Gene Expression Signatures Characterized by Longitudinal Stability and Inter-Individual Variability Delineate Baseline Phenotypic Groups with Distinct Responses to Immune Stimulation

Abstract

Human immunity exhibits remarkable heterogeneity among individuals, which engenders variable responses to immune perturbations in human populations. Population studies reveal that in addition to inter-individual heterogeneity systemic immune signatures display longitudinal stability within individuals, and these signatures may reliably dictate how given individuals respond to immune perturbations. We hypothesize that analyzing relationships among these signatures at the population level may uncover baseline immune phenotypes that correspond with response outcomes to immune stimuli. To test this we quantified global gene expression in peripheral blood CD4⁺ cells from healthy individuals at baseline and following CD3/CD28 stimulation at two time points 1 month apart. Systemic CD4⁺ cell baseline and post-stimulation molecular immune response signatures (MIRS) were defined by identifying genes expressed at levels that were stable between time points within individuals and differential among individuals in each state. Iterative differential gene expression analyses between all possible phenotypic groupings of at least 3 individuals using the baseline and stimulated MIRS gene sets revealed shared baseline and response phenotypic groupings, indicating the baseline MIRS contained determinants of immune responsiveness. Furthermore, significant numbers of shared phenotype-defining sets of determinants were identified in baseline data across independent healthy cohorts. Combining the cohorts and repeating the analyses resulted in identification of over 6000 baseline immune phenotypic groups, implying that the MIRS concept may be useful in many immune perturbation contexts. These findings demonstrate that patterns in complex gene expression variability can be used to define immune phenotypes and discover determinants of immune responsiveness.

Introduction

The vast immune heterogeneity that exists among individuals results in wide ranges of different responses to immune stimuli, such as immune modulatory therapies (1-3). This has spurred individualized medicine approaches, which aim to use specific information from particular patients to predict where in the range of responses to therapeutic agents patients will be and guide therapeutic decision-making to optimize patient outcomes (4, 5). Inter-individual immune variability, however, stands as one of the major challenges in the development of these approaches, and ways in which this variability can be used to inform individualized medicine is an area of intense study.

Next generation sequencing technologies such as RNA-Seq offer ways to quantify inter-individual immune heterogeneity at the level of global gene expression, which, with the exception of several immune cell gene networks whose expression is conserved among individuals (6), varies widely due to complex interactions among genetic and environmental variables (3, 7-14). Mounting population studies, however, have revealed that inter-individual immune heterogeneity is not without structure (15, 16). For example, data from population immunology studies support an elastic stability model of immunity, in which systemic immune signatures that vary widely among individuals remain relatively stable within individuals over time and return to that stable state after responding to perturbations such as vaccination or infection (7, 15, 17-19). These signatures are composed of many different parameters, which combined determine immune functionality and responsiveness. Furthermore, signatures from particular individuals are likely similar to those of other individuals in some respects, but different in others. Therefore, we hypothesize that defining the hierarchies of these relationships may elucidate groups of individuals sharing immune gene expression phenotypes, and that those groups will correspond with how those individuals will respond to various immune stimulations.

Peripheral blood, as a major immune cell pipeline, provides a rich source of information about systemic immune status. Circulating immune parameters, rather than resembling those in particular tissues (20, 21), represent cumulative host immune system interactions defining immune response potential, and as such systemic responses to immune perturbations can be definitively measured using peripheral blood (22-24). This, along with the need to understand how given individuals will respond to immune-based therapeutic (25) and prophylactic (26) treatments, has spurred searches for baseline immune determinants of immune response outcomes in peripheral blood. Vaccine studies using next-generation genomic and proteomic technologies have demonstrated the utility of this approach (18, 27).

To define immune signatures within individuals, we selected genes expressed at levels that are stable over time within individuals and differential among individuals based on RNA-Seq data from healthy donor peripheral blood CD4⁺ cells. Collectively, we refer to these sets of genes as molecular immune response signatures (MIRS). Since the MIRS captures inter-individual gene expression variability, we hypothesize that MIRS gene expression levels can be used to assemble individuals with similar immune statuses and response potentials into immune phenotypic groups that can inform how those individuals will respond to an immune stimulus. In this proof-of-concept study we show the capacity of the MIRS to define immune gene expression signatures in individuals, and that the resulting gene sets can be used to relate baseline gene expression to differential in vitro CD3/CD28 stimulation outcomes. Importantly, we demonstrate that the MIRS approach robustly identifies similar CD4⁺ cell gene expression phenotypes in independent healthy donor cohorts. We also show that the MIRS can be used to define thousands of baseline immune phenotypic groups, suggesting it can serve as a response determinant discovery tool in multitudes of immune perturbation contexts. The data from these initial experiments display the potential of the MIRS concept as a powerful tool for evaluating systemic immune gene expression signatures and inferring how individuals will respond to immune modulatory therapies prior to treatment.

