High dimensional model representation of log-likelihood ratio: binary classification with expression data.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Binary classification of high-dimensional data, common in bioinformatics, faces challenges with complex biological traits and interpretability.
• Existing methods struggle with gene interactions and the need for transparent 'glass-box' models.

Purpose of the Study:

• To develop interpretable low-dimensional approximations of the log-likelihood ratio using high dimensional model representation (HDMR).
• To account for individual gene effects and gene-gene interactions in classification models.
• To propose algorithms and a hypothesis test for identifying significantly dysregulated gene-gene interactions.

Main Methods:

• Utilizing high dimensional model representation (HDMR) theory.
• Developing two algorithms for approximating the second-order HDMR expansion.
• Implementing a hypothesis test based on HDMR for detecting dysregulated pairwise interactions.

Main Results:

• The proposed HDMR-based approach yields interpretable prediction rules with high accuracy.
• Successfully identifies significantly dysregulated gene-gene interactions from synthetic and real cancer gene expression data.
• Outperforms several state-of-the-art methods in synthetic data scenarios.

Conclusions:

• The HDMR approach provides a reliable classifier with the ability to explain gene and gene-gene interaction effects on classification.
• Enables identification of gene networks with dysregulated pairwise interactions, suitable for differential network analysis.
• Demonstrates effectiveness on both synthetic and real-world cancer datasets.