BEM: Mining Coregulation Patterns in Transcriptomics via Boolean Matrix Factorization.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Matrix factorization is crucial for analyzing transcriptomic data to understand gene coregulation.
• Existing methods lack clear bicluster structures and rely on linear assumptions, limiting their ability to capture complex patterns.

Purpose of the Study:

• To introduce a novel Boolean matrix factorization (BMF) algorithm, BEM, designed for improved analysis of transcriptomic data.
• To address limitations of current methods by providing clearer bicluster structures and capturing non-linear coregulation patterns.

Main Methods:

• Developed a new algorithm for Boolean matrix factorization (BMF) using expectation maximization (BEM).
• BEM is designed to align with molecular mechanisms of transcriptomic coregulation and offers scalability for large datasets.
• Utilized synthetic and real-world transcriptomic datasets (bulk RNA-seq, single-cell RNA-seq, spatial transcriptomics) for validation.

Main Results:

• BEM demonstrated superior performance in reconstruction error compared to other BMF methods in synthetic experiments.
• BEM successfully extracted biologically relevant coregulation patterns, correlating with disease subtypes, cell types, and spatial anatomy in real-world data.
• The algorithm is scalable to matrices with over 100 million data points.

Conclusions:

• BEM offers a robust and scalable approach for analyzing transcriptomic data, revealing intricate coregulation patterns.
• The algorithm's alignment with molecular mechanisms and its applicability across diverse transcriptomic data types make it a valuable tool for biological research.
• BEM provides clearer bicluster structures, enhancing the interpretability of transcriptomic analyses.