M3C

M3C applies a Monte Carlo–derived reference testing framework to consensus clustering to perform statistically rigorous unsupervised class discovery in genome-wide datasets by simulating null distributions of stability-based metrics to correct upward bias in the Monti consensus clustering algorithm and objectively select the number of clusters (K).

Key Features:

• Monte Carlo null reference: Generates a Monte Carlo–derived null reference system to model stability-score distributions under the null hypothesis of no underlying structure (K = 1).
• Null distribution simulation per K: Simulates null distributions of consensus clustering stability scores for each candidate K to enable comparative assessment.
• Statistical testing of stability: Compares observed stability metrics against simulated nulls to statistically test whether observed structure exceeds noise.
• Bias correction and false-positive control: Corrects the upward bias of standard consensus clustering and reduces inflated cluster counts and false-positive cluster detection.
• Compatibility with Monti algorithm: Operates on stability-based metrics derived from consensus clustering approaches such as the Monti consensus clustering algorithm.
• clusterlab simulation tool: Includes the companion clusterlab tool for generating multivariate Gaussian clusters for simulation-based validation.
• Benchmarking evidence: Demonstrates improved accuracy in selecting K and clearer separation of real versus noise-driven structure in simulated datasets and TCGA transcriptomics.

Scientific Applications:

• Unsupervised class discovery: Identification of reproducible classes in genome-wide datasets using statistically evaluated cluster stability.
• Transcriptomics analysis: Application to large-scale transcriptomics datasets, including analyses of The Cancer Genome Atlas (TCGA).
• Patient stratification and precision medicine: Stratifying patients into molecular subgroups for downstream translational and clinical research.
• Methodological benchmarking: Benchmarking and methodological research for clustering algorithms using simulated multivariate Gaussian clusters.

Methodology:

For each candidate K, M3C simulates Monte Carlo null distributions of consensus clustering stability scores under the hypothesis K = 1, compares observed stability-based metrics (e.g., from the Monti consensus clustering algorithm) to these nulls for statistical testing, and uses the clusterlab tool to generate multivariate Gaussian clusters for simulation-based validation on simulated datasets and TCGA transcriptomics.