Refined phenotyping of vaccine responses reveals transcriptomic determinants of neutralizing antibody heterogeneity

Abstract

Individuals exhibit considerable heterogeneity in vaccine-induced antibody responses, yet commonly used binary classifications may overlook intermediate patterns. More nuanced grouping could better capture inter-individual differences. Here, we analyzed longitudinal neutralizing antibody (NAb) trajectories in 73 adults after inactivated SARS-CoV-2 vaccination. Unsupervised analysis identified three phenotypes: Low-Delayed Responders (LR), with a modest, delayed NAb rise to day 30; Rapid-Stabilizing Responders (RS), peaking at day 7 and plateauing thereafter; and Continuous Increase Responders (CI), exhibiting sustained increases. Between days 0 and 7, these groups diverged in immune activation: LR showed limited pathway activation or cell shifts; RS exhibited early innate activation with reduced dendritic cells; CI mounted innate and adaptive responses with increased naive B cells. These differences culminated at day 7, when CI exhibited enhanced antigen presentation and Th1-related pathways, accompanied by higher IFN-γ and IL-2 T cell responses. CI also showed post-transcriptional regulation of innate signaling, including HLA-F/H splicing, 3’UTR shortening in IKBKE and HRAS, and biased IGHV4-59–IGHJ4 usage. Finally, we developed a baseline gene model accurately predicting LR individuals. Our work refines responder classification and provides molecular insights into antibody heterogeneity, laying groundwork for early stratification and personalized vaccination.

Introduction

Vaccines have been remarkably successful in preventing infectious diseases^1–3, yet individuals vary widely in their immune responses^4–8. This heterogeneity is particularly evident in antibody responses, with some individuals generating high levels quickly, while others exhibit delayed or lower responses^9,10. Although this phenomenon has been observed across various vaccines^11–13, the large-scale SARS-CoV-2 vaccination efforts have particularly highlighted the inter-individual variability in antibody responses^14–21, underscoring the need for a deeper understanding of this heterogeneity. Understanding these mechanisms is crucial for improving vaccine design and identifying individuals at risk of insufficient protection.

Previous studies on the heterogeneity of vaccine-induced antibody responses across various vaccine types have typically stratified individuals into high or low responders according to antibody titers following immunization^{9,15,16,22–25}. These studies have shown that high and low responders differ in immune cell composition^17,23, cytokine levels²⁴, BCR characteristics¹⁶, and transcriptomic signatures²⁵. Notably, transcriptomic analyses have identified splicing differences in HLA class I genes as a key factor contributing to the variability in NAb responses to SARS-CoV-2 mRNA vaccines¹⁵. However, such static classifications overlook the dynamic nature of antibody responses. Numerous studies have demonstrated that antibody generation is dynamic, with varying rates of production and persistence over time^26–28. This temporal complexity is not captured by binary classifications, which may miss intermediate or atypical response patterns. Studies integrating longitudinal antibody data with data-driven approaches to uncover this variability are still limited.

In this study, we established a longitudinal cohort of 73 healthy adults immunized with two doses of inactivated SARS-CoV-2 vaccine. We assessed NAb titers in serum samples collected before the first dose (day 0), on day 7, and on day 30 after the second dose. Bulk RNA sequencing was performed on Peripheral blood mononuclear cells (PBMCs) collected from day 0 and day 7. We applied unsupervised analysis to NAb trajectories across multiple time points and identified three distinct responder groups based on their patterns over time. Low-Delayed Responders (LR) showed a modest, delayed increase in NAb levels from day 0 to day 30, while Rapid-Stabilizing Responders (RS) exhibited rapid increases in NAb levels by day 7, followed by a plateau. Continuous Increase Responders (CI) showed a steady increase in NAb levels throughout the day 0 to day 30 period. Comprehensive profiling of immune signatures revealed group-specific patterns across gene expression, alternative splicing (AS), alternative polyadenylation (APA), and B cell receptor (BCR) repertoire. Notably, we observed that immune-related genes exhibited specific patterns of regulation, with differential splicing in HLA-F and HLA-H and differential polyadenylation in IKBKE and HRAS. To facilitate early identification of LR, we proposed a predictive model based on 8 gene expression features at day 0. In summary, we identify distinct patterns of NAb responses that could improve immune stratification and provide important insights for optimizing precision vaccination strategies.

Results

Vaccination induces three distinct patterns of neutralizing antibody responses

To investigate inter-individual variability in vaccine-induced NAb responses, we conducted a longitudinal study of 73 healthy adults vaccinated with a standard two-dose schedule of an inactivated SARS-CoV-2 vaccine (see Methods for details). We collected peripheral blood samples at three time points, including before the first dose (baseline, day 0), 7 days (day 7), and 30 days (day 30) after the second dose. Serum was systematically isolated from peripheral blood samples collected at three time points, yielding a total of 219 specimens for downstream serologic analyses. Using NAb titers quantified from these samples, we identified three distinct responder groups via unsupervised analysis. To further investigate gene regulatory mechanisms underlying the classified groups, we performed bulk RNA sequencing of PBMCs isolated from peripheral blood at day 0 and day 7 (n = 146 total samples). In parallel, we developed predictive models for vaccine-induced NAb responses based on pre-vaccination gene expression profiles (Fig. 1A).

Fig. 1. Vaccination induces three distinct patterns of neutralizing antibody responses.

