Correlates of protection and determinants of SARS-CoV-2 breakthrough infections 1 year after third dose vaccination

A cohort of 578 HCWs providing direct or indirect patient care at Hospital Clínic in Barcelona randomly selected was followed up for ≈3 years since the beginning of the pandemic in Spain. Questionnaires and sample collection were done over 7 visits: M0 (baseline, month 0, 28 March–9 April 2020, n = 578), M1 (month 1, 27 April–6 May 2020, n = 566), M3 (month 3, 28 July–6 August 2020) (n = 70), M6 (month 6, 29 September–20 October 2020, n = 507), M9 (month 9, 19 January–5 Feb 2021, n = 132), M12 (month 12, 12 February–30 April 2021, n = 414), and M24 (month 24, 14 March–25 April 2022, n = 393) (32% of initial participants were lost to follow-up). At M3 and M9, only participants with previous evidence of infection were invited to participate in the survey. Up to 17 January 2023 (end of the follow-up period), 114 participants have had the 4 vaccine dose (vaccination period ranged from October 2022 to January 2023).

Participant data were collected over the 7 study visits using a standardized electronic questionnaire through REDCap version 13.8.1 as previously described. Additionally, data on confirmed SARS-CoV-2 infections and COVID-19 symptoms, vaccination type, and dates and adverse effects up to 10 months after the M24 visit were collected at the Occupational Health and Preventive Medicine and Epidemiology departments at the Hospital Clinic. Molecular data regarding the SARS-CoV-2 variants was not collected routinely in the hospital. During the M0 and M1 visits, active detection of infection by real-time reverse transcription-polymerase chain reaction (rRT-PCR) was conducted. For subsequent visits, rRT-PCR or AgRDT (Antigen Rapid Diagnostic test) were performed whenever participants exhibited symptoms or had contact with an infected individual. In addition, individuals with a positive serology before vaccination or a positive N serology at any timepoint were also considered infected. When previous infection was detected by serology, the timepoint interval when this infection occurred was defined as the interval when a participant seroconverted for any of the antigens (for timepoints before vaccination) or for N antigen (for any timepoint).

IgA, IgG, and IgM levels (median fluorescence intensity, MFI) to S and N SARS-CoV-2 antigens were measured by Luminex-based assays (quantitative suspension array technology) in multiplex as previously described [19, 20]. The panel of antigens included the S full length, the subregion S1, and the receptor-binding domain (RBD) of S1, expressed in lentiviruses as described [21], as well as the subregion S2 (SinoBiological) and the nucleocapsid (N) protein (expressed in E. coli), all from the Wuhan strain. As variants of concern, the full-length S proteins from Delta and Omicron BA.1 (SinoBiological) were also included. Plasma samples were tested at 1:500 dilution for the 3 isotypes and additionally at 1:5000 for IgG to avoid saturated anti-S levels in the vaccinated participants. Optimal testing dilutions were previously assessed and samples were within the quantitative range of the assay. The investigators conducted the assays blinded.

MFI values were log10-transformed for analysis. Seropositivity was calculated with the values measured at 1:500 plasma dilution. The cutoffs for each isotype and antigen were calculated as 10 to the mean plus 3 standard deviations (SD) of the log10-transformed MFI from 128 pre-pandemic (negative) controls. Positive serology was defined as being positive for IgG, IgA, and/or IgM to any of the tested antigens, and serology was considered undetermined when the MFI levels for a specific isotype-antigen fell between the positivity threshold and 10 to the mean plus 4.5 SD of the log10-transformed MFIs of the negative controls, provided that no other isotype-antigen combination exceeded the positivity threshold.

The differences between groups at M24 were measured using a two-tailed Wilcoxon Rank- Sum test. The differences between timepoints within a group were measured with a Wilcoxon signed-rank test. MFI divided by the cutoff value was used when two timepoints were compared to address variations in MFI measures between timepoints. To account for multiple testing, both tests were adjusted using the Benjamini–Hochberg method [22]. All tests were performed two-sided. The strength of correlations between antibody levels was evaluated by Pearson’s correlation test.

