Prediction of SARS-CoV-2 exposure through T cell activation profiles

Abstract

In the ongoing effort to study the SARS-CoV-2 virus and COVID-19 disease, assessment of the T cell immune response has guided vaccine and therapeutic development. T cell immunity, as measured by T cell activation, is commonly assessed by the activation induced marker (AIM) assay. However, no concerted effort has been made to compare AIM pairs and identify those that best detect T cell activation. The emergence of SARS-CoV-2 provided unique access to both known naïve and COVID-19 convalescent donor samples for the comparison of 35 unique T cell AIM pairs using a 15-marker flow cytometry panel. Detailed comparative analysis identified top performing AIM pairs and informed the development of a machine learning algorithm that predictively classified samples as COVID-19 convalescent or naïve. This approach may be additionally applicable to the assessment and prediction of T cell antigen reactivity in cancer and autoimmune disease.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-48323-7.

Introduction

Despite vast efforts to fight and manage the COVID-19 pandemic since late 2019, SARS-CoV-2 is now endemic, with 777,825,189 confirmed cases of SARS-CoV-2 infection reported to WHO as of June 2^nd 2025¹. Several vaccines were developed and approved at a record speed to save millions of lives and reduce morbidity². However, the continued emergence of SARS-CoV-2 variants and waning immunity persistently decrease the effectiveness of the vaccines against symptomatic disease³. To enhance vaccine and therapeutic development against SARS-CoV-2 and other pathogens and to improve future pandemic preparedness, it is imperative to investigate the human host’s immune response to SARS-CoV-2 infection. Importantly, the assessment of immunity was aided in the case of SARS-CoV-2 by its novelty and pandemic nature, which provided a unique opportunity to measure the true delta in antigen-induced cellular activation between true naïve donor samples (collected prior to December 2019) and virus-exposed COVID-19 convalescent donor samples (collected prior to 2021, before COVID-19 vaccines became widely available).

The human immune response is fundamental to controlling any infection, including SARS-CoV-2, with B cells producing antibodies and presenting antigens and T cells playing a critical role in killing infected cells, coordinating the response of other immune cells, and creating memory T cells^4,5. As is the case with other respiratory viruses, B cells are important for producing neutralizing antibodies at an early stage of infection, with circulating antibody titers decaying within 6 months, making it hard to track the cellular immune response to infection long-term using antibody titers⁶. In contrast, T cell-mediated cellular immunity provides long-term protection from severe COVID-19 and a cell population for long-term immunity monitoring^6–10. Several approaches exist to monitor the T cell immune response, including Enzyme-Linked ImmunoSpot (ELISpot, to measure the frequency of cytokine-secreting T cells), intracellular cytokine staining (to measure cytokine production in T cells), next-generation sequencing (to profile functional T cell activation), whole blood stimulation (to study general immune response or specific T cell response) and activation induced marker (AIM) assay (to identify antigen-specific T cells independently of cytokine production).

Notably, the AIM assay offers a highly accessible and rapid means of analyzing antigen-specific CD4⁺ and CD8⁺ T cell response through measurement of the upregulation of antigen-responsive markers upon antigen restimulation. Specifically, recognition of peptide-major histocompatibility complexes by the T cell receptor (TCR) induces upregulation of activation markers, many of which can be detected within 0-24 h of antigen activation¹¹. In the context of SARS-CoV-2, the AIM assay has played a crucial role in determining which SARS-CoV-2 epitopes are most immunogenic^12–14, uncovering the development timeline and durability of cellular immunity^7,8,15, and identifying proposed mechanisms for reducing COVID-19 disease severity or reducing symptom length through modulation of the cellular immune response^13,16. Importantly, these antigen-responsive markers are produced and detected on a per cell basis^17–19, so, when a sufficiently large number of cells are analyzed, antigen responsive cells can be detected when they constitute <0.01% of the parent cell population.

Although studies have used AIM assays to detect increased populations of SARS-CoV-2-antigen responsive CD4⁺ and CD8⁺ T cells from COVID-19 convalescent donors^{6–10,12,15,16}, most of these studies only focused on a single AIM pair to identify antigen-specific activation for each of the CD4⁺ and CD8⁺ T cell subsets, and often the AIM pair was disparate between studies^7–10,12,15. SARS-CoV-2 antigen-responsive CD4⁺ T cell markers pairs have included CD25⁺OX40⁺⁷, OX40⁺CD137^+8,12, OX40⁺CD154⁺⁸, CD137⁺CD154⁺⁸, and CD154⁺CD69⁺,^9,10 while SARS-CoV-2 antigen-responsive CD8⁺ T cell markers pairs have included a single AIM pair, CD69⁺CD137^+7,9,12,15. Thus, there is limited analysis of the relative sensitivity of different AIM pairs for detecting antigen-responsive T cells or whether particular AIM pair combinations have an amplified ability to detect SARS-COV-2 antigen-responsive T cells.

To overcome these limitations and comprehensively characterize the true delta between naïve and convalescent donors’ cellular immunity, we developed and validated a comprehensive 15-color AIM panel, including immunophenotyping markers and 35 AIM pair combinations (15 for CD4⁺ T cells and 20 for CD8⁺ T cells). To compare T cell activation detection across the 35 AIM pairs, even for donors with different T cell “activatability”, we developed an intra-sample normalization method using negative and positive controls. After applying hierarchical clustering to the normalized samples, we identified a subset of 6 AIM pairs that best discriminated naïve and COVID-19 convalescent donor samples. These 6 pairs were used to train a machine learning model for automatic classification of the T cell antigen response as either activated (SARS-CoV-2 convalescent) or non-activated (SARS-CoV-2 naïve). Importantly, when applied to a set of test samples, our machine learning model correctly predicted the SARS-CoV-2 exposure status of the donor samples. These critical results significantly improve the ability to measure and understand the human CD4⁺ and CD8⁺ T cell responses following SARS-CoV-2 exposure, even when donor sample sizes are limited. More importantly, the combined approach of measuring T cell activation broadly, completing intra-sample normalization, selecting top AIM pairs, and developing a machine learning algorithm for classifying T cells as activated or non-activated may be broadly applicable to the prediction of a person’s reactivity to antigens, whether they be from an infectious agent or self-antigens (e.g., in autoimmune disease or cancer).