Materials and Methods

Experimental Design

The presented work is from an observational study involving human subjects. In it, our main objectives were to harness transcriptomic data to define baseline and stimulated systemic immune signatures, examine relationships in those signatures among individuals to identify phenotypic groupings of individuals, and test whether baseline signatures contain determinants of immune responsiveness. To those ends we obtained replicate peripheral blood samples from two independent cohorts (500 and 100 series, Table S1) of self-described healthy individuals, purified CD4⁺ cells from each sample, and performed RNA-Seq using baseline and in vitro CD3/CD28-stimulated cells. Replicate samples were obtained so we could measure temporal gene expression stability, and enough subjects to test several hundred possible phenotypic groupings were recruited. Individuals with outlier replicate samples, as determined by principal component analysis using all genes expressed ≥1 log2(RPM+1) in ≥6 samples at baseline, were prospectively excluded from analysis in each cohort. We used the RNA-Seq data to devise molecular immune response signatures (MIRSs) consisting of genes expressed stably over time within given individuals and differentially among the individuals in the stimulated and baseline states. We identified immune response and baseline phenotypic groups by assorting individuals into hundreds of distinct groupings and testing how many stimulated and baseline MIRS genes, respectively, were differentially expressed between each grouping. Gene sets describing phenotypic groups were characterized using pathway analysis. Identified baseline and immune response phenotypic groups were compared to see if the baseline MIRS contained determinants of immune responsiveness. The analyses were performed separately using each independent cohort to evaluate the reproducibility of the results and see if common phenotypes could be identified across cohorts, and then together to quantify the number of baseline immune phenotypic groups that could be defined using the largest possible number of individuals in our study.

Study Subjects

A cohort of 11 individuals (500 series, Table S1) was recruited to donate replicate peripheral blood samples, and informed consent was obtained in accordance with a protocol approved by the Mayo Clinic Institutional Review Board (IRB#12-002580). A twelfth individual from this cohort was prospectively excluded from analysis because one of the individual’s replicate samples was a clear outlier with respect to the other 23 samples in a principle component analysis, suggesting an undefined systematic technical issue occurred with that sample. All individuals self-identified as healthy (no immune-mediated disease or cancer) per questionnaire. Replicate peripheral blood samples from 9 individuals in a second independent cohort (100 series, Table S1) were obtained similarly, as previously described (6). Three additional individuals from this cohort, whose samples were collected on the same days, were prospectively excluded from analysis because their samples appeared to be mislabeled in a principle component analysis and there was no clear resolution of which samples belonged to which individuals.

CD4⁺ Cell Enrichment and Stimulation

Peripheral blood collection, CD4⁺ cell enrichment, and stimulation were performed in each cohort as described previously (6). Briefly, CD4⁺ cells were enriched from fresh peripheral blood using CD4 microbeads and MS columns (100 series, Table S1) or the autoMACS® Pro Separator (500 series, Table S1) (Miltenyi Biotec). Purity was assessed in each sample via flow cytometry, and lysates were prepared from freshly isolated cells at baseline and after stimulation with α-CD3/α-CD28 Human T-Activator Dynabeads® (Invitrogen) for 4 hours at 37°C using QIAzol (QIAGEN). Lysates were stored at −80°C for 3 to 8 months, and RNA was extracted using miRNeasy (QIAGEN) immediately prior to cDNA library preparation and RNA-Seq.

Library Preparation and RNA-Seq

Library preparation and RNA-Seq were performed by the Mayo Clinic Medical Genome Facility Gene Expression Core as previously described (6). Briefly, TruSeq libraries (RNA Prep Kit v2, Illumina) were sequenced as 51-bp paired-end reads on the HiSeq 2000 (Illumina) at depths of approximately 37 million reads. The reads were aligned to the hg19 reference genome using the Tophat aligner (28), and raw read counts of each annotated gene in GENCODE v12 (29) were calculated using HTSeq (http://www.huber.embl.de/users/anders/HTSeq/doc/count.html). Expression data is available in the NCBI Database of Genotypes and Phenotypes (dbGaP) using accession number phs001512.v1.p1 (https://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs001512.v1.p1) and in Table S2 as raw mapped counts.

RNA-Seq Data Normalization

After discarding genes with zero read in all samples, raw reads were converted to reads per million (RPMs) ((gene reads × 1000000)/total reads), and after adding 1 to all gene RPMs to rid of zeros the RPMs were log2-transformed (log2(RPM+1)). Non-transformed RPMs were used to calculate fold changes. The trimmed mean of M-values and voom normalization methods were employed when individuals from each cohort were analyzed all together (30, 31).

Statistical Analyses

Hierarchical Clustering

Hierarchical clustering was done using Genomics Suite® 6.6 (Partek®) and R. In each case log2(RPM+1) data was centered by subtracting mean gene expression from expression of each gene in each sample, and then scaled by dividing each centered expression value by the standard deviation of each sample. Dendrogram branch lengths represented Pearson dissimilarity coefficients ((1-Pearson correlation)/2) and distances between clusters were determined using average linkages.

