Human influenza virus challenge identifies cellular correlates of protection for oral vaccination

Summary

Developing new influenza vaccines with improved performance and easier administration routes hinges on defining correlates of protection. Vaccine-elicited cellular correlates of protection for influenza in humans have not yet been demonstrated. A phase 2 double-blind randomized placebo and active (inactivated influenza vaccine) controlled study provides evidence that a human adenovirus 5-based oral influenza vaccine tablet (VXA-A1.1) can protect from H1N1 virus challenge in humans. Mass cytometry characterization of vaccine-elicited cellular immune responses identified shared and vaccine type-specific responses across B and T cells. For VXA-A1.1, the abundance of hemagglutinin-specific plasmablasts and plasmablasts positive for integrin α4β7, phosphorylated STAT5, or lacking expression of CD62L at day 8 were significantly correlated with protection from developing viral shedding following virus challenge at day 90 and contributed to an effective machine learning model of protection. These findings reveal the characteristics of vaccine elicited cellular correlates of protection for an oral influenza vaccine.

eTOC blurb

Developing influenza vaccines with greater and broader efficacy and easier administration requires clear insights into correlates of protection. McIlwain & Chen et al. use mass cytometry to identify cellular immune responses elicited by an oral influenza vaccine tablet (VXA-A1.1) that correlate with protection from influenza virus challenge in humans.

Graphical Abstract

Introduction

Influenza viruses cause a major global annual burden of morbidity, mortality, and economic cost, as well as present a perpetual risk of causing catastrophic pandemics. Influenza vaccines with improved performance and easier routes of administration have the potential to save a tremendous number of lives. In particular, an effective oral influenza vaccine could greatly improve vaccination rates by reducing costs and logistic barriers of intramuscular immunization. Reducing such barriers can make a critical difference in the timely administration of vaccines to 100s of millions of people when resources are limited, such as in times of a major pandemic or in low-income geographic regions. A recent phase 2 study provided evidence for the ability of an adenovirus 5 (Ad5)-based oral influenza vaccine tablet (VXA-A1.1) to provide protection from H1N1 virus infection in humans (Liebowitz et al., 2020).

Establishing robust correlates of protection is important for ensuring consistent vaccine production, estimating susceptibility of populations post-immunization, and bringing new vaccines to market (Plotkin, 2010). Serum antibody assays such as hemagglutination inhibition (HAI), microneutralization (MN), and others, are classically used to evaluate influenza vaccine responses. However, these methods are unable to provide precise correlates of protection and are less relevant for non-traditional vaccine approaches that may rely to a greater extent on mucosal antibodies or cellular immunity for protection (Gianchecchi et al., 2019; Holmgren et al., 2017; Plotkin, 2020). The importance of cellular responses for protection provided by VXA-A1.1 is suggested by higher ELISPOT antibody-secreting cell (ASC) levels in protected versus unprotected individuals (Liebowitz et al., 2020) and elevated mucosal homing lymphocytes in circulation following vaccination with VXA-A1.1 (Brandtzaeg and Johansen, 2005; Kim et al., 2016; Liebowitz et al., 2020). In addition to classical serology techniques, ELISPOT, cytotoxicity, and cytokine assays have established utility for monitoring influenza vaccine immune responses. However, immunity to influenza and other respiratory viruses can be multifaceted with a complex interplay between antibodies and multiple immune cell types (Korenkov et al., 2018; Laidlaw et al., 2013; Plotkin, 2020; Wilkinson et al., 2012). Thus, a deeper exploration may expose other potential correlates of protection.

Mass cytometry is a high-parameter immune monitoring technique that has already advanced our understanding of responses to infectious diseases and immunizations (Kotliar et al., 2020; Lingblom et al., 2018; McElroy et al., 2020; Michlmayr et al., 2020; Rahil et al., 2020; Reeves et al., 2018; Subrahmanyam et al., 2020). Mass cytometry uses metal-tagged antibodies to capture single-cell data for more than 40 parameters without the fluorescence channel overlap issues encountered in flow cytometry, and unlike single-cell sequencing, for millions rather than thousands of cells per sample (Hartmann et al., 2020). Coupled with optimized workflows for rapid fixation of low volume whole blood samples (Rahil et al., 2020), mass cytometry is a powerful and efficient tool to broadly survey many immune subsets simultaneously for biomarker discovery in clinical settings.

Influenza virus challenge studies are well-established tools to evaluate susceptibility to infection in a highly controlled setting. The detection of virus shedding in individuals after experimental virus challenge is closely linked to the development of signs and symptoms of mild/moderate influenza disease. Furthermore, established correlates of protection, such as HAI, function similarly in naturally acquired infection and challenge studies (Balasingam and Wilder-Smith, 2016; Sherman et al., 2019).

Here we performed 43-parameter mass cytometry characterization of cellular immune responses using samples from 141 individuals during a phase 2 influenza vaccination and matched virus challenge study that compares the efficacy of an investigational oral tablet vaccine (VXA-A1.1), a marketed intramuscular influenza vaccine (IIV), and a placebo control. This study, through randomized administration of vaccination groups and challenge of all individuals with A/California/2009 H1N1 virus 90+ days post-vaccination, offered a rare opportunity to examine vaccine-elicited immune responses related to definitive outcomes of protection or susceptibility to developing infection in humans. Shared and vaccine-type specific responses across a wide breadth of immune subsets were explored using unsupervised and supervised analysis of the mass cytometry data. We identify previously unknown single and multi-variable cellular correlates of protection following oral vaccination, capable of preventing volunteers from developing virus shedding post-challenge.

Results

Study design

Using samples collected during a previously reported phase 2 clinical study (Liebowitz et al., 2020), peripheral blood was analyzed from 141 healthy volunteers immediately prior to (day 1) and seven days post (day 8) a double-blind randomized administration of an investigational Ad5-based oral tablet influenza vaccine (VXA-A1.1) + placebo intramuscular injection, a licensed intramuscular quadrivalent influenza vaccine (IIV) + placebo tablets, or a placebo-control (placebo tablets and placebo intramuscular injection). To determine vaccine efficacy, volunteers were challenged intranasally between study days 90–120 with a matched A/California/2009 H1N1 virus, monitored for signs and symptoms of disease, and viral shedding was quantified from nasopharyngeal swabs using qRT-PCR (Liebowitz et al., 2020). Individuals were considered to be viral shedders if they tested positive on two or more challenge days by qRT-PCR, the same criteria used previously for laboratory documented infection (Liebowitz et al., 2020). Although VXA-A1.1 offered more protection against virus shedding than IIV (Table 1), VXA-A1.1 elicited lower HAI and neutralizing antibody levels than IIV, suggesting that cellular responses may be of particular importance for the protection mechanism by VXA-A1.1 (Liebowitz et al., 2020).

