Bayesian Consensus Clustering

Employs a non-parametric Bayesian prior to automatically infer the optimal number of clusters (K) directly from the data. Unlike k-means, this eliminates the need for arbitrary pre-specification of patient subgroups. The Dirichlet Process allows model complexity to grow with the data, making it ideal for discovering novel disease subtypes in heterogeneous cancers where the true number of molecular strata is unknown.

Aggregates clustering results across Markov Chain Monte Carlo (MCMC) iterations into a probabilistic co-occurrence matrix. Each entry represents the posterior probability that two patients belong to the same cluster. This matrix serves as the foundational consensus structure, capturing the uncertainty of cluster assignments rather than forcing hard, deterministic boundaries that ignore borderline patient profiles.

Defines a separate likelihood function for each omics layer (e.g., Gaussian for mRNA, Multinomial for mutations) and multiplies them within a joint Bayesian hierarchical model. This principled probabilistic fusion respects the distinct statistical distributions of different data types, avoiding the information loss that occurs when naively concatenating normalized matrices from genomics, proteomics, and epigenomics.

Provides a full posterior distribution over cluster assignments, not just a point estimate. For each patient, the allocation probability vector quantifies the confidence of subtype membership. This allows clinicians to identify patients with ambiguous molecular profiles who may require further testing, directly supporting risk-stratified clinical decision-making with statistical rigor.

Incorporates Bayesian variable selection directly into the clustering process to identify which molecular features drive cluster separation. A spike-and-slab prior forces irrelevant genes or proteins to have zero weight, resulting in sparse, interpretable biomarker signatures. This prevents the consensus from being diluted by noise in high-dimensional omics datasets where features vastly outnumber samples.

Utilizes Gibbs sampling, an MCMC algorithm, to iteratively sample cluster assignments from their conditional posterior distributions. The algorithm cycles through patients, updating their cluster labels based on current assignments of all others. Convergence diagnostics like the Gelman-Rubin statistic ensure the chain has explored the posterior sufficiently to produce a stable consensus partition.

A nonparametric Bayesian model that allows the number of clusters to grow with the data, avoiding the need to pre-specify k. It uses a Dirichlet process prior over infinite mixtures, making it a natural foundation for Bayesian consensus clustering where the true number of patient subtypes is unknown. The Chinese restaurant process metaphor describes how new patients are probabilistically assigned to existing or new clusters.

The computational engine for approximating the posterior distribution in Bayesian consensus clustering. Gibbs sampling and Metropolis-Hastings algorithms iteratively draw samples from the joint distribution of cluster assignments and model parameters. Convergence diagnostics like the Gelman-Rubin statistic are critical to ensure the chain has explored the full parameter space before summarizing the consensus partition.

A precursor step often paired with Bayesian clustering. Similarity Network Fusion (SNF) constructs a fused patient similarity network from mRNA, methylation, and miRNA data. This fused network can serve as a distance-based prior for Bayesian consensus clustering, guiding the model toward biologically coherent groupings rather than relying solely on raw feature spaces.

A frequentist consensus method developed by The Cancer Genome Atlas (TCGA) consortium. COCA applies hierarchical clustering to the cluster assignments from individual omics platforms. Bayesian consensus clustering generalizes this idea by treating the individual platform clusterings as noisy observations generated by a single latent consensus, providing full uncertainty quantification.

The standard metric for evaluating consensus stability. ARI measures the similarity between two data partitions, corrected for chance. In Bayesian consensus clustering, the posterior similarity matrix—the probability that two patients co-cluster—is thresholded to produce a point estimate, and ARI quantifies agreement across MCMC samples or against known clinical subtypes.

A complementary latent variable model that decomposes multi-omics variation into a sparse set of latent factors. While Bayesian consensus clustering focuses on discrete patient subgroups, MOFA identifies continuous axes of variation. The two are often used in tandem: MOFA factors can inform prior distributions for clustering, or clusters can be interpreted by examining their factor profiles.