Evaluation of Different Classification Models to Extract Gene Signatures for Breast Cancer Recurrence Using Microarray Data

In this study, we aimed to improve the reliability and biological interpretability of gene signatures selected from microarrays by efficient usage of computational models and mathematical algorithms.
At the first step, a good model with high accuracy was chosen to predict cancer recurrence in microarray gene expression data on breast tumors. In this regard, microarray gene expression data of breast tumor in 1271 cancer patients (379 with recurrence and 892 people without recurrence) were utilized to construct an appropriate predictive model for recurrence by comparing the performance of multiple classifiers. In the pre-processing stage, different methods like correlation-based feature selection (CFS), principal component analysis (PCA), independent component analysis (ICA), and genetic algorithm as well as a random selection method were used to reduce the dimensions and choose the most appropriate genes (features).
Findings: A total of five gene signatures were selected by combining genetic algorithm, top scoring set (TSS), and random selection method, which showed the best results in most classification models. The final indicator genes were TRIP13, KIF20A, NEK2, RACGAP1 and TYMS, which had significant contribution in the structure of microtubules and spindle and also regulated the attachment of spindle microtubules to kinetochore.
By using hybrid models, we can avoid overfitting in training and achieve acceptable accuracy with biologically interpretable genes.