Discovering What Dimensionality Reduction Really Tells Us About RNA-Seq Data

Abstract

Biology is being inundated by noisy, high-dimensional data to an extent never before experienced. Dimensionality reduction techniques such as principal component analysis (PCA) are common approaches for dealing with this onslaught. Though these unsupervised techniques can help uncover interesting structure in high-dimensional data they give little insight into the biological and technical considerations that might explain the uncovered structure. Here we introduce a hybrid approach—component selection using mutual information (CSUMI)—that uses a mutual information—based statistic to reinterpret the results of PCA in a biologically meaningful way. We apply CSUMI to RNA-seq data from GTEx. Our hybrid approach enables us to unveil the previously hidden relationship between principal components (PCs) and the underlying biological and technical sources of variation across samples. In particular, we look at how tissue type affects PCs beyond the first two, allowing us to devise a principled way of choosing which PCs to consider when exploring the data. We further apply our method to RNA-seq data taken from the brain and show that some of the most biologically informative PCs are higher-dimensional PCs; for instance, PC 5 can differentiate the basal ganglia from other tissues. We also use CSUMI to explore how technical artifacts affect the global structure of the data, validating previous results and demonstrating how our method can be viewed as a verification framework for detecting undiscovered biases in emerging technologies. Finally we compare CSUMI to two correlation-based approaches, showing ours outperforms both. A python implementation is available online on the CSUMI website.