Digital cell quantification identifies global immune cell dynamics during influenza infection

Abstract

Hundreds of immune cell types work in coordination to maintain tissue homeostasis. Upon infection, dramatic changes occur with the localization, migration, and proliferation of the immune cells to first alert the body of the danger, confine it to limit spreading, and finally extinguish the threat and bring the tissue back to homeostasis. Since current technologies can follow the dynamics of only a limited number of cell types, we have yet to grasp the full complexity of global in vivo cell dynamics in normal developmental processes and disease. Here, we devise a computational method, digital cell quantification (DCQ), which combines genome-wide gene expression data with an immune cell compendium to infer in vivo changes in the quantities of 213 immune cell subpopulations. DCQ was applied to study global immune cell dynamics in mice lungs at ten time points during 7 days of flu infection. We find dramatic changes in quantities of 70 immune cell types, including various innate, adaptive, and progenitor immune cells. We focus on the previously unreported dynamics of four immune dendritic cell subtypes and suggest a specific role for CD103(+) CD11b(-) DCs in early stages of disease and CD8(+) pDC in late stages of flu infection.

Figure 1. Overview of the digital cell quantifier (DCQ) algorithm

Our DCQ method takes two gene expression datasets as input: First (top left), differential genome‐wide expression data from a complex tissue (here, lung), where rows are gene and columns are samples (here, time points during infection), high and low transcript level is color‐coded in red and blue, respectively. Second (top right), a precompiled compendium of prior information about the abundance of each cell surface marker in each immune cell type (rows—markers; columns—cell types). Immune cell types are illustrated together with their hierarchical hematopoietic cell lineages. DCQ provides as output a matrix (bottom) of predicted relative cell quantities for each immune cell type (row) in each sample (column). Increase or decrease in cell quantity is color‐coded in purple and green, respectively. Scatter plots (bottom right) exemplify the inferred amount of dendritic cells (y‐axis) during the time course of infection (x‐axis), where DC's quantity is reduced during the initial few time points and then elevated during latter time points. Standard deviations are calculated by DCQ based on an ensemble of alternative solutions (see Materials and Methods).

Figure 2. Digital cell quantification (DCQ) reconstruction of an in vitro‐defined complex cell mixture

1. Performance analysis on ten samples generated using an in vitro‐defined complex cell mixture (see Materials and Methods). The two matrices indicate the agreement among relative quantities that were inferred by DCQ (right) and the input relative cell quantities (left) for ten different experimental samples (rows), each of which involves five immune cell subsets (columns). Increase or decrease in cell quantity is color‐coded in purple and green, respectively.

2. A summary of DCQ's predicted relative cell quantities (y‐axis) and input relative cell quantities (x‐axis) across all ten samples from a. The plot indicates the high correlation in each of the cell types (color‐coded: black—B cells; brown—CD4⁺ T cells; green—CD8⁺ T cells; yellow—NK cells; cyan—CD11c⁺ DCs).

3. The effect of RNA sequencing depth on DCQ performance, tested on the dataset from a. Accuracy of DCQ predictions (y‐axis) are presented for various RNA sequencing depths (x‐axis). Accuracy is evaluated as correlation between predicted and input (“true”) relative cell quantities. Depicted are average of correlation and standard deviation over ten samples of each sequencing depth. The evident saturation with increasing depths implies that a sequencing depth of 2.5 million reads or higher is sufficient to provide high DCQ accuracy.

Figure 3. Digital cell quantification (DCQ) reconstruction of global immune cell dynamics during in vivo influenza infection

The immune dynamics map: Global dynamics in cell quantities (green/decrease, purple/increase in relative cell quantities) following influenza infection, predicted by DCQ at different time points (columns) for 213 different immune cell types (rows). Previously reported increase in cell types is marked in red (left, color bar). Each cell type heading is followed by the code of the tissue from which the cell type was isolated in the compendium. The box at the bottom right contains details for these abbreviations. Dendritic cells are shown at the top right panel and accompanied with four colored circles, indicating those subsets that were subject to FACS validations (see also Fig 5).

Figure 4. Digital cell quantification (DCQ) correctly predicts changes in most known increasing cell types during influenza infection

1. A

   Principle component analysis (PCA) of DCQ's predicted relative cell quantities. The PCA was applied on the profiles of predicted relative cell quantities for each cell type, at ten time points during influenza infection. Shown is a scatter plot of each cell type for the first two principle components PC1 and PC2. Red, cell types that were previously reported as increasing in quantity during infection; gray, the remaining cell types.

2. B, C

   Comparison of performance. False‐ (x‐axis) and true‐positive (y‐axis) rates of DCQ predictions. Rates are calculated for comparing predicted increase in cell quantities versus the known increasing cell types. In (B), we compare five alternative methods for selecting gene markers. In (C), we compare DCQ to several alternative computational cell quantification approaches, each consists of a different mathematical formulation and a different set of markers. The plots suggest the superiority of DCQ, and in particular its FACS‐based selection of markers, over extant methods.

Figure 5. Heterogeneity in DC dynamics highlights the distinct role of DC subtypes in influenza infection

1. Validation of cell quantity dynamics of four dendritic cell subtypes. Dendritic cell quantities (y‐axis) as predicted by digital cell quantification (DCQ) (blue diamonds) and measured by FACS (red diamonds) for two mice at each time point (x‐axis). The plot indicates the high correlation between DCQ predictions and FACS validations.

2. PCA analysis of the genome‐wide transcriptional responses to influenza infection. The PCA was applied on RNA‐Seq expression values that were profiled on isolated DCs (normalized by the respective values before infection). Shown are three time points (1, 3, and 5 days) for each of the DC subsets (blue, CD8⁺ pDC; green, CD8⁻ pDC; red, CD103⁻ CD11b⁺ cDCs; brown, CD103⁺ CD11b⁻ cDCs; black, lung tissue).

3. Expression of CCL17 (y‐axis) at four time points during influenza infection (x‐axis) for four DC subsets (color‐coded as in B).

4. Gene expression profiles of four DCs populations during influenza infection in lung. Left: Shown are the (log2 ratio) expression levels of selected genes (rows) at three time points (columns) for four isolated subsets of DCs relative to control subsets before infection; Z‐normalized per row. Previously reported inflammatory and anti‐viral genes in dendritic cells are marked in left (blue, Amit et al, 2009; Gat‐Viks et al, 2013). Cluster C–I is highlighted. Right: average expression (y‐axis) at each time point (x‐axis) of genes in clusters C‐I, for the four isolated DC subsets (color‐coded as in B).