Table 1.

Shedding outcomes of individuals following influenza challenge with A/California/09 wild-type virus.

Treatment	Shedding Outcome	Count (%)	p value*
VXA-A1.1 (n = 58)	Non-Shedder	38 (65.5%)	0.008
	Shedder	20 (34.5%)
IIV (n = 54)	Non-Shedder	32 (59.3%)	0.044
	Shedder	22 (40.7%)
Placebo (n = 31)	Non-Shedder	11 (35.5%)	-
	Shedder	20 (64.5%)

The number of individuals of each treatment arm who were identified as viral non-shedders and shedders. Viral shedders are individuals whose NP swabs tested positive by qRT-PCR on two or more days throughout the course of the challenge.

^*p values show comparison with placebo (two-tailed fisher’s exact test). This data was originally reported by Liebowitz et al. (Liebowitz et al., 2020).

To identify vaccine type-specific circulating immune cell changes and early cellular correlates of protection from future virus shedding, mass cytometry was applied to fixed whole blood samples stained with a custom 43-antibody panel recognizing surface and intracellular epitopes, including a Y98F hemagglutinin (HA) probe (Whittle et al., 2014) (Fig. 1a, Table S1). This mutant probe allows for detection of influenza virus HA-specific plasmablasts while disrupting the HA sialic acid binding site, reducing nonspecific cell labeling (Whittle et al., 2014). Single-cell data was explored using an unsupervised clustering approach to group all cells from all samples based on marker expression similarity. Dimensionality reduction visualization by uniform manifold approximation and projection (UMAP) revealed distinct groupings for clusters recognizable by expression of canonical markers of major immune cell populations: B cells, T cells, NK cells, and monocytes/dendritic cells (Fig. 1b,c). To resolve additional immune subsets, each of these major immune cell groupings was further sub-clustered (See Unsupervised Single-Cell Analysis in Methods, Figure S2a). We confirmed the presence of minimal batch effects by observing batches to be well-distributed across UMAP plots, and observed low variation for internal controls run across batches (Figure S1). The fidelity of this approach is demonstrated by accurate mapping of annotated cell clusters to corresponding populations defined by supervised hierarchical gating (Fig. S2b–3).

Fig. 1. Study design and visualization of mass cytometry by UMAP.

(a) Study schematic. Whole blood samples were analyzed from healthy volunteers at study day 1 prior to receiving indicated vaccine/placebo and at day 8. Samples were stained with a 43-antibody mass cytometry panel to identify immune cells. Mass cytometry data were clustered, visualized using dimensionality reduction methods, and input into a random forest machine-learning algorithm to define correlates of protection from day 90 virus challenge. (b) All gated CD66-CD45+ cells (n=27M cells) from all samples were clustered. Clusters were manually annotated based on marker expression and visualized by UMAP. Cells initially separated into four major clusters annotated as B cells, T cells, natural killer (NK) cells, and monocytes (MC) and dendritic cells (DC) (center UMAP plot). These groupings were then sub-clustered to allow annotation of additional minor immune subsets (outer UMAP plots). (c) Heatmap showing median expression of canonical phenotyping markers across major cluster groups confirming appropriate assignment of annotations.

Plasmablast correlates of protection following VXA-A1.1 immunization

Among the B cell clusters, a distinct grouping of cells with plasmablast features (high levels of CD27 and CD38 and low levels of CD19 and CD20) were captured by UMAP (Fig. 1b, 2a–c). Cells from volunteers treated with VXA-A1.1 and IIV grouped separately at day 8 (Fig. 2a), but not on day 1 (Fig. S4a), indicating the existence of vaccine-type specific plasmablast responses. Day 8 VXA-A1.1-specific responses occupied UMAP regions with prominent β7 integrin expression and low CD62L expression, consistent with mucosal homing responses triggered by the oral vaccine (Fig. 2a,c) (Nolte and Margadant, 2020; Toapanta et al., 2014). Plasmablasts also segregated due to different levels of several other markers including IgM, IgA, IgG, and phosphorylated STAT5 (pSTAT5) (Rahil et al., 2020), with pSTAT5+ cells appearing to be restricted to a subset of IgG+ plasmablasts (Fig. 2c, S4a–c). Projection of plasmablasts on UMAP plots pointed towards distinct features between viral shedders versus non-shedders which may be associated with protection (Fig. 2b). On day 8, but not day 1 (Fig. S4a–c), cells from VXA-A1.1 non-shedders were concentrated in regions with high β7 integrin, high pSTAT5, and low CD62L (Fig. 2b,c). Consistently, normalized day 8 abundances of manually gated α4+β7+ (α4β7+) plasmablasts, CD62L− plasmablasts, pSTAT5+ plasmablasts, HA+ plasmablasts, and total plasmablasts (Fig S3a) were each significantly correlated with protection from virus shedding in VXA-A1.1 recipients, but not IIV or placebo groups (Fig. 2d).

Fig. 2: Plasmablast responses correlate with protection for VXA-A1.1.

(a-c) UMAP of plasmablasts at (a) day 8, colored by treatment arm (VXA-A1.1, IIV, or Placebo), (b) day 8, colored by shedding outcome (shedder, non-shedder) for each treatment arm, or (c) by expression values for HA reagent, IgA, IgG, IgM, pSTAT5, β7 integrin, or CD62L. For (a-b), plots are overlaid on all other cells in the plasmablast cluster from day 1 and 8 (grey). (d) Box plots display baseline normalized day 8 abundance for indicated gated cell populations in each treatment arm (placebo, green; IIV, blue, VXA-A1.1, orange). Dark-colored boxes show non-shedders, light-colored boxes show shedders. (p-values were FDR-adjusted using the Benjamini-Hochberg method; n.s. = not significant; *,**,***,**** = p-values of <0.05, 0.01, 0.001, 0.0001 respectively; ANOVA and Tukey HSD were used to evaluate differences in means between treatment groups and a Wilcoxon signed-rank test was used to compare shedders and non-shedders within treatment groups; Box plots: upper, lower, and center box lines represent upper quartile, lower quartile, and mean, respectively. Whiskers represent 1.5x IQR. n = number of individuals). (e) Heatmap showing baseline normalized abundances of β1 and β7 integrin subsets of gated B cell populations. CSM (CD27+IgM−), Naïve (CD27−IgM+), NCSM (CD27+IgM+). See Figure S3a–e for gating strategy and S3f,g for HA probe specificity.

