Study design

A nested case-control study was designed within a large prospective cohort of frontline workers from eight locations in the US during early Omicron predominance (defined here as December 2021 through September 2022). Breakthrough cases were defined as those with a first-time SARS-CoV-2 infection at least 14 days following a third dose of COVID-19 origin strain WA-1 monovalent mRNA vaccine, and controls were defined with the same criteria but without a breakthrough infection.

Participants

Beginning in July 2020, frontline workers were followed in prospective cohorts through the Arizona Healthcare, Emergency Response, and Other Essential workers Study (HEROES) and the Research on the Epidemiology of SARS-CoV-2 in Essential Response Personnel (RECOVER) sites in Arizona, Florida, Minnesota, Oregon, Texas, and Utah [10,11]. Briefly, eligible participants included adults who worked at least 20 hours per week in occupations requiring frequent direct contact with non-household members (i.e., healthcare workers, first responders, and other public-facing frontline workers). Upon enrollment, participants completed a survey on sociodemographic characteristics, occupation, health status, health-related behaviors, and prior SARS-CoV-2 infection. COVID-19 vaccination information was obtained through surveys that were regularly sent to participants when they became eligible for vaccination according to evolving vaccine guidelines and validated through electronic health records or vaccine registries. Additional surveys were conducted upon infection, and information on mask use and exposures was collected quarterly, upon infection, or upon onset of illness symptoms.

Participants provided a mid-turbinate nasal specimen weekly and with onset of any illness symptoms. These specimens were tested for SARS-CoV-2 using reverse transcription-polymerase chain reaction (RT-PCR) at the Marshfield Clinic Laboratory (Marshfield, WI). Additionally, blood samples were collected at the following time points: (1) upon enrollment; (2) quarterly; and (3) approximately 14–60 days after any immunity-conferring event, such as SARS-CoV-2 infection or COVID-19 vaccination. All study protocols were reviewed and approved by the Institutional Review Boards at each site, and participants provided informed consent for all study activities.

Eligibility criteria and case definition

Omicron breakthrough infections were defined as SARS-CoV-2 infections caused by the Omicron variant occurring at least 14 days following a third dose of COVID-19 mRNA vaccine in individuals with no history of previous infection. In-study infections were determined using the self-collected mid-turbinate sample provided weekly and with any onset of symptoms. Prior infection history was ascertained via self-report upon study enrollment or a negative qualitative antibody result from the baseline blood draw using a locally-developed and validated semi-quantitative enzyme-linked immunosorbent assay (ELISA), which measured antibody binding to the receptor binding domain (RBD) and S2 subunit domain (S2) of the SARS-CoV-2 Washington-1 spike protein, as previously described [15]. Variant of infection was confirmed by whole-genome sequencing of eligible specimens or estimated based on the state-specific predominant variant at the time of infection according to Centers for Disease Control and Prevention data [16].

Participants were eligible for inclusion if they had received three doses of an mRNA COVID-19 vaccine, did not withdraw from HEROES/RECOVER prior to Omicron becoming the site-specific dominant variant, had a blood draw at any time after the third vaccine dose but prior to any additional COVID-19 vaccine or any SARS-CoV-2 infection, and had no history SARS-CoV-2 infection at the time of the blood draw. Among those included individuals, a case was then defined as a participant who tested positive for SARS-CoV-2 infection after the blood draw and prior to any additional vaccine dose (Fig 1).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Flowchart showing inclusion and exclusion criteria for nested case-control studies within the HEROES/RECOVER cohort of frontline workers: 181 cases of first-time post-vaccination Omicron infection following a third dose of COVID-19 origin strain WA-1 monovalent mRNA vaccine) were matched 1:1 to controls based on study site, sex, race/ethnicity, number of chronic conditions, age, and time between 3^rd vaccine dose and blood draw.

