Serum samples were collected from healthy vaccine recipients by Feinstein-Northwell Institute for Medical Research, Manhasset, NY (Institutional Review Board #20-1007) and by the National Institutes of Health’s Occupational Safety and Health Office located at Ft. Detrick, MD under the Research Donor Protocol (RDP). RDP participants were healthy NCI-Frederick employees and other NIH staff that donated blood samples for in-vitro research at the NCI-Frederick laboratories. The protocol is listed under NIH protocol number OH99CN046 and NCT number NCT00339911. Blood donors ranged in age from 25-76 years and included 67 females and 65 males; demographics are presented in Table 1 . Study participants were sampled 61-377 days after administration of the second dose of the same primary series vaccine (either BNT162b2 or mRNA-1273) serving as a pre-boost timepoint. Participants were then boosted with a homologous third dose, and then sampled at 1-month (15 to 45 days post-third vaccination) and at 6-months (165-195 days post-third vaccination). Blood samples were processed at the collection sites, sera were frozen and stored at -80°C and then shipped on dry ice to the NCI-Frederick Repository until requested for testing at the Vaccine, Immunity, and Cancer Directorate (VICD).

SomaLogic’s SomaScan v4.1 7K Assay platform was used to evaluate serum protein content longitudinally for significant changes in abundance at 1-month and 6-months following homologous third vaccination. Protein quantitated in sera collected pre-third dose set baseline values for each study participant. The SomaScan assay was performed on a Tecan Fluent 780 high throughput system according to manufacturer’s instructions at the NIH Center for Human Immunology. A complete list of the 7,289 targets analyzed is available at https://menu.somalogic.com. The data are reported in relative fluorescence units (RFU), a surrogate for protein concentration. The SomaScan Assay data were normalized using SomaLogic’s standardization procedure (26). In short, data were first normalized to correct for well-to-well variation in microarray hybridization steps. This was then followed by intraplate median signal normalization to correct for sample-to-sample differences that may be introduced due to technical assay effects. Plate scaling was applied to normalize global signal differences between plates, and calibration was applied to adjust for SOMAmer reagent-specific differences between tests. Finally, median signal normalization was performed via Adaptive Normalization by Maximum Likelihood (ANML) to harmonize data across multiple assay plates.

Serum IgG antibody to SARS-CoV-2 spike concentrations are reported as geometric mean binding antibody units per milliliter (BAU/mL) based on established serological WHO standards (27).

Proteomic data were stratified based on vaccine received and sex assigned at birth for these analyses. In total, data analyses were performed on 12 data subsets as follows: 2 vaccines assessed at 3 timepoints and against 2 sex assigned at birth stratifications.

Serum antibody levels and avidity are expressed as geometric mean concentrations (GMC) and geometric mean avidity levels (GMA), respectively, with corresponding 95% confidence intervals. A Mann-Whitney U test (Wilcoxon Rank Sum test) was used when comparing antibody content or avidity values between vaccine type, or between sex assigned at birth. A Wilcoxon Signed Rank test was used when serological tests were evaluated between timepoints. These tests were chosen because they do not have the requirement of a normal distribution of values, as in an unpaired or paired t-test, and work for smaller sample sizes; p<0.05 was considered significant.

Serum proteins were quantitated using the SomaScan v4.1 7K Assay platform (28, 29). The geometric mean value for each detectable protein within each cohort was quantitated in relative fluorescent units (RFU), and the magnitude of protein abundance change was expressed with the following formulas: log2(1-month value/pre-third-dose value) and log2(6-month value/pre-third-dose value). A 2-tailed t-test was performed to determine if the calculated geometric mean values were significantly different between the 2 timepoints. p-values in this assessment provided significance thresholds. Proteins having a p-value of <0.05 and a log protein abundance difference of >0.20 (significant increase) or <-0.20 (significant decrease) were identified as proteins that demonstrated significant differences in content.

Biological processes impacted by vaccination were evaluated through KEGG and REACTOME pathway functional enrichment analyses. Enrichment ratio and false discovery rate (FDR) were calculated for significant proteins using the WEB-based GEne SeT AnaLysis Toolkit (WebGestalt) 2 (30). EntrezGene was used to map proteins to gene identifiers. Background gene sets for enrichment analyses were created by selecting all unique EntrezGene names from all protein analytes detected in the SomaScan Assay.