Although the overall abundance of total plasmablasts and a subset of HA+ plasmablasts were greater at day 8 for IIV than VXA-A1.1 or placebo groups, neither feature was significantly correlated with protection following IIV vaccination (Fig. 2d). Across multiple B cell subsets, activation (CD71) and proliferation (Ki67) markers trended upwards at day 8 for both vaccines, along with opposing patterns of elevated β1 and β7 integrin expression for IIV and VXA-A1.1, respectively (Fig. 2e, Fig S3d). Importantly, an α4β7+ gate applied to total B cells (including plasmablasts) was significantly associated with protection for VXA-A1.1 recipients, indicating that this correlate of protection may be cheaply and easily quantifiable in future studies with the use of a small, fluorescent antibody panel. For VXA-A1.1 recipient non-shedders, the majority of cells in this gate on day 8 were plasmablasts (Figure S4d).

Vaccine elicited T cell responses for VXA-A1.1 and IIV

A cluster of T cells with CD4+ memory features (high levels of CD3, CD4, and CD45RO, and low levels of CD45RA) was also examined in depth by unsupervised clustering followed by UMAP dimensionality reduction (Fig. 3a–c). Similar to plasmablast results, cells from volunteers immunized with VXA-A1.1 and IIV grouped separately at day 8 (Fig. 3a), but not day 1 (Fig. S4e–g), indicating the presence of vaccine-type specific T cell responses. Both vaccines, but not placebo, elicited cells on day 8 expressing ICOS, CD38, and PD-1, characteristics of T follicular helper (Tfh) cells which play a central role in generating influenza antibody responses (Bentebibel et al., 2013; Herati et al., 2014, 2017; Ueno, 2019). However, unlike IIV, VXA-A1.1-specific responses occupied UMAP regions with high β7 integrin and CCR9 expression, indicative of enhanced mucosal homing (Fig. 3a–c). Manual gating confirmed significant increases in the abundance of day 8 total aTfh cells (CD3+CD7−CD4+CD38+ICOS+) for both VXA-A1.1 and IIV relative to placebo and an expansion of α4β7+ Tfh cells in VXA-A1.1 recipients (Fig. 3d, Fig S3a).

Fig. 3: Differential T cell responses to VXA-A1.1 and IIV.

(a-c) UMAP of CD4 memory T cells at (a) day 8, colored by treatment arm (VXA-A1.1, IIV, or Placebo), (b) day 8, colored by shedding outcome (shedder, non-shedder) for each treatment arm, or (c) by expression values for HLA-DR, Ki67, CD39, PD-1, CD161, β7, or CCR9. For (a-b), plots are overlaid on all other cells in the CD4 memory T cell cluster from day 1 and 8 (grey). (d) Box plots display baseline normalized day 8 abundance for indicated gated cell populations in each treatment arm (placebo, green; IIV, blue, VXA-A1.1, orange). Dark-colored boxes show non-shedders, light-colored boxes show shedders. (p-values were FDR-adjusted using the Benjamini-Hochberg method; n.s. = not significant; *,**,***,**** = p-values of <0.05, 0.01, 0.001, 0.0001 respectively; ANOVA and Tukey HSD were used to evaluate differences in means between treatment groups and a Wilcoxon signed-rank test was used to compare shedders and non-shedders within treatment groups; Box plots: upper, lower, and center box lines represent upper quartile, lower quartile, and mean, respectively. Whiskers represent 1.5x IQR. n = number of individuals). (e) Heatmap showing abundance normalized β1 and β7 integrin subsets of gated CD4 T cell populations (upper panels) or CD8 T cell populations (lower panels). aTFH (CD3+CD7−CD4+CD38+ICOS+), CM (CD62L+CD45RA−), EM (CD62L−CD45RA−) Naïve (CD62L+CD45RA+), Exhausted (CD8+PD-1+CD39+), TEMRA (CD62L−CD45RA+). See Figure S3a–e for gating strategy.

Across CD4 T cell subsets, mutually exclusive trends of elevated β1 or β7 integrin following IIV or VXA-A1.1 respectively were evident for Tfh, central memory, and effector memory cells (Fig. 3e). Markers of activation/proliferation also tended to be elevated across these cell types post-vaccination, most prominently for IIV where a subset of gated CD27+ CD4 central memory T cells expressing CD38 and Ki67 was significantly increased at day 8 post-vaccination with IIV but not VXA-A1.1 or placebo (Fig. 3d, Fig S3a–c). β7 integrin and activation/proliferation markers also tended to increase among central end effector CD8 T cell subsets for VXA-A1.1 recipients, reflected by a significant increase in the relative abundance of α4β7+ CD8 effector memory cells for VXA-A1.1 compared to placebo and IIV recipients on day 8. Although variations in levels of T cell features discussed above were evident between shedder and non-shedder groups, neither of these T cell features met the criteria for univariate statistical significance after multiple hypothesis correction (Fig. 3d). Statistics for all cell subsets defined by manual gating and clustering are listed in Supplementary Tables S2 and S3.

