Cuckoo search optimisation for feature selection in cancer classification: a new approach

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning in Oncology

Background:

• Microarray gene expression data presents challenges in cancer classification due to high dimensionality (thousands of genes) and limited samples.
• Effective feature selection is crucial for identifying relevant genes to enhance classification accuracy and reduce computational complexity.

Purpose of the Study:

• To apply the Cuckoo Search (CS) optimization algorithm for informative gene selection in cancer classification using microarray data.
• To evaluate the performance of CS-based feature selection in improving the accuracy of cancer classification models.

Main Methods:

• Genes were initially ranked using statistical measures: T-statistics, Signal-to-Noise Ratio (SNR), and F-statistics.
• The Cuckoo Search (CS) algorithm was employed to select informative genes from the top-ranked gene subsets.
• The k-Nearest Neighbour (kNN) classification accuracy served as the fitness function for the CS algorithm.

Main Results:

• The proposed CS-based feature selection method was evaluated on ten diverse cancer gene expression datasets.
• Achieved 100% average classification accuracy for several datasets, including DLBCL Harvard, Lung Michigan, Ovarian Cancer, AML-ALL, and Lung Harvard2.
• Demonstrated superior performance compared to existing techniques on DLBCL outcome and prostate cancer datasets.

Conclusions:

• Cuckoo Search is a highly effective optimization algorithm for feature selection in cancer classification from high-dimensional gene expression data.
• The CS approach significantly enhances classification accuracy, offering a promising tool for identifying key genes in cancer research.
• This method provides a robust and accurate solution for analyzing complex genomic data in oncology.