MOSCATO: A Supervised Approach for Analyzing Multi-Omic Single-Cell Data

Advancements in genomic sequencing continually improve personalized medicine in complex diseases. Recent breakthroughs generate multiple types of signatures (or multi-omics) from each cell, producing different data ‘omic’ types per single-cell experiment. We introduce MOSCATO, a technique for selecting features across multi-omic single-cell datasets that relate to clinical outcomes. For example, we leverage penalization concepts often used in multi-omic network analytics to accommodate the high-dimensionality where multiple-testing is likely underpowered. We organize the data into multi-dimensional tensors where the dimensions correspond to the different ‘omic’ types. Using the outcome and the single-cell tensors, we perform regularized tensor regression to return a variable set for each ‘omic’ type that forms the clinically-associated network. Robustness is assessed over simulations based on available single-cell simulation methods. Real data comparing healthy subjects versus subjects with leukemia is also considered in order to identify genes associated with the disease. The flexibility of our approach enables future extensions on distributional assumptions and covariate adjustments. This algorithm may identify clinically-relevant genetic patterns on a cellular-level that span multiple layers of sequencing data and ultimately inform highly precise therapeutic targets in complex diseases. Code to perform MOSCATO and replicate the real data application is publicly available on GitHub at https://github.com/lorinmil/MOSCATO and https://github.com/lorinmil/MOSCATOLeukemiaExample .