Univariable and multivariable linear regression models were fitted to assess factors associated with antibody responses to SARS-CoV-2 at M24 among exposed, naive, and all individuals. Separate models were also fitted for individuals vaccinated with at least 2 doses and for those vaccinated with 3 doses. The regression coefficients (β) obtained from each model were converted into percentage values to facilitate interpretation. The transformed β value (%) was calculated using the formula ((10^β) − 1) × 100. This indicates the percentage difference in the dependent variable associated with a 1-unit increase in the corresponding independent variable (for continuous variables) or the percentage difference in the dependent variable between the reference group and the study group (for categorical variables). Causal assumptions for each of the analyses are reported in directed acyclic graphs (DAGs), which can be found in the supplementary material (Additional File 1: Fig. S15 - S18).

The relationship between antibody levels and risk of SARS-CoV-2 breakthrough infection was modeled using logistic generalized additive models (GAM) with cubic splines, for examining the non-linear relationship between continuous predictor and the response.

To investigate the effect of antibody levels on the risk of SARS-CoV-2 breakthrough infection, we conducted a survival analysis using both Kaplan–Meier analysis and multivariate Cox regression modeling. Individuals with at least two doses were considered at risk of infection from M24 visit until the occurrence of the first reported episode of SARS-CoV-2 breakthrough infection, the receipt of the fourth dose of the vaccine, or the last day of study follow-up established for the analysis reported here (17 January 2023). Causal assumptions for each analysis are reported in DAGs, which can be found in the supplementary material (Additional File 1: Fig. S19 - S20).

To identify the most influential antibody predictors of breakthrough infection, we employed a Cox regression with LASSO regularization. This method helps to address collinearity by penalizing and shrinking the coefficients of highly correlated predictors, resulting in a more robust model. Prior to analysis, each antibody variable was log₁₀-transformed and scaled to have mean 0 and standard deviation 1. One thousand replicates of nested fivefold cross validation using λ equal to lambda.min (the value which minimized leave-one-out cross validation misclassification) were performed to assess model stability. For final regression, lambda.min and lambda.1se values were obtained using leave-one-out cross-validation, with models shown for the complete path of λ values.

Hierarchical clustering was performed using the complete linkage method on the Euclidean distances of the normalized variables. Heatmap was plotted using the R package pheatmap (v1.0.12).

Out of the 578 HCW recruited at the beginning of the pandemic and visited at 6 time points over 2020–2021 [17, 19, 23, 24], 393 participants came to the month (M)24 visit between 14 March and 25 April 2022 (32.0% lost to follow-up) (Fig. 1). At M24, most of the participants (n = 370) had received the primary series of vaccination (67.0% BNT162b2 and 33.0% mRNA-1273) and 287 received the 1st booster dose (0.7% BNT162b2 and 99.3% mRNA-1273) (Table 1). The mean time since the 2nd vaccine dose was 381 days (SD 90.7), while the mean time since the 3rd dose (booster) was of 115 days (SD 15.8). At M24, 99.7% of the individuals (n = 392) were seropositive. Up to M24, 61.8% of the participants (n = 243) had been previously infected by SARS-CoV-2 according to rRT-PCR (77.0%) or only serology data (23.0%), and 38.1% (n = 150) had no evidence of infection. The mean time since last infection was 172 days (SD 105.6). Of those without evidence of previous infection, 1 had one vaccine dose (0.7%), 20 (13.3%) had two vaccine doses, and 129 (86.0%) had three doses, while among those with infection, 11 (4.5%) were not vaccinated, 11 (4.5%) had 1 dose, 63 (25.9%) had 2 doses, and 158 (65.0%) had 3 doses. Among those with 3 doses (n = 287), 69 (24.0%) had evidence of infection previous to primary vaccination, 15 (5.2%) after primary vaccination but before booster, and 71 (24.7%) post-booster. Most of the study participants were females (73.0%) and had a mean age of 46.7 (SD 11.2) years (Table 1). Around 29.7% had underlying comorbidities, and among those, 78.6% were under chronic medication (Table 1). Among the 370 participants vaccinated with at least two doses, 109 (29.4%) had vaccine breakthroughs post-M24 detected by rRT-PCR from 14 March 2022 (M24) up to 17 January 2023 (Table 1). Seropositivity of study participants at M24 is described on Additional File 1: Table S1.