Results

Donor selection

We compared samples from two donor groups, the first group comprising frozen healthy donor PBMCs collected between March 2018 and October 2019 (COVID-19 naïve donor samples, n = 7), and the second group comprising frozen PBMCs collected between May 2020 and March 2021 from donors who tested PCR positive for SARS-CoV-2 three months prior to sample collection (COVID-19 convalescent donor samples, n = 9) and who predominantly tested serologically positive for SARS-CoV-2 antibodies the day of blood collection (n = 7 of 9). These samples were unique, as this was the first pandemic respiratory virus where there was widespread testing, tracking and collection of virus-exposed samples and stores of naïve donor samples with limited T cell cross-reactivity to existing pandemic viruses^1,9,20,21.

AIM assay development and optimization for the detection of SARS-CoV-2 specific T cell response

We designed and optimized a flow cytometry-based AIM assay to detect SARS-CoV-2 specific T Cell response after 18 h of incubation with SARS-CoV-2-derived synthetic peptide pools (pp, Fig. 1a, 1b). Following stimulation, cells were labeled with a new 15-color flow cytometry panel that included three immunophenotyping markers (CD3, CD4, CD8), two markers for exclusion of monocytes and B cells (CD14, CD19), an exhaustion marker (LAG-3), and a viability dye to assist with gating (Fig. 2a, Supplementary Table 1, Supplementary Fig. 1). Activation was detected using eight early T cell activation markers, including CD25, CD69, OX40 (CD134), 4-1BB (CD137), CD38, CD40L (CD154), CD107a and PD-L1 (CD274)^11,12,16. The broad labeling of eight activation-induced markers provided a comprehensive identification of early T cell activation, enabling investigation of 15 AIM pairs specific to CD4⁺ T cells and 20 AIM pairs specific to CD8⁺ T cells (Supplementary Table 2).

Fig. 1.

Schematics of AIM assay workflow in combination with cytokine analysis. (a) Cell preparation and peptide stimulation workflow. Frozen PBMCs were thawed and rested at 37 °C in an incubator for 6 h before cell plating and peptide pool or vehicle control addition. Cells were incubated with peptide pools or vehicle control for 18 h at 37 °C before sample preparation. Cell samples for ELISpot analysis were plated in a separate filter plate and incubated with peptide pools or vehicle control for 18 h at 37 °C. (b) Peptide pool or vehicle control treated cell samples were stained for flow cytometry analysis and fixed prior to same day analysis. Cell sample supernatant was collected and reserved for bead capture analysis and the ELISpot plate was stained and developed for spot counting.

Fig. 2.

Comparison of naïve and COVID-19 convalescent donor T cell activation for select AIM pairs. (a) Flow cytometry plots comparing naïve and COVD-19 convalescent donor PBMCs with or without SARS-CoV-2 peptide pool stimulation. Five CD4⁺ T cell (b) and five CD8⁺ T cell (c) AIM pairs discriminate naïve (circle) and COVID-19 convalescent (square) donors. (b, c) Graphs indicate significant differential expression between naïve and COVID-19 convalescent donors, at 5% FDR (*), ns = not significant.

Sensitivity and specificity of the assay were optimized through rigorous attention to multiple assay steps. For example, serum-free media was used to decrease background activation, since testing indicated this media reduced the number of activated cells detected in vehicle control samples when compared to media containing non-human or human serum. Relatedly, we found and others have shown that additives, particularly DMSO, can impact T cell activation (data not shown),²² so we used water as the diluent for all cell treatments and as the vehicle control (Supplementary Fig. 2). Additionally, several SARS-CoV-2 proteins and peptide pools were compared for their ability to activate COVID-19 convalescent samples more readily than naïve donor samples. The peptide pools, pp A and pp B, that provided the largest percent activated cells in COVID-19 donor samples with the least activation of naïve donor samples (Supplementary Fig. 3) were chosen for the study and included peptides previously shown to activate CD8⁺ T cells²³. Both pp A and pp B were selected for their high variability of HLA coverage and inclusion of both major histocompatibility complex (MHC) class I and class II activating peptides. While pp A contained 88 SARS-CoV-2 peptides, of which 63 were MHCI-restricted and 25 were MHCII-restricted, pp B contained 47 SARS-CoV-2 peptides derived from the S, N, and M proteins, of which 8 were MCHI-restricted¹³. Careful attention was also given to the positive control cell treatment, where CEF and CEFTA peptide pools were combined to generate CD4⁺ and CD8⁺ T cell responses in all donors while avoiding a SARS-CoV-2 antigen response through exclusion of SARS-CoV-2 antigens. In a further optimization step, we measured at least 350,000 cells/sample to increase the sensitivity of the assay to smaller populations of activated cells. Finally, to monitor sample preparation and analysis consistency, we analyzed 5 technical replicates for each treatment group. Notably, this AIM assay is expected to be applicable to PBMCs freshly isolated from whole blood, as initial testing included this sample type.

While the main goal of this study was to compare the ability of AIM pairs to discriminate COVID-19 naïve and convalescent donors, we also compared the AIM assay results to two cytokine release assays: 1) ELISpot completed in parallel with the AIM assay and 2) cytokine capture beads using the supernatant taken from the cells analyzed by the AIM assay (Fig. 1b).

