Leveraging latent patterns in the study of living systems

Abstract

The development of high-throughput techniques to observe and perturb biological systems has led to remarkable progress in the last several decades. From the tremendous amounts of data being accumulated, new opportunities have emerged, including the possibility of finding latent patterns in high-dimensional variables that are reflective of underlying biological processes. While these methods have led to countless discoveries and innovations, it is clear there is much more we could learn by measuring and perturbing at far greater scales. Here, I advance methods to understand and utilize latent patterns in new types of high-dimensional data. I devise a method of analyzing networks of 'frequency interactions' in 16S/18S time series data, showing that these can be used to identify microbial communities and associated environmental factors.

Then, as part of a highly collaborative project, I show how latent patterns in single cell RNA-Seq can be used together with optimal transport analysis to identify cell types and cell type trajectories, regulatory pathways, and cell-cell interactions in a time-course of developmental reprogramming. I then step back to ask a fundamental question: how do we choose which observations and perturbations to make, and how many of each are necessary? I approach this question on the basis of the inherency of latent structure in biology, and on foundational mathematical results concerning the analysis of highly-structured data. I present the beginnings of a framework to formalize how random composite experiments can make biological discovery more efficient by leveraging latent patterns. I first show how to recover individual genomes using covariance patterns in a series of composite (meta-) genomic data.

I then describe how random composite measurements and compressed sensing can be used to make gene expression profiling more efficient. Finally, I apply this idea to in situ imaging transcriptomics, demonstrating how many individual gene images can be efficiently recovered from a small number of composite gene images.