Infection or booster vaccination after dose 2 (between M12 and M24) (Fig. 2a, b) were associated with an increase in IgA levels against the receptor-binding domain (RBD), spike (S), and S2 antigens between the two study visits, whereas naive individuals vaccinated with 2 doses had a non-significant increase for RBD or S antigens (Fig. 2b). However, IgG levels waned or were maintained from M12 to M24 regardless of infection or booster vaccination after dose 2 (Fig. 2b). At M24, infection after dose 2 was associated with higher levels of IgA to S1 and RBD, and IgG to RBD, than booster vaccination (Fig. 2c). This was also the case for IgG and IgA against Omicron S (Additional File 1: Fig. S1). Regarding IgM, levels of anti-RBD and anti-S antibodies decreased from M12 to M24 in all three groups, while levels of anti-nucleocapsid (N) antibodies increased in the three groups, and levels of anti-S2 antibodies only decreased significantly for those naive vaccinated with 3 doses (Additional File 1: Fig. S2). In addition, at M24, anti-S1 IgM levels were higher in individuals with two doses and infection than in those naive vaccinated with 2 doses. These findings indicate that infection may induce a higher antibody response than vaccination for these antibodies.

Post-booster infection resulted in increased or stable IgG levels from M12 to M24 (Additional File 1: Fig. S3a, b). Conversely, pre-primary vaccination infections and post-primary pre-booster infections resulted in stable or decreased antibody levels from M12 to M24 (Additional File 1: Fig. S3b). Moreover, post-booster infection was positively associated with an increase in IgA antibody levels from M12 to M24, particularly for antibodies targeting N and RBD. In naïve vaccinated individuals, there was also an increase in IgA levels from M12 to M24 due to the booster dose.

At M24, post-booster infection was associated with elevated levels of both IgA and IgG to N and S antigens (Additional File 1: Fig. S3a, c). Compared to pre-primary infection, post-booster infection was associated with higher antibody levels to RBD, S, S1, and N antigens. Notably, post-booster infection led also to higher IgG levels compared to post-primary pre-booster infection (Additional File 1: Fig. S3c). Regarding IgM, there were some differences in magnitude, but the vast majority of participants had levels below the seropositivity cutoff (Additional File 1: Fig. S4). Post-booster infection was also associated with higher levels of IgA and IgG to S from Delta and Omicron variants compared to the other 3 groups, except for IgA against S from Delta variant, which was only higher than those naive and infected pre-primary vaccination. (Additional File 1: Fig. S5).

Among infected participants with 3 doses, a longer time since the last infection was associated with lower IgG levels for all antigens and variants assessed in non-adjusted linear models (Additional File 1: Fig. S6). Along these lines, post-booster infections were strongly associated with higher antibody levels, particularly to N, in models adjusted for age, comorbidities, chronic medication, infection post-primary pre-booster vaccination, infection pre-primary vaccination, sex, smoking, primary vaccine type, and adverse events (AEs) after doses 1, 2, and 3 (Fig. 3). In contrast, infections occurring before primary vaccination and post-primary pre-booster vaccination were only associated with higher anti-N but not anti-S antigens antibody levels (in models adjusted for age, comorbidities, chronic medication, sex, smoking, and primary vaccine type, or age, comorbidities, chronic medication, infection pre-primary vaccination, sex, smoking, primary vaccine type, AEs after doses 1 and 2, respectively). These findings indicate the timing of infection greatly influences the impact of previous infection on antibody levels against Wuhan, Delta, and Omicron variants.

Having 3 compared to 2 doses was positively associated with IgG levels against S, S1, RBD, and N from Wuhan and S from Delta variant in multivariable analysis (Additional File 1: Table S2). However, no significant association was detected with IgG against S from Omicron. Additionally, having 3 doses was negatively associated with anti-N IgG levels, reflecting lower incidence of infection in those vaccinated with 3 doses.

