Self-amplifying COVID-19 mRNA vaccine boosts antibody function over time in a Phase 3 trial

A phase III non-inferiority trial comparing responses to COVID-19 boosters was performed (26–28; see Methods), and a random subset of specimens were analyzed through systems serology. All participants had received a 2-dose primary vaccination series with mRNA COVID-19 vaccines and a booster dose of BNT162B2 vaccine at least three months before enrollment. Participants were randomized into two groups: those receiving a booster dose of BNT162b2 and those receiving a booster dose of the self-amplifying mRNA ARCT-154^26–28. Blood draws were taken at day 1 (baseline, day of booster dose), day 29 post-boost, day 91 post-boost, day 181 post-boost, and day 361 post-boost.

Evaluation of Ig isotypes (IgG, IgA, IgM), subclasses (IgG1-4), and low-affinity Fc-gamma receptor binding (FcγR) antibodies to on-target (WT Spike), target-related (Delta Spike, BA.2 Spike, BA.5 Spike, XBB.1.5 Spike, and KP.3 Spike), and non-target (influenza H1N1 A/Wisconsin/22 hemagglutinin (HA), human cytomegalovirus (HCMV) Glycoprotein B (GB), and Ebolavirus Glycoprotein (GP)) controls was assessed (Fig. 1). Influenza H1N1 HA and HCMV GB served as consistency non-target controls, which should be present in the vast majority of participants, but not responsive to the boosters, while Ebola GP served as a non-target negative control. Additionally, antibody-dependent phagocytosis by monocytes (ADCP) and antibody-dependent natural killer cell activation (ADNKA) to target (WT), target-related (BA.5), and non-target (Ebola GP) were assessed. During the observation period of this clinical study, BA.5 was the predominant circulating SARS-CoV-2 strain.

Overall antibody binding and functional architecture revealed stimulation at 29 days post-vaccination for both study groups (BNT162B2 in red, ARCT-154 in blue) with responses waning thereafter. This stimulation and waning showed antibody isotype, subclass, and FcγR specificity, with not all features showing similar trends, as expected.

To identify if the overall antibody profile could be separated between the two boosted groups, we performed an exploratory partial least squares discriminant analysis (PLSDA) model of antibody binding. A least absolute shrinkage and selection operator (LASSO) regularization was used to down-select the number of antibody features for the PLSDA fit^46,47 (see Methods). This was done for each draw during the observation period (Fig. 2A–D and Supplementary Fig. 1). No model could be built distinguishing recipients of BNT162B2 and ARCT-154 for day 1, indicating that baseline humoral profiles were not significant between groups. However, multivariate clustering through this method could statistically separate out vaccinated groups (BNT162B2 in red, ARCT-154 in blue) at days 29-, 91-, and 181-post-vaccination (for each, p < 0.04 between the model and random features, p < 0.02 between the model and permuted labels). A model could not be built, however, at 361 days post-boost (p = 0.1 between the model and random features, p = 0.07 between the model and permuted labels) (Fig. 2E–H).

The LASSO-selected features contributing to separation at the multivariate scale were plotted along latent variable axes 1 and 2 (LV1 and LV2, respectively). FcγR-binding antibody responses to target WT Spike were selected in the ARCT-154 recipients at each timepoint, while antibody binding features to Omicron lineages were selected in the BNT162B2 arm (Supplementary Fig.). These exploratory results suggested that the profiles were moving at distinctly different rates after the boost and before the end of the observation period. Also, even though ARCT-154 and BNT162B2 were directed to WT/D614G Spike, there were differences noted in the multivariate models. 1

It was previously reported that recipients of the sa-mRNA vaccine ARCT-154 had a more durable neutralizing antibody phenotype compared to other mRNA-based vaccines^27,28. Our data supported that rates of change to target and target-related Spikes between the two treatment arms may be driving separation at the multivariate level. We thus took the first derivative (d) of the collection periods to quantify how rates of change of antibody levels may differ between sa-mRNA and mRNA booster recipients. We were particularly interested in how activating FcγR-binding antibodies moved with time, given the previously reported difference in neutralization between an mRNA and a sa-mRNA vaccination platform^26–28. Our primary analysis was thus post-peak responses (after day 29) to the target (WT Spike) and target-related (Delta, BA.2, BA.5, XBB.1.5, and KP.3) Spikes during this window.

