Critical limitations of consensus clustering in class discovery

Area of Science:

• Genomics
• Computational Biology
• Statistical Analysis

Background:

• Consensus clustering (CC) is widely used for unsupervised class discovery in genomic studies.
• CC relies on pairwise consensus rates from repeated clustering runs to assess cluster stability and estimate the optimal number of clusters (K).
• The sensitivity and specificity of CC have not been systematically evaluated.

Purpose of the Study:

• To systematically assess the sensitivity and specificity of consensus clustering (CC).
• To introduce and evaluate a new metric, the proportion of ambiguously clustered pairs (PAC), for more reliable estimation of K.
• To propose an improved approach using realistic null distributions and PAC for robust cluster number estimation.

Main Methods:

• Simulations were performed on both randomly generated unimodal data and data with known structures.
• Evaluated the performance of common CC implementations in identifying the true K.
• Developed and applied the proportion of ambiguously clustered pairs (PAC) metric.
• Incorporated realistic null distributions based on gene-gene correlation structures.

Main Results:

• CC can identify apparent clusters in data lacking true structure, suggesting it may report chance partitions.
• Common CC implementations performed poorly in identifying the true K for structured data.
• The PAC metric demonstrated equal or superior reliability compared to other methods for inferring K in simulated data.
• The proposed approach using realistic null distributions and PAC improved K estimation accuracy.

Conclusions:

• Consensus clustering (CC) should be applied and interpreted with caution due to potential for false positives.
• The proportion of ambiguously clustered pairs (PAC) offers a more reliable alternative for estimating the optimal number of clusters (K).
• The developed approach enhances the accuracy of cluster number estimation in genomic studies.