Enhancing CITE-seq analysis: a comprehensive pipeline leveraging machine learning techniques 3817

Description Cellular Indexing of Transcriptomes and Epitopes by Sequencing (CITE-seq) technology enables the concurrent assessment of gene expression and cell surface proteins in individual cells using oligonucleotide-tagged antibodies, providing crucial insights into cell identity, state, functions, and signaling pathways. However, challenges such as sample demultiplexing, doublet detection, batch effects, and discrepancies between surface protein expression and mRNA levels complicate the downstream analysis. In this study, we propose a comprehensive pipeline that incorporates various advanced techniques for quality control, including filtering, doublet and multiplet identification, sample integration, and thresholding the levels of protein marker expression. Additionally, we utilize the random forest algorithm to identify a list of top informative features and apply non-negative matrix factorization for dimensionality reduction to improve clustering outcomes. Our approach was validated on two distinct datasets: human bone marrow and blood samples from patients with and without melanoma and mouse-derived samples from melanoma tumors, lymph nodes, spleen, blood, and bone marrow. The results demonstrate a significant improvement in downstream analysis, characterized by distinct separation of immune cell populations on the Uniform Manifold Approximation and Projection (UMAP) and cluster identification enabling effective subclustering and group comparison of immune cells.