BARTMAP Biclustering for High-Speed Gene Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current clustering algorithms face challenges in efficiently performing biclustering on large data sets, particularly in biological or biomedical research, due to high computational complexity and the need for simultaneous clustering of genes and samples without prior information, which limits their ability to establish accurate local relations.
Innovation Solution
A neural-based classifier, specifically a modified Fuzzy Adaptive Resonance Theory MAP (BARTMAP) system, is used to perform biclustering by creating clusters in both gene and sample subspaces, allowing for unsupervised learning and automatic feature selection, thereby establishing relationships between gene and sample clusters efficiently.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional clustering algorithms are used to perform biclustering on large data sets, then the computational complexity is high and processing speed is slow, but the algorithms can still handle the data without prior information
Solution Approach 1:
The patent segments the biclustering problem into two separate clustering tasks: clustering genes based on expression patterns and clustering samples based on gene expressions. This is achieved through two independent clustering modules that operate separately, reducing the overall computational complexity compared to simultaneous biclustering while maintaining the ability to discover local relationships between genes and samples.
Solution Approach 2:
The patent introduces a new dimension of analysis by creating separate clustering spaces for genes and samples. Instead of operating in a single high-dimensional space, the system creates two distinct clustering dimensions that can be independently processed and then integrated, effectively reducing computational complexity while preserving biclustering capabilities.
2Measurement precision
If simultaneous clustering of genes and samples is performed without prior information, then accurate local relations can be established, but the computational burden increases significantly
Solution Approach 1:
The patent divides the simultaneous clustering task into sequential independent clustering operations. First, genes are clustered based on their expression patterns across samples. Then, samples are clustered based on their gene expression profiles. This segmentation maintains clustering accuracy by allowing each module to focus on its specific dimension while reducing computational burden through independent processing.
Solution Approach 2:
The patent performs preliminary clustering of genes before clustering samples. The results from the first clustering operation are used to inform and guide the second clustering operation. This preliminary action allows the system to establish accurate local relations incrementally, reducing the overall computational burden compared to attempting simultaneous clustering of all data points.
3Measurement precision
If iterative clustering is performed to build local relationships, then clustering quality improves, but processing time increases
Solution Approach 1:
The patent segments the iterative process into two independent iterative clustering loops: one for gene clustering and one for sample clustering. Each loop iterates independently to improve its respective clustering quality, and the iterations are fewer and more efficient than simultaneous biclustering iterations. This segmentation maintains high clustering quality while reducing total processing time.
Data Source
AI summary
Methods and systems for improved unsupervised learning are described. The unsupervised learning can consist of biclustering a data set, e.g., by biclustering subsets of the entire data set. In an example, the biclustering does not include feeding know and proven results into the biclustering methodology or system. A hierarchical approach can be used that feeds proven clusters back into the biclustering methodology or system as the input. Data that does not cluster may be discarded. Thus, a very large unknown data set can be acted on to learn about the data. The system is also amenable to parallelization.