Random forest models predict VXA-A1.1 vaccine conferred protection

To explore the full extent of mass cytometry data collected, we created networks showing marker expression similarity between cell subsets (clustering and manual gating) and the relationship of each cell subset to virus shedding for different vaccination groups (Fig. 4a; S5a,c). These plots support the notion of the greater importance of cellular immune responses for protection following VXA compared to IIV and suggest that a variety of cell subsets may contribute to VXA conferred protection. Because of the wide breadth of these responses, we set out to define a high-dimensional cellular correlate of protection that would make use of the entire dataset to generate random forest classifiers for each treatment group to predict future virus shedding outcomes from data seven days post-vaccination. To establish the robustness of this approach, we created and compared the performance of separate models using cell abundance data from either unsupervised clustering or manual gating. To enrich features dynamically responding to vaccination, we first calculated the magnitude of differences in all features between day 1 and day 8 and selected for model building only those whose fold difference was in the upper or lower quartile of change within each vaccine/placebo group and irrespective of virus shedding. For each treatment group, a matrix of features and future viral shedding outcomes was used to create random forest classifiers through bootstrapping (by randomly sampling the data) to increase robustness. We report the performance of these models as the median test set value from 100 repetitions. The model-predicted shedding probabilities were compared to the true shedding outcomes using a Wilcoxon Test and by calculating the area under the receiver operating characteristic curve (AUC).

Fig. 4: Random forest models predict VXA-A1.1 vaccine conferred protection.

(a) Correlation networks depicting marker similarities between cell subsets and the predictive value of cell subsets for distinguishing shedding for IIV (left) and VXA-A1.1 (right) (nodes = unsupervised cell clusters; node diameter = −log₁₀ FDR adjusted Wilcoxon p-values between subset abundances of shedders versus non-shedders; edge weight = spearman correlation of marker similarities between subsets). (b) Performance of random forest models built using unsupervised clusters to predict virus shedding for IIV recipients (left) or VXA-A1.1 recipients (right). Box plots show the distribution of model coefficients (model prediction) for non-shedder and shedder groups with shedding probabilities compared to the true shedding outcomes using a Wilcoxon Test (left, center right). ROC plots display model performance with inlaid area under the curve (AUC) (center left, right). n = number of individuals (c) Correlation plots showing relationships between random forest model coefficients, previously collected immunogenicity data, and challenge outcome data for IIV recipients (left) and VXA-A1.1 recipients (right). Gated random forest model (Gated RF), cell cluster random forest model (Cluster RF), day 8 T cell ELISPOT (ELISPOT GrzB, ELISPOT IFNγ), day 8 ASC ELISPOT (ELISPOT IgA, ELISPOT IgG), day 90/day 1 microneutralization titer (MN), day 90/day 1 hemagglutination inhibition titer (HAI), maximum post-challenge viral titer by qRT-PCR (Viral load), area under the curve for the number of post-challenge symptoms by Flu-PRO (Symptoms). TPR, true positive rate. FPR, false-positive rate. Spearman correlation values with p-values less than 0.01 following FDR adjustment are displayed.

Random forest models of the VXA-A1.1 treated group could distinguish those individuals who were later protected versus those who remained susceptible to virus using datasets from both unsupervised clustering approaches or manual gating (p-value of 0.00001 and AUC of 0.86 for unsupervised clustering, p-value of 0.00002 and AUC 0.84 for manual gates, both subject to cross-validation) (Fig. 4a,b; S5a,b). Model specificity was demonstrated by poor prediction of virus shedding for models built using placebo group data (Fig. S5c–e). For a visualization of the univariate predictive value of cell subsets for distinguishing shedding, see Fig. 4a; S5a,c. For the ranked importance of features driving random forest predictions, see Tables S2 and S3. Interestingly, the same approach was unable to significantly predict outcomes for IIV recipients using either version of the mass cytometry dataset, underscoring the greater importance of the measured cellular responses for the protection offered from VXA-A1.1 vs. IIV (Fig. 4a,b; Fig. S5a,b).

We examined these results in relation to previously reported immunogenicity data from this study including T cell and B cell ELISPOT, MN, and HAI assays (Liebowitz et al., 2020). Consistently for IIV, previously collected outcome data for post-challenge symptoms and post-challenge peak virus levels correlated with HAI titer; however, interestingly for VXA-A1.1, these outcome data correlated with cellular random forest models (Fig. 4c). We also tested whether HAI and MN titer outcomes could be predicted based on cellular responses using a random forest approach similar to what we used to predict shedding outcomes. We found that the likelihood of VXA-A1.1 recipients having a greater than 4 four-fold rise in HAI or MN titer by day 90 could be predicted using a random forest model built on manually gated cellular data (p-value of 1.2×10⁻³ and AUC of 0.76 for HAI; p-value of 5.2×10⁻⁴ and AUC of 0.77 for MN) (Fig. S5f–h). Consistently, among VXA-A1.1 non-shedders, normalized day 8 HA+ plasmablast abundance was also correlated with normalized HAI (r=0.61, p=1.40×10⁻²) and MN titers (r=0.72, p=4.07×10⁻⁴) (Table S2).

Discussion

This study examines vaccine responses in a human influenza challenge study using mass cytometry. Guided by inspection of UMAP plots, analysis of data generated from samples pre/post-vaccination, with IIV, VXA-A1.1, or placebo, identified shared and vaccine-type-specific T cell, B cell, and plasmablast subset responses to vaccination. Among these responses, we determine that the abundance of α4β7+, CD62L−, pSTAT5+, HA+, and total plasmablast subsets are each significantly correlated with protection from developing viral shedding post-challenge in oral vaccine recipients, but not in intramuscular vaccine or placebo recipients when challenged after day 90 with A/California/2009 H1N1 virus.

Prior studies established that the Ad5-based oral vaccine platform elicits plasmablasts with mucosal homing properties including α4β7 expression (Kim et al., 2016; Liebowitz et al., 2020). The current study suggests that these cells are indeed correlates of protection and can be defined by either their expression of α4β7 or by their downregulation of CD62L, an effect that enhances mucosal homing (Brandtzaeg and Johansen, 2005; Nolte and Margadant, 2020). This data aligns well with the association between oral polio vaccine-elicited α4β7 expressing, circulating ASCs, and protection from poliovirus (Dey et al., 2016) and more generally with the utility of circulating ASC levels to determine the level of vaccine response (Fink, 2012). Patterns of enhanced mucosal homing and activation markers observed in this study across various circulating B cell subsets may point to additional targets to evaluate vaccine response.