Total IgG levels were first analyzed for rates of change between the collections. The ARCT-154 treatment arm exhibited a sustained total IgG stimulation to target WT Spike in the time interval of day 29 to day 91 (median d > 0), whereas recipients of BNT162B2 displayed a sharp contraction in total IgG to the target antigen (median d < 0). This was statistically significant between the two treatment arms (false-discovery rate, or FDR, corrected p value < 0.05). Subsequent time intervals showed a median d < 0 for both treatment arms, indicating contracting total IgG levels. The d Total IgG at day intervals 91–181 and 181–361 were not statistically significant (Fig. 3A). Interestingly, this same trend was observed for each target-related Spike (Fig. 3B–F). For these antigenically drifted Spike variants, the sa-mRNA ARCT-154 recipients showed a median d Total IgG ≥ 0 in the day 29–91 time interval, indicating that median waning was not occurring for the group until after day 91. This was in contrast to BNT162B2, which showed a d Total IgG < 0 at the day 29–91 time interval, and either further contracted or showed a flat rate of Total IgG levels thereafter. There were no changes in d Total IgG for HA A/Wisconsin/22 or HCMV GB, indicating that non-target, high-exposure antigen responses were neither stimulated nor contracted (Fig. 3G, H). Lastly, no changes were observed for non-target, negative control Ebola GP (Fig. 3I).

The same approach was used for IgG subclasses 1-4. A similar trend for IgG1, IgG2, and IgG4 was observed for the sa-mRNA ARCT-154 recipients with sustained antibody levels through the day 29–91 time interval. Only minor IgG3 rates of change were observed, and they were exclusive to Delta and KP.3 Spike in the day 29–91 time interval (Supplementary Figs.–). These results indicate that sa-mRNA and mRNA vaccinations elicit differing stimulation and decay dynamics for binding antibodies to target (WT Spike) and target-related (Delta, BA.2, BA.5, XBB.1.5, and KP.3) antigens, but have minimal non-target responses. 2 5

Similarly, the derivative plots for the Spike proteins showed discrepancies for activating FcγRIIIA-binding antibodies between the two treatment arms for the target WT Spike (Fig. 4A). Surprisingly, this d FcγRIIIA between the two treatment groups remained statistically significant at the day 91–181 time interval, indicating that the rate of contraction for recipients of ARCT-154 was lower than for BNT162B2 at extended time periods. The two groups both approached d FcγRIIIA WT Spike = 0 at the day 181–361 time interval, with BNT162B2 recipients having a higher d FcγRIIIA WT Spike at that interval (Fig. 4A). Sustained FcγRIIIA-binding levels to target-related Spikes also showed statistically significant differences in durability for recipients of the sa-mRNA ARCT-154 at the day 29–91 interval, and a few even showed statistically significant differences at the day 91–181 interval. These rates eventually flatline at days 181–361 (Fig. 4B–F). There were no changes in d FcγRIIIA-binding antibodies observed in our control antigens, which indicates that there were negligible non-target binding fluctuations occurring (Fig. 4G–I).

