G-Forest: An ensemble method for cost-sensitive feature selection in gene expression microarrays.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Microarray gene expression profiling is vital for cancer research but suffers from the curse of dimensionality.
• Existing feature selection methods overlook genomic data variance and gene costs, leading to inefficient profiling.

Purpose of the Study:

• To introduce G-Forest, a novel cost-sensitive feature selection algorithm for gene expression microarrays.
• To address the limitations of existing methods by incorporating feature costs and handling rare variances.

Main Methods:

• G-Forest is an ensemble, cost-sensitive feature selection algorithm based on Random Forest.
• It embeds gene costs into the selection process, prioritizing low-cost, informative features.
• Features are randomly selected with probabilities inversely proportional to their associated costs during population construction.

Main Results:

• G-Forest demonstrated superior effectiveness and robustness compared to state-of-the-art algorithms.
• The algorithm achieved up to a 14% improvement in accuracy.
• G-Forest reduced gene profiling costs by an average of 56%.

Conclusions:

• G-Forest offers an effective solution for cost-sensitive feature selection in gene expression data.
• The algorithm successfully balances the selection of informative genes with cost reduction.
• This approach enhances the utility of microarray data for cancer diagnosis, prognosis, and treatment.