摘要：SummaryWe developed Miscell, a self-supervised learning approach with deep neural network as latent feature encoder for mining information from single-cell transcriptomes. We demonstrated the capability of Miscell with canonical single-cell analysis tasks including delineation of single-cell clusters and identification of cluster-specific marker genes. We evaluated Miscell along with three state-of-the-art methods on three heterogeneous datasets. Miscell achieved at least comparable or better performance than the other methods by significant margin on a variety of clustering metrics such as adjusted rand index, normalized mutual information, andV-measure score. Miscell can identify cell-type specific markers by quantifying the influence of genes on cell clusters via deep learning approach.Graphical abstractDisplay OmittedHighlights•We presented a deep learning approach Miscell to dissecting single-cell transcriptomes•Miscell achieved high performance on canonical single-cell analysis tasks•Miscell can transfer knowledge learned from single-cell transcriptomes to bulk tumorsBiological sciences; Neural networks; Transcriptomics
