Recovering Single-cell Heterogeneity Through Information-based Dimensionality Reduction

Dimensionality reduction is crucial to summarizing the complex transcriptomic landscape of single cell datasets for downstream analyses. However, current dimensionality reduction approaches favor large cellular populations defined by many genes, at the expense of smaller and more subtly-defined populations. Here, we present surprisal component analysis (SCA), a technique that leverages the information-theoretic notion of surprisal for dimensionality reduction, and demonstrate its ability to improve the representation of clinically important populations that are indistinguishable using existing pipelines. For example, in cytotoxic T-cell data, SCA cleanly separates the gamma-delta and MAIT cell subpopulations, which are not detectable via PCA, ICA, scVI, or a wide array of specialized rare cell recovery tools. We also show that, when used instead of PCA, SCA improves downstream imputation to more accurately restore mRNA dropouts and recover important gene-gene relationships. SCA’s information-theoretic paradigm opens the door to more meaningful signal extraction, with broad applications to the study of complex biological tissues in health and disease.