Semantic biclustering for finding local, interpretable and predictive expression patterns.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Analyzing gene expression data presents challenges in identifying local patterns of coherent gene expression across experimental conditions.
• Identifying these patterns is crucial for understanding underlying biological processes.
• Concise characterizations of genes and conditions are needed to enhance pattern interpretability.

Purpose of the Study:

• To propose semantic biclustering for detecting interpretable rectangular patterns in binary data matrices.
• To ensure detected biclusters are described by semantic annotations of both genes and samples.
• To explore two distinct strategies for finding interpretable biclusters.

Main Methods:

• Semantic biclustering integrates existing biclustering algorithms with semantic annotations.
• A machine learning approach based on rule and tree learning is also explored.
• Both strategies aim to find homogeneous submatrices with joint semantic descriptions.

Main Results:

• Experiments on Drosophila melanogaster gene expression datasets demonstrate the detection of compact biclusters.
• These biclusters possess semantic descriptions that generalize well to unseen data.
• One strategy emphasizes generalization but has higher description complexity; the other offers simpler descriptions.

Conclusions:

• The proposed semantic biclustering methods successfully detect interpretable biclusters with good generalization.
• The choice between strategies involves a trade-off between description complexity and generalization performance.
• These findings advance the analysis of gene expression data for biological insight.