Primary vaccination with BNT162b2 followed by mRNA-1273 booster was associated with an increase in IgA levels to all antigens from M12 to M24, while those vaccinated with mRNA-1273 in primary and booster vaccination showed an increase only for antibodies to N and S2 (Additional File 1: Fig. S7a). IgG levels waned for RBD and S2 regardless of type of vaccination, while IgG levels to S increased for those primary vaccinated with BNT162b2. At M24, primary vaccination with BNT162b2 followed by mRNA-1273 booster was associated with higher levels of both IgA and IgG antibodies to S2 compared to 3 doses of mRNA-1273 (Additional File 1: Fig. S7b) and higher anti-S IgA levels. For Omicron and Delta S, levels of IgA to both variants and IgG to Delta variant were also higher in the heterologous vaccinees at M24 (Additional File 1: Fig. S8). In non-adjusted linear models, primary vaccination with BNT162b2 followed by mRNA-1273, compared to homologous vaccination with mRNA-1273, was associated with higher levels of IgG to S antigens for all variants (for all participants and for those infected) (Fig. 3, Additional File 1: Fig. S6).

Systemic AEs compared to local or no AEs after third dose were associated with increased levels of IgGs to S antigens from the three variants (for all of participants and for those naive) (Fig. 3, Additional File 1: Fig. S9) in models adjusted for age, comorbidities, chronic medication, sex, smoking, and primary vaccination type. Systemic AEs compared to local or no AEs after a second dose were only linked to increased levels of anti-S IgG in naive individuals with 3 doses (Additional File 1: Fig. S9), while systemic AEs after first dose were negatively associated with IgG levels to Wuhan S antigens in infected individuals (Fig. 3, Additional File 1: Fig. S6) in models adjusted for age, comorbidities, chronic medication, sex, smoking, and primary vaccination type. Age was negatively associated with anti-N IgG levels and positively associated with IgG to S from Delta variant among participants with 3 doses in non-adjusted linear models (Fig. 3). Smoking also exhibited a negative association with IgG levels to S1, RBD, and N from Wuhan variant and S from Omicron variant, among participants with 3 doses and with IgGs to all S antigens in infected individuals with 3 doses in models adjusted by age and sex (Fig. 3, Additional File 1: Fig. S6).

IgG levels to all antigens and variants tested at M24 were lower in those individuals with a post-M24 infection compared to the non-infected. (Fig. 4). This was also the case for IgA to S antigens (Additional File 1: Fig. S10) but not IgM levels, which were not correlated with post-M24 infection (data not shown).

Stratifying anti-S antibody levels in quartiles, we found that the risk of breakthrough infections for the lowest and highest quartiles began to diverge from the risk for highest quartile after around 50 days of follow-up for IgG and around 80 days for IgA (Fig. 5). Along these lines, when we examined the relationship between risk of breakthrough infection and antibody levels using a GAM model, the risk significantly decreased with increasing levels of anti-S and anti-N IgG and IgA antibodies (p < 0.01) (Fig. 6, Additional File 1: Fig. S11).

Multivariable models (adjusted for age, comorbidities, chronic medication, sex, smoking, infection, time since last infection, vaccination with 3 doses (ref: 2 doses), and time since last dose) estimated that higher levels of IgA and IgG to S antigens (RBD, S, S1, S2) were associated with protection against breakthrough infection (Table 2, Additional File 1: Table S3). Hazard ratios associated with each tenfold increase in antibody levels was 0.25 (95% CI, 0.11–0.57) and 0.57 (95% CI, 0.38–0.86) for IgG and IgA against Wuhan S, respectively. IgG and IgA levels to both Omicron and Delta S variants were also found to be associated with protection: HR Omicron 0.06 (95% CI, 0.01–0.26) for IgG, 0.20 (95% CI, 0.07–0.56) for IgA, HR Delta 0.30 (95% CI, 0.14–0.66) for IgG, 0.61 (95% CI, 0.39–0.96) for IgA (Table 2, Additional File 1: Table S3).

Altogether, these findings suggest that higher anti-S and anti-N IgG and IgA levels were associated with a reduced risk of a breakthrough infection during a period (2022) when Omicron was the dominant infecting variant in Spain.

