Study design - highly multiplexed measurement of cytokine secretions from single CD19 CAR-T cells upon antigen-specific stimulation
To capture the full spectrum of complex T cell functions within the heterogeneous CAR-T population, we analyzed CAR-T cytokine production at the single cell level using a 16-plex panel. The 16-plex cytokine assay panel includes the key immune functions of T cells (e.g. effector, stimulatory, regulatory and inflammatory) (Fig. 1b & Additional file 3) and has been validated at both a population level and a single-cell level (Additional files 4 and 5). CAR-T cells generated from 4 different healthy donors were analyzed for cytokine secretion at the single cell level. The enriched CAR+ T cells were stimulated with anti-CAR beads or control IgG beads, stained with anti-CD4-PE and anti-CD8-AF647 and loaded into SCBC microchips for single cell cytokine analysis (Fig. 1a). Each of SCBC microchip device is comprised of a polydimethyl siloxane (PDMS) microchamber layer and a glass slide (see schematic in Fig. 1c
(i)). The supporting glass slide is surface patterned with a miniaturized 16-element antibody microarray (Fig. 1c). The design of the SCBC permits optical microscopy inspection of the individual microchambers (Fig. 1c
(ii)), including number of cells in each well and the cellular phenotypes (e.g., CD4+, CD8+, etc.) based on surface marker staining. The fluorescence based cytokine data is merged with the microscopy imaging data to generate the final data set. The number of cells, the cell phenotype and the cytokine production are therefore specified for each microchamber. We further adopted and/or developed advanced informatics tools (in IsoPlexis’ IsoSpeak software package) for not only statistical analysis of SCBC data sets but also the dissection of functional subsets (Fig. 1c
(iii)). To evaluate the specificity of anti-CAR bead stimulation, CAR-T cells were stimulated with anti-CAR or IgG control beads in wells of 96-well plate for 24 h and the supernatant was analyzed by ELISA. Anti-CAR stimulation showed approximately 1000-fold increase of IFN-γ compared to control IgG beads stimulation (Fig. 1d), indicating the good specificity of in vitro anti-CAR bead stimulation. In addition, the increased cytokine secretion levels in CD19 CAR-T cells upon anti-CAR bead stimulation were observed across 4 donors compared to IgG control bead stimulation at a single-cell level, further demonstrating the stimulation specificity (Additional files 3 & 6).
Increased polyfunctional heterogeneity of activated CD19 CAR-T cells dominated by an effector cytokine profile upon anti-CAR bead stimulation
Given that the frequency of polyfunctional effector T cells correlates with potency of anti-tumor or anti-virus T cell immunity [24,25,26], we assessed the polyfunctionality of our CAR-T cell products. First, we computed the percentage of polyfunctional cells regardless of the combination of cytokines co-produced. As shown in Fig. 2a, under 5% of cells with IgG bead stimulation exhibited polyfunctionality. By contrast, the anti-CAR bead-stimulated CAR-T cells of donors 1–4 respectively showed a 6-, 2-, 47-, and 11-fold increase in polyfunctional cell counts for 4 donors, highlighting both the polyfunctionality and variability of CAR-T products across donors. In all 4 donors, we also noted that anti-CAR bead-stimulated CD8+ T cells were more polyfunctional than CD4+ T cells. Moreover, we used the polyfunctional strength index (PSI) described previously to quantify the collective impact of polyfunctional T cells [27]. The PSI of a sample is defined as the percentage of polyfunctional cells multiplied by the average signal intensity of the cytokines secreted by these cells. We further broke down PSI by cytokine function (Fig. 2b) – effector, stimulatory, regulatory, and inflammatory – to highlight the contribution of each group to the overall polyfunctionality of the sample. The PSI breakdown in Fig. 2b further revealed inter-donor heterogeneity of polyfunctional CAR-T cells. While effector and stimulatory cytokines contribute the most to polyfunctionality across all 4 donors, a small portion of regulatory and inflammatory cytokines were observed in donor 1 and donor 4. Furthermore, the observed regulatory and inflammatory response was mainly from the CD4+ T cells, consistent with the notion that both regulatory and helper T cells are subsets of CD4+ T cells, further proving the specificity of the SCBC assay. The donor-to-donor variability in polyfunctionality could be caused by a population-level shift of cytokine profiles, or the alteration of specific polyfunctional subpopulations. To better understand this variation, we first sought to use a conventional heatmap to visualize cytokine production from all single cells across four donors. The heat map data visualization in Fig. 2c gives a high-level indication that the analyzed single CAR-T cells exhibit significant differences in the combinations and intensities (red = low, green = high) of secreted proteins. However, it remains difficult to clearly visualize many functional subsets from heterogeneous CAR-T cells across donors.