Other FcγR-binding antibodies showed differing rates of antibody levels during the observation period. FcγRIIA-binding antibodies showed a trend toward enrichment for ARCT-154 recipients at the day 29–91 interval, with a few Spikes showing statistical significance compared to BNT162B2 recipients, but not nearly to the same degree as FcγRIIIA (Supplementary Fig. 6). FcγRIIB showed a statistically significant enhancement in d for all Spikes at the day 29–91 time interval for ARCT-154 recipients, but not at any other time (Supplementary Fig. 7). Lastly, no differences in d FcγRIIIB-binding antibodies were observed between the two treatment groups (Supplementary Fig. 8). Importantly, there were no changes in d FcγIIA, d FcγRIIB, FcγRIIIA, or d FcγRIIIB for any of our non-target control antigens. These data indicate that the strongly activating FcγRIIIA was disproportionately sustained in recipients of the sa-mRNA vaccine boosted. This response was specific to on-target and related-target Spikes.

We then profiled the functionality of the humoral response longitudinally over time as Fc-mediated effector functions have been shown to confer protection against SARS-CoV-2^48–51. Antibody-dependent cellular phagocytosis by primary-derived monocytes (ADCP—see Methods and Supplementary Fig. 9A) to target WT Spike and target-related BA.5 Spike had similar initial responses for the two groups. Baseline values were well above our lower limit of detection since the participants all had previous immunity through vaccination. Both groups showed a waning period from days 181 to 361 post-boost. Although trends to elevated ADCP were observed for recipients of ARCT-154, the two groups’ responses were not statistically significant after FDR correction to WT or BA.5 Spike (Fig. 5A, left and middle). Neither group had any detectable ADCP to our negative exposure control antigen Ebola GP (Fig. 5A, right).

Cellular cytotoxic activity was quantified through an antibody-dependent natural killer cell activation (ADNKA) assay in which percent positivity of degranulation marker CD107a, macrophage inflammatory protein 1 beta (MIP-1β), and pro-inflammatory cytokine interferon gamma (IFN- γ) were measured (Supplementary Fig. 9B). ADNKA activity to the target WT Spike for ARCT-154 recipients showed a significant stimulation at day 29 compared to BNT162B2 recipients for CD107a and MIP-1β. A persistence of the CD107a marker (cellular degranulation) was noted until day 181 for the ARCT-154 treatment arm. A similar trend, albeit not statistically significant after FDR correction, was observed for IFN-γ production in response to target WT Spike exposure. (Fig. 5B, top row). It is important to note that both treatment arms had high pre-existing immunity to the target WT Spike.

We next quantified ADNKA activity to the target-related BA.5 Spike. Here, both treatment arms exhibited a strong acute response for all three ADNKA outputs, indicating that the breadth of responses was increasing, consistent with previously reported observations^39,44,52. Similar to other assays, ADNKA responses were relatively sustained in the ARCT-154 treatment arm. For CD107a expression, recipients of ARCT-154 exhibited a significantly higher response to the target-related BA.5 Spike compared to the BNT162B2 recipients for the entire observation period. All three ADNKA markers showed enrichment for ARCT-154 recipients at day 91 post-boost (Fig. 5B, middle row). Lastly, ADNKA responses were quantified to an off-target antigen, Ebolavirus GP. Here, neither treatment arm had detectable baseline ADNKA activity, nor were they stimulated by vaccination at any timepoint (Fig. 5B, bottom row).

Taken together, our functional assays closely tracked with binding antibody profiles for recipients of a BNT162B2 or ARCT-154 booster after receiving three doses of an mRNA vaccine. Notably, ADNKA remained elevated for recipients of the sa-mRNA ARCT-154 booster, which supported the slower decay dynamics of FcγRIIIA-binding antibodies observed. Also, antibody effector functions were specific to SARS-CoV-2 for both treatment arms.