A Schematic of the study design. Blood samples were collected from 73 participants at D0, D7, and D30 after receiving two doses of inactivated SARS-CoV-2 vaccine. Serum and PBMCs were isolated from these blood samples. NAb titers were measured via competitive ELISA at all time points. PBMCs from D0 and D7 underwent whole-transcriptome sequencing. Unsupervised analysis of longitudinal NAb levels identified three distinct response groups. Differential transcriptomic analyses and baseline gene expression-based predictive modeling were performed. Graphical elements in this panel were created using FigDraw (www.figdraw.com) under authorized use. Credit: by figdraw.com. B T-SNE plot shows dimensionality reduction of all participants (N = 73) based on NAb changes (D7–D0, D30–D0, D30–D7). Each dot represents a biologically independent sample (N = 29 in red; N = 29 in blue; N = 15 in green). C Boxplots show NAb levels at each time point per group, with line charts depicting median trends over time. Each dot represents an independent sample colored by group. Boxplots depict the distribution of clinical characteristics across response groups (LR, RS, CI), including sex (D), age (E), BMI (F), vaccine dose interval (G) and vaccine product combination (H). Statistical significance was assessed by Fisher’s exact test for sex (D) and by Kruskal–Wallis tests for age (E), BMI (F), and vaccine dosing interval (G). Fisher’s exact test was also used to evaluate vaccine product combination distribution (H), with a significance threshold of P < 0.05. D0, D7, and D30 indicate day 0 (baseline), day 7, and day 30 after the second vaccine dose, respectively. PBMCs indicates peripheral blood mononuclear cells; NAb indicates neutralizing antibody; BMI indicates body mass index; BBIBP indicates two-dose BBIBP-CorV; CoronaVac indicates two-dose CoronaVac; Mixed represents to one dose of BBIBP-CorV and one dose of CoronaVac, in any order.

Baseline levels of immunoglobulin G (IgG) and immunoglobulin M (IgM) antibodies specific to the receptor-binding domain (RBD) of the wild-type SARS-CoV-2 spike protein were generally low across all participants, indicating minimal pre-existing immunity (Supplementary Fig. 1A and Supplementary Data 1). Individual NAb titers against the wild-type SARS-CoV-2 RBD were measured at days 0, 7, and 30 using a competitive inhibition assay, as previously described¹⁸, to evaluate both response magnitude and temporal dynamics. We observed distinct temporal patterns in NAb titers across participants. While some individuals showed early peaks that subsequently declined, others displayed sustained increases or minimal changes, reflecting substantial inter-individual heterogeneity (Supplementary Fig. 1B and Supplementary Data 1). To further characterize this variation, we performed t-SNE based on changes in NAb titers (day 7 - day 0, day 30 - day 0, day 30 - day 7), which revealed three distinct response clusters (Fig. 1B). Additionally, hierarchical clustering with k = 3, supported by gap statistic analysis, also identified the same three groups (Supplementary Fig. 1C, D). The concordant results from these two independent methods further support the robustness and consistency of the identified response patterns. The three response groups exhibited distinct NAb dynamics, with differences in onset speed, peak timing, and overall magnitude (Fig. 1C). Accordingly, we defined them as Low-Delayed Responders (LR), Rapid-Stabilizing Responders (RS), and Continuous Increase Responders (CI). LR individuals showed a modest and delayed increase in NAb titers, with little change by day 7 and only a limited rise by day 30. RS participants mounted a rapid NAb response that peaked by day 7 and then gradually plateaued through day 30. In contrast, CI individuals exhibited a gradual and continuous increase from day 0 to day 30.

We next examined whether these response patterns were associated with demographic or clinical differences. Stratification was independent of sex, age, BMI, and vaccine dose interval. Sex distribution was similar among LR (44.8% female), RS (55.2% female), and CI (60.0% female) groups (Fig. 1D; P = 0.65, Fisher’s exact test). No significant differences were observed for age (Fig. 1E; P = 0.44, Kruskal–Wallis test), BMI (Fig. 1F; P = 0.79), or vaccine dose interval (Fig. 1G; P = 0.55). Although the response groups were defined independently of vaccine product, the distribution of vaccine product combination differed significantly across groups (Fig. 1H; P < 0.05, Fisher’s exact test), suggesting a potential confounding effect. Specifically, 55.2% of LR participants received two doses of BBIBP-CorV, whereas 65.5% of RS and 53.3% of CI participants received two doses of CoronaVac. Therefore, to account for potential confounding, all comparisons of post-vaccination responses across responder groups included vaccine product combination as a covariate.

To further support these distinct temporal patterns, we conducted quantitative comparisons of NAb changes across all measured intervals (day 0–7, day 0–30, and day 7–30). After adjustment for vaccine product combination, significant differences were observed at each interval (Supplementary Fig. 1E; FDR < 0.05). RS and CI increased more than LR between day 0 and day 7, with RS rising fastest. By day 30, CI showed the greatest overall increase, while LR had the smallest. From day 7 to day 30, RS increased less than LR, but CI remained significantly higher than both. Together, these findings quantitatively validate the classification of LR, RS, and CI groups based on their distinct temporal NAb response patterns.