Limitations of conventional PCA and other standard visualizations in dissecting high-dimensional, single-cell, proteomic data of CAR-T cells
We further explore other standard bioinformatics tools in visualizing this high-dimensional data set. In this standard bar graph visualization of the functional groups secreted by the four donors’ CD4+ CAR-T cells (Fig. 3a), the dimensionality of the data makes it cumbersome to see which are the major functional groups being secreted by each donor, as well as the largest fold differences across donors. An alternative approach to visualizing high-dimensional datasets is to first reduce the dimensionality of the data, while retaining as much of the original information as possible. PCA (principal component analysis) is a common dimensionality reduction technique which uses an orthogonal transformation to convert the original dataset of possibly correlated variables into a set of linearly uncorrelated principal components, where the number of components is smaller than the number of original variables. The transformation is defined in such a way that the first principal component has the largest possible variance (accounting for as much variability as possible within the dataset), followed by the second component, and so on. While reducing the dimensionality to two principal components may still result in some loss of information, the benefit is that the transformed data points can then be visualized on a two-dimensional scatterplot. Additionally, key differences within the transformed data should be magnified through this transformation. Figs. 3b-c displays the results of applying PCA to the 4-donor CD4+ CAR-T secretion dataset. Each cell’s secretions (signal intensity of each cytokine) are log transformed prior to dimensionality reduction. Fig. 3b shows a scatterplot of the transformed data color-coded by donor, while Fig. 3c shows the same data color-coded by some of the individual cytokines. The combination of these graphs does reveal additional information about donor response differences and expressed polyfunctional subsets, such as the lower overall polyfunctionality of donor 2, or the higher Granzyme B + MIP-1α + polyfunctionality of donor 1. A similar pattern of cytokine secretions was seen in the CD8+ CAR-T cells of all 4 donors (Additional file 7). However, two key issues still exist with this visualization: (1) it is difficult to infer which high-dimensional cytokine subsets are driving the polyfunctionality of each sample, and (2) it is unclear what are the more granular polyfunctional differences across the analyzed donor samples. Other methods such as viSNE (Additional file 8) that map high-dimensional cytometry data onto two dimensions, yet conserve the high-dimensional structure of the data, improve the ability to distinguish cell subpopulations [28, 29]. However, the polyfunctional breakdown of the samples remains unclear when this visualization is used. In the context of polyfunctional T cell analysis, the routine practice has been to manually enumerate the major polyfunctional subsets by quantifying the cell count in each of the possible cytokine combinations, which serves well the ability to distinguish polyfunctional T cell subsets but loses the relationship between subsets or the hierarchical structure of the population. We hereby propose two alternative visualizations, polyfunctional heat map and polyfunctional activated topology principal component analysis (PAT PCA) which in tandem work to solve these problems.
Polyfunctional heat map to highlight distinct polyfunctional T cell subsets and the heterogeneity within a cell population
To distinguish all polyfunctional subsets within a sample and dissect the population architecture, we developed a new polyfunctional heat map visualization. This visualization, shown in Fig. 4a for CD4+ CAR-T cells and Fig. 5a for CD8+ CAR-T cells, displays the major functional subsets secreted across the 4 donor samples. The heat maps are color-coded from light to dark, depending on the frequency of the polyfunctional cell subsets. The four rows of squares correspond to the four donors; each column corresponds to a polyfunctional group of cytokines that was expressed in at least one of the four samples. To condense the large number of functional groups arising from the high dimensionality of the data set, we use agglomerative hierarchical (complete linkage) clustering to attain a condensed set of functional groups that still faithfully represent the overall secretion profile of the donors. Using a Euclidean distance to measure similarity between functional groups, each unique group is input into the clustering algorithm as a 16-dimensional vector of 1 s and 0 s, corresponding to the presence or absence, respectively, of each cytokine in the group. We define the minimum permitted similarity value to perform a clustering operation, to ensure that clusters do not contain functional groups that are too distinct from each other. The resulting clusters and their frequencies are displayed in the functional heat map, with the size of the cytokine dots below each column representing the frequency of the corresponding cytokine in the cluster. As seen in Fig. 4a, donor 1, closely followed by donor 4, has the highest frequencies of most expressed functional groups. Donor 3 is less polyfunctional, while donor 2 has the least polyfunctional groups. The group GM-CSF, Granzyme B, IL-13 and TNF-α is expressed exclusively by the CD4+ CARs of donors 1 and 4, but not by the CARs of donor 2 or donor 3. Similarly, the 7-plex group containing GM-CSF, Granzyme B, IFN-γ, IL-8, IL-13, MIP-1α, and TNF-α is unique to these two donors, revealing functional differences of these CAR cells relative to those of donors 2 and 3 at a high degree of granularity. Functional groups not containing GM-CSF or IL-13 are expressed at similar frequencies by donor 3 as they are by donors 1 and 4. As seen in Fig. 5a, the CD8+ CAR cells similarly show donors 1 and 4 to be significantly more polyfunctional than donors 2 and 3, and also uniquely secrete the 7-plex group containing GM-CSF, Granzyme B, IFN-γ, IL-8, IL-13, MIP-1α, and TNF-α. Donor 1 has a higher number of unique polyfunctional groups than donor 4, particularly groups not containing IL-8. The only polyfunctional groups secreted by all four donors contain Granzyme B, MIP-1 α with smaller amounts of IFN- γ.
