Identifying clusters in genomics data by recursive partitioning.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Genomics studies often use clustering to group molecular data.
• Traditional methods like K-means and hierarchical clustering struggle to determine the optimal number of clusters.
• Identifying substructures within existing clusters is biologically significant but often overlooked.

Purpose of the Study:

• To introduce a novel algorithm, Partitioning Algorithm based on Recursive Thresholding (PART), for uncovering subgroups within data clusters.
• To address the challenge of outliers in high-dimensional genomics data that can obscure substructure.
• To provide a robust method for identifying both global and local structures in genomic datasets.

Main Methods:

• The Partitioning Algorithm based on Recursive Thresholding (PART) recursively identifies subgroups.
• The algorithm incorporates tentative cluster splitting to manage outliers, preventing premature termination of recursion.
• Performance is evaluated on simulated and real gene expression microarray data with known cluster structures.

Main Results:

• PART demonstrates superior performance over established global methods on simulated data, especially when subclusters are present with large or varying variances.
• On real gene expression datasets, PART combined with hierarchical clustering shows improved overall performance compared to global methods.
• The algorithm successfully isolates outliers, allowing for the detection of substructures that might otherwise be missed.

Conclusions:

• The PART algorithm is an effective tool for discovering hidden substructures within genomic data clusters.
• Its ability to handle outliers makes it particularly suitable for high-dimensional genomics data.
• The clusterGenomics R package provides accessible implementation of this advanced clustering technique.