The primary study was a double-blind, multicenter, randomized, controlled, phase 3, non-inferiority trial, conducted at 11 outpatient clinical sites in Japan (the Japan Registry for Clinical Trials, jRCT 2071220080). It enrolled healthy adults aged at least 18 years who had previously been immunized with two doses of an mRNA COVID-19 vaccine (BNT162B2 [Comirnaty, Pfizer/BioNTech] or mRNA-1273 [Spikevax; Moderna]) followed by a third dose of BNT162B2 at least 3 months before enrollment. The original study protocol was approved by the institutional review boards (IRB) at all participating sites. The study was conducted in accordance with the ethical principles outlined in the Declaration of Helsinki and Good Clinical Practice guidelines. The results for primary and secondary endpoints have been published previously^26–28. The multi-center study was approved by the Hakata Clinic IRB (Fukuoka, site name Hakata CL), IRB of Shinanozaka Clinic (Tokyo, site name Shinanozaka CL), P-One Clinic Keikokai Medical Corp IRB (Tokyo, site names Higashi-Shinjuku CL and P-One CL), Medical Corporation Heishinkai OPHAC Hospital IRB (Osaka and Tokyo, site names Osaka Pharmacology Clinical Research Hospital and ToCROM), Kobori Central Clinical Research Ethics Committee IRB (Hokkaido, Kanagawa, and Kumamoto, site names Shin-Sapporo HP, LUNA, and Medimessee Sakurajyuji, respectively), Fukushima Medical University Hospital IRB (Fukushima, site name Fukushima Medical HP), and the Juntendo University Hospital IRB (Tokyo, site name Juntendo HP).

All participants provided informed consent and agreed to comply with the study requirements, including attendance at all study visits, provision of blood samples, and use of approved contraception from 28 days before study vaccination through the end of the study duration. Detailed inclusion and exclusion are described in the primary study publication²⁶. For this post-hoc research, we randomly selected two subsets comprising 30 participants from each vaccine group (ARCT-154 and BNT162B2, total n = 60) who were seronegative at baseline and all subsequent analysis timepoints for SARS-CoV-2 nucleocapsid protein, which is an indicator of recent SARS-CoV-2 infection⁶⁸, and had samples available at days 1, 29, 91, 181, and 361. Samples were de-identified and processed in a blinded manner. A sample size of 30 participants per group is deemed adequate to assess potential differences in humoral immune responses between the two vaccine groups. This post-hoc testing and analysis were determined as not meeting human subjects research as determined by the Harvard IRB. All measurements were obtained from prealiquoted samples from each participant at each timepoint in each group.

This project was determined as not human subjects research by the Harvard IRB (IRB24-1684). The clinical trial number for the Phase III trial is jRCT2071220080 registered on December 16, 2022.

Antigen-specific antibody subclass/isotype levels and Fc-gamma receptor (FcγR) binding profiles were measured using a custom multiplex Luminex assay, as previously described^69–71. Carboxylated Luminex MagPlex microspheres (DiaSorin) were covalently linked to target antigens via NHS-ester linkages by the addition of EDC and Sulfo-NHS (Thermo Fisher). Immune complex formation between serum samples diluted in 1X PBS (1:500 for Total IgG, IgG1, and IgG2; 1:250 for IgG3, IgG4, IgA1, and IgM; 1:2000 for FcγRIIA, FcγRIIB, FcγRIIIA, and FcγRIIIB) and the antigen-coupled microspheres occurred overnight at 4 °C in 384-well plates, shaken at 750 rpm in 1X Assay Buffer (1X PBS, pH = 7.4, 0.1% BSA, 0.02% Tween-20). After immune complex formation, the plates were washed with 1X Assay Buffer. Ig isotypes and subclasses were detected with diluted Phycoerythrin (PE)-conjugated secondary antibodies (SouthernBiotech, see Supplementary Table 1) for 1 h at room temperature. Complexes were then washed with 1X Assay Buffer for three washes and resuspended in 1X Luminex Sheath Buffer.

FcγR-binding antibodies were detected by Avi-tagged human FcγRs (Duke Human Vaccine Institute), which were biotinylated and coupled to Streptavidin-PE (Agilent Technologies). The Strep-PE-FcγRs were incubated with the immune complexes for 1 h at room temperature in 1X Assay Buffer. Complexes were then washed with 1X Assay Buffer three times and resuspended in 1X Luminex Sheath Buffer.

