Multimodal Single-Cell Integration Across Time, Individuals, and Batches

In the past decade, the advent of single-cell genomics has enabled the measurement of DNA, RNA, and proteins in single cells. These technologies allow the study of biology at an unprecedented scale and resolution. Among the outcomes have been detailed maps of early human embryonic development, the discovery of new disease-associated cell types, and cell-targeted therapeutic interventions. Moreover, with recent advances in experimental techniques it is now possible to measure multiple genomic modalities in the same cell. While multimodal single-cell data is increasingly available, data analysis methods are still scarce. Due to the small volume of a single cell, measurements are sparse and noisy. Differences in molecular sampling depths between cells (sequencing depth) and technical effects from handling cells in batches (batch effects) can often overwhelm biological differences. When analyzing multimodal data, one must account for different feature spaces, as well as shared and unique variation between modalities and between batches. Furthermore, current pipelines for single-cell data analysis treat cells as static snapshots, even when there is an underlying dynamical biological process. Accounting for temporal dynamics alongside state changes over time is an open challenge in single-cell data science. Generally, genetic information flows from DNA to RNA to proteins. DNA must be accessible (ATAC data) to produce RNA (GEX data), and RNA in turn is used as a template to produce protein (ADT data). These processes are regulated by feedback: for example, a protein may bind DNA to prevent the production of more RNA. This genetic regulation is the foundation for dynamic cellular processes that allow organisms to develop and adapt to changing environments. In single-cell data science, dynamic processes have been modeled by so-called pseudotime algorithms that capture the progression of the biological process. Yet, generalizing these algorithms to account for both pseudotime and real time is still an open problem.