Comparison of Clustering Metrics and Unsupervised Learning Algorithms on Genome-Wide Gene Expression Level Data

Sonia M. Leach, Lawrence Hunter, David LandsmanAAAI 1999

Introduction

The advent of high-throughput DNA sequencing has presented biologists with a significant "data firehose" challenge. While it is now possible to monitor the expression levels of thousands of genes simultaneously, the functional roles of many of these genes remain unknown. A primary strategy for addressing this knowledge gap is the principle of "guilt by association." This approach posits that genes exhibiting similar expression patterns across various conditions or time points are likely to participate in the same biological pathways or functional modules.

However, identifying these patterns is not straightforward. The biological insights derived from gene expression data are heavily dependent on the choice of similarity metrics and clustering algorithms. Historically, research has often relied on a single combination of these tools without a systematic evaluation of their relative performance. This paper provides a comparative analysis of different mathematical "yardsticks" and unsupervised learning algorithms to determine which combinations most effectively group genes with shared biological functions.

Methodology: Similarity Metrics and Clustering Algorithms

The study evaluates a matrix of approaches to identify the most reliable methods for gene expression analysis, focusing on how different mathematical models interpret the relationship between gene trajectories.

Similarity Metrics

The choice of metric determines how the "distance" or similarity between two gene expression profiles is calculated:

  • Euclidean Distance: A standard geometric distance. While intuitive, it is often insufficient for biological data as it fails to capture negative correlations (e.g., when one gene's expression increases while another's decreases).
  • Correlation Metric: Specifically employed to capture both positive and negative linear relationships between gene trajectories.
  • Mutual Information (MI): A more sophisticated metric designed to detect non-linear dependencies. MI is particularly valuable for identifying complex biological relationships where genes may be positively correlated during one phase of a process but negatively related in another.

Clustering Algorithms

The authors compare three distinct unsupervised learning paradigms:

  • k-means Clustering: An efficient algorithm that partitions data into a pre-specified number of clusters ($k$). Its primary drawback is the requirement for the user to define $k$ in advance, which can introduce human bias into the discovery process.
  • Hierarchical Agglomerative Clustering: This method builds a tree-like structure (dendrogram) of relationships. While it does not require a pre-set number of clusters, it often necessitates manual intervention and expert biological knowledge to determine where to "cut" the tree to define functional groups.
  • AutoClass: A Bayesian approach that models clusters as Gaussian distributions. AutoClass is stochastic and can automatically determine the most likely number of clusters based on the data. However, this mathematical rigor comes at a higher computational cost compared to simpler algorithms.

Data Pre-processing and Experimental Setup

Genomic data is frequently "noisy" and contains missing values due to measurement errors. The authors propose two primary strategies for handling these issues: weighting results based on the number of valid data points and using linear interpolation to fill gaps. The study emphasizes that the method of handling "missingness" is as critical to the final result as the choice of clustering algorithm itself.

The evaluation was conducted across three diverse biological contexts:

  1. Rat Spinal Cord Development: A time-course dataset focusing on neural development.
  2. Human Cell Growth Model: A time-course dataset examining cellular proliferation.
  3. Yeast (S. cerevisiae): An aggregated dataset spanning multiple experimental conditions.

Key Insights and Results

The comparative analysis revealed several critical trade-offs in unsupervised learning for genomics:

  • The Dynamic Nature of Correlation: Standard metrics often assume a constant relationship between genes. However, biological systems are dynamic; genes might switch from being partners to competitors depending on the phase of a biological process. The authors highlight that Mutual Information is better equipped to handle these non-linear nuances.
  • Automation vs. Manual Tuning: There is a clear trade-off between the speed of hierarchical clustering and the objectivity of Bayesian methods. While AutoClass reduces human bias by automatically determining cluster numbers, its complexity can make it prohibitively slow for very large genome-wide datasets.
  • Metric Sensitivity: The findings suggest that Euclidean distance is generally less effective than correlation-based or information-theoretic metrics because it ignores the directionality of gene regulation (positive vs. negative correlation).

Conclusion and Future Work

The study establishes a systematic framework for evaluating unsupervised learning in the context of functional genomics. While sophisticated methods like AutoClass offer a robust, automated alternative to traditional clustering, their scalability remains a challenge for the scale of modern genomic data.

The ultimate validation of any clustering method remains its biological relevance—whether the clusters match known biological functional classes. Future research must continue to test the sensitivity of these algorithms to noise and missing data, ensuring that the discovered gene groups represent genuine biological pathways rather than mathematical artifacts.

Summary updated

BibTeX
@inproceedings{leach1999comparison,
  author = {Sonia M. Leach and Lawrence Hunter and David Landsman},
  title = {{Comparison of Clustering Metrics and Unsupervised Learning Algorithms on Genome-Wide Gene Expression Level Data}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1999},
}
Figure