Background subtraction analysis identifies AIM pairs that differentiate COVID-19 convalescent and naïve donors

We initially compared AIM assay data from naïve and COVID-19 convalescent donors through a traditional background subtraction analysis. Briefly, the percent positive cells for each of the AIM pair combinations was extracted, the technical replicates were averaged for each treatment group, and the average percent of positive cells detected in the vehicle control group were subtracted from the treatment groups. Comparison of background subtracted AIM assay values (percent of cells positive for each of the AIM pairs) using nonparametric analysis identified five CD4⁺ T cell AIM pairs and five CD8⁺ T cell AIM pairs that significantly differentiated COVID-19 convalescent donors from naïve donors following treatment with pp A and pp B (Fig. 2b-c). For CD4⁺ T cells, CD25⁺4-1BB⁺, CD25⁺CD38⁺, CD25⁺CD40L⁺, CD69⁺OX40⁺, and CD69⁺CD40L⁺ positive cells were significantly increased (Fig. 2b), and for CD8⁺ T cells, CD25⁺4-1BB⁺, CD25⁺CD38⁺, CD25⁺OX40⁺, CD274⁺CD38⁺, and CD274⁺OX40⁺positive cells were significantly increased (Fig. 2c) in COVID-19 convalescent donor samples. Both naïve and COVID-19 convalescent donor samples had an increased percentage of positive cells when stimulated with positive control peptide pools across all AIM pairs (data not shown). Interestingly, another set of AIM pairs were significantly increased following stimulation with a single peptide pool, pp B. For CD4⁺ T cells this included AIM pair CD25⁺OX40⁺ and for CD8⁺ T cells this included AIM pairs CD274⁺4-1BB⁺, CD274⁺CD25⁺, and OX40⁺4-1BB⁺ (Supplementary Fig. 4). Thus, the developed AIM assay identified differential early activation of T cells from COVID-19 and naïve donors when exposed to SARS-CoV-2 antigens, and the pattern of differential activation varied between the two curated SARS-CoV-2 antigen pools, with pp B significantly activating more marker pairs in COVID-19 convalescent donors.

Identification of AIM pairs with the largest activation signal in convalescent donors compared to naïve donors

We next assessed which of the 35 AIM pairs exhibited the greatest signal-to-noise, with the COVID-19 convalescent donor samples representing signal and the naïve donor samples representing noise. Using vehicle control subtracted data, the percent positive T cells from individual COVID-19 convalescent donor PBMC samples were compared to the median of the percent positive T cells from the naïve donor PBMC samples. Two separate analysis methods, a difference analysis (subtraction) or a fold-change analysis (division), identified AIM pairs with the greatest difference and greatest fold-change, respectively, for CD4⁺ and CD8⁺ T cell subsets following treatment with SARS-CoV-2 peptide pools (Fig. 3). This analysis revealed that for CD4⁺ T cells (Fig. 3a-b), the AIM pairs that performed best in the difference analysis also performed well in the fold-change analysis and included CD25⁺4-1BB⁺, CD25⁺CD38⁺, CD25⁺OX40⁺, CD69⁺OX40⁺, and OX40⁺4-1BB⁺. The fold-change analysis additionally highlighted several AIM pairs containing CD40L, including CD25⁺CD40L⁺. The AIM pairs with CD40L had some of the highest fold-change versus the naïve donors; however, the variation across donors was higher than other best fold-change AIM pairs (Fig. 3a-b). Interestingly, the best performing AIM pair for fold-change with the lowest variability across donors was CD25⁺CD38⁺ (Fig. 3a-b), which showed significance for both peptide pools in the background subtraction analysis (Fig. 2b).

Fig. 3.

Comparison of naïve and COVID-19 convalescent donor T cell activation using fold-change and difference analyses. (a) Difference analysis for AIM pairs identified from the CD4⁺ T cell population; wherein, the median background subtracted activation marker pair signal from the naïve donor samples was subtracted from the background subtracted activation AIM pair signals detected from each COVID convalescent donor sample. (b) Fold-change analysis for marker pairs identified from the CD4⁺ T cell population; wherein, the background subtracted activation AIM pair signals detected from each COVID convalescent donor sample were divided by the median background subtracted activation marker pair signal from the naïve donor samples. (c) Difference analysis for AIM pairs identified from the CD8⁺ T cell population, analyzed similarly to the CD4⁺ T cell population. (d) Fold-change analysis for marker pairs identified from the CD8⁺ T cell population, analyzed similarly to the CD4⁺ T cell population. Each data point represents cells from a single COVID-19 convalescent donor treated with a single SARS-CoV-2 peptide pool in comparison to the median signal from the naïve donors. For all graphs, each unique symbol color represents data from a unique COVID-19 convalescent donor.

The CD8⁺ AIM pairs exhibited a different trend from the CD4⁺ AIM pairs (Fig. 3c-d), showing less overlap of AIM pairs that performed well in both the difference and the fold-change analyses. While both analyses identified top performing AIM pairs containing CD69 and CD274, the difference analysis also highlighted one pair with CD107a and the fold-change analysis highlighted pairs with CD25, again including CD25⁺4-1BB⁺ and CD25⁺CD38⁺ (Fig. 3c-d). The inclusion of a CD107a containing AIM pair in the difference analysis was due to downregulation of CD107a in naïve donor samples treated with SARS-CoV-2 peptides, as opposed to an upregulation of CD107a in the COVID-19 convalescent donor samples (Supplementary Fig. 3). As with the CD4⁺ T cell population (Fig. 3b), the CD25⁺CD38⁺ AIM pair again exhibited a large fold-change (Fig. 3d) with less donor-to-donor variability and less peptide pool variability than any other AIM pair.

Interestingly, analysis of the CD8⁺ cell population displayed a disparity in signal-to-noise between the two SARS-CoV-2 peptide pools, with pp A intermittently exhibiting a larger fold-change than pp B (Fig. 3d). This could be a result of uneven distribution of CD8 specific peptides compatible with a variety of HLA types as we didn’t have access to the HLA information on the donor samples used in this study. Additionally, this trend may result from the smaller number of MHC class I-restricted peptides in pp B (8), as compared to pp A (63). Nonetheless, the best performing CD8⁺ AIM pair was CD274⁺CD38⁺, when considering only pp A stimulation (Fig. 2c). The average fold-change for this AIM pair and peptide pool combination was 2.84-fold and the COVID-19 convalescent donor-to-donor variability was relatively low. Moreover, the AIM pair CD69⁺OX40⁺ had a similar donor-to-donor variability and only a slightly lower fold-change than CD274⁺CD38⁺ with pp A stimulation (Fig. 3d).

