Supervised classification of array CGH data with HMM-based feature selection.

Area of Science:

• Genomics
• Bioinformatics
• Cancer Research

Background:

• Understanding molecular mechanisms in tumorigenesis requires advanced analytical methods.
• Copy Number Variations (CNVs) are crucial in tumorigenesis, detectable via Comparative Genomic Hybridization (CGH).
• High-resolution array CGH generates large datasets necessitating sophisticated mathematical approaches.

Purpose of the Study:

• To develop and validate a computational framework for classifying cancer subtypes using CNV data.
• To identify key chromosomal regions associated with different cancer subtypes.
• To improve the accuracy and efficiency of cancer subtyping.

Main Methods:

• Utilized recurrent hidden Markov Models (HMMs) to identify characteristic chromosomal regions in array CGH data.
• Integrated univariate feature selection methods to reduce data dimensionality.
• Employed Weighted Least Squares Support Vector Machines (LS-SVM) for supervised classification, accounting for data imbalance.

Main Results:

• Successfully classified cancer subtypes using CNV data with high accuracy (88-95.5% cross-validation).
• Identified specific chromosomal regions distinguishing patient groups across multiple datasets, including ovarian cancer.
• Demonstrated the effectiveness of the integrated HMM and LS-SVM approach.

Conclusions:

• The combination of recurrent HMMs and LS-SVM provides a robust and novel method for cancer classification based on CNVs.
• This approach effectively identifies and narrows down critical chromosomal regions for accurate subtyping.
• The methodology aids in a deeper understanding of molecular mechanisms driving tumorigenesis.