As serological antibody responses to the 2 vaccines were largely analogous irrespective of vaccine or sex assigned at birth ( Tables 2 , 3 ), and because of the limited sizes of the vaccine recipient cohorts, the entire database was evaluated without segregation by sex assigned at birth or vaccine to identify pre-third-dose markers predicting 6-month antibody response levels. Attempts to develop predictive models specific for sex assigned at birth or vaccine were unsuccessful lacking sufficient predictive accuracy; larger sample sizes would be necessary to develop segregated models.

A pilot machine learning model was developed ( Figure 1 ) using Random Forest (RF) to predict antibody response levels at 6-months. Thirty-nine subjects were identified as higher responders (IgG > 5000 BAU/mL) and 44 subjects were identified as lower responders (IgG < 2000 BAU/mL) based on IgG levels at 6-months. A 2-tailed t-test was performed to determine if the calculated geometric mean values of the protein levels at pre-third dose for the 2 groups were significantly different from zero. One hundred and seven statistically significant markers were identified based on comparative changes between the extreme (higher or lower) responders (p<0.05). The markers were reduced to an 85-feature set by removing redundant predictors, or co-linear markers. The RF model was trained with data from 28 higher responders and 31 lower responders, and the model was tested on the remaining 11 higher responders and 13 lower responders and achieved a predictive accuracy of 79.17% on test data with an area under the curve (AUC) of 86.71%.

Assessments of vaccination responses were confounded with the observation that 18 mRNA-1273 and 19 BNT162b2 recipients had antibody reactive to SARS-CoV-2 nucleocapsid protein, despite no self-reports of infection indicating coronavirus infections or potential exposure. Moreover, 14 mRNA-1273 and 7 BNT162b2 vaccine recipients had positive nucleocapsid tests at 1-month and 6-month timepoints, possibly indicating subclinical infections. To evaluate the priming influence of the infections, the dataset was re-examined excluding vaccine-recipient samples that tested positive for coronavirus nucleocapsid. Evaluation of nucleocapsid-antibody negative sera did not significantly affect pre-third-dose predictive marker sets or pathways associated with higher or lower 6-month serology. Analyses of the nucleocapsid naïve dataset did produce 1 additional cellular process - mitophagy (selective degradation of defective mitochondria) -which contributed to model development (Data not shown).

The development of a predictive model of 6-month IgG antibody response levels from vaccination through proteomic assessments of sera required desegregated analyses due to study size. Consequently, we evaluated in-depth both the quality and robustness of the IgG responses by vaccine received and sex assigned at birth to assure validity of our desegregated assessment. Serological responses were comparable across both vaccine types and sexes assigned at birth at both 1-month and 6-months post-third vaccinations ( Figures 2A, B ; Tables 2 , 3 ); the only significant differences were detected in pre-third-dose sera (mRNA-1273 primary vaccine recipients demonstrated statistically higher antibody and avidity levels to spike compared to recipients of BNT162b2, irrespective of sex assigned at birth [antibody levels: p=0.0006 vaccines, p=0.0092 males, p=0.0436 females; avidity: p=0.0247]) ( Figures 2A, C, D ; Table 2). Additionally, the mean time between primary series vaccination and third dose of vaccine was different in the 2 vaccine groups: 294 days (164-377) for mRNA-1273 recipients and 265 days (61-377) for BNT162b2 recipients (p<0.0001, Mann Whitney).

