LAUSR

To the Editor — The proliferation of single-cell datasets from mass cytometry has driven the need for an open registry capable of hosting data from different laboratories, correcting for batch effects, and interrogating a broad array of markers that can aid in inferring the underlying mechanistic significance of cell subsets identified under certain conditions. Here, we present EPIC (https://epicimmuneatlas. org), a web-based analytical and discovery platform for hosting and analyzing mass cytometry data from immune cells in a standardized manner. Upon request for access, a unique user identifier and password will be assigned to each investigator. A file-based user and password authentication system is used and passwords are encrypted for further protection. The platform provides a pipeline for the construction of immune maps that integrate thematically grouped mass cytometry data with clustering outputs, cell-type annotations, demographic metadata and predictive models to facilitate multidimensional data analysis (Supplementary Fig. 1b). EPIC functionality is tailored to the needs and perspective of the user (Fig. 1). Our immune map construction pipeline is complemented by a web-browser-based data mining app dubbed ‘Sci-Atlas Miner’ (Single-Cell Immune Atlas Miner tool; see Supplementary Methods) that enables exploration of EPIC and of uploaded mass cytometry data (Fig. 1). Sci-Atlas Miner provides a multilayer cell-type classification structure for data presentation at different cartographical levels of detail, from a major cell lineage to specific cell subsets. Using Sci-Atlas Miner, users can explore the whole architecture of the human immunome regarding distribution and representation of immune subsets and functions along and within any age gradient. Sci-Atlas Miner also enables supervised analyses of specific cell types (Fig. 1a). We implemented our system using the R statistical programming language, including the Shiny package for creating web-based data analytics applications. This open-source environment enables us to constantly review and integrate state-of-the-art data analytics methods in the areas of machine learning and data visualization, taking into account any contribution from the scientific community. For benchmarking, users can upload their dataset to EPIC for comparison and mapping with an internally generated and curated (177 individuals) single-cell high-dimensional dataset that represents the healthy human immunome across a spread of ages (Supplementary Fig. 1b; see Supplementary Methods for details on data upload). Data are obtained from two antibody panels (A and B), which have been carefully selected and validated to encompass a total of 63 individual non-redundant functional and phenotypical variables per cell (Supplementary Fig. 1a). The provision of preannotated cell types (Fig. 1a,b) enables a user not only to carry out their own data mining experiments, but also to use the dataset as a reference tool, providing a high-quality standardized healthy reference to explore how diseases or other events, such as vaccination or therapeutic intervention, may alter the architecture of the healthy immunome. Users can also upload their own datasets generated with antibody panels that are only partially matched to EPIC for reclustering with the user-selected dataset from EPIC database (Fig. 1d). This unsupervised approach (EPIC Discovery Tool) (Fig. 1d) has two important implications: first, it allows a high degree of freedom in the design of specific panels to address research questions, while still fully exploiting EPIC analytical pipeline and reference data; and second, it enables discoveries of rare or even unknown cell subsets, which could not be captured in the supervised analysis provided by the Sci-Atlas Miner tool. Overall, EPIC provides a high-quality, standardized dataset encompassing a broad span of ages. Data can be mined as desired in both a supervised or unsupervised manner. EPIC enables the healthy immune landscape to be holistically depicted across two planes as age-related changes and cross-sectional age-specific immune profiles in a three-dimensional space (Supplementary Fig. 2). The x, y, and z axes represent the age, immune cell subset and cell frequency (proportion of total CD45+ peripheral blood mononuclear cells (PBMC)), respectively. In the immune landscape constructed with the panel A dataset, we present two examples of the age-related changes of two immune cell subsets with prominent peaks at the younger ages: naive CD4+ T cells not expressing any bioactive marker and naive IL8+ CD4+ T cells (Supplementary Fig. 2a). The potential of this approach is self-evident as the developmental gradient of any immune cell subset can be contextualized within the entire immune landscape by following visually the changing patterns across the ages while maintaining the cross-sectional perspective of the immune profile at a specific age (Supplementary Fig. 2a). Of note, EPIC outputs are consistent with data from published literature, including, for instance, a pattern of reduction of IL8+ CD4+ T cells and naive CD4+ T cells with age1,2. EPIC has the inherent functionality for any cell-type developmental trajectories to be selected for comparison either across the entire available age spectrum or aggregated age ranges (Supplementary Fig. 3). Distinct developmental undulations shaping the immune landscape are also evident in the panel B dataset (Fig. 2a and Supplementary Fig. 2b). As an example, one can easily observe that the age-related reduction in the CD4+ CD38+ CD62L+ T-cell subset is paired with a reciprocal increase in the CD4+ CD38– CD62L– T-cell subset (Fig. 2a,b,d). CD38 is highly expressed on neonatal T cells; its frequency and reduction with age is consistent with current knowledge2–4. When the immune architecture of cord blood and newborn is compared directly to that of adults (18 to 60 years old), this reduction in the CD4+ CD38+ CD62L+ T-cell subset is observed in the older age group (Fig. 2b). With EPIC, we can inspect the CD38 expression pattern across different cell clusters. Indeed, CD38 distribution, although most prominent, is not restricted to CD4+ T cells but is present on CD8+ T cells, CD19+ B cells and CD3– CD56+ natural killer (NK) cells (Fig. 2c). With a targeted analysis, which is available in the pipeline, the data can be explored at finer granularity with the selection of various subsets for comparison. Key summary statistics, such as the sample size, cell frequency and data spread (as