Elevated levels of phosphorylated STAT5 (pSTAT5) in plasmablasts following vaccination have not been previously reported, however, an upregulation in pSTAT5+ plasmablasts was observed in a prior H1N1 virus challenge study in humans, suggesting similar responses exist between vaccination and H1N1 infection (Rahil et al., 2020). B cell STAT3 and STAT5 phosphorylation can be activated in vitro by cytokines supporting their differentiation including IL-2, IL-4, IL-7, and IL-21 (Berglund et al., 2013; Kwakkenbos et al., 2016). In a recent study differentiation of B cells in vitro via a combination of BCR crosslinking, CD40L, IL-4, and IL-21 resulted in early induction of pSTAT5+ in a CD27+CD38+ ASC population (Marsman et al., 2020). pSTAT5 has been variably shown to either promote or repress BCL-6, an important survival factor for B cells (Diehl et al., 2008; Heizmann et al., 2020; Scheeren et al., 2005; Walker et al., 2007). The role of plasmablast pSTAT5 in vivo remains to be defined, but the elevated presence of pSTAT5+ plasmablasts following immunization and their correlation with VXA-A1.1 protection suggests that pSTAT5 may be an important marker of plasmablast function.

Abundances of multiple T cell subsets increased significantly post-vaccination in both IIV and VXA-A1.1 vaccines, with the expansion of α4β7+ populations restricted to VXA-A1.1 recipients. Tfh cells play a critical role in class switching antibody responses in germinal centers and were found to be expanded after both vaccines. Expansion of the α4β7+ and CCR9+ Tfh cells specifically following VXA-A1.1 suggests effective mucosal priming of these immune responses during oral vaccination (Bentebibel et al., 2013; Herati et al., 2014; Holmgren and Czerkinsky, 2005). This priming appears to extend to cytotoxic T cells with α4β7+ CD8 effector memory T cells, also specifically elevated following VXA-A1.1. CD8 T cell responses are critical for protection from severe disease during influenza infection, but the importance of CD8 T cell responses following vaccination is less understood (Korenkov et al., 2018; Koutsakos et al., 2019). Although none of these T cell populations met significance thresholds for univariate correlation with protection, T cell subsets meaningfully contributed to multivariate models. Future studies using peptide restimulation or peptide MHC multimer binding to distinguish antigen specificity among vaccine-elicited T cell subsets may help define additional correlates of protection. While peripheral blood assays are highly desirable from an accessibility standpoint, investigating cellular responses in the gut and respiratory mucosal sites may be especially relevant for better understanding mechanisms of protection conferred by oral vaccination.

Maximizing the utility of the high parameter mass cytometry data, we successfully defined a robust random forest-based multivariate cellular correlate of protection for VXA-A1.1 using both manual gating and unsupervised clustering approaches. Consistently, the top 10 features of greatest importance for driving the random forest prediction in both manual gating and unsupervised clustering approaches included subsets of B cells, plasmablasts, NK/NKT cells, intermediate monocytes, and CD4+ T cells (see Tables S2,3). These included specific pSTAT5+ and CD62L− plasmablast populations explored earlier in the manuscript. Application of this model to future cohorts will help establish whether this multivariate approach can provide a stronger indication of protection than individual features alone and whether this model may be extendable to additional Ad5 based oral vaccine targets.

Our results are consistent with the notion that circulating antibody responses are more relevant for IIV-protection, whereas circulating cellular responses are more relevant for VXA-A1.1-protection (Liebowitz et al., 2020). Unlike VXA-A1.1, neither individual cellular features nor multivariate cellular random forest models correlated with IIV protection. When examining previously collected immunogenicity data from this study, outcomes of post-challenge symptoms and post-challenge peak virus levels were significantly correlated with HAI titer for IIV, whereas these same outcomes were correlated with cellular random forest models for VXA-A1.1. Furthermore, specific cellular responses and random forest models were predictive of HAI and MN titers in VXA-A1.1 but not IIV recipients. Bringing the next generation of influenza vaccines to market requires an improved understanding of how vaccine-conferred cellular responses contribute to immunity. Using mass cytometry, we identify cellular correlates of protection for an investigational oral influenza vaccine with demonstrated phase 2 efficacy. These correlates of protection can be measured in blood seven days post-vaccination, significantly earlier than the classic markers of immunity such as HAI or MN typically measured after four or more weeks.

Limitations of the study

The custom-designed mass cytometry panel used here may be better at detecting cellular changes resulting from VXA-A1.1 than IIV, and although we report data from the largest influenza virus challenge study on record, the ability of these correlates to extend to community-acquired infections with greater diversity in demographics, levels of pre-existing immunity, and virus strain will require examination of significantly larger cohorts. We profile samples from a study where vaccination provided significant protection from virus shedding, but not from influenza positive illness, an outcome incorporating virus both shedding and symptom reports (Liebowitz et al., 2020). Future studies are needed to determine whether or not the cellular correlates of protection from virus shedding identified here may extend to correlates significant protection from symptomatic shedding.

STAR Methods

Resource Availability

Lead Contact

Further information and requests for resources and reagents should be directed to David McIlwain (mcilwain@stanford.edu).

Materials Availability

This study did not generate new unique reagents.

Data and Code Availability

• Data reported in this paper will be shared by the lead contact upon reasonable request

• This paper does not report original code

• Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon reasonable request.

Experimental Model and Subject Details

Human subjects and clinical trial

The aim of this study was to investigate cellular correlates of protection from a previously reported double-blinded, placebo-controlled and active-controlled phase 2 clinical trial that evaluated the safety and efficacy of an oral tablet H1N1 HA-Adenoviral-vector based seasonal influenza vaccine with a dsRNA adjuvant (VXA-A1.1) compared to a licensed quadrivalent inactivated influenza vaccine (IIV, Fluzone 2015–2016 year formulation) administered via intramuscular injection (Liebowitz et al., 2020). Also see details provided at ClinicalTrials.gov, identifier NCT02918006. Briefly, a total of 179 healthy adult volunteers provided informed consent and were enrolled at WCCT Global’s challenge facility in Costa Mesa, CA and randomized to receive one of three treatment arms: 1) a single dose of an IM IIV vaccine and an oral placebo, 2) a single dose of VXA-A1.1 and a placebo IM injection, or 3) a placebo IM injection and an oral placebo. At 90 (or up to 120 days) days post-vaccination, subjects were challenged with influenza virus. Following influenza challenge, subjects remained in an isolation unit for 6 to 9 days, where they were monitored for influenza signs and symptoms. Vital signs were measured and collected every four hours during waking hours, and symptomatic responses were assessed using the influenza patient-reported outcome questionnaire (Flu-PRO). Nasopharyngeal swabs were collected twice daily post-challenge and were evaluated for viral shedding by qRT-PCR at the University of Cincinnati Children’s Hospital, Cincinnati, OH, USA. The same endpoint definition from Liebowitz et al. for ‘laboratory documented infection’ was used to define viral shedders in this study (Liebowitz et al., 2020). ‘Virus shedders’ were defined as participants with laboratory-documented infection with two or more positive results for influenza shedding by qRT--PCR, individuals not meeting these criteria were classified as ‘non-shedders’ (Liebowitz et al., 2020).