Positive correlation between AIM assay data with gold standard method for detection of antigen-specific T cell response

Interferon-γ (IFN-γ) is one of the key cytokines released after SARS-CoV-2 T cell activation²⁴ and IFN-γ measurement using the ELISpot assay following SARS-CoV-2 antigen stimulation is used as a gold standard method for vaccine response validation²⁵. When performed in parallel with the AIM assay, the ELISpot assay exhibited significantly more IFN-γ spots for COVID-19 convalescent donor samples treated with pp B (Fig. 4a). In addition, we evaluated the secretion of IFN-γ, IL-2, TNF-α, IL-10, IL-4 and IL-6 cytokines using a flow cytometry-based cytokine capture bead assay. Our data showed a significant increase in IFN-γ, IL-2, IL-6, and IL-10 secretion in supernatants from the COVID-19 convalescent donor samples, following exposure to both SARS-CoV-2 peptide pools (Fig. 4b). Interestingly, COVID-19 convalescent donor samples exhibited significantly higher TNF-α cytokine release only after stimulation with pp A (Supplementary Fig. 5a-c). We assessed the correlation (Spearman) between cytokine secretion and AIM pairs exhibiting significant activation following both pp A and pp B stimulation from Figure 2b and c. Overall, the pp A-stimulated cytokine release exhibited a stronger correlation with the background subtraction AIM assay results than the pp B-stimulated cytokine release. IL-4 concentrations were non-detectable in all samples and thus were removed from the study. Both donor groups showed increased IFN-γ, IL-2, TNF-α, IL-6 and IL-10 secretion when stimulated with control peptides (data not shown). Overall, selective activation of COVID-19 donor T cells was confirmed by both ELISpot and the cytokine capture bead assays (Fig. 4).

Fig. 4.

Cytokine production is increased in COVID-19 convalescent donors following stimulation with SARS-CoV-2 peptide pools. (a) IFN-γ cytokine production is increased in COVID-19 convalescent donor samples following stimulation with SARS-CoV-2 pp B as monitored via an Enzymatic 2-step ELISpot assay. Shown are a graph and a representative image of ELISpot results. Note: due to a technical issue, pp A failed to elicit an ELISpot response in 2 experiments, even for positive controls, which impacted the naïve to COVID-19 convalescent comparison for the pp A ELISpot data. (b) IFN-γ, IL-2, IL-6, and IL-10 cytokine production is increased in COVID-19 convalescent donor samples following SARS-CoV-2 peptide stimulation as monitored via a bead-based flowcytometry assay. (a, b) Graphs indicate significant differential expression between naïve and COVID-19 convalescent donors, at 5% FDR (*), ns = not significant. (c) Spearman correlation analysis of AIM pairs described in Fig. 2 that significantly discriminated naïve and COVID-19 convalescent samples with cytokine release assay and ELISpot.

In another comparison, the top AIM pairs identified via the fold-change and difference analyses were analyzed for their correlation with both the ELISpot and cytokine bead assay results (Supplementary Fig. 6). Surprisingly, in pp B stimulated CD8⁺ T cells, the CD274⁺CD69⁺ AIM pair showed the poorest correlation to the cytokine data (Supplementary Fig. 6), despite decent performance in the fold-change and difference analyses (Fig. 3c-d), perhaps due to the high variability in AIM signal for the CD8⁺CD274⁺CD69⁺ AIM pair. Similarly, the CD4⁺ T cell CD69⁺4-1BB⁺ AIM pair that performed well in the difference analysis (Fig. 3a-b) exhibited lower correlation with the cytokine results (Supplementary Fig. 6). Notably, CD69 is a common marker in these two AIM pairs. Regarding CD8⁺ T cells, AIM pairs that included CD107a displayed the lowest correlation with cytokines (data not shown). As with the background subtraction analysis, we hypothesize that the downregulation of CD107a in naïve donor samples is driving this result (Supplementary Fig. 3). When comparing the fold-change AIM pairs to the difference AIM pairs, the fold-change AIM pairs correlated better with the cytokine data overall (Supplementary Fig. 6), suggesting the fold-change analysis better identifies activated T cells that are both upregulating activation markers as well as downstream cytokine release.

Intra-donor sample normalization technique enables comparison of AIM pairs with varying signal-to-noise

To enable a comparison of CD4⁺ and CD8⁺ T cell activation in COVID-19 convalescent and naïve donors across all 35 AIM pairs, we devised a method for intra-donor sample normalization. For normalization to the negative control samples, the activation signal produced for a given AIM pair in response to the SARS-CoV-2 pp was first tested against the signal for the same AIM pair when the donor sample was treated with the negative, vehicle control. If the SARS-CoV-2 signal was statistically higher than the vehicle control signal, the vehicle control signal was subtracted from the SARS-CoV-2 signal. If not statistically higher, the SARS-CoV-2 signal was set to zero. For normalization to the positive control samples, the SARS-CoV-2 signal for each AIM pair was compared to the signal of the same donor cells treated with the common pathogen pp. If the SARS-CoV-2 signal was statistically higher than the common pathogen pp signal, the SARS-CoV-2 signal was set to a value of 1. All SARS-CoV-2 signals not higher than the common pathogen pp signals were then normalized to the common pathogen pp signals. This technique provided a robust method for intra-donor sample normalization, which does not require reference to any other donor sample. More importantly, this normalization technique provides a new method to determine whether a single donor’s sample, when referenced to a threshold response level, contains T cells that have been activated by the stimulatory pp.

After completing the intra-sample normalization and combining the activation by both SARS-CoV-2 peptide pools, our data revealed that 22 of the 35 total AIM pairs demonstrated a significantly higher signal in the COVID-19 convalescent donor samples as compared to the naïve donor samples (Supplementary Table 3). All of the AIM pairs containing CD274 showed significance in this analysis, while the CD107a AIM pairs did not show significance in this analysis. In addition, CD40L⁺CD38⁺ AIM pair for CD4⁺ did not show significance. Intriguingly, one CD8⁺ AIM pair, CD25⁺CD38⁺, which exhibited significant difference in the background subtraction analysis and was a top pair in the fold-change analysis, did not exhibit a significant difference between donor groups in this analysis (Supplementary Table 3).