Stimulated MIRS

The stimulated MIRS was constructed as in Fig. 2A. Only genes expressed at ≥1 log2(RPM+1) in ≥6 samples in stimulated CD4⁺ cells were considered for inclusion. Y chromosome genes were excluded to eliminate explicitly gender-based contributions to the signature. Stimulation responses were measured for each gene by calculating absolute value fold changes between stimulated and baseline counts for each replicate blood sample and averaging those absolute value fold changes. Absolute values were calculated to ignore directions in gene expression changes in response to stimulation, and genes with average absolute value fold changes >2 were retained for analysis (Fig. 2A, left). Intra-individual temporal expression stability was measured for each gene by calculating absolute value fold changes between stimulated B and A blood sample replicates for each individual and averaging those absolute value fold changes. Absolute values were calculated to ignore directions between arbitrary A and B replicate samples, and genes with average absolute value fold changes <1.5 were retained for analysis (Fig. 2A, right). Differential gene expression was performed between all pairwise individual comparisons using the individuals’ replicate blood samples with the limma package in R (32), and all comparisons were combined into an F-test from which F-statistics and p-values were derived for all genes. P-values were adjusted to FDRs for each gene using the Benjamini-Hochberg method (33). Genes with FDR<0.01 were retained for analysis (Fig. 2A, right).

Fig. 2. Defining and characterizing a CD3/CD28 stimulation response MIRS.

(A) Expression of an example stimulated MIRS gene, IL17A. Connected dots represent unstimulated and stimulated expression from the same sample (left) and stimulated expression from replicate samples (right). (B) Hierarchical clustering of stimulated (S) replicate samples (A and B, color-coded) using stimulated MIRS gene expression levels (1147 genes). (C) Hierarchical clustering of stimulated replicate samples using expression levels of genes expressed at ≥1 log2(RPM+1) in ≥6 samples post-stimulation (13943 genes). (D) Mean replicate sample Pearson dissimilarity coefficients using the stimulated MIRS, all genes expressed at ≥1 log2(RPM+1) in ≥6 samples, and 10000 randomized sets of 1147 genes.

Baseline MIRS

The baseline MIRS was constructed as shown in the right panel of Fig. 2A (response to stimulation was not considered), except baseline gene expression data was used. Only genes expressed at ≥1 log2(RPM+1) in ≥6 samples at the baseline state (excluding Y chromosome genes) were considered for inclusion. A batch variable in which independent cohorts were considered batches was included in linear modeling when the cohorts were combined to make a baseline MIRS from 20 individuals.

Phenotypic Group Analysis

Phenotypic groups represented all possible combinations of 3 and 8, 4 and 7, and 5 and 6 individuals in the cohort of 11, all combinations of 3 and 6 and 5 and 4 individuals in the cohort of 9, and all possible combinations of 3 and 17, 4 and 16, 5 and 15, 6 and 14, 7 and 13, 8 and 12, 9 and 11, and 10 and 10 individuals when the cohorts were combined. For the 2 smaller cohorts differential MIRS gene expression analysis between phenotypic groups was performed iteratively with all possible combinations of single blood samples from each individual with the limma package in R (32). For the combined cohort analysis, 1000 random combinations of single samples from each individual were used to limit the computational load required to complete the analysis. Titration analysis showed that randomly sampling 1000 single sample combinations provided highly similar results compared to using all possible single sample combinations. P-values were adjusted to FDRs using the Benjamini-Hochberg method (33) and relatively upregulated and downregulated genes were considered separately to select differential genes expressed in consistent directions. These methods resulted in two statistical thresholds for consideration: FDR and percent of iterative tests in which chosen FDRs were met. Multiple threshold combinations were tested. When ≥5 differentially expressed genes resulted in a given threshold combination the accuracy with which the genes described the given group was evaluated by clustering all individuals’ replicate samples using their expression of those genes. Groups that clustered discretely using expression of differential genes, meaning all replicate samples from the first group of individuals formed one cluster and all replicate samples from the second group formed another cluster, were considered phenotypic groups. Groups were defined as phenotypic groups at all statistical threshold combinations after the most stringent combination at which the phenotypic group criteria were met, where stringency was defined by numbers of differential genes each statistical threshold combination produced on a per group basis. Independent cohorts were considered batches in the combined cohort analysis (Fig. S1C), and the removeBatchEffect function from the limma package was used prior to clustering to ensure batch effects did not influence clustering analyses (32).

Pathway Analysis

Pathway analyses were performed using the Core Analysis feature in Ingenuity® Pathway Analysis (IPA®, QIAGEN Redwood City, http://www.qiagen.com/ingenuity). Gene sets were analyzed at the most stringent statistical thresholds resulting in ≥100 genes and discrete group clustering, and only gene sets with ≥100 genes were analyzed per the developer’s instructions. All enrichment p-values were calculated using Fisher’s exact tests, and all IPA data was visualized using the heatmap.2 function from the gplots package in R (34).

Phenotypic Group Age and Gender Association Analyses

Phenotype descriptive gene sets were analyzed at the same statistical thresholds as for pathway analysis. To determine age associations the number of phenotype descriptive genes whose expression correlated with age (p<0.05) per linear regression was determined for each phenotypic group. Ages were then randomly assigned among individuals 10000 times, and numbers of correlative genes were determined using linear regression each time. P-values represent proportions of tests with numbers of correlative genes greater than or equal to the numbers of genes correlating with actual ages. To assess gender associations for each phenotypic gene set linear regression and ANOVAs were performed for each gene. One-way ANOVAs between groups of males and females were conducted for each gene, and the resulting sums of squares for each gene were added together in each gene set to derive total sums of squares for each gene set. Total sums of squares for all possible groupings of the same sizes as the male and female groups were then calculated for each gene set, and p-values represent the proportions of groupings with total sums of squares greater than or equal to those for the male and female groupings.