To determine whether these response patterns were influenced by baseline immunity, we first compared anti-RBD IgG and IgM levels across the groups at day 0. Overall, no significant differences were observed (Supplementary Fig. 1F and Supplementary Data 1), indicating similar baseline antibody levels prior to vaccination. We then examined the relationship between antigen-specific immunoglobulin responses and NAb titers. Correlations between anti-RBD IgG/IgM and NAb were calculated within each responder group. Overall, anti-RBD IgG showed positive correlations with NAb in most groups and intervals (Supplementary Fig. 1G, H), consistent with its predominant contribution to neutralizing antibody responses following inactivated SARS-CoV-2 vaccination²⁹. Notably, LR group showed anti-RBD IgM exhibited a slightly higher correlation with NAb from day 0 to day 7 (Supplementary Fig. 1H; R = 0.63 for IgM vs. 0.55 for IgG). This suggests that early neutralizing activity in LR individuals may be partially supported by IgM. In contrast, CI group displayed IgG exhibited no significant correlation with NAb from day 0 to day 30 or from day 7 to day 30 (Supplementary Fig. 1G). This implies that NAb responses in this group may depend more on antibody quality, such as affinity or affinity changes, rather than solely on IgG quantity. To complement the analysis of humoral responses, we evaluated wild-type SARS-CoV-2–specific T cell responses at day 30, which corresponds to the stage when robust T cell responses are detectable after inactivated vaccination³⁰. Interferon-gamma (IFN-γ), Interleukin-2 (IL-2), and IFN-γ⁺IL-2⁺ T cells were quantified using FluoroSpot assays. Across the three responder groups, CI individuals tended to have higher T cells responses than both LR and RS groups, although these differences did not reach statistical significance (Supplementary Fig. 1I and Supplementary Data 1).

Taken together, we identified three distinct responder phenotypes, each characterized by a unique pattern of NAb response dynamics. Notably, LR individuals exhibited both the slowest onset and the lowest overall NAb titers throughout the observation period, suggesting a subdued or delayed humoral activation.

Dynamic immune signaling and transcription factors activity drive divergent neutralizing antibody responses

Prior to downstream analyses, sequencing depth and read alignment quality were assessed across samples. The relationship between deduplicated reads and detected genes showed a tendency to plateau, indicating adequate sequencing depth (Supplementary Fig. 2A and Supplementary Data 2). We next assessed the potential impact by calculating red blood cell (RBC) enrichment scores using xCell³¹. Overall, the RBC enrichment scores were very low, ranging from 0.00 to 0.011 (median = 0.00), indicating minimal RBC contamination (Supplementary Fig. 2B and Supplementary Data 2). With these quality controls in place, we analyzed PBMC transcriptomes collected at days 0 and 7 to characterize early transcriptional programs underlying phenotypic heterogeneity in vaccine responses. Differential expression between these timepoints revealed distinct response patterns across the responder groups. The LR group showed 10 significantly differentially expressed genes (DEGs), including 4 upregulated and 6 downregulated. The RS group demonstrated a broader expression shift, featuring 72 DEGs, including 22 upregulated and 50 downregulated. In contrast, the CI group only showed 1 downregulated gene (Fig. 2A, left panel; |log₂FC | > 0.5, FDR < 0.05; Supplementary Data 2). To determine whether these divergent responses arose from baseline differences, we compared day 0 gene expression profiles across the groups. Overall, no significant differences were observed. This indicates comparable transcriptional baselines, with a minor exception between LR and CI, where 3 genes (SLFNL1, PTGS2, and NR4A3) were significantly downregulated in LR (Fig. 2A, middle panel; |log₂FC | > 0.5, FDR < 0.05; Supplementary Data 2). Despite overall baseline similarities, no significant DEGs were observed between responder groups at day 7 after adjusting for vaccine product combination (Fig. 2A, right panel).

Fig. 2. Dynamic immune signaling and transcription factors activity drive divergent neutralizing antibody responses.

A DEGs identified between D7 and D0 within groups (left), and between groups at D0 and D7 (middle, right). Red dots indicate significant DEGs (|log₂FC | > 0.5, FDR < 0.05). B Bar plots show the number of immune pathways significantly enriched by GSEA ( | NES | > 1, FDR < 0.05), comparing D7 vs D0 within groups (left) and among groups at D7 (right). C Bubble plots depict representative immune pathways selected based on consistent trends between groupwise ssGSEA scores (median) and GSEA-derived NES across response groups. Dot size reflects |NES | , color shows enrichment direction, “X” marks non-significance. D Boxplots display immune cell proportion changes estimated by CIBERSORT within (left/middle) and across groups (right). Dots represent independent samples, color-coded by group. Flow cytometry analysis of selected immune cell subsets in CI (E) and RS (F) at D0 and D7. Boxplots show paired proportions across all individuals, with lines connecting paired measurements. Representative gating is shown in the figure for reference. Dots represent independent samples. G Heatmap shows correlations between WGCNA modules and NAb titer changes. Red and blue indicate positive and negative correlations, respectively; “X” marks non-significance. H Scatterplots show associations between TF expression and NAb changes (D7–D0); shaded areas are 95% confidence intervals. I Bubble plots show GO enrichment of biological processes for hub genes from key NAb-related modules. Bubble size reflects gene count, color indicates FDR-value. Statistical significance was assessed by two-tailed paired Wilcoxon rank-sum test for (D, left and middle panels), (E), and (F), and by Pearson correlation for (H) (|R | > 0.3, FDR < 0.05). For group comparisons in panel D (right panel), a linear model adjusting for vaccine product combination was applied, followed by pairwise comparisons using emmeans. BH method corrected multiple tests. *FDR < 0.05; **FDR < 0.01; ***FDR < 0.001; ****FDR < 0.0001; ns = not significant. D0, D7 indicate day 0 (baseline), day 7 after the second vaccine dose, respectively. DEGs indicates differentially expressed genes; NAb indicates neutralizing antibody; NES indicates normalized enrichment score; TF indicates transcription factors.