In detail, the geometric mean anti-SARS-CoV-2 spike IgG antibody levels in mRNA-1273 and BNT162b2 recipients at 1-month were 8655 (95% CI: 7074-10591) BAU/mL and 7635 (95% CI: 6542-8910) BAU/mL, respectively; p=0.3008 ( Figure 2A ; Table 2 ). The geometric mean antibody avidity levels at 1-month for mRNA-1273 and recipients BNT162b2 were 5.5 (95% CI: 5.4-5.6) M and 5.4 (95% CI: 5.2-5.5) M, respectively; p=0.3640 ( Table 2 ). The 6-month geometric mean anti-SARS-CoV-2 levels for mRNA-1273 and BNT162b2 recipient sera were 3074 (95% CI: 2382-3966) BAU/mL and 2823 (95% CI: 2222-3587) BAU/mL, respectively; p=0.5872 ( Figure 2A ; Table 2 ). The geometric mean antibody avidity levels at 6-months for mRNA-1273 and BNT162b2 recipients were 5.3 (95% CI: 5.1-5.5) M and 5.3 (95% CI: 5.1-5.5) M, respectively; p=0.8964 ( Table 2 ).

When serology responses were analyzed according to sex assigned at birth, male and female recipients had geometric mean serum antibody levels at 1-month of 8124 (95% CI: 6757-9768) BAU/mL and 8088 (95% CI: 6804-9614) BAU/mL, respectively; p=0.7382 ( Figure 2B ; Table 3 ). At 6-months, male and female recipients had geometric mean serum antibody levels of 3175 (95% CI: 2423-4161) BAU/mL and 2747 (95% CI: 2193-3440) BAU/mL, respectively; p=0.6434 ( Figure 2B ; Table 3 ). There were also no statistically significant serum avidity differences based on sex assigned at birth (data not shown).

Further analyzing according to both sex assigned at birth and vaccine, male recipients of mRNA-1273 and BNT162b2 at 1-month demonstrated geometric mean serum antibody levels of 9400 (95% CI: 6936-12739) BAU/mL and 7173 (95% CI: 5707-9016) BAU/mL, respectively; p=0.1940 ( Figure 2C ; Table 3 ). Female recipients of mRNA-1273 and BNT162b2 at 1-month had geometric mean serum antibody levels of 8050 (95% CI: 6070-10675) BAU/mL and 8126 (95% CI: 6535-10103) BAU/mL, respectively; p=0.9851 ( Figure 2D ; Table 3 ). At 6-months, male recipients of mRNA-1273 and BNT162b2 had geometric mean serum antibody levels of 3879 (95% CI: 2570-5855) BAU/mL and 2682 (95% CI: 1856-3875) BAU/mL, respectively; p=0.2403 ( Figure 2C ; Table 3 ). Finally, at 6-months female recipients of mRNA-1273 and BNT162b2 had geometric mean serum antibody levels of 2540 (95% CI: 1839-3510) BAU/mL and 2963 (95% CI: 2133-4115) BAU/mL, respectively; p=0.6582 ( Figure 2D ; Table 3 ).

Proteomic profiles in serum from each vaccine recipient were assessed with the SomaScan v4.1 7K Assay platform before the third vaccination (after completion of 2-dose primary vaccine series), and then at 1-month and 6-months post-third vaccination; summary of changes in protein marker expression profiles of the respective cohorts are presented in Supplementary Table S1A (downregulated markers) and Supplementary Table S1B (upregulated markers) and Figure 3 . Significant proteins were selected based on p-value and average log abundance differences ( Figure 4 ; Supplementary Tables S1A, B ).

Sera collected from the male recipients of mRNA-1273 1 month post-third vaccination demonstrated upregulation of 11 markers and downregulation of 14, while sera from female recipients showed upregulation of 14 markers and downregulation of 472 ( Figure 3 ; Supplementary Tables S1A, B ). By 6 months, sera from the male recipients of mRNA-1273 showed upregulated 232 protein markers and downregulated 29, while the sera from the female recipient group demonstrated upregulated 102 markers and downregulated 255 ( Figure 3 ; Supplementary Tables S1A, B ).

The sera from the male recipients of BNT162b2 showed upregulation of 9 markers at 1-month and downregulation of 218 markers post-third vaccination, while sera from the female recipient cohort demonstrated upregulated 38 markers and downregulated 362 proteins ( Figure 3 ; Supplementary Tables S1A, B ). At 6-months post-third vaccination, sera from male recipients showed upregulated 83 markers and downregulated 220, while sera from female recipients demonstrated upregulated 172 markers and downregulated 229 ( Figure 3 ; Supplementary Tables S1A, B ).