Gene Set Overlap Statistical Tests

Statistical thresholds for phenotype-describing gene sets were chosen as for pathway analysis prior to overlapping. To calculate probabilities of overlapping gene numbers between gene sets occurring randomly (Fig. 5B-C), appropriate numbers of corresponding MIRS genes were randomly selected 10⁵ times for each gene set and gene sets were overlapped each time. In Fig. 5B response regulators were overlapped with genes randomly selected from the baseline MIRS for each phenotype. This iterative randomized overlapping was performed 10 times, and p-values represent mean proportions of random overlaps resulting in overlapping gene numbers greater than or equal to actual overlapping gene numbers.

Fig. 5. Identifying determinants of immune responsiveness in the baseline MIRS.

(A) Venn diagram of baseline and response phenotypic groups where response groups are defined at FDR≤0.15 in 70% of tests and baseline groups are defined at FDR≤0.15 in 70% of tests (left) FDR≤0.2 in 50% of tests (right). (B) Presence of predicted response regulators (rows) in baseline gene sets of phenotypic groups that arise at the baseline and stimulated states (columns). Predicted response regulators for each phenotype are upstream regulators with enrichment p-values<0.05 in stimulated gene sets describing each phenotype as determined using IPA (significant –log10-transformed enrichment p-values shown in purple, p<0.05 = −log10(p)>1.3). Upward arrows designate increased expression of indicated regulators in response to CD3/CD28 stimulation, downward arrows designate decreased expression of indicated regulators in response to CD3/CD28 stimulation. Probabilities of given numbers of response regulators appearing in baseline gene sets randomly were calculated for each phenotype. *p<0.05. (C) Numbers of genes shared between baseline and stimulated gene sets for each phenotypic group (light green) as well as genes unique to each gene set for each group (purple and orange, respectively). P-values represent probabilities of given numbers of shared genes occurring randomly. ***p<0.001.

Group Overlap Statistical Test

To calculate probabilities of observed overlaps between response and baseline phenotypic groups occurring randomly (Fig. 5A, Fig. 6C), appropriate numbers of groups were randomly selected 10⁶ times and overlapped each time. This test was performed 10 times, and p-values represent mean proportions of random overlaps that resulted in numbers of overlapping groups equaling or exceeding the actual numbers of overlapping groups.

Fig. 6. Comparing phenotypic groups across cohorts.

(A) Number of genes with FDR≤0.2 in ≥50% of iterative tests in all possible groups of 3 and 6 (teal) and 4 and 5 (lavender) using baseline data from a second independent cohort. Baseline phenotypic groups are numbered. (B) Number of genes with FDR≤0.2 in ≥40% of iterative tests as in A but using stimulated data. Phenotypic groups with red numbers are associated with age (p<0.05). (C) Venn diagram of baseline and response phenotypic groups, ***p<0.001. (D) Baseline phenotypic gene sets from the first cohort (rows) corresponding to baseline phenotypic groupings from the second cohort (columns). Phenotypic gene sets were defined at FDR≤0.2 in ≥50% of iterative tests for each cohort. Purple denotes phenotype descriptive gene sets from the first cohort that cluster samples from the second cohort into phenotypic groups identified in A. *p<0.05.

Cross-Cohort Phenotype Correspondence Statistical Test

To quantify the probability that 6 of 9 baseline phenotypic groupings from the second cohort would cluster using baseline phenotypic gene sets from the first cohort randomly (Fig. 6D), each baseline phenotypic gene set from the first cohort was randomized using the baseline MIRS from that cohort 10⁴ times. Gene set sizes identified in the initial analysis were maintained. Individuals from the second cohort were clustered using their expression data for the randomly-selected gene sets, and the number of baseline phenotypic groupings from the second cohort that formed using the random gene sets were recorded each time. To account for the fact that relationships among groups of baseline phenotypic gene sets from the first cohort (rows with purple boxes in the same columns in Fig. 6D) would be lost upon gene set randomization, allowing more opportunities for random gene sets to form baseline phenotypic groupings in the second cohort, one baseline phenotypic gene set from each group was chosen for randomization in each iterative test (19 gene sets, rather than 28, were tested each time). The iterative analysis of 10⁴ trials was performed 10 times, such that the resulting p-value represents the average proportion of tests in which ≥6 baseline phenotypic groupings from the second cohort were formed using gene sets selected randomly from the baseline MIRS of the first cohort.

Results

Defining and characterizing a CD3/CD28 stimulation response MIRS

Replicate peripheral blood samples were obtained from 11 healthy individuals (500 series, Table S1) approximately 1 month apart (Fig. 1). CD4⁺ cells were immediately positively selected from each sample using magnetic beads, and RNA from both unstimulated cells and cells stimulated with α-CD3 α-CD28 antibody-coated beads was isolated for RNA-Seq. Gene expression data from the stimulated CD4⁺ cells was used to model immune response outcomes.

Fig. 1. Data collection strategy.

Replicate peripheral blood samples were obtained from healthy individuals approximately 1 month apart, and CD4⁺ cells were enriched from whole blood by positive magnetic bead separation each time. Lysates were prepared from freshly isolated cells at baseline and after stimulation with α-CD3/α-CD28 antibody-coated beads for four hours in vitro and frozen. RNA was prepared from lysates immediately prior to cDNA library preparation and RNA-Seq.

