A mixture model with a reference-based automatic selection of components for disease classification from protein...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Bioinformatics data analysis commonly employs linear mixture models to represent samples as additive mixtures of components.
• Blind matrix factorization methods can extract these components from mixture samples.
• Automatic selection of components for classification remains a challenge.

Purpose of the Study:

• To develop a novel additive mixture model for feature extraction in bioinformatics.
• To enable automatic selection of components for disease prediction without using label information.
• To improve the accuracy and interpretability of component-based analysis in cancer datasets.

Main Methods:

• Proposed a sample-by-sample additive mixture model utilizing sparseness-constrained factorization.
• Decomposed each sample into control-specific, case-specific, and neutral components.
• Determined the number of components via cross-validation and assigned features based on sample-specific thresholds.
• Enabled automatic component selection for classification without relying on prior label information.

Main Results:

• Applied the method to ovarian, prostate, and colon cancer datasets for disease prediction.
• Achieved high average sensitivities (e.g., 96.2% for ovarian cancer) and specificities (e.g., 99% for prostate cancer).
• Demonstrated improved prediction accuracy through reduced component complexity and automatic feature allocation.

Conclusions:

• The proposed sample-by-sample factorization method offers an alternative to simultaneous dataset factorization.
• Automatic component selection allows for direct use in classification, unlike standard methods.
• Disease-specific components can be interpreted as sub-modes for biomarker identification.
• The method enhances prediction accuracy by refining component selection and reducing complexity.