https://doi.org/10.1371/journal.pone.0347602.g001

Matching

Subjects were matched into pairs of case and control individuals in a 1:1 ratio using a greedy matching algorithm based on the following criteria: study site (Tucson, AZ; Phoenix, AZ; Other location, AZ; Temple, TX; Portland, OR; Duluth, MN; Salt Lake City, UT), sex, race/ethnicity (non-Hispanic/White; non-Hispanic/Black; non-Hispanic/Asian; Hispanic; other race), number of chronic conditions (0, 1, 2, or 3 + chronic conditions), age (±3 years), and time between third vaccine dose and blood draw. If no exact match was found, then the case was excluded from the analysis. Self-reported chronic conditions included asthma, chronic lung disease, cancer, diabetes, heart disease, hypertension, immunosuppression, kidney disease, liver disease, neurologic or neuromuscular disease, and autoimmune disease. To ensure a close match on the time between a third vaccine dose and blood draw, the control selected was the one with the smallest difference in the number of days between the vaccination and blood draw dates as compared to the case. For case individuals in which the blood draw occurred less than 150 days after vaccination, this difference must have been less than 21 days. Although immunosuppressed individuals were a minority population, the Fisher’s exact test was used to ensure immunosuppression was not overrepresented in either case or control groups and would not significantly confound findings related to immune activity between the experimental groups.

Sample preprocessing and proteomics

Blood samples were collected in 8.5 mL Vacutainer tiger-top serum separator tubes (SST) and allowed to clot for at least an hour at room temperature. Samples were then centrifuged at 1300 rpm for 15 minutes at 4°C and stored at 4°C for up to 24 hours on weekdays or at −20°C on weekends before receipt at the University of Arizona or other site research laboratories for preprocessing. After receipt, if frozen, the sample was thawed at room temperature for 1 hour, and the serum was divided into 1.8 mL aliquots. Individual aliquots were stored at −80°C until further analysis. Serum aliquots were sent to SomaScan Assay platform [12] for proteomics profiling in 7k format. Relative abundance, measured in relative fluorescent units (RFU), was returned for each SOMAmer (tentatively annotated protein) per sample.

Data preprocessing and quality control

To remove assay and sample bias from proteomics signals, the standard SomaScan v4.1 data standardization pipeline was applied. The steps are as follows: 1) “hybridization control normalization” using twelve hybridization control sequences added to each sample to adjust for biases arising from hybridization, 2) “intraplate median signal normalization” using all SOMAmer reagents on a sample to remove systematic biases between samples within a plate, 3) “plate scaling” using a pooled calibrator sample across plates of identical sample matrices to adjust for inter-plate variation, 4) “calibration” using a pooled calibrator to correct for SOMAmer-level variation, and 5) “adaptive normalization to a reference” using a healthy cohort’s SOMAmer RFUs as reference to standardize each sample’s overall signal. A final check is performed with replicate quality control samples that were run alongside clinical samples to assess the accuracy of SOMAmer reagents’ signal compared to the quality control samples.

The acceptance criteria for steps 1–3 require scale factors to be between 0.4 and 2.5, and between 0.8 and 1.2 for step 4. Step 5 required that at least 30% of the sample value lies between two standard deviations of the healthy population reference.

Statistical modeling

We utilized the Wilcoxon signed-rank test to evaluate the between-group differences for the number of chronic conditions, Fisher’s exact test for the number of immunosuppressed versus non-immunosuppressed individuals, and two-sample t-test for other continuous characteristics, including age, days from third dose to case infection, and average hours exposed to COVID per week. Necessary transformations were applied to ensure the normality of the data distribution.

Given the high-dimensional nature of the proteomics data, a robust feature selection algorithm was needed to identify key proteins of interest and reduce the risk of model overfitting. Further, to handle the matched pair structure and the binary response variable, we chose to utilize elastic net regularized conditional logistic regression, hereon referred to as EN-CLR, as our model. In general, penalized regression has the property of “selecting” important features by reducing the coefficients of less relevant features to zero. We opted to use the elastic net penalty instead of the LASSO or Ridge penalty as the LASSO penalty is liable to select only one covariate out of a set of highly correlated but equally “important” covariates, which is an undesirable property for identifying biologically important features, and the Ridge penalty tends to introduce algorithmic instability for the penalized conditional logistic regression [17]. The elastic net penalty has been shown to generate better performing models in assessing datasets with many correlated features compared to LASSO [18,19] – as is the case with our dataset (S1 Fig).