A CD3/CD28 stimulation response MIRS was devised by assembling a set of genes expressed stably over time within individuals, yet differentially among individuals. Specifically, genes included in the MIRS were those whose expression changed by an average absolute value of more than 2-fold between the baseline and stimulated states for each blood sample (Fig. 2A, left). Genes whose expression levels varied by less than an average absolute value of 1.5-fold among paired replicate samples post-stimulation were then retained (Fig. 2A, right). Finally, to identify genes whose expression levels varied among individuals in response to stimulation, differential gene expression analysis was performed between all pairwise individual comparisons, and genes whose expression differentiated individuals at false discovery rates (FDR)<0.01 were retained (Fig. 2A, right).

To test how well these methods defined individualistic responses to CD3/CD28 stimulation among individuals, the samples were organized in hierarchical clusters using expression of the MIRS genes (1147 genes). All replicate samples clustered together as nearest neighbors (Fig. 2B), indicating that the MIRS genes were indeed consistently uniquely expressed among individuals in response to CD3/CD28 stimulation. This was opposed to hierarchical clustering of the samples using expression data from all genes expressed ≥1 log2(reads per million (RPM)+1) in ≥6 samples in CD4⁺ cells post-stimulation (13943 genes), in which several replicates failed to cluster together as nearest neighbors (Fig. 2C). Furthermore, the dendrogram branches connecting replicate samples, which represent Pearson dissimilarity coefficients, were shorter on average in the MIRS clusterings compared to the all expressed genes clusterings or clusterings using expression levels of random subsets of 1147 genes (Fig. 2D). Therefore, the stimulated MIRS contains genes whose expression defines individualistic responses to CD3/CD28 stimulation.

Identifying response phenotypes using the stimulated MIRS

To test the hypothesis that stimulated MIRS genes could be used to distinguish phenotypic groups arising in response to CD3/CD28 stimulation, differential gene expression analyses were performed between all possible groups of 3 and 8 individuals, 4 and 7 individuals, and 5 and 6 individuals, totaling 957 possible groupings in all (Fig. 3A). Groups that were defined by at least five differentially expressed genes were interrogated using hierarchical clustering to determine how well the gene expression measures defined the groups. Only groups that clustered discretely, with all replicates of individuals in one group separating from all replicates of individuals in the second group, were considered response phenotypic groups. Additionally, to take advantage of having replicate samples from each individual but avoid using them simultaneously in each differential gene expression analysis, we tested for differential MIRS gene expression between the groups iteratively using all possible combinations of single samples from each individual. This method resulted in two statistical thresholds: FDR and proportions of iterative analyses in which MIRS genes were below that FDR. The analyses were performed at multiple thresholds to evaluate a range of differential gene numbers for each defined phenotypic group.

Fig. 3. Identifying response phenotypes using the stimulated MIRS.

(A) Response phenotypic group identification strategy, where n is the number of possible combinations of individuals for each group size. (B) Numbers of genes with FDR≤0.15 in ≥70% of iterative differential gene expression tests. Group sizes are color-coded as in A, and response phenotypic groups are numbered by the order in which they were tested. (C) Hierarchical clustering of response phenotypic group 333, where rows represent phenotypic gene expression in the indicated samples and columns represent each gene in the phenotypic gene set. (D) Comparison of 10 of the 50 most enriched pathways in response phenotypes defined by ≥100 genes per IPA (left). Shades of purple represent significant –log10-transformed enrichment p-values (p<0.05 = −log10(p)>1.3). Downward arrows indicate inhibition of pathways (z-score≤−2) in the smaller group of individuals in each phenotype, black dashes signify no significantly discernable pattern of pathway activation or inhibition (−2<z-score<2), and white dashes indicate that patterns of activation and inhibition cannot be determined in IPA.

This method yielded several response phenotypic groups differentiated by expression of various numbers of stimulated MIRS genes (numbered peaks, Fig. 3B; Table S2), with minimal background numbers of differentially expressed genes identified for the vast majority of groupings (Fig. 3B). An example of discrete group clustering is shown in response phenotypic group 333, where individuals 503, 505, 510, and 512 cluster discretely from the rest of the individuals in the cohort (Fig. 3C). Gene sets defining the response phenotypic groups were enriched for distinct profiles of biological pathways according to Ingenuity Pathway Analysis (IPA, Fig. 3D). Additionally, some of the biological pathways were inhibited in the smaller groups of some phenotypes but not others (Fig. 3D, downward arrows). None of the phenotypes significantly correlated with age or gender in linear regression analyses. Overall, these findings demonstrate that analyzing relationships in CD3/CD28 stimulation response MIRS gene expression patterns in cohorts of individuals delineates multiple distinct immune response phenotypic groups.

The baseline MIRS and baseline gene expression phenotypes

Next, a MIRS was constructed using baseline CD4⁺ cell expression data to test whether the MIRS concept could identify baseline immune phenotypic groups in a cohort of individuals. In an analysis mirroring the one that defined the stimulated MIRS, baseline gene expression data was used to identify a set of genes expressed stably within individuals, yet in a manner discriminant among them (as in Fig. 2A, right). Expression of the baseline MIRS genes (3384 genes) clustered replicate samples together as nearest neighbors (Fig. 4A), and shortened mean replicate sample dissimilarity coefficients compared to clustering samples using all genes expressed ≥1 log2(RPM+1) in ≥6 samples in baseline CD4⁺ cells (Fig. 4B, 15669 genes). Therefore the MIRS platform also successfully described individualized gene expression patterns in unstimulated CD4⁺ cells.