All binding measurements were done using the xMAP Intelliflex (Luminex) with readouts as median fluorescence intensity (MFI) and reported in arbitrary units (AU). All samples were run in duplicate, and the means of the technical replicates were reported.

Human monocytes were obtained from leukopacks (StemCell). The leukopacks were initially processed for red blood cell depletion through the EasySep RBC Depletion Reagent (StemCell). This involved adding EDTA and diluting the Leukopacks at a 1:1 ratio with EasySep Buffer (StemCell). The diluted Leukopacks then underwent a series of magnetic separations with the addition of RBC Depletion Reagent (StemCell). The cells are spun down and washed using EasySep Buffer (StemCell). Total cells were then counted using the Cell Countess, and a total of 5 × 10^5 cells/well were seeded onto a 96-well round-bottomed plated with R10 media (RPMI-1640 (Sigma-Aldrich) supplemented with 10% fetal bovine serum (FBS) (Sigma-Aldrich), 5% penicillin/streptomycin (Corning, 50 μg/mL), 5% L-glutamine (Corning, 4 mM), 5% HEPES buffer (pH 7.2) (Corning, 50 mM)).

Antigen target proteins were biotinylated using the EZ-link Sulfo-NHS-LC-LC-Biotin kit (Thermo Fisher). The biotinylated antigens were then coupled to neutravidin beads (Thermo Fisher, F8776). Next, the bead-antigen conjugates were incubated with 1:100 diluted serum overnight at 4 °C to form immune complexes. The unbounded antibodies were removed using a wash buffer (1x PBS). The immune complexes were then incubated with PBMCs for 1.5 h at 37 °C and 5% CO2. The cells were then stained with CD14 Pacific blue antibody cocktail at a 1:100 dilution and incubated for 30 min in the dark at room temperature. The stained cells were subsequently washed and fixed in 4% paraformaldehyde (PFA). The cells are washed a final time and resuspended in 1X PBS. The antigen-specific PhagoScore was determined as (% cells positive × Median Fluorescent Intensity of positive cells). Flow cytometry was performed with an iQue (IntelliCyt) instrument, and population measurements were conducted using IntelliCyt ForeCyt (v8.1). The reagents and materials used are listed in Supplementary Table, and the gating for ADCP is shown in Supplementary Fig.. 1 9A

Human natural killer (NK) cells were isolated from leukopacks (StemCell). The NK cells were enriched from the leukopacks using the EasySep Human NK Cell Isolation Kit. This is a negative selection procedure where unwanted cells are labeled with antibody complexes and magnetic rapidspheres (StemCell). The unwanted magnetically tagged cells are then separated from the desired NK cells through a series of separations using the EasySep magnets (StemCell). Donors were anonymized. Cell concentration and viability was quantified on the Cell Countess. NK cells were resuspended in RPMI-1640 (Sigma-Aldrich) supplemented with 10% fetal bovine serum (FBS) (Sigma-Aldrich), 5% penicillin/streptomycin (Corning, 50 μg/mL), 5% L-glutamine (Corning, 4 mM), 5% HEPES buffer (pH 7.2) (Corning, 50 mM) in 96-well plates. The NK cells were activated overnight with the addition of 1 ng/mL IL-15 (StemCell Technologies) at 37 °C and 5% CO2.

