Multi-category Classification (CDMC) system. This feature selection method combines filter and wrapper based methods. A set of highly differentially expressed genes (features) are pre-selected using the F statistic for each dataset as a filter for selecting the most meaningful features. MRBPSO and CDMC function as a wrapper to select desirable feature subsets for each class and classify the samples using those chosen class-dependent feature subsets. The performance of the proposed methods is evaluated on eight real cancer datasets. The results indicate that the class-dependent approaches can effectively identify biomarkers related to each cancer type and improve classification accuracy compared to class independent feature selection methods.