When evaluated at the cohort level, sera from male recipients of either vaccine at 1-month post-third vaccination demonstrated 1 common upregulated marker (UB2D1/PolyUbiquitin K48), and 1 common downregulated marker (CXCL8, interleukin-8) ( Figure 3 ). In sera from the female recipient cohorts, 7 markers were upregulated (UBE2D1|UBB, CHAC1, LEP, CST5, CST2, INS) and 342 markers downregulated ( Figure 3 ). By 6-months, sera from male recipients of either vaccine upregulated 61 markers and downregulated 18, while sera from female recipients of either vaccine demonstrated upregulation of 86 markers and downregulation of 177 ( Figure 3 ; Table 4 ; Supplementary Tables S1A, B ). Upregulation of UB2D1/PolyUbiquitin K48 was common to sera from all recipient cohorts at 1-month after third vaccination, regardless of sex assigned at birth or vaccine received ( Figure 3 ).

Evaluation of the 10 most significant (highest statistical significance/abundance change) proteomic markers upregulated in sera at 1-month after the third dose of vaccine demonstrated that both vaccine groups modulated sets of vaccine-type associated markers, regardless of sex assigned at birth. Specifically, sera from both mRNA-1273 recipient cohorts showed upregulated UB2D1/PolyUbiquitin K48 (UBE2D1|UBB), Insulin (INS) and Band 4.1-like protein1 (EPB41L1), while sera from both BNT162b2 recipient cohorts showed upregulated UB2D1/PolyUbiquitin K48 (UBE2D1|UBB), Glutathione-specific gamma-glutamylcyclotransferase 1 (CHAC1), and Cancer/testis antigen 1 (CTAG1A|CTAG1B) ( Table 4 ).

Biological pathways were assessed based on observed proteomic marker changes according to vaccine received and sex assigned at birth. Proteins with significant changes between pre-third dose and 1-month or 6-months after third dose of vaccine were assessed through REACTOME and KEGG databases to evaluate differential pathways and cellular processes impacted (31, 32). A complete accounting at the cohort level of pathways up- and downregulated in 6-month sera compared to pre-third-dose sera measurements are listed in Table 5 .

Sera from male recipients of the mRNA-1273 vaccine demonstrated no significantly upregulated or downregulated pathways at 1-month, while sera from female recipients demonstrated an upregulation of 21 (including TLR signaling cascades, fibrin clot formation, and platelet aggregation) and downregulation of 132 pathways, (including membrane and vesical trafficking, platelet activation, and RNA transport) ( Figure 5 ; Table 5 ; Supplementary Table S2 ).

At 6-months, sera from male recipients of mRNA-1273 demonstrated upregulation of 51 pathways (including class I antigen processing, metabolism of RNA, eucaryotic translation and peptide chain elongation) and downregulation of 4 (including complement activation, coagulation cascades, and peptide ligand receptors). Sera from female recipients of mRNA-1273 demonstrated upregulation of 30 pathways at 6-months (including class I antigen processing, metabolism of RNA, and peptide chain elongation) and downregulation of 25 (including endocytosis, membrane and vesicle trafficking and platelet activation) ( Table 5 ; Figure 5 ; Supplementary Table S2 ).

One month after booster, the sera from male recipients of BNT162b2 demonstrated upregulation of 11 pathways (including APC-Cdc20 degradation and TICAM1 and RIP1 signaling) and downregulation of 30 (including platelet activation, membrane trafficking, and RNA transport), while sera from the female recipients demonstrated upregulation of 4 pathways (including TICAM1 and RIP1 mediated signaling) and downregulation of 67 (including membrane and vesical trafficking, platelet activation, and RNA transport) ( Table 5 ; Figure 5 ; Supplementary Table S2 ). Upregulated pathways in sera from both male and female recipients of BNT162b2 included TICAM1, RIP1 mediated IKK signaling ( Figure 5 ; Supplementary Table S2 ).

