Assessing Dissimilarity Measures for Sample-Based Hierarchical Clustering of RNA Sequencing Data Using Plasmode...

Area of Science:

• Bioinformatics
• Computational Biology
• Statistical Genetics

Background:

• Hierarchical clustering is vital for analyzing high-throughput gene expression data.
• RNA sequencing (RNA-seq) data presents unique challenges for clustering due to discrete, over-dispersed, and skewed read counts, plus sequencing depth variations.
• Existing methods for assessing dissimilarity measures, like parametric simulations, may offer limited scope.

Purpose of the Study:

• To evaluate the adequacy of various dissimilarity measures for sample-based hierarchical clustering of RNA-seq data.
• To introduce plasmode datasets as a robust method for simulating realistic RNA-seq experimental conditions.
• To identify reliable dissimilarity measures that accurately reflect hierarchical structures in RNA-seq data.

Main Methods:

• Generation of plasmode datasets simulating realistic RNA-seq data characteristics and experimental conditions.
• Assessment of multiple dissimilarity measures, including Euclidean-based, Poisson-based, and rank correlation-based methods.
• Comparison of clustering results against expected hierarchical structures derived from plasmode data.

Main Results:

• Dissimilarity measures relying solely on normalization or standardization of Euclidean distance were unreliable.
• Poisson-based dissimilarity, rank correlation-based dissimilarity, and appropriate data transformations yielded dendrograms consistent with expected structures.
• Plasmode datasets generated from diverse RNA-seq experiments yielded consistent results, validating their utility.

Conclusions:

• Plasmode datasets provide a powerful and flexible approach for evaluating dissimilarity measures in RNA-seq hierarchical clustering.
• Effective dissimilarity measures for RNA-seq data include Poisson-based, rank correlation-based, and those employing appropriate data transformations.
• The choice of a specific measure may depend on software availability within a given analysis pipeline.