Fig. 4. The baseline MIRS and baseline gene expression phenotypes.

(A) Hierarchical clustering of baseline (B) replicate samples using expression of baseline MIRS genes (3384 genes) (B) Hierarchical clustering of baseline replicate samples on genes expressed at ≥1 log2(RPM+1) in ≥6 samples in the baseline state (15669 genes). (C) Numbers of genes with FDR≤0.15 in 70% of iterative differential expression tests for baseline groups as in Fig. 3B. Baseline phenotypic groups are numbered by order in which they were tested. Phenotypic groups with red numbers are associated with age, groups with light blue numbers are associated with gender (p<0.05). (D) Hierarchical clustering of baseline phenotypic group 333 as described in Fig. 3C. (E) Comparison of 10 of the 50 most enriched canonical pathways in baseline phenotypic groups defined by ≥100 genes as in Fig. 3D. Upward arrows represent activation of pathways (z-score≥2) in the smaller group of individuals in the phenotype, downward arrows indicate inhibition of pathways (z-score≤−2) in the smaller group of individuals in each phenotype, black dashes signify no significantly discernable pattern of pathway activation or inhibition (−2<z-score<2), and white dashes indicate that patterns of activation and inhibition cannot be determined in IPA.

To delineate baseline immune phenotypic groups, differential expression tests using baseline MIRS data between the same groups tested with the stimulated MIRS (Fig. 3A) were performed. This analysis also identified phenotypic groups that were differentiated by diverse numbers of MIRS genes (Fig. 4C-D, Table S2). Once again the identified phenotypic gene sets were enriched for distinct profiles of biological pathways, and some pathways were activated in the smaller groups of some phenotypes but inhibited in the smaller groups of other phenotypes (Fig. 4E, upward and downward arrows, respectively). Unlike the stimulated analysis the baseline analysis resulted in several groups whose differential gene expression correlated with gender and age (Fig. 4C, colored numbers). Taken together, assessing relationships in expression data from genes consistently expressed at unique levels in CD4⁺ cells in our baseline MIRS can be used to define baseline immune phenotypic groups.

Identifying determinants of immune responsiveness in the baseline MIRS

To test whether baseline determinants of immune responsiveness were captured in the baseline MIRS, the identified baseline phenotypic groups were compared to those found following stimulation. Using identical statistical thresholds to define phenotypic groups both before and after stimulation, 4 of 8 response phenotypic groupings were identical to groupings also formed among 18 identified baseline phenotypic groups (Fig. 5A, left, p<0.001). Furthermore, using a looser threshold to define baseline phenotypic groups, 7 of the 8 response phenotypic groupings were identified among 28 baseline phenotypic groups (Fig. 5A, right, p<0.001). The concordance of 7 of 8 response phenotypic groups with pre-stimulation groupings supports our hypothesis that stably expressed molecular determinants specifying differential response outcomes are present and identifiable prior to stimulation.

To characterize the determinants of immune responsiveness in the baseline MIRS, we examined the relationships between the baseline and stimulated gene sets defining each of the 7 phenotypes that arose at both states. One possibility is that the baseline MIRS contains master regulators of immune responsiveness that, while not responding to stimulation themselves, regulate genes that do respond to stimulation. To test this possibility, upstream regulator genes were determined for each of the 7 response phenotypes using IPA core analyses. Each response regulator gene set was then overlapped with the gene set that defined the same phenotypic groups at baseline. This analysis showed that predicted response regulator genes were differentially expressed at baseline in 5 of the 7 phenotypes and significantly enriched at baseline in 2 of the phenotypes (Fig. 5B), indicating some immune response determinants in the baseline MIRS regulate genes that respond to stimulation. Expression of some predicted regulators was upregulated in response to CD3/CD28 stimulation, while expression of others was downregulated (Fig. 5B, upward and downward arrows, respectively). Another possibility is that the baseline MIRS contains genes that directly respond to stimulation and are expressed at unique levels post-stimulation. To test this we overlapped the genes that differentiated each grouping pre- and post-stimulation. In each case, with the exception of phenotype 317, significant numbers of genes were differentially expressed both pre- and post-stimulation (Fig. 5C), indicating that some immune response determinants present in the baseline MIRS directly participate in the response to stimulation. Altogether these results support the hypothesis that analyzing relationships in baseline MIRS gene expression data among individuals unveils molecular determinants of immune responsiveness, and demonstrate that the determinants modulate the response to CD3/CD28 stimulation in different ways.

Comparing phenotypic groups across cohorts

To evaluate the robustness of the MIRS concept we performed similar analyses on a second independent cohort of 9 healthy individuals (100 series, Table S1). Baseline and stimulated MIRS gene sets constructed using gene expression data from this cohort once again identified genes whose expression consistently differentiated individuals from one another (Fig. S1A, Fig. S1B, respectively).

