Exhaustive biclustering driven by self-learning evolutionary approach for biomedical data

Abstract

Background and Objective: Biclustering is a key data analysis technique that identifies submatrices with coherent patterns, widely applied in biomedical fields such as gene co-expression analysis. Despite its importance, in the context of evolutionary algorithms, traditional partial representations in biclustering algorithms face significant limitations, such as redundancy and limited adaptability to domain-specific objectives. This study aims to overcome these challenges by introducing MOEBA-BIO, a new evolutionary biclustering framework for biomedical data. Methods: MOEBA-BIO is designed as a flexible framework based on the evolutionary metaheuristics scheme. It includes a self-configurator that dynamically adjusts the algorithm’s objectives and parameters based on contextual domain knowledge. The framework employs a complete representation, enabling the integration of new domain-specific objectives and the self-determination of the number of biclusters, addressing the limitations of traditional representations. The source code is available through the following git repository: https://github.com/AdrianSeguraOrtiz/MOEBA-BIO. Results: Experimental results demonstrate that MOEBA-BIO overcomes the limitations of classical partial representations. Furthermore, its application to simulated and real-world gene expression datasets highlights its ability to specialize in specific biological domains, improving accuracy and functional enrichment of biclusters compared to other state-of-the-art techniques. Conclusions: MOEBA-BIO represents a significant advancement in biclustering applied to bioinformatics. Its innovative framework, combining adaptability, self-configuration, and integration of domain-specific objectives, addresses the main limitations of traditional methods and offers robust solutions for complex biomedical datasets.