Intricacies of single-cell multi-omics data integration

Identifying cell-to-cell correspondences between unpaired datasets from different single cell protocols promises to provide a more comprehensive view of cellular states. Integration of unpaired data from multiple modalities is more complicated than single-omics integration due to a lack of feature correspondence across modalities and ground truth information about biological differences between modalities. Retention of biological variation during multi-omic data integration has been insufficiently addressed to date, but is essential to leverage complementary information from different omics layers. Ground truth data can now be provided by new paired multi-omics assays. This will inform robust associations between features of different modalities and reveal modality-specific biological patterns that may also help to improve methods for multimodal integration of unpaired data. A wealth of single-cell protocols makes it possible to characterize different molecular layers at unprecedented resolution. Integrating the resulting multimodal single-cell data to find cell-to-cell correspondences remains a challenge. We argue that data integration needs to happen at a meaningful biological level of abstraction and that it is necessary to consider the inherent discrepancies between modalities to strike a balance between biological discovery and noise removal. A survey of current methods reveals that a distinction between technical and biological origins of presumed unwanted variation between datasets is not yet commonly considered. The increasing availability of paired multimodal data will aid the development of improved methods by providing a ground truth on cell-to-cell matches. A wealth of single-cell protocols makes it possible to characterize different molecular layers at unprecedented resolution. Integrating the resulting multimodal single-cell data to find cell-to-cell correspondences remains a challenge. We argue that data integration needs to happen at a meaningful biological level of abstraction and that it is necessary to consider the inherent discrepancies between modalities to strike a balance between biological discovery and noise removal. A survey of current methods reveals that a distinction between technical and biological origins of presumed unwanted variation between datasets is not yet commonly considered. The increasing availability of paired multimodal data will aid the development of improved methods by providing a ground truth on cell-to-cell matches. a low-dimensional representation of the high-dimensional data. a quantifiable characteristic of a cell. For example, in the context of scRNA-seq, the expression level of each gene is a feature. For scATAC-seq, the features are the accessibilities of defined genomic regions. features from two or more datasets refer to the same entities (e.g., genes). a matrix that aggregates quantitative genome-level data (e.g., chromatin accessibility or DNA methylation data) to the gene level. process in which the information encoded in genes is transformed into functional gene products, such as proteins or functional RNA molecules. In the context of single-cell analysis it often refers to steady-state mRNA levels in the cell measured by scRNA-seq (i.e., an intermediate step of the gene expression process). a parameter that specifies a part of the method setting and often needs to be selected by the user. combining data from different sources into a unified view. transferring cell or cluster labels to a different dataset based on similarities to the source dataset. a topological space that preserves the neighborhood structure of a dataset. A manifold can be used to represent high-dimensional biological data in a lower-dimensional space that is easier to analyze while maintaining the original dataset information. a mode in which the cell exists (i.e., gene expression space or chromatin accessibility space). The term modality is often used to refer to different data types that assay these very modes. a specific aspect of the cell’s molecular biology that is represented by a set of biomolecules or their state. Examples of molecular layers include the chromatin state, gene expression levels, and protein levels. involving information from two or more modalities. Also see modality. data where different modalities are measured in the same single cell. a linear dimensionality reduction technique that reduces the number of features of a dataset while preserving most of the variation in the original dataset. a training strategy for machine learning where at least a small amount of labeled data is required. a function that provides a similarity measure between vectors (i.e., gene expression vectors for two cells). methods that profile the entire gene expression profile of individual cells. methods for the genome-wide profiling of open chromatin regions in individual cells. a statistic that describes feature importance for a specific sample (i.e., how important a particular open chromatin region is for a specific cell). a nonlinear dimensionality reduction technique that reduces the number of features of a dataset while preserving the similarity between data points from the original dataset. involving information from a single modality. Also see modality. data where different modalities are measured in distinct cells. a training strategy for machine learning that only uses unlabeled data.