Information regarding volunteer demographics, enrollment criteria, and study design have been previously reported (Liebowitz et al., 2020). Specifically, Liebowitz et al. report sex, age, and race/ethnicity of the volunteer study cohort. Enrollment was restricted to healthy volunteers who met eligibility criteria including HAI titer <1:20 (Liebowitz et al., 2020). Volunteers were not involved in previous procedures and were naïve to the therapeutics being tested upon enrollment (Liebowitz et al., 2020).

Viral challenge

Influenza virus challenge was conducted using a matched wild-type strain of A/California/2009 influenza that has 99.4% and 99.6% hemagglutinin cDNA identity to A/California/04/2009 (H1N1) and A/California/07/2009 (H1N1), respectively (Watson et al., 2015). Note: VXA-A1.1 contains a hemagglutinin sequence for A/California/04/2009 (H1N1) (Liebowitz et al., 2015) and the IIV used contained A/California/07/2009 (H1N1) (Liebowitz et al., 2020).

Method Details

Mass cytometry

Blood samples collected at day 1 (immediately prior to vaccination) and day 8 (7 days post-vaccination) by routine phlebotomy from volunteers who completed both the vaccination and influenza virus challenge phases of the study were used for mass cytometry (n=58 VXA-A1.1, n=54 IIV, and n=31 placebo; total n=143 volunteers, Table 1). Of the 143 individuals, one individual withdrew from the study following virus challenge (Liebowitz et al., 2020) and samples from a single volunteer (C117) were collected but excluded from analysis due to high baseline plasmablast levels unrelated to vaccine treatment, resulting in a total of 141 individuals analyzed by mass cytometry.

For initial processing, 750uL of heparinized blood was mixed with 1050uL of Smart Tube Proteomic Stabilizer (Smart Tube Inc., cat #PROT1), incubated for 11 minutes at room temperature, then flash-frozen on dry ice and stored at −80°C. Frozen samples were shipped on dry ice for subsequent analysis.

Samples were transferred from −80°C storage in an ice water bath and resuspended and washed twice in 1X Thaw-Lyse Buffer (Smart Tube Inc., cat # THAWLYSE1) for red blood cell lysis. Following red blood cell lysis, cells were counted, and 1.5×10⁶ cells were manually arranged in a 96-well block. Samples were barcoded in a 20-plex scheme with palladium metal by a robotics system that was previously described (Bjornson-Hooper et al., 2019; Zunder et al., 2015). Subsequent staining steps on the barcoded samples were performed manually. Barcoded cells were treated at room temperature with Fc-block (Human TruStain FcX, Biolegend cat# 422302) for 10 minutes and followed by an additional 30-minute incubation with biotinylated recombinant hemagglutinin (HA) from the A/California/07/2009/H1N1 strain (mutation Y98F) (a gift from Barney Graham VRC/NIAID/NIH). Samples were then washed once with cell staining medium (PBS with 0.5% BSA and 0.02% sodium azide) and stained with surface antibodies (Table S1) for 30 minutes in cell staining media. Following surface antibody staining, cells were permeabilized with ice-cold 100% methanol (Thermo Fisher, cat# A412–4), washed, and stained with intracellular antibodies (Table S1) for 60 minutes. After intracellular staining, cells were washed and resuspended in a solution containing iridium intercalator (Fluidigm, cat # 201192B) and 1.6% paraformaldehyde (Thermo Fisher, cat# 50–980-487). Prior to sample analysis on the mass cytometer, samples were washed, resuspended in 1X four-element normalization beads (140/142Ce, 151/153Eu, 165Ho, 175/176Lu) (Fluidigm, cat# 201078). Collected data were normalized across all barcoded samples and debarcoded as previously described (Finck et al., 2013). Samples were thawed in 12 batches of up to 72 samples each and stained in plates as described above. Sets of up to 20 samples were barcoded and pooled into single tubes for CyTOF analysis also as described above. Day 1 and day 8 samples from the same donor were always included in the same barcoded tube along with a control aliquot created from blood drawn from a single healthy volunteer for identification of possible plate-specific batch effects (Table S2,3,4; Fig. S1).

Unsupervised single-cell analysis

Normalized and debarcoded FCS files were uploaded to https://cellengine.com (Primity Bio) for analysis. Normalization beads were gated out from downstream analysis by their co-expressions of Eu140, Eu151, Lu175, and Ho165. Red blood cells and granulocytes were excluded by their positive expression of high CD235a and CD66, respectively. Mononuclear leukocytes were identified by their expression of CD45 and low CD66 and were exported from donors (27 million events) at day 1 and day 8 and pooled together for clustering and UMAP embedding. Marker expressions were arcsinh transformed with a cofactor of 5, scaled between the 1st to 99th percentiles, then normalized on a 0 to 1 scale. Cells were clustered by their expressions of all available markers using FastPG (Bodenheimer et al., 2020), and clusters were annotated using canonical phenotypic markers into major cell types. In the first iteration of clustering, clusters were broadly identified as predominantly containing B cells (CD19+CD20+CD3−CD7−), T cells (CD3+CD19−CD20−), NK cells (CD3−CD7+CD19−CD20−), dendritic cells, and monocytes (combinations of HLA-DR+, CD14+, CD16, CD11c+, CD123+), or an unclassified group with inconclusive marker expressions (Fig. 1b,c). Each of these primary clusters were subsequently re-clustered using the same process to generate secondary cluster subsets and annotated as described below.