By 6-months post-third vaccination, sera from the male recipients of BNT162b2 showed upregulation of 17 pathways or processes associated with protein synthesis (mRNA processing and translation), regulation of cellular function (SLIT – ROBO pathway), and suppression of RIG-1 signaling pathway (DDX58/IFIH1) and downregulation of 19 (including membrane trafficking, endocytosis, and EGFR, RTK, MET, CD28, SCF-KIT, and ESCRT signaling). The sera from the female recipients of BNT162b2 demonstrated upregulation of 20 pathways and processes (including innate immune system activation, class I antigen processing, glycolysis and neddylation) and downregulation of 14 (including membrane and vesical trafficking, endocytosis, and platelet activation) ( Table 5 ; Figure 5 ; Supplementary Table S2 ).

We evaluated common trends detected in pathways impacted by vaccination and found 17 pathways, including membrane trafficking, endocytosis and EGFR signaling pathways, were downregulated in sera from male recipients of BNT162b2 at 1-month and at 6-months, and 1 upregulated (metabolism) ( Figure 6 ; Table 5 ). Likewise, sera from female recipients of BNT162b2 demonstrated 12 common pathways downregulated at 1-month and 6-months, most notably platelet activation, vesicle trafficking, endocytosis and signaling by RAB ( Figure 6 ; Table 5 ). No comparisons were possible between sera from the male recipients of BNT162b2 and mRNA-1273 at 1-month because no significant pathway changes were noted in the mRNA-1273 group even though there were significant marker changes ( Figure 6 ; Table 5 ). There was 1 common pathway upregulated in sera from female recipients of either vaccine at 1-month post-third vaccination: synthesis, secretion, and deacylation of Ghrelin ( Figure 6 ; Table 5 ). Sera from recipients of either vaccine, regardless of sex assigned at birth (except male recipients of mRNA-1273), demonstrated upregulation of translation, peptide chain elongation and RNA metabolism and processing at 6-months ( Figure 6 ; Table 5 ).

The overarching goal of this study was to investigate the possibility of identifying proteomic markers in pre-boost sera that are predictive of humoral response robustness to vaccination at later timepoints. Predictive modeling was performed as a pilot effort to investigate the utility of machine learning to identify markers or develop models of vaccine responsiveness ( Figure 1 ). Change in anti-spike IgG antibody levels (BAU/mL) in pre-third vaccination and 6-month post-third vaccination sera were analyzed to identify vaccine-recipient samples with either “higher” or “lower” antibody content. The predictive potential of proteomic markers of avidity responses were not analyzed due to the limited range of AI measurements. The sera with the lowest and highest quartile of antibody titers were identified as “lower” and “higher” responders ( Figure 7 ). Specifically, 44 vaccine-recipients with 6-month serum IgG anti-SARS-CoV-2 Spike antibody less than 2000 BAU/mL (7 male and 11 female recipients of mRNA-1273, 13 male and 13 female recipients of BNT162b2) were identified as relatively “Lower” responders, and 39 vaccine-recipients with sera containing greater than 5000 BAU/mL (9 male and 10 female recipients of mRNA-1273, and 8 male and 12 female recipients of BNT162b2) were identified as relatively “Higher” responders ( Figures 1 , 7 ; Table 6 ). The demographics of these 2 populations were well-balanced concerning sex assigned at birth and vaccine-received ( Table 6 ).

Random Forest (RF) modeling, which combines outputs of multiple “decision trees” to reach a single result, was used for predictive modeling. Attempts to develop predictive models restricted by vaccine or sex assigned at birth did not return results with sufficient power due to the limitations of small cohort sizes. However, evaluation of the entire dataset as a single cohort returned a productive model that was statistically different from random selection ( Figure 1 ).

An RF model with 85 marker values from the pre-third vaccination sera could predict higher (>5,000 BAU/ml) and lower (<2,000 BAU/mL) responders at month-6 with 79.17% accuracy ( Figure 1 ). The associated markers and comparative changes are listed in Supplementary Table S3 ( Figure 7 ). Protein markers with the highest predictive power (i.e. markers that contributed the most to the model) were associated with complement cascade and activation, signaling by interleukins, tumor necrosis factor receptor (TNFR) apoptotic signaling, IL-17 signaling and phosphoinositide 3-kinase (PI3K) signaling ( Table 7 ).