Building on these results, we performed Gene Set Enrichment Analysis (GSEA) to assess whether coordinated immune pathway activity distinguishes the responder phenotypes. The LR group showed no significant enrichment of immune pathways from day 0 to day 7. In contrast, the RS and CI groups displayed distinct immune response profiles, as reflected by the enrichment of 14 and 46 pathways, respectively (Fig. 2B, left panel; |NES | > 1, FDR < 0.05; Supplementary Data 3). Notably, although no significant DEGs were detected at day 7, pathway-level differences remained evident among the responder groups (Fig. 2B, right panel; |NES | > 1, FDR < 0.05; Supplementary Data 3). Pairwise comparisons revealed 2, 24 and 39 significantly enriched immune pathways in the LR vs. RS, LR vs. CI and RS vs. CI comparisons, respectively.

To visualize representative immune pathways, we selected a subset of antibody-related terms that exhibited consistent enrichment trends in both GSEA and single-sample Gene Set Enrichment Analysis (ssGSEA) analyses. Seven such pathways were significantly enriched in at least one intergroup or intragroup comparison (Fig. 2C). We next examined how these pathways changed from day 0 to day 7 within each group. The LR group did not show significant enrichment of immune-related pathways. In contrast, the RS and CI groups exhibited marked activation of pathways such as “activation of innate immune response” and “B cell receptor signaling pathway”. Notably, only the CI group exhibited significant enrichment in “antigen processing and presentation” (Fig. 2C, left panel; |NES | > 1, FDR < 0.05; Supplementary Data 3). These findings were corroborated by ssGSEA, which revealed increasing enrichment scores from day 0 to day 7 in the RS and CI groups (Supplementary Fig. 2C and Supplementary Data 4). We also assessed intergroup differences in the enrichment of these pathways at day 7. Compared to the LR group, both RS and CI exhibited significant enrichment in “B cell receptor signaling pathway”. The CI group also exhibited significantly higher enrichment of “regulation of antigen receptor mediated signaling pathway” relative to both LR and RS. In addition, CI showed significantly higher enrichment of “positive regulation of type 2 immune response” and “T helper 1 type immune response” compared with RS, and significantly higher enrichment of “regulation of B cell receptor signaling pathway” compared with LR (Fig. 2C, right panel; |NES | > 1, FDR < 0.05; Supplementary Data 3). These observed differences in immune pathway activity across groups were further supported by ssGSEA, which revealed higher enrichment scores in RS and CI for the corresponding immune pathways (Supplementary Fig.2D and Supplementary Data 4). Importantly, the CI group exhibited significantly higher enrichment scores for the “positive regulation of type 2 immune response” relative to RS (Supplementary Fig. 2D; FDR < 0.05; Supplementary Data 4). Together, these results reveal distinct immune pathway activation patterns across responder groups, with the LR group showing a marked lack of activation.

We next assessed changes in immune cell populations across responder groups by estimating the proportions of 22 immune cell types per sample using a deconvolution algorithm (Supplementary Data 5). The LR group showed no significant changes in immune cell populations from day 0 to day 7 (Supplementary Fig. 2E, top panel). By comparison, both the RS and CI groups exhibited significant changes in immune cell composition (Fig. 2D, left and middle panel; FDR < 0.05; Supplementary Fig. 2E, middle and bottom panel). Specifically, the RS group exhibited decreased proportions of activated dendritic cells and mast cells, alongside an increase in naive B cells and mast cells resting. Conversely, the CI group showed increased proportions of naive B cells, naive CD4 T cells, macrophages M2, and regulatory T cells (Tregs), as well as reduced frequencies of activated mast cells and follicular helper T cells. We also observed marked variation in immune cell composition across groups at day 7. Notably, the LR group exhibited significantly reduced proportions of activated dendritic cells and regulatory T cells (Tregs) compared to RS (Fig. 2D, right panel; FDR < 0.05; Supplementary Fig. 2F). Furthermore, the CI group exhibited significantly increased proportions of activated dendritic cells and regulatory T cells (Tregs) compared with RS, as well as higher proportions of Tregs relative to LR. To empirically validate the CIBERSORT deconvolution results, flow cytometry was performed on five immune cell subsets in a subset of RS (N = 10) and CI (N = 10) participants (Supplementary Fig. 2G; Supplementary Data 5). The subsets included three showing prominent group-specific differences (naive B cells in RS, naive B cells, naive CD4 T cells, and Tregs in CI) and two without significant differences as controls (CD8 T cells and memory B cells), allowing direct comparison with the CIBERSORT estimates. Flow cytometry results partially validated the CIBERSORT predictions. Specifically, naive B cells and memory B cells in CI showed directionally consistent trends with the deconvolution estimates, although none reached statistical significance (Fig. 2E; Supplementary Fig. 2H). CD8 T cells in RS exhibited similar directional concordance (Fig. 2F; Supplementary Fig. 2H). In summary, responder groups exhibit varied immune cell dynamics, with limited cellular changes in the LR group.

