Identifying Cell Types in Single-Cell Multimodal Omics Data via Joint Embedding Learning

Emerging single-cell technologies profile different modalities of data in the same cell, providing opportunities to study cellular population and cell development at a res-olution that was previously inaccessible. The first and most fundamental step in analyzing single-cell multimodal data is the identification of the cell types in the data using clustering analysis and classification. However, combining different data modalities for the classification task in multimodal data remains a computational challenge. We propose an approach for identifying cell types in multimodal omics data via joint dimensionality reduction. We first introduce a general framework that extends loss based dimensionality reduction methods such as nonnegative matrix factorization and UMAP to multimodal omics data. Our approach can learn the relative contribution of each modality to a concise representation of cellular identity that enhances discriminative features and decreases the effect of noisy features. The precise representation of the multimodal data in a low dimensional space improves the predictivity of classification methods. In our experiments using both synthetic and real data, we show that our framework produces unified embeddings that agree with known cell types and allows the predictive algorithms to annotate the cell types more accurately than state-of-the-art classification methods.