Previous infection had a strong negative association with breakthrough infection (HR, 0.13; 95% CI, 0.06–0.29) in models adjusted for age, comorbidities, chronic medication, sex, smoking, and vaccination with 3 doses (ref: 2 doses) (Table 3, Additional File 1: Table S4). However, the analysis also revealed that the time since the last infection was negatively associated with the level of protection (HR associated to 30 days, 1.1, 95% CI, 1.03–1.20) (Table 3, Additional File 1: Table S4). In those participants vaccinated with 3 doses, previous infection following booster vaccination was also strongly associated with protection against breakthrough infection (HR, 0.043, 95% CI, 0.18–0.01). Nevertheless, infections occurring before primary vaccination and infections following primary vaccination but prior to booster vaccination were not associated with protection. These findings emphasize the importance of time since the last infection in the level of protection.

When the impact of previous infection was analyzed using a model adjusted for antibody levels, which are mediators of the protective effect, the HR associated with previous infection was notably reduced (HR at day 0, 0.27, 95% CI, 0.75–0.1). Nevertheless, the results still indicate that previous infection confers a protective effect independent of antibody levels.

Furthermore, in non-adjusted models, vaccination with 3 doses as compared with 2 doses was not found to be associated with protection against breakthrough infection (HR 1.54, 95% CI, 7.7–0.31, non-significant) (Table 3). This was also the case for models adjusted for previous infection (HR 1.45, 95% CI, 2.7–0.77, non-significant).

A strong degree of correlation was found within IgG and IgA levels to S antigens, while a moderate correlation was observed between IgG and IgA levels against S antigens, and IgG and IgA levels against N (Additional File 1: Fig. S12). Antibodies to N and S antigens were poorly correlated (Additional File 1: Fig. S12).

To identify the most influential antibody predictors of breakthrough infection while addressing their collinearity, we employed a LASSO penalized Cox regression model (Additional File 1: Fig. S13). This approach allowed us to effectively select the antibody variables that had the strongest association with the risk of breakthrough infection, while mitigating the impact of multicollinearity among the predictors. In the penalized regression analysis, which incorporated the levels of all 14 different antibodies as independent variables, only IgA to Omicron S and IgG to Wuhan N and Omicron S were selected (Additional File 1: Fig. S13). Those antibody variables consistently maintained a negative coefficient and were included in the model in all of the cross-validation rounds, reinforcing their robust association with breakthrough infection.

To further explore the relationship of antibody levels with subsequent breakthrough infection, and to find potential patterns (combined effect of antibodies) associated with breakthrough infection outcome, we conducted a hierarchical clustering analysis which included levels of IgG and IgA to Omicron S and IgG and IgA to N (Additional File 1: Fig. S14). We observed two distinct clusters (Additional File 1: Fig. S14), the first consisting predominantly of individuals who experienced breakthrough infections, while the second mainly comprised non-infected individuals. Within the predominantly infected cluster, most individuals had low levels of IgA to both S and N. Additionally, IgG levels to N were generally moderate to low, while IgG levels to S were relatively high for the majority of individuals in this cluster. In contrast, in the predominantly non-infected cluster, IgA levels to N were mostly low (except for a subset of participants), while IgA levels to S and IgG levels to N were notably higher for most individuals. Moreover, IgG levels to S were also higher in this cluster compared to the predominantly infected cluster. Notably, individuals who displayed high levels of all four antibodies had the lowest incidence of breakthrough infections. These findings suggest that a combination of elevated IgA and IgG against S and IgG against N may contribute to a reduced risk of breakthrough infections, where the association with elevated anti-N antibody levels may be a surrogate for additional immune responses related to recent infections.

To assess the individual contributions of IgG against S and N to the risk of breakthrough infection, we constructed an adjusted Cox regression model incorporating antibody levels of both IgG and IgA against N from Wuhan variant and against S from Omicron variant (Table 4) and additional covariates (age, comorbidities, chronic medication, sex, smoking, vaccination with 3 doses (ref: 2 doses), and type of primary vaccination). In this model, IgG to Omicron S (HR, 0.12, 95% CI, 0.027–0.50) and to Wuhan N (HR, 0.59, 95% CI, 0.37–0.93) were found to be independently associated with breakthrough infection (Table 4). No significant interactions were detected between IgG and IgA antibodies targeting S or between IgG and IgA targeting N.