Keywords: computational biology, graph signal processing, genomics
TL;DR: A novel graph signal processing framework for quantifying the effects of experimental perturbations in single cell biomedical data.
Abstract: Single-cell RNA-sequencing (scRNA-seq) is a powerful tool for analyzing biological systems. However, due to biological and technical noise, quantifying the effects of multiple experimental conditions presents an analytical challenge. To overcome this challenge, we developed MELD: Manifold Enhancement of Latent Dimensions. MELD leverages tools from graph signal processing to learn a latent dimension within the data scoring the prototypicality of each datapoint with respect to experimental or control conditions. We call this dimension the Enhanced Experimental Signal (EES). MELD learns the EES by filtering the noisy categorical experimental label in the graph frequency domain to recover a smooth signal with continuous values. This method can be used to identify signature genes that vary between conditions and identify which cell types are most affected by a given perturbation. We demonstrate the advantages of MELD analysis in two biological datasets, including T-cell activation in response to antibody-coated beads and treatment of human pancreatic islet cells with interferon gamma.