Boolean Reasoning-Based Biclustering for Shifting Pattern Extraction

TL;DR

This paper introduces a Boolean reasoning-based biclustering method to effectively identify shifting patterns in data, especially useful for analyzing fluctuating biological data like gene expression, with promising experimental results.

Contribution

It extends Boolean reasoning to detect more general shifting patterns in biclustering, including the mathematical foundation and an effective methodology.

Findings

Successfully identifies biclusters with shifting patterns in real data

Achieves low mean squared residue (MSR) scores close to zero

Demonstrates potential for biological data analysis

Abstract

Biclustering is a powerful approach to search for patterns in data, as it can be driven by a function that measures the quality of diverse types of patterns of interest. However, due to its computational complexity, the exploration of the search space is usually guided by an algorithmic strategy, sometimes introducing random factors that simplify the computational cost (e.g. greedy search or evolutionary computation). Shifting patterns are specially interesting as they account constant fluctuations in data, i.e. they capture situations in which all the values in the pattern move up or down for one dimension maintaining the range amplitude for all the dimensions. This behaviour is very common in nature, e.g. in the analysis of gene expression data, where a subset of genes might go up or down for a subset of patients or experimental conditions, identifying functionally coherent categories. Boolean reasoning was recently revealed as an appropriate methodology to address the search for constant biclusters. In this work, this direction is extended to search for more general biclusters that include shifting patterns. The mathematical foundations are described in order to associate Boolean concepts with shifting patterns, and the methodology is presented to show that the induction of shifting patterns by means of Boolean reasoning is due to the ability of finding all inclusion--maximal {\delta}-shifting patterns. Experiments with a real dataset show the potential of our approach at finding biclusters with {\delta}-shifting patterns, which have been evaluated with the mean squared residue (MSR), providing an excellent performance at finding results very close to zero.