Gene-expression-based cancer subtypes prediction through feature selection and transductive SVM

Area of Science:

• Bioinformatics
• Cancer Genomics
• Machine Learning in Medicine

Background:

• Microarray technology enables gene expression profiling for cancer outcome prediction.
• Classifying cancer subtypes and identifying diagnostic gene markers are crucial but challenged by small sample sizes.
• Traditional supervised learning requires fully labeled data, often excluding valuable unlabeled microarray datasets.

Purpose of the Study:

• To propose a novel approach integrating feature (gene) selection with transductive support vector machines (TSVMs).
• To identify potential gene markers for cancer subtypes.
• To enhance cancer classification accuracy using semisupervised learning.

Main Methods:

• A forward greedy search algorithm based on consistency and signal-to-noise ratio was used for gene selection.
• Selected genes were utilized to design a TSVM classifier.
• The proposed method was compared against standard inductive SVMs (ISVMs) and low-density separation.

Main Results:

• The study successfully identified potential gene markers associated with cancer subtypes.
• TSVMs demonstrated improved prediction accuracy compared to ISVMs.
• The proposed semisupervised technique proved effective for cancer classification and gene-marker identification.

Conclusions:

• Combining gene selection with TSVMs offers a powerful strategy for cancer classification, especially with limited labeled data.
• The developed method enhances diagnostic capabilities through accurate gene-marker identification.
• This approach effectively leverages unlabeled microarray data, overcoming a significant bottleneck in cancer research.