To explore coordinated gene regulatory networks underlying the observed transcriptional and cellular changes, Weighted Gene Co-expression Network Analysis (WGCNA) was applied to day 7 expression profiles. A soft-thresholding power of 12 was selected to approximate scale-free topology (Supplementary Fig. 2I), and 22 co-expression modules were subsequently identified (Fig. 2G and Supplementary Data 6). Module-trait correlation analysis identified 4 NAb-associated modules (ME-Red, ME-Brown, ME-Yellow, ME-Pink) significantly associated with early NAb changes from day 0 to day 7 (Fig. 2H; |R | > 0.3, FDR < 0.05). Among these, ME-Red, ME-Brown, and ME-Yellow were further highlighted as key modules based on significant correlations between module membership and gene significance (Supplementary Fig.2J; |R | > 0.3, P < 0.05). From these key NAb-associated modules, we identified 57 (ME-Red), 388 (ME-Brown) and 683 (ME-Yellow) hub genes (|cor.Weighted | > 0.3, q-weighted <0.001; Supplementary Data 6). We then predicted transcription factors (TFs) targeting hub genes in each module to explore upstream regulatory mechanisms. A total of 33 TFs were associated with ME-Brown, 5 with ME-Yellow and 80 with ME-Red. Six of these TFs were immune-related and showed significant correlations with changes in NAb titers from day 0 to day 7 (Fig. 2H; |R | > 0.3, P < 0.05). Specifically, 5 TFs were positively correlated, including ARNT, ELF2, HIVEP2, SMAD2 and SMAD4. In contrast, KLF15 showed a negative correlation. Subsequently, we examined TF activity to characterize regulatory differences among the groups. The LR group tended to exhibit reduced activity of ELF2 and ARNT, but increased activity of HIVEP2, KLF15, SMAD2, and SMAD4 relative to RS and CI (Supplementary Fig. 2K; FDR > 0.05). Overall, these analyses suggest a trend for a distinctive transcriptional profile in the LR group.

To further explore the biological roles of these key NAb-associated modules, Gene Ontology (GO) enrichment analysis was performed. Our results revealed significant enrichment in post-transcriptional processes, including “RNA splicing” and “mRNA 3’-end processing”. We also observed enrichment in immune activation pathways, such as “immune response-activating signaling pathway” (Fig. 2I; FDR < 0.05; Supplementary Data 6).

Group-specific alternative splicing and polyadenylation shape divergent immune signaling in vaccine-induced neutralizing antibody responses

Considering post-transcriptional enrichment in key NAb-associated modules, we next examined alternative splicing (AS) as a potential mechanism underlying group-specific NAb differences. A total of 160,367 splicing events were identified at day 7 across all 23 chromosomes (Supplementary Fig. 3A), showing similar distributions among the three groups (Supplementary Fig. 3B). Despite this, differential AS analysis uncovered group-specific differences. The LR vs. RS comparison revealed 65 differential alternative splicing events (DASEs) with 30 upregulated and 35 downregulated. LR vs. CI showed 60 DASEs (28 upregulated, 32 downregulated), while RS vs. CI had 14 (7 upregulated, 7 downregulated) (Fig. 3A; |ΔPSI | > 0.1, FDR < 0.05). Furthermore, these three comparisons identified largely non-overlapping DASEs (Fig. 3B). The genes corresponding to these DASEs were enriched in immune-related pathways, notably “antigen processing and presentation via MHC class Ib” (Fig. 3C; FDR < 0.05; Supplementary Data 7). Within the MHC class Ib pathway, HLA-F and HLA-H exhibited significantly higher percent spliced-in (PSI) values in the CI group compared with RS, respectively (Fig. 3D-3E; PSI > 0.1, FDR < 0.05). Although this pathway is distinct from the “regulation of antigen receptor–mediated signaling pathway” enriched in the CI group (Fig. 2C; |NES | > 1, FDR < 0.05), both are functionally linked to antigen presentation–related immune processes. Together, these results suggest that CI individuals exhibit coordinated transcriptional and post-transcriptional features associated with antigen presentation–related processes.

Fig. 3. Group-specific alternative splicing and polyadenylation shape divergent immune signaling in vaccine-induced neutralizing antibody responses.

A DASEs were identified across groups at D7. Red indicates significant events (|ΔPSI | > 0.1, FDR < 0.05). B UpSet plot illustrates the overlap and group-specific DASEs across different comparison groups. C Bar plot shows GO-BP enrichment of genes with DASEs (FDR < 0.05). Density plots show PSI distributions for HLA-F (D: RS vs. CI) and HLA-H (E: RS vs. CI). Dashed lines indicate group means. F DAPAEs were identified across groups at D7. Red indicates significant events (|ΔPDUI | > 0.1, FDR < 0.05). G UpSet plot illustrates the overlap and group-specific DAPAEs across different comparison groups. H Bar plot shows KEGG enrichment of genes with DAPAEs (FDR < 0.05). I Bar plot shows usage patterns of proximal (blue) and distal (red) poly(A) sites across groups. J Density plot shows PDUI distribution of DAPAEs for IKBKE (LR vs. CI). K Density plot shows PDUI distributions for HRAS in RS vs. CI. Dashed lines indicate group means. D7 indicates day 7 after the second vaccine dose; DASEs indicates differential alternative splicing events; PSI indicates percent spliced-in; DAPAEs indicates differential APA events; PDUI indicates percentage of distal polyA site usage index.

