Integrative network modeling of bulk and single-cell sequencing data to characterize multi-scale cell architecture

Abstract Single cell sequencing technology enabled us to identify the aberrant molecular alterations in diseased cells presenting altered signaling pathways and emergence of disease-associated cell populations, leading to multi- scale cell architectures in disease tissue micro-environment. However, drawbacks such as low sequencing depth per cell, expensive costs and loss of cross-cell signaling information have hindered its broader applicability to large-scale cohorts. On the contrary, clinically well-defined, multi-modal and high-depth bulk- based sequencing data are abundantly available in public domain, and can be utilized to assemble robust molecular models of the genetic diseases, and infer cell population abundances in the samples. Especially, network biology approaches have been effective for integrating large-scale and diverse biomedical datasets in complex human diseases, and dissect the disease mechanisms and novel therapeutic strategies. Thus, a systems approach to synergistically utilize these complementary aspects of bulk and single-cell sequencing data is urgently needed to construct the robust molecular models of disease mechanisms while addressing the multi-scale nature of cell architectures in diseased tissues. Firstly, we will systematically investigate multi-scale cell architectures by developing a novel unsupervised cell clustering approach, single-cell recursive multi-scale clustering via local embedding (scRECIEM). Within scRECIEM, a novel cell-cell network construction algorithm will be developed by embedding each cell with its nearest neighboring cells on topological sphere, and yield computation complexity that linearly scales with the number of cells when parallelized. This will be accompanied by a top- down divisive clustering approach that adaptively utilizes informative features at each split, which is guided by network compactness measure, υ(α). These will identify a hierarchy of cell clusters captured at different resolutions. Secondly, we will develop integrative multi-scale network analysis (iMUSNET) framework to construct data-driven and mechanistic network models of disease etiology by utilizing the context-matched bulk samples. Within iMUSNET, the context-matched pairs of bulk and single-cell cohorts will be systematically collected, and we will construct multi-scale gene interaction networks capturing diverse co-expressed modules at different resolutions. These gene modules will be tested for enrichments with a compendium of clinico- genomic gene signatures curated within the bulk cohort. Key driver analysis will systematically look for potential up-stream regulators of the clinic-genomic signatures by leveraging the network model topology. Further, we will infer abundances of the context-matched single-cell clusters with high accuracy by utilizing the scRECITE-inferred cell phylogeny, and these will inform relevant disease associated cell populations in the bulk cohort. Overall, iMUSNET will generate a number testable hypotheses as potential regulators and subnetworks underlying the disease of interest.