Unsupervised hierarchical clustering using 35 AIM pairs near-perfectly classifies naïve and convalescent donor samples

Next, we performed unsupervised hierarchical clustering on the intra-sample normalized data, for all donors, across all 35 CD4⁺ and CD8⁺ AIM pairs, and for both SARS-CoV-2 peptide pools. Naïve and convalescent donor samples had different responses to the peptide pools, with nearly all convalescent samples showing higher T cell activation than naïve samples (Fig. 5). One convalescent sample treated with pp A clustered with the naïve samples. The cells from this donor had a notable lack of signal for most CD8⁺ AIM pairs and the same donor’s cells clustered with the other COVID-19 convalescent donors when treated with pp B. Importantly, no naïve donors clustered with the COVID-19 convalescent donors, indicating the SARS-CoV-2 peptide pools did not activate naïve T cells to the extent that they activated COVID-19 convalescent T cells.

Fig. 5.

Unsupervised hierarchical clustering of all 35 CD4⁺ and CD8⁺ T cell AIM pair combinations. Using data from normalized samples, unsupervised hierarchical clustering of all 35 CD4⁺ and CD8⁺ T cell marker pair combinations separated naïve and COVID-19 convalescent donor samples, except for a single COVID-19 convalescent donor sample stimulated with SARS-CoV-2 pp A (blue arrow). The same donor sample stimulated with SARS-CoV-2 pp B clustered with the COVID-19 convalescent donor samples. Obtaining this clear separation of naïve and convalescent donor samples by chance is very unlikely, p < 2^-10.

For the AIM pair clustering, our data revealed two main clusters, one with higher signals in the COVID-19 convalescent donor group and the other with lower signals in the COVID-19 convalescent donor group (Fig. 5). Of the 20 CD8⁺ AIM pairs, 12 clustered in the group with lower convalescent donor signals, while only 2 of the 15 CD4⁺ AIM pairs clustered in this group. Thus, a higher proportion of the CD4⁺ AIM pairs are key to detecting an activation signal that can classify naïve and COVID-19 convalescent donors. Eight CD8⁺ AIM pairs grouped with the 13 CD4⁺ AIM pairs that show higher signals in the COVID-19 convalescent donor group. Principal component analysis (PCA) indicated a tight clustering of naïve donor samples, as expected, and a looser clustering among the COVID-19 convalescent donor samples (Fig. 6).

Fig. 6.

PCA dimensionality reduction analysis of the CD4⁺ and CD8⁺ T cell AIM pair combinations indicate distinct separation of naïve and COVID-19 convalescent donor samples. Naïve donor samples cluster tightly together while COVID-19 convalescent donor samples exhibit a broader distribution.

AIM pair subset perfectly classifies naïve and convalescent donor samples

Next, we investigated whether a smaller, targeted AIM pair subset could perfectly discriminate naïve and COVID-19 convalescent donor samples. To identify top performing AIM pairs, the significance of an AIM pair’s activation signal was graphed against the AIM pair’s importance score, determined via Random Forest modeling (Fig. 7a). A new subset of six AIM pairs, CD4⁺: CD69⁺4-1BB⁺, CD25⁺4-1BB⁺, CD69⁺OX40⁺, CD25⁺OX40⁺; CD8⁺: CD274⁺4-1BB⁺, CD25⁺4-1BB⁺, demonstrated high significance, Mann-Whitney corrected p < 10^-3 (Supplementary Table 4), as well as high importance scores, >0.06. Surprisingly, three of the AIM pairs relevant in this analysis, CD4⁺: CD69⁺4-1BB⁺, CD25⁺OX40⁺ and CD8⁺: CD274⁺4-1BB⁺, did not show a significant difference between naïve and convalescent donors in the background subtraction analysis (Fig. 2b-c, Fig. 7a).

Fig. 7.

Subset of six AIM pairs perfectly stratifies naïve and COVID-19 convalescent donor samples. (a) A subset of six high-performing CD4⁺ and CD8⁺ T cell AIM pair combinations were identified by selecting the AIM pairs with the highest significant differential activation (pp A and pp B results combined, corrected Mann-Whitney q < 10^-3) that also had a high importance score from Random Forest modeling. (b) Hierarchical clustering using the six high-performing AIM pair subset perfectly separated naïve and COVID-19 convalescent donors.

The identified six AIM pair subset was used for an additional hierarchical clustering analysis, which outperformed the unsupervised hierarchical clustering analysis by perfectly separating naïve and COVID-19 convalescent donor samples (Fig. 7b), including the single donor sample stimulated with pp A that previously clustered with the naïve donors. Notably, this six-AIM pair subset included AIM pairs from only the cluster in the unsupervised clustering approach with the larger activation signals in COVID-19 convalescent donors. Of the six-AIM pair subset, the CD8⁺CD25⁺4-1BB⁺ AIM pair has the greatest distance from the other AIM pairs, while all other AIM pairs clustered closely together.

Trained machine learning model predictively classifies test samples as naïve or COVID-19 convalescent

Finally, the targeted six-AIM pair subset was used to develop a machine learning model for classification of samples as naïve or COVID-19 convalescent. Using 75% (n = 24) of the samples as a training set, a Random Forest model was developed and tested for accuracy in sample classification for the remaining 25% (n = 8) of samples not used to train the model. The trained Random Forest model correctly predicted whether each donor sample was naïve (non-activated) or COVID-19 convalescent (activated). Importantly, multiple separations of the samples into unique training and testing subsets resulted in a similar accuracy of sample classification by the trained model.

Discussion

Measuring the antigen specific T cell response against specific pathogens has been adopted as a tool to estimate long term immunity and is key to vaccine development^26,27. The most popular read out of T cell response is quantification of excreted cytokines following peptide stimulation. Despite broad adoption, one disadvantage is lack of immunophenotyping, a crucial step for understanding which T cell populations are responsive to an antigen. To address this, immunophenotyping flow cytometry analysis can be combined with intracellular cytokine staining (ICS); however, ICS requires lengthy sample preparation. As described herein, and alternative is to combine an AIM assay with extracellular cytokine analyses^19,28,29. To date, there has been limited analysis of the relative sensitivity of different activation induced AIM pairs for detecting antigen-responsive T cells, and it is unknown whether certain combinations of AIM pairs have an enhanced capability to detect a response to SARS-CoV-2. In this study, we developed a workflow for a broad flow cytometry-based AIM panel and robust statistical analysis to provide a deep characterization of COVID-19 naïve and convalescent samples across 15 CD4⁺ and 20 CD8⁺ T cell early AIM pairs following a short, 18 h, activation with SARS-CoV-2 antigens.