ELISA plates were coated with 3 μg/mL of target antigen and incubated for 2 h at 37 °C. The coated plates were then washed with PBS and blocked with blocking buffer consisting of 5% bovine serum albumin (BSA) in 1X PBS, pH = 7.4, for 1 h at room temperature. The ELISA plates were washed with PBS, and 1:80 diluted serum samples were added to the plates to form immune complexes overnight at 4 °C. After another wash, NK cells at a concentration of 5 × 10^5 in R10 media supplemented with anti-CD107a–phycoerythrin (PE)–Cy5 (BD Biosciences), brefeldin A (5 μg/mL) (Sigma-Aldrich), and GolgiStop (BD Biosciences) was added to the plates. The NK cells were incubated with immune complexes for 5 hours at 37 °C and 5% CO2. The incubated NK cells were subsequently washed and stained for cell surface markers with anti-CD3 Pacific Blue (BD Biosciences), anti-CD16 allophycocyanin (APC)-Cy5 (BD Biosciences), and anti-CD56 PE-Cy7 (BD Biosciences) for 20 min at room temperature. The washed NK cells were then fixed with PermA (Life Technologies) for another 15 minute. The cells were washed once more and then permeabilized with PermB (Life Technologies) supplemented with anti-MIP-1β PE (BD Biosciences) and anti-IFNγ FITC for 20 min at room temperature. Fluorescent intensity was measured using the iQue Cytometer (Intellicyt). NK cells were gated as CD56 + /CD16 + /CD3-. NK activation was evaluated as the percentage of NK cells positive for CD107a, IFNγ, or MIP-1β. All assays were performed with at least two healthy donors with one male and one female. All reagents and materials used are listed in Supplementary Table, and the gating strategy for ADNKA is shown in Supplementary Fig.. 1 9B

All antibody effector function assays were quantified for statistical significance between groups (BNT162B2 and ARCT-154) using a two-sided Wilcoxon Test followed by a false discovery rate adjustment.

The rate of change was calculated based on the difference of the MFI within a certain period of time, as shown here:. The calculation of the rate of change was paired based on the subjects; if there was no participant on certain timepoint, the data point for the time interval was not included in quantifying the group median or comparisons across groups. A two-sided Wilcoxon Test followed by a false discovery rate adjustment was used for all statistical comparisons. Rate = Δ MFI timepoint timepoint 2 1 − Δ Timepoint Days

A partial least squares discriminant analysis (PLSDA) was used to identify features separating vaccine arms post vaccination on day 1 (baseline), 29, 91, 181, and 361. Initially, all data were background subtracted, log 10 transformed, and scaled through z-scoring. Data were split into 5 folds, and a least absolute shrinkage and selection operator (LASSO) was used to identify the minimal features required to explain separation using 4 of the folds; the 1 withheld fold was used to test the LASSO penalty (λ). This cross-validation process was repeated 5 times to identify the best-performing λ, which was then fit into the final LASSO model. This process was repeated 100 times. Features were selected for as being present in at >80% of the 100 iterations. This regularization and downselection of the humoral features was used as a way to prevent statistical overfitting, given the coordinated nature of antibody responses⁴⁶, particularly for a cohort with existing immunity. The LASSO-selected datasets were then used to train the PLSDA. The LASSO-selected features were projected onto latent variables 1 (LV1) and LV2 (Supplementary Fig. 1).

For cross-validations, a separate script was run with the first stage being the same as above in building the LASSO using five internal folds of the data. This dataset was used to train the PLSDA and predict accuracy, ranging from 0 to 1, where 1 indicates 100% model accuracy. This model was then compared to one set of randomly selected features, in our case, 30 random features, per cross-validation. The randomly selected features were then used to train the PLSDA and predict the accuracy of the model, again represented from 0 to 1. A separate comparison was made using permuted (shuffled) labels. Here, the permuted labels were used to train the LASSO, with a total of 30 label permutations per cross-validation, and an internal 5-fold cross-validation to train the LASSO and PLSDA. The resulting model accuracy was represented from 0 to 1.

Model performance (original data, random features, and permuted labels) was compared using an exact p-value obtained for model comparison through the tail probability of the null distribution. This approach has been previously validated elsewhere^46,47,67,72.

Raw data used for this paper has been deposited on the HSPH Systems Serology GitHub page under the accession number QW20251028 (https://github.com/HSPHSystemsSerology/QW20251028↗). No unique code was generated for this study. All data generated or analyzed during this study are included in this article.

No previously unreported code or scripts was used to generate results in the manuscript. All coding was done in R Studio V. 2023.06.0.