To identify baseline and response phenotypic groups in the 9 individuals from the second cohort, we used the baseline and stimulated MIRSs, respectively, from that cohort to perform differential gene expression analysis in all possible combinations of 3 and 6 and 4 and 5 individuals, totaling 210 possible combinations. As before this method resulted in robust identification of baseline (Fig. 6A, Table S2) and response (Fig. 6B, Table S2) phenotypic groups. Unlike the first cohort none of the baseline phenotypic groups from the second cohort significantly correlated with age or gender, while 2 response phenotypic groups significantly correlated with age (Fig. 6B, red numbers). Once again we observed significant overlap between phenotypic groups at the baseline and stimulated states (p<0.01, Fig. 6C). This further supported the hypothesis that the baseline MIRS harbors determinants of immune responsiveness.

Finally, we evaluated whether similar baseline phenotypic groups were identified in the first and second cohorts. 6 of the 9 baseline phenotypic groupings from the second cohort (Fig. 6D, columns) were formed by clustering individuals from the second cohort using their expression data for baseline phenotypic gene sets from the first cohort (Fig. 6D, rows). In each case multiple gene sets from the first cohort corresponded to a single baseline phenotypic grouping from the second cohort, indicating that groups of gene sets from the first cohort may be related even though each gene set is enriched for unique profiles of biological pathways (Fig. 4E). Taking those relationships into account, the finding that 6 of 9 phenotypic groups from the second cohort correspond to phenotypic gene sets from the first cohort was significant (p=0.03). Thus, the baseline MIRS can be used to robustly define determinants of immune response phenotypes, which are shared across independent human cohorts.

Quantifying baseline immune phenotypic groupings in a larger healthy cohort

The discovery of 28 baseline immune phenotypic groupings using a baseline MIRS derived from 11 individuals (Fig. 5A, right) indicates that the MIRS concept may have the capacity to identify baseline determinants of immune responsiveness to a variety of immune perturbations. The number of possible group combinations increases with increasing cohort size. To further characterize the diversity of immune phenotypic groupings that can be identified using the MIRS, we combined our healthy individual cohorts into a cohort of 20 individuals, generated a baseline MIRS using their combined gene expression data, and used that MIRS to define baseline immune phenotypic groupings as before. Batch effects were present between the cohorts (Fig. S1C) and were accounted for in linear modeling, and the MIRS once again successfully described longitudinally stable, individualistic baseline gene expression patterns among individuals (Fig. S1D). Differential MIRS gene expression was tested between all possible groups of 3 vs. 17, 4 vs. 16, 5 vs. 15, 6 vs. 14, 7 vs. 13, 8 vs. 12, 11, vs. 9, and 10 vs. 10 individuals, totaling 524076 possible group combinations. Using the same criteria for phenotypic groupings described previously the analysis resulted in identification of 6692 baseline immune phenotypic groupings at the least stringent statistical thresholds (Fig. 7), suggesting that the MIRS concept, rather than serving as a context-specific biomarker discovery tool, may provide a platform for identifying determinants of immune responsiveness to a wide variety of potential immune perturbations.

Fig. 7. Quantifying baseline immune phenotypic groupings in a larger healthy cohort.

Iterative differential gene expression was performed using baseline MIRS genes derived from a combination of two independent healthy cohorts (Fig. S1D, 20 individuals, 5880 genes) between all possible groups of 3 and 17, 4 and 16, 5 and 15, 6 and 14, 7 and 13, 8 and 12, 9 and 11, and 10 and 10 individuals. Analysis was performed at multiple statistical thresholds as shown above, and phenotypes were determined as described in Materials and Methods.

Discussion

Heterogeneity is a key trait of the immune system. While affording advantages at the population level (35), it provides challenges in predicting how given individuals will respond to immunotherapies (1, 2). Accumulating population studies demonstrate that immune status is highly variable among individuals yet consistent across time within individuals (15, 16). These findings suggest the presence of stable immune signatures whose parameters might dictate individual immune statuses and response potentials. To test this possibility at the transcriptional level we used global gene expression from baseline and CD3/CD28-stimulated healthy individual CD4⁺ cells to construct molecular immune response signatures (MIRS) containing genes consistently expressed at unique levels at each state. By exploring relationships in MIRS gene expression patterns among individuals, we delineated baseline phenotypic groups that were in part defined by determinants of immune responsiveness in two independent cohorts, displaying the potential of the MIRS concept as a powerful tool for identifying baseline determinants of immune responsiveness.

The relatively unbiased global gene expression measurements that RNA-Seq offers were critical to our approach. Although polymorphisms and variable gene expression have been described for single or relatively small numbers of genes in immune phenotypes and diseases (36, 37), using relatively small numbers of genes to characterize immune phenotypes fails to appreciate the vast interplays that exist among those genes and many others (38). Similarly, we analyzed data from pooled CD4⁺ cells rather than single cells or whole blood to capture all the complexity of a defined, undiluted component of systemic immunity. Further evaluation of MIRS-derived phenotypes of interest using other platforms, such as mass cytometry and functional assays, will be important for further characterizing the phenotypes and their biological implications.