To determine whether post-transcriptional regulation also involved alternative polyadenylation (APA), we analyzed APA dynamics across all three groups. A total of 12,072 APA events were detected at day 7 (Supplementary Fig. 3C), showing similar percentage of distal polyA site usage index (PDUI) value distributions across the three groups (Supplementary Fig. 3D). Despite these overall similarities, differential APA analysis revealed pronounced group-specific differences. In pairwise comparisons, the LR vs. RS comparison identified 41 differential APA events (DAPAEs), including 11 upregulated and 30 downregulated events. LR vs. CI showed 49 DAPAEs (34 upregulated, 15 downregulated), whereas RS vs. CI exhibited 32 events (26 upregulated, 6 downregulated) (Fig. 3F; |ΔPDUI | > 0.1, FDR < 0.05; Supplementary Data 8). Consistent with the AS analysis, DAPAEs identified in each comparison showed minimal overlap (Fig. 3G). Genes corresponding to these DAPAEs were significantly enriched in innate immune receptor signaling pathways, with prominent enrichment in “C-type lectin receptor signaling pathway” (Fig. 3H; FDR < 0.05; Supplementary Data 8). Overall, across these differential APA events, RS tended to exhibit longer 3’UTRs compared with LR, whereas CI showed a bias toward 3’UTR shortening relative to both LR and RS (Fig. 3I). Notably, CI individuals exhibited significant 3’UTR shortening in innate immune signaling genes, with IKBKE shortened relative to LR and HRAS relative to RS (Fig. 3J, K; ΔPDUI < −0.1, FDR < 0.05). Collectively, APA patterns showed group-specific differences, with CI exhibiting notable 3’UTR shortening in genes related to innate immune signaling.

B cell receptor features associate with variation in vaccine-induced neutralizing antibody responses

To determine whether differences in transcriptional and post-transcriptional regulation were accompanied by changes in B cell responses, we next examined the B cell receptor (BCR) repertoire across groups. We reconstructed and analyzed immunoglobulin heavy chain (IGH) repertoires from day 7 transcriptome data using TRUST4³². After excluding individuals with fewer than 100 IGH sequences³³, the analysis included 70 participants (LR = 27, RS = 28, CI = 15), yielding 41,746 IGH transcripts and 31,558 clonotypes (Supplementary Data 9). IGH repertoire profiles exhibited both shared and group-specific features across the three groups. Consistently, clonal expansion rates were similar across groups (LR: 17.4%; RS: 15.4%; CI: 15.8%), with singletons (contig count = 1) predominating (Fig. 4A, Supplementary Fig. 4A). Clonal diversity measured by Shannon index was also comparable (LR: 8.15; RS: 8.33; CI: 8.36) (Fig. 4B). Isotype usage showed similar patterns across groups, with IgA1 being the most abundant subclass (~30%) (Supplementary Fig. 4B). This observation is consistent with previous BCR repertoire studies of inactivated SARS-CoV-2 vaccines reporting an IgA1-enriched isotype distribution¹⁶. Likewise, V and J gene usage was similar across groups. For example, IGHJ4 was the most frequently used J gene (~45%) (Supplementary Fig. 4C). The IGHV3, IGHV1, and IGHV4 families together accounted for approximately 85% of V gene usage (Supplementary Fig. 4D). Moreover, frequently used V genes such as IGHV3-23, IGHV1-18, and IGHV4-59 were commonly shared across groups (Supplementary Fig. 4E). Additionally, Complementarity-Determining Region 3 (CDR3) lengths usage patterns were similar across groups, with peak frequencies at 15-17 amino acids (aa) (Supplementary Fig. 4F). Despite these overall similarities, subtle differences in CDR3 motifs of 15-aa sequences emerged among the responder groups (Supplementary Fig. 4G). For example, the consensus CDR3 motifs were “CARDGGGGGYYFDYW” for LR, “CARDGGGGSGYFDYW” for RS and “CARGGGGGGGYFDYW” for CI. Position 4 of CDR3s was occupied by aspartic acid (D) in LR and RS but glycine (G) in CI. At position 10, LR predominantly used tyrosine (Y), while RS and CI favored glycine (G).

Fig. 4. B cell receptor features associate with variation in vaccine-induced neutralizing antibody responses.