Statistical analyses, including traditional background subtraction technique, fold-change and difference analyses, hierarchical clustering, and machine learning identified the AIM pairs best able to discriminate naïve and COVID-19 convalescent donors for both CD4⁺ and CD8⁺ T cell subtypes. Overall, the best performing AIM pair was CD25⁺4-1BB⁺, which detected activation of both CD4⁺ and CD8⁺ T cells in response to stimulation by both antigen peptide pools. This AIM pair was identified as a top marker in the background subtraction analysis, the fold-change analysis, and the hierarchical clustering analysis. This CD25⁺4-1BB⁺ AIM pair is a newly identified pair which, to our knowledge, has not been used in previous SARS-CoV-2 activation studies despite CD25⁺ and 4-1BB⁺ having been used as single components in other AIM pairings^7,8,12,15. Notably, both of these activation markers play an important role in building memory T cells following viral infection^4,5, which may make these activation markers better at identifying T cell activation, particularly when used together.

Of the top 6 AIM pairs identified following intra-sample normalization through statistical significance and Random Forest importance score, four were CD4⁺ AIM pairs and two were CD8⁺ AIM pairs, indicating both CD4⁺ and CD8⁺ T cell activation assisted with naïve and COVID-19 convalescent donor classification. Notably, of the top six pairs, only one had been previously used in SARS-CoV-2 AIM assays, CD4⁺: CD25⁺OX40⁺⁷ indicating activated T cells have been going undetected in AIM assays when top AIM pairs we identified are not used. When populations of activated T cells are small to begin with, labeling as many as possible through use of the optimal activation markers is crucial for detecting T cell activation.

Interestingly, our data revealed another AIM pair, CD25⁺CD38⁺, for detecting both CD4 and CD8 T cell activation following stimulation with both peptide pools, suggesting the possibility of analyzing this AIM pair, as well as CD25⁺4-1BB⁺, in mixed CD4⁺ and CD8⁺ populations. This AIM pair’s importance was moderated during selection of the top 6 AIM pairs due to a lower Random Forest importance score, although its ability to significantly discriminate naïve and convalescent donors for the CD4⁺ cell population remained. Notably, CD38 positively correlated with exacerbated immune response during COVID-19 infections³⁰, which may increase its utility for detecting T cell activation for SARS-CoV-2.

Key to development of a machine learning Random Forest model that could predict the donor SARS-CoV-2 antigen exposure status was our unique intra-sample normalization technique; whereby, an individual donor’s T cell activation response was normalized both to its response to vehicle control, ensuring capture of true positive responses to SARS-CoV-2 antigens, and normalized to the T cell response to non-SARS-CoV-2 antigens, ensuring the SARS-CoV-2 response could be compared across AIM pairs and across donors. Importantly, this approach provides measurement of T cell activation on a sample-by-sample basis, making it a potential approach for longitudinal monitoring of a single donor to detect the degree of activation to a specific antigen or group of antigens over time, irrespective of changes to the donor’s baseline level of T cell activation, e.g., as a result of new infections, chemotherapy treatment, or aging. Notably, the efficacy of the intra-sample normalization technique relies on careful design of the negative and positive control samples. The vehicle control, including culture media and any additives, must limit T cell activation, for example, through use of a serum-free media and water vehicle to dissolve the SARS-CoV-2 antigens. The positive antigen control must include antigens from a broad group of pathogens to ensure T cell activation and must activate T cells with broad HLA subtypes, while excluding any antigens present in the antigen test pool, in this case, SARS-CoV-2 antigens. In future studies, it may be interesting to compare this approach to multiplexed activation marker analysis³¹. While some of the AIM markers that performed well in our analysis overlapped with results from Lemieux et. al., we didn’t see the same dominance of the CD69⁺4-1BB⁺ AIM pair for CD8⁺ T cells and we found CD25 and CD274 to be important markers across our analyses, including for hierarchical clustering and predictive modeling.

In addition to revealing key AIM pairs for detecting T cell activation, our data highlighted common activation markers that were not consistently or robustly upregulated by SARS-CoV-2 antigens, CD40L, and CD107a. Notably, CD40L has been used in several studies of SARS-CoV-2 antigen specific T cell activation^8,10. While CD40L performed well when paired with CD25 and CD69 in our background subtraction and fold-change analyses, our data indicates a better activation response could have been detected if CD40L was replaced with one of the highest-performing activation markers identified in this study, CD25, 4-1BB, CD69, or OX40. Based on these results and removal of the LAG-3 exhaustion marker, the initial 15-marker panel could be reduced to a 12-marker, 11-color flow cytometry AIM panel for the analysis of early-stage T cell activation. This cost-effective panel is amenable to 3-laser (and some 2-laser) flow cytometer instruments, making it easy to monitor early-stage T cell activation comprehensively and sensitively in most lab settings. Going forward, we anticipate broader adoption of larger activation panel AIM assays like the one described herein. This adoption will be driven by reproducibility of flow cytometry assays across laboratories, relative affordability, instrument and panel flexibility, and parallel cell-by-cell immunophenotyping.

One limitation of this study was the relatively small number of naïve (7) and COVID-19 convalescent (9) donors. Indeed, since the global spreading of COVID-19 and the large administration of SARS-CoV-2 vaccines, getting access to true naïve and true convalescent donors is a very rare opportunity. Follow-on studies could be designed with larger sample sizes to determine whether the identified top AIM pairs remain consistent.

Additionally, our AIM panel, intra-sample normalization technique, and machine learning model may also be used to determine the exposure status of donor samples to other viruses and other pathogens^28,32, even in immune depressed individuals³³, or to monitor the T cell response status to auto-antigens involved in autoimmune disease or cancer. Notably, the opportunities for T cell response analysis in cancer and autoimmune diseases are growing, whereby these assays may be used to measure T cell response to cancer antigens following immunotherapy or cancer vaccine treatments to monitor therapeutic efficacy^34–40. In addition, this assay and data analysis workflow may be valuable when trying to detect very small populations of cells, e.g., blood circulating cancer cells, non-activated Memory T cells, and Naïve T cells.