We specifically examined CD4⁺ cells in this proof-of-concept study because they interact with many other immune cell types to orchestrate many facets of immune responses (39). Owing to those interactions, we hypothesize that analyzing CD4⁺ cell gene expression data and gene expression from other circulating immune cell types using the MIRS platform would result in identification of some of the same phenotypic groups and additional unique groupings. Testing this hypothesis will be critical for furthering our understanding of immune variability patterns among individuals, and will inform how the MIRS platform is used to predict individual responses to particular immune modulatory therapies. For example, it may be important to analyze gene expression from CD8⁺ T cells, which serve as crucial effectors in anti-cancer immunity and are therefore the targets of many cancer immunotherapeutic strategies (40, 41), while employing the MIRS concept to predict how cancer patients will respond to a given cancer immunotherapy. Similarly, analyzing gene expression in tissue-resident T cells, while more difficult to obtain from humans than circulating T cells, would likely provide additional perspectives on human immune phenotypes (20, 21). Phenotypic relationships between circulating and tissue-resident T cells could be analyzed using the MIRS approach by seeing if the phenotypic groups identified using circulating T cells are also identifiable in tissue-resident T cells.

The immune phenotypic groups derived using the MIRS gene sets could have arisen through multiple distinct mechanisms. Some groups may result from different proportions of CD4⁺ cell subsets across individuals and/or differential gene expression in the same cell types across individuals. Both scenarios dictate individualistic immune status (15), and our approach takes both of them into account. The cell types likely include multiple CD4⁺ T cell subsets, such as Th1, Th2, and Th17 cells, but other cell types, such as CD4⁺ monocytes (42), dendritic cells (43), and others (44) may contribute to the phenotypes as well. The majority of phenotypic groups that formed, however, are likely CD4⁺ T cell-based, as evidenced by flow cytometry data showing the majority of CD4⁺ cells are CD3⁺ (first cohort: 91.7±5.6% CD3⁺CD4⁺ cells, second cohort: 93±4.8% CD3⁺CD4⁺, see Materials and Methods). Additionally, gene expression analysis showed that genes reported to be expressed at higher levels in CD4⁺ T cells relative to other CD4⁺ cells, including subsets of B cells, dendritic cells, and monocytes, are indeed expressed at significantly higher levels than genes reported to be expressed more highly in the other cell types relative to CD4⁺ T cells (45-47) in our data (Fig. S2).

Age and gender, which also influence immune status (48, 49), correlated with several baseline and response phenotypic groups (Fig. 4C and Fig. 6B, respectively). The majority of phenotypic groups, however, did not correlate with these intrinsic factors. This was underscored by the observation of significant numbers of corresponding baseline phenotypic groups between independent cohorts even though the first cohort was generally older than the second (Table S1). A relatively broad age range was investigated in this study (24–57 years of age), but examining phenotypes at additional age ranges, for example older ages at which responses to immune perturbations such as vaccines tend to wane (50), would be informative as well.

The phenotypic groups identified by analyzing relationships in MIRS gene expression among individuals are distinct, as evidenced by enrichment of unique biological pathway profiles for each phenotype (Fig. 3D, Fig. 4E). There may be degrees of relatedness among the phenotypic groups as well, as indicated by the fact that groups of phenotypic gene sets from the first cohort corresponded to single phenotypic groups from the second cohort (Fig. 6D). These observations highlight the complexity of systemic immune phenotypes, where various phenotypes can arise from a number of factors, even in particular immune cell compartments. In this view of systemic immune phenotypes individuals may be similar with respect to some factors but differ with respect to others.

While this proof-of-concept study displayed the capacity of the MIRS concept to define determinants of immune responsiveness, further studies using larger sample sizes are needed to further solidify the approach. The robustness of the MIRS concept was exemplified by the fact that it identified corresponding baseline phenotypes in two independent cohorts (Fig. 6D), but selecting MIRS genes using larger sample sizes would result in more comprehensive gene sets capable of describing immune signatures in larger populations of individuals. It would also allow for analysis of more complex individual groupings, which will be important because many immune phenotypes may bin individuals into more than two groups or into continuous distributions (16, 51). Phenotypes that segregate individuals into more than two groups can be investigated using the MIRS approach by performing differential MIRS gene expression analysis among three or more groups. Continuous phenotypes are likely identifiable using the MIRS platform as well, because, like continuous phenotypes, the phenotypes delineated using the MIRS approach can be inter-related (multiple rows correspond to the same columns in Fig. 6D) and could represent gradations along a spectrum. For example, a continuum of baseline phenotypic groups may correspond to a single clinical response to an immunotherapeutic agent, and each of those groups will likely be delineated within the relatively large number of phenotypic groups identifiable using the MIRS platform (Fig. 7).

The translational application of the MIRS concept is to measure systemic transcriptional immune signatures prior to treatment with immune modulation-based therapies and use those measurements to predict patient responses to those therapies. To that end future applications of the MIRS platform would involve deriving baseline MIRS gene sets and using them to define baseline phenotypic groups in a patient training cohort. Of the thousands of baseline phenotypic groups likely to be identified, groups of interest for a given immunotherapeutic will be identified by determining which groups contain patients with the same clinical response to that immunotherapeutic. The ability of each baseline phenotypic gene set to successfully predict how patients will respond to the immunotherapeutic would then be quantified and validated in patient test cohorts.

Abbreviations

MIRS

molecular immune response signature

IPA

Ingenuity Pathway Analysis

RPM

reads per million