Automated calibration of consensus weighted distance-based clustering approaches using sharp.

Area of Science:

• Bioinformatics
• Computational Biology
• Data Mining

Background:

• Consensus clustering combines algorithms and subsampling for stable cluster detection.
• Existing methods suggest consensus clustering surpasses native algorithms in performance.

Purpose of the Study:

• Extend consensus clustering to incorporate attribute weighting for pairwise distance calculations.
• Introduce a novel stability score, the 'sharp score,' for calibrating the number of clusters and regularization parameters.
• Evaluate the performance of weighted consensus clustering and sharp score calibration.

Main Methods:

• Implemented attribute weighting within regularized consensus clustering frameworks.
• Developed the sharp score for direct calculation from consensus clustering outputs.
• Conducted simulation studies comparing sharp score calibration with existing methods.
• Applied the enhanced consensus clustering approach to lung tissue gene expression data.

Main Results:

• Weighted consensus clustering demonstrated superior performance compared to unweighted approaches, especially with non-informative features.
• Calibration using the sharp score yielded better clustering performance than existing calibration scores.
• The method successfully identified distinct clusters corresponding to lung cancer subtypes in real-world data.
• The sharp score offers a computationally efficient alternative for parameter calibration.

Conclusions:

• Attribute weighting and sharp score calibration significantly enhance consensus clustering.
• This approach provides a robust and computationally efficient tool for cluster analysis.
• The method has practical applications in identifying subtypes within complex biological datasets, such as lung cancer.