Introduction
Analyzing microarray data presents a significant challenge because genes often exhibit correlated patterns across specific experiments while remaining independent in others. Traditional clustering methods typically fall into two categories: distance-based methods (like K-means) that require a pre-specified number of clusters ($K$), and flat model-based methods (like standard Dirichlet Process models) that treat clusters as independent entities.
However, biological systems are inherently hierarchical. A broad process like the "cell cycle" may contain specific sub-groups of genes handling distinct phases. Standard algorithms often fragment these related groups or fail to recognize their shared functionalities because they lack a mechanism to model these nested relationships.
Methodology
The authors propose a clustering algorithm based on the Hierarchical Dirichlet Process (HDP). Unlike a standard Dirichlet Process, the HDP uses a base measure that is itself a Dirichlet Process, creating a "clustering among clusters" effect. This allows the model to automatically determine the optimal number of clusters while capturing multi-level relationships.
The Chinese Restaurant Metaphor
To explain the HDP, the authors utilize the "Chinese Restaurant Metaphor." In this visualization, a restaurant contains an infinite number of tables (representing first-level clusters). Customers (data points) sit at these tables, and each table orders a specific dish (representing a second-level cluster). Because multiple tables can order the same dish, different clusters can share statistical properties, effectively modeling the hierarchy.
Technical Implementation
- Granular Modeling: Instead of treating a gene's expression across $M$ experiments as a single fixed vector, the model associates latent variables with individual expression points. This makes the approach more robust against noise in specific experiments.
- Inference: Since the posterior distribution is mathematically intractable, the authors developed a Gibbs sampling algorithm to estimate cluster memberships.
Experimental Evaluation
The model was evaluated on several synthetic and real-world biological datasets, including the Yeast Galactose, Yeast Cell Cycle, and Human Fibroblasts Serum datasets. The HDP model consistently achieved higher Rand Index (RI) and Silhouette Index (SI) scores compared to popular algorithms like SVM, K-means, and MCLUST.
- Yeast Galactose: HDP achieved an RI of 0.973, significantly outperforming MCLUST (0.903) and LDA (0.942).
- Yeast Cell Cycle: The HDP identified 10 clusters and discovered 126 genes with functional roles (such as protein synthesis and cell fate) that were overlooked in previous standard studies.
- Structural Discovery: In synthetic tests, the HDP correctly segmented networks into distinct, nested clusters, whereas traditional SVMs produced coarse results that missed sub-hierarchies.
Discussion and Limitations
By building hierarchy into the statistical prior, the HDP naturally consolidates fragmented data. It avoids the "overfitting vs. underfitting" struggle by scaling the number of clusters based on the inherent complexity of the data.
However, the reliance on Gibbs sampling introduces a significant computational cost, as Markov Chain Monte Carlo (MCMC) methods are slower to converge on very large datasets compared to simple heuristics. Future work may explore deeper hierarchical structures beyond two levels and the use of non-Gaussian distribution families to better represent diverse biological data types.
Summary updated