To ensure stability in our results, rather than directly applying the bootstrap method to the original dataset, we generated 1,000 bootstrap resampled datasets and applied EN-CLR to each bootstrap. We performed bootstrapping by resampling entire strata with replacement with random seeds 1–1000. The same seeds were used for downstream analyses involving nondeterministic outcomes. For each bootstrap dataset, the parameter determining the strength of regularization was fitted using 10-fold cross-validation, and the model with the lowest deviance was selected. Each resulting model contains a set of proteins with non-zero coefficients yielding a total of 1,000 sets of selected proteins. Additionally, we recorded the frequencies that each protein was selected and denoted the frequency set as.

The proposed method is closely related to the conventional stability selection as originally characterized by Meinshausen and Bühlmann [20] but differing in three regards: 1) elastic net penalty instead of LASSO penalty, 2) bootstrapping instead of subsampling, and 3) cross-validation to obtain the regularization parameter instead of predefining an array of, then selecting the value maximizing a predictor’s selection probability. Meinshausen and Bühlmann demonstrated that bootstrapping and subsampling exhibited similar behaviors while the selection of was not typically a strong determinant of selection outcome.

All statistical analyses were performed using R version 4.2.2. Specifically, the elastic net regularized conditional logistic regression models were fit using the clogitL1 R package with the participant outcome (breakthrough infection/control) as the binary response variable and all proteins (7,289 total) as independent variables [17]. Unregularized conditional logistic regression models were fit using the survival R package with case (breakthrough infection) versus control status as the binary response variable and the log₂ transformed relative fluorescence units of proteins resulting from stability selection (28 total) as independent variables. Each package provided cross-validation functionalities, and each sample remained with their stratum during cross-validation. Data for each protein was scaled to a mean of 0 and standard deviation of 1 prior to model fitting.

Pathway analysis

To better understand the functional implications of the highly selected (HS) proteins, we utilized two methods of pathway analysis: 1) over-representation analysis (ORA) which leverages the hypergeometric distribution to assess how well a set of proteins of interest is represented in a pathway compared with random chance, and 2) Signaling Pathway Impact Analysis (SPIA) [21] which combines ORA with information about the total accumulated perturbation of protein abundance within pathways between groups, to identify pathways of importance.

We conducted overrepresentation analyses using the clusterProfiler R package [22], with all the available proteins as the background. P-values for significance of overrepresentation were false discovery rate (FDR)-adjusted for multiple comparisons to avoid false positives. For SPIA, we first calculated the log₂ fold change of each protein. Because SPIA’s internal KEGG pathways were outdated, we used KEGG’s API to retrieve the most recent pathway versions and processed them with SPIA. SPIA reports and combines p-values for both pathway overrepresentation and the accumulated perturbation of protein abundance between conditions. This combined p-value is then FDR adjusted and reported in Results.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. EN-CLR on 1000x bootstrap resamples selects 28 proteins in a majority of models.

A) Red and blue bar colors indicate lower and higher abundance of protein in cases compared with controls, respectively. B) Volcano plot showing mean log₂ effect sizes versus -log₁₀ p-values of two-sample paired t-tests on log2 RFU between conditions. Highly selected proteins are labeled and shown in red.

https://doi.org/10.1371/journal.pone.0347602.g002

Participant characteristics and matching

Prior to matching, 385 and 371 participants were eligible for the study as cases and controls, respectively. Their characteristics are summarized in Supplemental S3 Table. Following the matching process, a total of 362 participants (181 cases and controls) was included in the final matched dataset, and their characteristics are summarized in Table 1. The average age for cases and controls was 43 years (SD = 10 years). Most participants were non-Hispanic and white (97%), female (74%), and had no self-reported chronic conditions (70%). The average number of days between the third vaccine dose and blood draw was 28 days for cases and 26 days for controls. No significant differences were observed between case and control groups in any of the measured characteristics, including variables matched with tolerances (e.g., age and number of chronic conditions). The final match cohort remains representative of the broader eligible population.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Demographics and health characteristics for frontline workers participating in the study. Participants were matched on site, age, race/ethnicity, gender, number of chronic conditions, and days between third vaccine dose and blood draw. Matching differences in age and number of conditions are tolerated within ±3 years and ±1 condition, respectively.

