Identifying simple discriminatory gene vectors with an information theory approach

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning in Medicine

Background:

• Traditional cancer classification methods often use individual genes, leading to redundant information and complex models due to overlapping gene expression patterns.
• The principle of Occam's razor favors simpler models when prediction performance is comparable.
• High-dimensional gene expression data presents challenges for feature selection and classifier design.

Purpose of the Study:

• To develop a new method for learning accurate and low-complexity classifiers from gene expression profiles for cancer classification.
• To address the redundancy and complexity issues inherent in single-gene feature selection methods.
• To leverage gene vectors for improved class distinction in cancer samples.

Main Methods:

• Utilized mutual information to measure the relationship between gene vectors (sets of genes) and sample class attributes.
• Employed gene vectors in higher-dimensional spaces to capture more diverse and informative patterns compared to individual genes.
• Validated the proposed method on three gene expression datasets.

Main Results:

• The proposed gene vector method demonstrated comparable or superior prediction performance to existing classification methods.
• Achieved significantly lower model complexity compared to traditional single-gene selection approaches.
• Effectively utilized the richer information contained in gene vectors for class distinction.

Conclusions:

• Gene vectors, analyzed using mutual information, offer a more effective and efficient approach to cancer classification feature selection.
• The developed method successfully balances prediction accuracy with model simplicity.
• This approach provides a promising alternative for building robust and interpretable cancer classifiers from gene expression data.