PAT PCA to further visualize distinct polyfunctional T cell subsets and a complex landscape of CD19-specific immune response
In addition to the heat map visualization, we propose a modified PCA visualization named PAT PCA illustrated in Additional file 2. We use this visualization to help reveal a complex landscape of polyfunctional subsets of CAR-T cells in response to antigen-specific stimulation across donors. Figs. 4b and 5b show PAT PCA visualizations of the same single-cell-resolution CD4+/CD8+ CAR results as the heat maps in Figs. 4a and 5a, respectively. Similarly for the hierarchical clustering input, we apply PCA on a binarized dataset (0 = no secretion, 1 = secretion), to focus on visualizing combinatorial differences, rather than intensity differences. In the resulting scatterplot visualization, each color-coded dot represented a single-cell from one of the four donors, respectively, in which the larger color-coded circles represented a unique polyfunctional subset. The colors of the subsets identified the sample where this subset was most frequent (i.e., largest as a percentage of the sample). The functional groups (columns) in the heat map are represented by a circle in the corresponding PAT PCA graph. The size of the circles corresponds to the frequency of the group; the number of cells (dots) within each group indicates how frequently the sample of the corresponding color secreted the group. The makeup of the two principal components PC1 and PC2 is indicated by the listed cytokines, with the composition of PC1 indicating the main drivers of polyfunctionality. In both graphs, PC1 is a combination of the seven dominant secretors of the anti-CAR stimulated cells: effector cytokines Granzyme B, MIP-1α, IFN-γ, and TNF-α; stimulatory cytokines GM-CSF and IL-8; and the regulatory cytokine IL-13. As a result, the highly-polyfunctional groups are farther along (on the right-side of) this axis. PC2 differed between CD4 and CD8 cells. In the CD4 case, groups towards the top tend to have more Granzyme B, IL-8 and TNF-α, while groups near the bottom more often contain IL-13, IFN-γ and MIP-1α. In the CD8 case, groups near the top more commonly secreted IFN-γ, IL-13 and GM-CSF, while groups near the bottom more commonly secreted Granzyme B, IL-8 and MIP-1α. Plotting the polyfunctional subsets in such a manner allows overall similarities and differences in the donor profiles to emerge. The lack of donor 2 (orange) subsets indicated the lower polyfunctionality of this sample, while the presence of numerous donor 1 (blue) and donor 4 (green) groups in the right area of the graph indicated the highly-polyfunctional makeup of these two samples. Donor 3 has fewer polyfunctional subsets mostly comprised of combinations of Granzyme B, MIP-1α, IL-8, and TNF-α but lacking IFN-γ, IL-13, and in the case of CD4+ cells, GM-CSF. Donor 4 largely spans the polyfunctional profiles of both donors 1 and 3, which can also be seen in the heat maps. Capturing the full polyfunctional landscape of each sample is critical to effectively analyzing this data, as well as identifying the major subsets contributing to differences between samples/donors. The presented polyfunctional heat map and PAT PCA graphs are effective at achieving these objectives, both individually and together, and provide the first landscape of effector function phenotypes of CD19 CAR-T cells in response to antigen-specific challenge.