Gene Expression Cartography

Abstract

Massively multiplexed sequencing of RNA in individual cells is transforming basic and clinical life sciences[1-4]. In standard experiments, however, tissues must be first dissociated. Thus, crucial information about spatial relationships between cells and the tissue-wide expression patterns they confer is lost. This poses a fundamental problem for elucidating collective function of tissues and mechanisms of cell-to-cell communication[5,6]. Considerable efforts to overcome this challenge include experimental methods that are either technically challenging, or have limited resolution or throughput[5,7,8]. Existing computational approaches predict spatial positions by comparing each sequenced cell, independently, to an imaging-derived spatial gene expression database for that tissue and rely on prior knowledge of spatial expression patterns which often does not exist, or is difficult to construct[9,10]. Here, we explore a radically different idea. We postulate that spatially-proximal cells share more similar transcriptional profiles than cells farther apart. We validate this hypothesis for several complex biological systems. Consequently, we seek to find spatial arrangements of sequenced cells on tissue space which optimally preserve this principle. We cast this hard optimization problem as a generalized optimal transport problem for probabilistic embedding, for which we derived an efficient iterative algorithm. We successfully reconstruct the mammalian liver, intestinal epithelium, fly and zebrafish embryos, cerebellum sections and kidney. We use the reconstructed tissues to infer spatially informative genes directly from single-cell data. Our results demonstrate that we have identified a spatial expression organization principle in animal tissues which can be used to infer meaningful spatial position probabilities for individual cells. Our framework (“novoSpaRc”) is flexible, can incorporate prior spatial information and is compatible with any single-cell technology. We envision novoSpaRc to be valuable in collaborative efforts to characterize various tissues[11,12], and that additional principles underlying spatial organization of gene expression can be formulated and tested using our approach.