A Stacked bar plot shows the distribution of IGH clone sizes per sample at D7 (LR: N = 27; RS: N = 28; CI: N = 15). Clones are categorized as single (contig count = 1), duplicated (contig count = 2), or clonal (contig count ≥ 3), and color-coded accordingly. B Boxplot shows Shannon diversity indices across groups. Each dot represents a biologically independent sample, colored by group. Boxes indicate medians and interquartile ranges; whiskers show the full range. C Venn diagram shows shared and unique clonotypes across groups. D Alluvial plot illustrates shared clonotype proportions across groups. Band width reflects mean clonotype proportion; colors distinguish clonotypes. E Stacked bar plot illustrates the distribution of clone sizes for group-specific clonotypes across groups. IGHs are color-coded as single (contig count = 1), duplicated (contig count = 2), or clonal (contig count ≥ 3). F Line chart depicts the mean frequency of CDR3 amino acid lengths in group-specific clonotypes with ≥ 2 IGHs across groups. Groups are color-coded: LR (red), RS (blue), and CI (green). G Sequence logo highlights motifs in CDR3 aa sequences (length = 20) of group-specific expanded clonotypes (contig count ≥ 2) across groups. Red arrows indicate motif positions showing inter-group divergence. H Bar plot shows constant (C) gene usage in group-specific clonotypes with ≥ 2 IGHs across groups (LR: red, RS: blue, CI: green). I Chord diagram shows the top 10 IGHV-IGHJ pairings in group-specific expanded clonotypes (contig count ≥ 2) across groups. Groups are color-coded (LR: red, RS: blue, CI: green). IGH represents Immunoglobulin Heavy chain; D7 indicates day 7 after the second vaccine dose; CDR3 indicates Complementarity-Determining Region 3.

Despite overall similarities in IGH repertoire features, group-specific patterns were evident. We observed minimal clonotype overlap between groups. Only 7 clonotypes were shared across all three groups, while the majority were unique to each group (LR: 12,159; RS: 12,645; CI: 6,618) (Fig. 4C). Expansion patterns among shared clonotypes differed notably. LR predominantly expanded clonotype41149 (IGHV3-48-IGHJ4), RS favored clonotype51960 (IGHV3-74-IGHJ4) and CI was dominated by clonotype58257 (IGHV3-23-IGHJ4) (Fig. 4D). Additionally, most group-specific clonotypes were singletons (contig count = 1), reflecting the overall repertoire pattern (Fig. 4E). However, within expanded clonotypes (contig count ≥ 2), distinct patterns emerged in CDR3 structure and IGHV-IGHJ pairing across groups. LR showed enrichment for CDR3 lengths of 15-17 aa, RS for 17-18 aa and CI for 14, 16 aa (Fig. 4F). Focusing on CDR3s of length 20 aa, each group displayed distinct sequence motifs. LR exhibited “CARDGGGSGSGSYYYYFDVW”, RS showed “CARDPGGYGSGSGYYYFDYW” and CI presented “CARDGGYGSGGYYYYGFDVW” (Fig. 4G). Notably, position 16 was dominated by tyrosine (Y) in LR and RS but glycine (G) in CI. Position 19 was predominantly valine (V) in LR and CI, while RS favored tyrosine (Y). While CDR3 structures varied across groups, isotype usage remained consistent with IgA1 predominating (Fig. 4H). In contrast, IGHV-IGHJ pairing showed group-specific preferences among the top 10 combinations. LR was enriched for IGHV1-69-IGHJ2, RS for IGHV3-53-IGHJ4, and CI for IGHV4-59-IGHJ4 (Fig. 4I). We next examined whether these clonotypes were already present prior to vaccination. IGH repertoires at day 0 were reconstructed using the same criteria as day 7, covering 68 participants (LR = 27, RS = 27, CI = 14) and yielding 37,360 transcripts and 27,106 clonotypes (Supplementary Data 9). At day 0, clonotype size distributions across samples were dominated by singleton clonotypes, indicating the absence of widespread clonal expansion before vaccination (Supplementary Fig. 4H). Notably, fewer than 3% of the group-specific expanded clonotypes identified at day 7 were detectable at day 0 (LR: 53 of 2138; RS: 24 of 2083; CI: 17 of 878) (Supplementary Fig.4I). Among these detectable clonotypes at day 0, most were present as singletons, with only a small fraction exhibiting clonal expansion (contig count ≥ 2) (Supplementary Fig. 4J). Consistently, the top 10 IGHV–IGHJ pairings at day 7 showed minimal overlap with the most frequent IGHV–IGHJ pairings among expanded clonotypes at day 0 (Supplementary Fig. 4K). Specifically, only 2, 1, and 2 pairings were shared between day 7 and day 0 in LR, RS, and CI, respectively. Together, these results indicate that the group-specific IGH expansions observed at day 7 are largely absent at day 0 and predominantly emerge following vaccination.

Machine learning model identifies low responders from baseline gene expression

Based on the distinct transcriptomic profile of the LR group, we aimed to predict LR individuals to guide personalized vaccination strategies. We focused on distinguishing LR from CI individuals due to the latter exhibiting the highest NAb responses. Given the limited baseline differences between LR and CI, genes with |log₂FC | > 0.5 at day 0 were chosen as the candidate feature set for predictive modeling. We applied two feature selection methods, including LASSO regression and random forest. LASSO yielded 24 candidate genes (Fig. 5A, B, Supplementary Fig. 5A). Random forest, in contrast, prioritized 70 genes (Fig. 5C, Supplementary Fig. 5B).

Fig. 5. Machine learning model identifies low responders from baseline gene expression.

A LASSO coefficient paths of candidate genes are shown as a function of log(lambda) during 10-fold cross-validation. Twenty-four genes had non-zero coefficients at the minimum lambda, representing the selected optimal biomarkers distinguishing LR and CI groups. B Binomial deviance is plotted as a function of log(lambda) during 10-fold cross-validation. C Random Forest feature selection identified 70 important genes contributing to group classification. D Venn diagram shows 8 overlapping genes identified by both LASSO and Random Forest. E ROC curve illustrates the classification performance of the candidate genes in distinguishing LR from CI. F Lollipop plot displays the relative contribution of the overlapping genes to XGBoost model predictions, with importance represented by a gradient color scale.

