Clustering gene expression data with a penalized graph-based metric

Area of Science:

• Bioinformatics
• Computational Biology
• Data Science

Background:

• Clustering microarray datasets is crucial for -omic sciences.
• Handling manifold structures in high-dimensional data presents a significant clustering challenge.
• Gene expression datasets often exhibit complex, non-compact data structures.

Purpose of the Study:

• Introduce a novel distance metric for clustering high-dimensional data with manifold structures.
• Enhance the performance of existing clustering algorithms for -omic data analysis.
• Provide a user-friendly tool for accurate and interpretable data clustering.

Main Methods:

• Developed the Penalized k-Nearest-Neighbor-Graph (PKNNG) based metric.
• PKNNG metric employs a two-step graph construction with penalized edge weights.
• Evaluated the metric's performance on public gene expression datasets and simulated data.

Main Results:

• The PKNNG metric demonstrated promising clustering results across various datasets.
• It effectively handles data with manifold structures, improving clustering accuracy.
• PKNNG metric-based clustering achieved performance comparable to advanced algorithms.

Conclusions:

• The PKNNG metric offers a significant improvement for pairwise-distance based clustering methods.
• Researchers can integrate PKNNG with existing algorithms like hierarchical clustering for better dendrograms.
• This approach enhances interpretability and accuracy in high-dimensional data analysis without requiring new methods.