https://doi.org/10.1371/journal.pone.0347602.t001

Protein markers for Omicron breakthrough infection cases

We applied the proposed EN-CLR method to the proteomics data. Out of the 7289 candidate proteins, 3113 were selected in at least one bootstrap model, reinforcing concerns of overreliance on a single model. We therefore only considered the proteins selected in the majority of the models (selection frequency) as our final set of proteins. These proteins, referred to as “highly selected proteins” (HS proteins), were used for downstream functional analyses (Fig 2A). A total of 28 proteins met this requirement and were subsequently used to fit a classical conditional logistic regression model to determine the expected difference in protein abundance between cases and controls (S1 Table). While multiple regression in high dimensions and penalized regression coefficient values do not easily lend themselves to interpretation, the directionality of HS proteins’ coefficients was consistent across every bootstrap resample, and 10 of the HS proteins were found to be lower in abundance in case individuals compared with controls while 18 were found to be higher. Among those HS proteins, four belong to the chemokine family (CXCL2, CXCL3, CCL19, CCL23), which are essential for virus control and are directly involved in the COVID 19 infection [23]. Additionally, the T-cell development cytokine IL-7 showed higher serum levels in breakthrough infection cases.

We compared our results to a univariate method by selecting proteins of interest based on a threshold fold change and p-value derived from the paired two-sample t-test performed on each protein. While many proteins were nominally significant, adjusting for FDR suggested insufficient evidence to assign importance to any protein, confirming the need for more advanced feature selection methods (Fig 2B). The selection of HS proteins differed considerably from a threshold fold change and p-value approach, suggesting sensitivity to higher order relationships.

Enrichment of immune-related pathways from pathway analysis

To perform ORA, we focused on two Gene Ontology (GO) gene sets: Biological Process (BP) and Molecular Function (MF). Within the BP category, 13 pathways related to immune cell migration and chemokine response were enriched at FDR < 0.15 with key proteins from the CXC chemokine ligand family (CXCL2 and CXCL3) and CC chemokine family (CCL19 and CCL23). Among these, 3 pathways (“chemokine-mediated signaling pathway,” “response to chemokine,” and “cellular response to chemokine”) showed enrichment at FDR < 0.05 (Fig 3). For MF, 6 pathways associated with chemokine/cytokine activity and immune receptor binding were found to be enriched (FDR < 0.15), with 2 pathways (“chemokine activity” and “chemokine receptor binding”) achieving FDR < 0.05. To perform SPIA, we used the latest version of the Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway database: the only database supported by the tool. Overall, 5 pathways with at least 2 proteins within each pathway were enriched at the FDR < 0.15, among which 3 pathways had FDR < 0.05. All enriched pathways were cytokine/chemokine-related. Each pathway exhibited downregulation in case individuals compared to controls. S2 Table provides a detailed list of proteins enriched in each pathway.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Dot plot showing pathway enrichment in immune-related pathways.

The pathway dataset is encoded as dot color while number of highly selected proteins found in each pathway is encoded as dot size. Over-representation analysis was performed using the Gene Ontology-Biological Process (GO-BP) and Molecular Function (GO-MF) databases, while signaling pathway impact analysis (SPIA) was performed using KEGG pathways. Though SPIA reports the activation/inhibition of pathways, all shown KEGG pathways were downregulated in case individuals compared with controls.

https://doi.org/10.1371/journal.pone.0347602.g003

Notably, many pathways were enriched by the same core sets of highly selected chemokines (CXCL2, CXCL3, CCL19, CCL23), suggesting that these proteins function within interconnected chemokine response processes rather than representing isolated findings. Additional selected proteins, including IL-7, LGALS1, HAVCR2, SELE, and PLVAP, mapped to complementary pathways involving immune exhaustion and regulation, endothelial activation, and leukocyte adhesion, further support the immune signaling network. Moreover, cytokine/chemokine-related pathways remain enriched over a range of threshold selection frequencies (and) used to define the HS protein set. Together, these results demonstrate that our agnostically selected set contains proteins converge on biologically coherent immune pathways that are relevant to breakthrough infections.