In conclusion, we used a broad activation marker and immunophenotyping panel to identify new, high-performing AIM pairs not previously used to study SARS-CoV-2, which have better sensitivity in detecting activated T cells than previously used AIM pairs. Application of intra-sample normalization and machine learning to the acquired data provided a model able to predict whether the donor of a sample had previous SARS-CoV-2 exposure. Broader utilization of this AIM panel and data analysis method may be instrumental to identifying AIM pairs that are generally sensitive to T cell activation, independent of antigen identity. Additionally, this approach may be broadly applicable to the prediction of a person’s reactivity to antigens, whether their source is a pathogen or self.

Methods

Donor selection and PBMC sourcing

Frozen PBMCs isolated from peripheral blood were purchased from StemCellXpress (PBMNC100C). This study utilized previously collected, de-identified frozen PBMC samples, which are exempt from regulations concerning human samples.

Briefly, 7 Naïve donor PBMCs were collected between March 2018 and October 2019, and 9 COVID-19 convalescent donor PBMCs were collected between May 2020 and March 2021. The convalescent donors tested PCR positive for SARS-CoV-2 three months prior to blood collection. Seven of 9 convalescent donors tested serologically positive for SARS-CoV-2 on the day of blood collection.

PBMC preparation for AIM, cytokine bead, and ELISpot assays

Donor PBMC vials containing 100 M PBMCs/donor were stored in liquid nitrogen until use. In preparation for the T cell assays, Frozen PBMC vials were thawed in a water bath at 37 °C and the contents were transferred to a 50 mL conical tube containing 45 mL of thawing media (AIM-V media, Gibco supplemented with 100 µg/mL of DNAse I, #07900, Stem Cell technologies). After gently mixing with the thawing media, the cells were centrifuged at 300 × g, 20 °C for 5 min, and the supernatant was removed. The cells were resuspended in 45 mL of AIM-V media and allowed to rest in a CO₂ incubator at 37 °C for 6-8 h.

Peptides for T cell activation

Curated SARS-CoV-2 peptide pools (130-127-309, Miltenyi and 3622-1, Mabtech) containing 9-22 mer peptides were used to activate SARS-CoV-2 antigen-responsive CD4⁺ and CD8⁺ T cells. The Miltenyi peptide pool (pp A) contained 88 peptides, 63 MHCI-restricted peptides and 25 MHCII-restricted peptides. The peptides originated from structural proteins (spike, S; membrane, M; nuclei, N; and envelope, E) as well as non-structural proteins. The Mabtech peptide pool (pp B) contained 47 SARS-CoV-2 peptides derived from the S, N, and M proteins as well as the open reading frame (ORF)-3a and ORF-7a proteins. These curated peptide pools were selected by the manufacturers due to their stimulation of interferon-gamma (IFN-γ) by T cells, as measured by Interferon-Gamma Release Assay (IGRA) or ELISpot.

As a positive control, a combination of MHC class I- and class II-restricted peptide pools, CEF and CEFTA (Mabtech, 3616-1 and 3617-1, respectively) were combined during dilution. Each peptide pool was diluted in water at 200 μg/mL. T cell receptors (TCRs) were chemically activated using 1mg/ml of Phytohemagglutinin (PHA, Roche) and water (distilled, GIBCO) was used as a negative, vehicle control.

T cell activation for AIM and cytokine bead assays

After resting, PBMCs were centrifuged at 200 × g for 5 min, resuspended in 500 µL AIM-V media, and counted. Cell counts were used to determine the volume of AIM-V media to add to achieve a 10 M cell/mL concentration. Aliquots of 1 M cells in 100 µL were distributed into tissue culture treated 96-well U-bottom plates (Nuclon, Thermo). The cells were incubated with vehicle, SARS-CoV-2 peptides (1 μg/mL), control MHC I and MHC II peptides (1 μg/mL), or PHA (2 μg/mL) for 18 h at 37 °C and 5% CO₂ before supernatant collection for cytokine bead analysis and cell surface staining for AIM analysis.

Flow cytometry panel for AIM assay

A 15-color flow cytometry panel was designed comprising immunophenotyping and early-stage activation induced markers. Immunophenotyping markers included CD3, CD4, CD8, and a CD14/CD19 dump channel. The early T cell activation markers included CD25, CD69, OX40, 4-1BB (CD137), CD38, CD40L, CD107a, and CD274. A viability marker, live/dead NIR (Thermo), and an exhaustion marker, LAG-3, were also included. The antibodies were titrated to determine their optimal signal-to-noise concentration (Supplementary Table 1). The flow cytometry gating strategy included gating on 1) singlets, 2) lymphocytes in the SSC X FSC, 3) live cells, 4) CD3⁺CD14^-CD19^-, 5) T CD4⁺ and CD8⁺ positives and 6) AIM pairs (Fig. 2a). For precise gate area determination of the fluorescence positive area, unstimulated controls, FMOs, and positive controls were used.

AIM assay compensation samples and FMOs

Single-stained homogenous bead controls were used to correct for spillover through a compensation process during data analysis. To preserve donor PBMCs, aliquots of Ultracomp Beads (#01-2222-42, Thermo) were incubated with each antibody for 30 min on ice, washed with 200 µL PBS, and analyzed by flow cytometry. For the viability dye single-stain control, the Arc Amine reactive kit (#A10628, Thermo) was used according to manufacturer specifications. Fluorescence minus one controls (FMOs) were made for all antibodies in the panel to determine the cut-off point between background fluorescence and the fluorescence generated by positive populations within a sample of cells. FMOs were created for each sample run for markers that exhibited overlap. Unstained PBMCs with identical buffer treatment as fully stained samples were used to control for background autofluorescence inherent in specific cell types.

AIM assay cell staining and fixation

Following peptide stimulation, PBMCs were centrifuged at 200 × g for 5 min and washed in 200 µL FACS buffer (PBS containing 3% bovine serum albumin, Sigma, and 0.01% NaN₃, Sigma). After washing, PBMCs were stained for 30 min on ice with 200 µL of the AIM panel antibodies diluted in FACS buffer according to their titrated concentrations. After staining, the PBMCs were washed with 200 µL PBS and stained with the viability dye for 20 min, followed by a 200 µL PBS wash and fixation in 150 µL 0.5% of formaldehyde in PBS. The cells were kept at 4 °C (for up to 24 hours) until flow cytometry analysis.