Secondary cluster subsets were manually annotated based on putative cell type composition using canonical marker expressions visualized in Fig. S2a. Briefly, within the B cell cluster, subset annotations were driven by the expression of CD27 and CD38 (Plasmablasts), IgA, IgG, IgM, or an unresolved B cell subset. T cell annotations were first based on CD4 (Th) or CD8 (Tc) expression, then by CD45RA and CD45RO for naive and memory subset discrimination, respectively. Eight groupings of subsets were evident in the T cell cluster: Tc memory (CD8+CD45RO+), Tc naive (CD8+CD45RA−), Tc RO−RA− (CD8+CD45RA−CD45RA−), Th memory (CD4+CD45RO+), Th naive (CD4+CD45RA+), Th RO+RA+ (CD4+CD45RA+CD45RO+), CD4−CD8−, and CD4+CD8+ cells. NK cluster subsets were characterized by their expressions of CD161, CD16, CD56, Ki67, or an inconclusive NK cell subset. Within the monocyte and dendritic cell cluster group, monocyte subsets were identified as classical monocytes (cMC, HLA-DR+CD14+CD16−), non-classical monocytes (ncMC, HLA-DR+CD14−CD16+), or intermediate monocytes (intMC, HLA-DR+CD14+CD16+). Myeloid dendritic cell (mDC) and plasmacytoid dendritic cell (pDC) subsets were classified by the expression of HLA-DR and by their expressions of CD11c and CD123, respectively, while a basophil cluster was identified by lack of HLA-DR, but high expression of CD123. Each secondary cluster subset was further reclustered to generate numbered tertiary cluster subsets used for downstream analysis (Table S3).

Cells and clusters were visualized on uniform manifold approximation and projection (UMAP) (Becht et al., 2019) embeddings generated using UWOT, the R implementation. Heatmaps of marker expressions were created using the heatmap.2 function from the gplots package (Warnes et al., 2020).

Manual gating

Cell Engine software (Primity Bio) was used for manual gating. After exclusion of beads, red blood cells, and granulocytes, and positive selection of mononuclear leukocytes, populations were manually identified as depicted in gating strategies (Fig. S3). Abundances of each population were exported from Cell Engine and enumerated as a percentage of the gated CD66-CD45+ population.

Random forest classifier

The probability of each donor being a viral shedder or non-shedder was calculated based on a random forest classifier using cell abundance data from either unsupervised clustering or manual gating population as features. For each treatment arm, features that were dynamically responding to vaccination were enriched by calculating the magnitude of differences in features prior to and after vaccination, and post-vaccination features whose fold differences were in the upper or lower quartile were selected for the generation of a random forest classifier (Table S2,3). Bootstrapping was implemented by randomly pulling approximately 64% of the data as the training group and applying the classifier to the remaining 36% testing data (Efron and Tibshirani, 1997; Stanley et al., 2020) to predict a patient’s shedding outcome based on their vaccine response 7 days post-vaccination. The performance is reported as the median value from 100 repetitions and was evaluated by calculating the true positive rates, false-positive rates, and the AUC of the resulting ROC as previously described (Robin et al., 2011). The predicted shedding probabilities were then compared to the true patient shedding outcomes post-virus challenge using a Wilcoxon test. For each of the random forest iterations, the importance of a given feature that contributed to the model was extracted using the importance function within the randomForest package in R, and the final importance of the feature was calculated as the mean importance across all repetitions. For random forest prediction of HAI and MN titer, we used the same approach, but trained and tested models for the indicated binary outcomes of HAI/MN titer level changes rather than binary shedding outcomes.

Quantification and Statistical Analysis

All statistical details can be found in figures and figure legends. Unless otherwise stated, plots and statistical analysis relate to the following group sizes: VXA-A1.1 n = 58 (20 shedders, 38 non-shedders); IIV n = 53 (22 shedders, 31 non-shedders); placebo n = 31 (20 shedders, 11 non-shedders). n = number of individuals.

Baseline normalization and comparison between treatment arms and shedding outcomes

To investigate the differences between treatment groups and between shedders and non-shedder, each individual was normalized to their own day 1 baseline. Abundances of unsupervised cell clusters and manually gated populations at day 1 and day 8 were first log10(x+1)-transformed to unskew the data, and differences between day 8 and day 1 were calculated and plotted. A one-way analysis of variance (ANOVA) was performed to compare the day 8 normalized abundances of clusters and populations between VXA-A1.1, IIV, and placebo groups. A Tukey honest significant difference (HSD) test was then performed to identify significant differences between treatment groups. Within a given treatment group, differences between shedders and non-shedders were evaluated using a one-sided Wilcoxon signed-rank test. All p-values were corrected using the Benjamini-Hochberg (BH) procedure to control for the false discovery rate (FDR), and adjusted p-values less than 0.05 were considered significant for all statistical tests.

Correlation networks

Correlation networks were generated in R using Rtsne with cell subsets (clusters or manual gates) as nodes, node diameter proportional to -log10 FDR corrected Wilcoxon p-values for differences in normalized cell subset abundance between shedders and non-shedders, and edges proportional to Spearman correlation of marker similarities between cell subsets.

Box Plots

In all box plots, the center line represents the mean value; upper and lower box limits, upper and lower quartiles, respectively; whiskers, 1.5x interquartile range; and points, all data points. Statistical comparisons were performed as indicated in figure legends, and results are displayed above box plots.

Additional Resources

Clinical Registry

This study investigates cellular correlates of protection using samples from a previously reported clinical trial (Liebowitz et al., 2020). Details of that clinical trial are provided at ClinicalTrials.gov under identifier NCT02918006.

KEY RESOURCES TABLE.

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Mouse Monoclonal Anti-CD38	Biolegend	Biolegend Cat#303535, RRID:AB_2562819
Mouse Monoclonal Anti-CD45	Biolegend	Biolegend Cat#304045, RRID:AB_2562821
Mouse Monoclonal Anti-CD25	Biolegend	Biolegend Cat#356102, RRID:AB_2561752
Mouse Monoclonal Anti-STAT3	BD Biosciences	BD Biosciences Cat#610926, RRID:AB_398241
Rat Monoclonal Anti-CD66	Invitrogen	Invitrogen Cat#MA5-17003, RRID:AB_2538475
Mouse Monoclonal Anti-IkBa	CST	CST Cat#4814, RRID:AB_390781
Mouse Monoclonal Anti-CD11c	Biolegend	Biolegend Cat#337202, RRID:AB_1236381
Mouse Monoclonal Anti-CD27	Biolegend	Biolegend Cat#356401, RRID:AB_2561786
Mouse Monoclonal Anti-Biotin	Biolegend	Biolegend Cat#409006, RRID:AB_10695474
Mouse Monoclonal Anti-CD161	Biolegend	Biolegend Cat#339938, RRID:AB_2564141
Mouse Monoclonal Anti-CCR9 (CDw199)	Biolegend	Biolegend Cat#358902, RRID:AB_2562298
Mouse Monoclonal Anti-CD20	Biolegend	Biolegend Cat#302302, RRID:AB_314250
Mouse Monoclonal Anti-Granzyme B	BD Biosciences	BD Biosciences, Custom Purified
Mouse Monoclonal Anti-STAT6	BD Biosciences	BD Biosciences Cat#611567, RRID:AB_399013
Mouse Monoclonal Anti-CD24	Biolegend	Biolegend Cat#311102, RRID:AB_314851
Mouse Monoclonal Anti-CD14	Biolegend	Biolegend Cat#301802, RRID:AB_314183
Mouse Monoclonal Anti-PD-1 (CD279)	BD Biosciences	BD Biosciences Cat#562138, RRID:AB_10897007
Mouse Monoclonal Anti-Alpha4 (CD49d)	BD Biosciences	BD Biosciences Cat#555502, RRID:AB_395892
Rat Monoclonal Anti-Beta7	BD Biosciences	BD Biosciences Cat#555943, RRID:AB_396240
Mouse Monoclonal Anti-CD8	Biolegend	Biolegend Cat#301053, RRID:AB_2562810
Mouse Monoclonal Anti-CD4	Biolegend	Biolegend Cat#317402, RRID:AB_571963
Mouse Monoclonal Anti-CD3	BD Biosciences	BD Biosciences Cat#551916, RRID:AB_394293
Mouse Monoclonal Anti-IgA	Biolegend	Biolegend Cat#411502, RRID:AB_2650697
Mouse Monoclonal Anti-CD7	BD Biosciences	BD Biosciences Cat#555359, RRID:AB_395762
Mouse Monoclonal Anti-CD28	BD Biosciences	BD Biosciences Cat#348040, RRID:AB_400367
Mouse Monoclonal Anti-CD39	Biolegend	Biolegend Cat#328202, RRID:AB_940438
Mouse Monoclonal Anti-CD123	BD Biosciences	BD Biosciences Cat#554527, RRID:AB_395455
Mouse Monoclonal Anti-CD45RO	Biolegend	Biolegend Cat#304202, RRID:AB_314418
Mouse Monoclonal Anti-CD16	Biolegend	Biolegend Cat#302051, RRID:AB_2562814
Mouse Monoclonal Anti-STAT5	BD Biosciences	BD Biosciences Cat#611965, RRID:AB_399386
Mouse Monoclonal Anti-CD45RA	Biolegend	Biolegend Cat#304143, RRID:AB_2562822
Rat Monoclonal Anti-Fox P3	Thermo Fisher	Thermo Fisher Cat#14-4776-82, RRID:AB_467554
Mouse Monoclonal Anti-Beta1 (CD29)	BD Biosciences	BD Biosciences Cat#555442, RRID:AB_395835
Rat Monoclonal Anti-Ki67	Thermo Fisher	Thermo Fisher Cat#14-5698-82, RRID:AB_10854564
Mouse Monoclonal Anti-IgG	Milipore Sigma	Milipore Sigma Cat#CBL101, RRID:AB_92830
Armenian Hamster Anti-ICOS (CD278)	Biolegend	Biolegend Cat#313502, RRID:AB_416326
Mouse Monoclonal Anti-CD62L	Thermo Fisher	Thermo Fisher Cat#BMS1015, RRID:AB_10596353
Mouse Monoclonal Anti-CD19	Beckman-Coulter	Beckman-Coulter, Custom Order
Mouse Monoclonal Anti-IgM	BD Biosciences	BD Biosciences Cat#555780, RRID:AB_396115
Mouse Monoclonal Anti-CD56	BD Biosciences	BD Biosciences Cat#559043, RRID:AB_397180
Mouse Monoclonal Anti-HLA-DR	Beckman-Coulter	Beckman-Coulter, Custom Order
Mouse Monoclonal Anti-CD235	Biolegend	Biolegend Cat#306615, RRID:AB_2562825
Mouse Monoclonal Anti-CD71	BioXcell	BioXcell Cat#BE0023, RRID:AB_1107669
Biological samples
Peripheral blood from participants of: Phase 2 Influenza A Challenge Study Following Oral Administration of an H1N1 HA Ad-Vector Seasonal Flu Vaccine	Liebowitz et al., 2020	ClinicalTrials.gov identifier NCT02918006
Chemicals, peptides, and recombinant proteins
Biotinylated recombinant hemagglutinin A/California/07/2009/H1N1 (mutation Y98F)	Gift from Barney Graham VRC/NIAID/NIH	Gift from Barney Graham VRC/NIAID/NIH
Critical commercial assays
Smart Tube Proteomic Stabilizer	Smart Tube Inc.	Cat#PROT1
10X Thaw-Lyse Buffer	Smart Tube Inc.	Cat#THAWLYSE1
Fc Block: Human TruStain FcX	Biolegend	Cat#422302
100% Methanol	Thermo Fisher	Cat#50-980-487
Iridium DNA Intercalator	Fluidigm	Cat#201192B
16% Paraformaldehyde	Thermo Fisher	Cat#50-980-487
Four Element Normalization Beads	Fluidigm	Cat#201078
Software and algorithms
Cell Engine	Primity Bio	https://cellengine.com
Single Cell Debarcoder	Nolan Lab	https://github.com/nolanlab/single-cell-debarcoder
Bead Normalization	Nolan Lab	https://github.com/nolanlab/bead-normalization
R	R Foundation	https://r-project.org
Fast PG	Bodenheimer et al., 2020	https://github.com/sararselitsky/FastPG
UWOT (v0.1.9)	Becht et al., 2018	https://github.com/jlmelville/uwot

Highlights.

• Mass cytometry characterizes immune responses elicited by influenza vaccination

• VXA-A1.1 elicits mucosal homing cell subsets that correlate with protection

• Random forest models can predict VXA-A1.1 protection from influenza challenge

Data Availability

Raw data, processed data, and source code for the reproduction of the results are available upon request.