AIM assay data collection and analysis

Activated and stained PBMCs were plated in 96-well U-bottom plates (Thermo # 143761) and analyzed on a Novocyte Quanteon flow cytometer equipped with a NovoSampler Q for automated sampling (Agilent). The acquisition settings were the following: 100 µL sample volume acquired; plate mixing at 1,500 rpm for 10 s prior to each well injection, medium flow rate of 35 µL/min, FSC> 100,000 threshold, and 1 rinse cycle between each well to reduce carryover. After acquisition, data were exported as FSC files and analyzed using FlowJo (FlowJo LLC). The gating strategy is described in Supplementary Fig. 1. The percentage value in the positive gate area of AIM⁺ cells was exported to an excel table and used for downstream statistical analyses.

Cytokine bead-based detection assay

Supernatants collected during peptide stimulation of PBMCs were assessed for cytokine secretion using the LegendPlex HU Essential Immune Response Panel (#740930, BioLegend). IL-2, TNF-α, IFN-γ, IL-4, IL- 6, and IL-10 cytokines were selected for detection. Sample preparation strictly followed the manufacturer’s guidelines and utilized the filter plate option. Plate washes were completed using a vacuum manifold and house vacuum. Samples were acquired using NovoCyte Quanteon flow cytometer equipped with a NovoSampler Q and the data were analyzed using the NovoExpress software (Agilent).

IFN-γ ELISpot assay

An enzymatic single-color IFN-γ ELISpot assay (CTL) was also performed. Briefly, rested PBMCs were plated in triplicate onto PVDF-membrane containing plates (Immunospot by CTL) at 4.0 × 10⁵ cells/well for vehicle and peptide stimulation and 2.5 × 10⁴ cells/well for PHA stimulation. Plated PBMCs were treated with vehicle, peptide pools, or PHA and incubated at 37 °C, 5% CO₂ for 20 h. After stimulation, the PBMCs were removed from the PVDF membrane via forceful manual pipetting. Subsequent plate washing steps were automated using a BioTek EL406 washer dispenser (Agilent). Plate preparation steps followed the manufacturer’s guidelines and included washing with PBS and PBS + 0.05% Tween 20, incubation with the detection antibodies, washing with PBS + 0.05% Tween 20, and washing with water. Then, the diluted developing reagent was added to the wells and development was stopped after 8-12 min, based on visual inspection with a microscope. After final water washes, the plates were dried overnight before imaging with a Cytation 7 Cell Imaging Multimode Reader (Agilent) using the upright microscope. Spot counting was accomplished using Gen5 software (Agilent). The data were graphed in GraphPad Prism (Dotmatics). For each sample, the number of spots were counted, triplicate measurements were averaged, the number of background spots (vehicle) were subtracted, and the number of spots/million cells were plotted.

Intra-sample normalization of AIM data

Median values were used to represent repeated measurements. If the SARS-CoV-2 signal was statistically higher than the vehicle control signal, the vehicle control signal was subtracted from the SARS-CoV-2 signal. If not statistically higher, the SARS-CoV-2 signal was set to zero. For the positive control normalization, the SARS-CoV-2 signal for each AIM pair was compared to the signal of the same donor cells treated with a common pathogen pool. If the SARS-CoV-2 signal wasn’t statistically lower than the common pathogen pool signal, the SARS-CoV-2 signal was set to a value of 1. All SARS-CoV-2 signals not higher than the common pathogen pool signals were then normalized to the common pathogen pool signals. Let S_negative, S_covid, and S_positive be a set of measurements for the negative control, COVID-19 stimuli, and positive control stimuli, respectively. Let P_negative, and P_positive be the t-test p-values comparing S_negative to S_covid, and S_covid to S_positive, respectively. If P_negative is not significant (i.e. the average of the S_covid measurements is not significantly higher than the average of the S_negative measurements (p-value < 0.05)) we set the normalized value to 0. If P_negative is significant, but P_positive is not significant (i.e. the average of S_covid measurements is not significantly lower than the average of the S_positive measurements) we set the normalized value to 1. In case both P_positive and P_negative are significant, we set the normalized value to minimum {1, (median(S_covid)-median(S_negative)) / (median(S_positive)- median(S_negative))}. This provided, for each AIM pair, a value between 0 and 1 that represents its activity compared to the positive control.

Unsupervised hierarchical clustering of AIM data

Unsupervised hierarchical clustering was done through the clastmap function in seaborn library⁴¹.

Random Forest modeling of AIM data for sample classification

Classification of samples was done using the RandomForestClassifier from sklearn library⁴². 75% of samples were used for training with 10-fold cross validation, and 25% of samples were used for validation.

Selection of top AIM pairs

Importance scores for all AIM pairs were collected from a Random Forest classifier that classified the COVID-19 convalescent samples. The top 6 AIM pairs with high importance scores, that also have a Mann-Witney significant differential expression to negative control measurements, were selected for subsequent analysis.

Author contributions

G.C.V.d.B. and T.P.S. equally contributed to the paper, they conceptualized the project and assays, performed experiments analyzed the results, wrote the manuscript and prepared figures. G.C.V.d.B. and T.P.S. both reserve the right to list their name first on their resume or curriculum vitae. I.S. provided conceptual advice, analyzed the results, wrote the manuscript and prepared figures. B.P., K.B. and K.K. conceptualized the project and provided technical expertise. All authors contributed to data interpretation and provided feedback on the manuscript. All authors approved the final version of the manuscript.

Data availability

Due to proprietary reasons, the raw data related to this study cannot be made openly available.

Code availability

Due to proprietary reasons, code developed for sample normalization cannot be made openly available.

Declaration

Competing interests

The authors declare the following competing interests: This research was supported by Agilent Technologies, Inc. and all authors were employed by Agilent Technologies, Inc. during completion of this work. Agilent Technologies, Inc. has filed a patent application, PCT/US2024/011684, that covers elements of activation-induced marker assays. G.C.V.d.B., T.P.S., and I.S. are the named inventors.