To identify robust predictors, we examined the overlap between the two methods. Notably, both methods identified 8 overlapping genes (Fig. 5D), including RHOB, CSNK1E, ENDOG, MIR4295, MIR7113, NR4A3, TAS2R7, SLFNL1. These genes collectively highlight the role of baseline immune features in shaping vaccine responsiveness and underscore their potential as predictive biomarkers. When evaluated individually, these genes achieved classification accuracies of 65.9%, 77.3%, 59.1%, 81.8%, 65.9%, 46.2%, 76.5%, and 79.5%, respectively (Supplementary Fig. 5C–J). We then used the baseline expression levels of these 8 genes to build a final classifier, which achieved an improved accuracy of 92.4% (AUC = 0.924) in distinguishing LR individuals (Fig. 5E). Among these, only 5 genes (CSNK1E, RHOB, ENDOG, SLFNL1, TAS2R7) showed non-zero importance in the model. Notably, CSNK1E and RHOB together contributed approximately 85% of the overall predictive power. In addition, the remaining 15% was contributed by the other genes (Fig. 5F). CSNK1E encodes a kinase that regulates WNT/β-catenin signaling³⁴ and p53 signaling³⁵. RHOB modulates key macrophage effector functions, including phagocytosis³⁶, inflammatory cytokine and nitric oxide production³⁷, and adhesion and migration³⁸, highlighting its central role in innate inflammatory responses. Together, these findings suggest that even subtle baseline gene expression differences can accurately predict LR individuals.

Discussion

In this study, we identified three distinct phenotypes of NAb responses following inactivated SARS-CoV-2 vaccination: LR, RS, and CI. This classification highlights the heterogeneity of NAb kinetics across individuals, offering a more refined understanding of NAb variability post-vaccination. Through multi-layered transcriptomic analysis, we uncovered immune signatures and regulatory features that differentiate the three response groups. In addition, we developed a predictive model capable of identifying LR individuals prior to vaccination. These findings provide new insights into the immunological mechanisms underlying vaccine-induced NAb responses and establish a framework for optimizing personalized vaccination strategies.

Previous studies have typically categorized antibody responses into binary groups such as high or low responders^{9,15,16,22–25}, which may not fully capture the complexity of antibody response kinetics. Indeed, the temporal dynamics of antibody responses have been increasingly recognized as important for understanding vaccine efficacy and long-term protection^39,40. By measuring antibody levels at day 0, 7, and 30, we identified three NAb response phenotypes, including LR, RS, and CI. While LR and CI resemble previously described low and high responders, similar longitudinal response patterns have been reported^16,17. Our analysis further revealed an RS group characterized by an early NAb peak followed by a plateau, highlighting previously unrecognized heterogeneity among high responders. These findings underscore the importance of longitudinal sampling to capture the full dynamics of immune responses, extending beyond the binary responder classification.

Despite response groups being defined independently of vaccine product, the distribution of vaccine product combination significantly differed across groups. Although BBIBP-CorV and CoronaVac are both inactivated vaccines based on wild-type SARS-CoV-2, prior studies have reported differences in the magnitude of NAb responses between them^29,41,42, which partially contribute to variability in NAb responses. Importantly, significant differences in NAb dynamics across groups persisted throughout all measured intervals after adjustment for vaccine product combination, indicating stable differences in the kinetics of NAb responses. Taken together, this phenotypic stratification underscores the clinical relevance of NAb response differences. Although our study did not directly measure clinical protection, prior evidence indicates that reduced NAb levels are associated with lower protection against SARS-CoV-2 infection after vaccination^40,43. For example, in a phase 3 trial of CoronaVac, NAb titers measured 28 days after the second dose were inversely associated with the risk of symptomatic SARS-CoV-2 infection⁴³. In this context, LR individuals with delayed and weaker NAb responses may have reduced immune protection, suggesting potential benefit from modified vaccine schedules or enhanced formulations. Notably, additional vaccine doses have been shown to substantially boost NAb levels in low responders across multiple platforms, such as inactivated vaccines¹⁸ and mRNA vaccines⁴⁴. These findings support the utility of targeted booster strategies in improving immune protection in this group. On the other hand, RS individuals who show a rapid but transient antibody response might require booster doses to sustain long-term immunity. In contrast, CI individuals demonstrate a robust and sustained antibody response and are likely to achieve durable protection under standard vaccination schedules. These findings underscore the importance of dynamic NAb profiling to tailor vaccination strategies, ensuring that each individual receives the most effective regimen based on their immune response profile.

To investigate the immunological mechanisms underlying divergent NAb response phenotypes, we performed multi-dimensional transcriptomic analysis across the three groups. LR individuals exhibited the weakest immune responses by day 7, with minimal transcriptional activation and no significant immune cell shifts, despite comparable baseline profiles. One possible explanation for this attenuated response is an altered regulatory landscape in LR individuals. Compared to RS and CI, LR tended to exhibit reduced activity of TFs previously implicated in early immune activation (e.g., ARNT, ELF2) and relatively higher activity of TFs associated with immunoregulatory and immunosuppressive signaling (e.g., HIVEP2, SMAD2, SMAD4, KLF15)^45–53. These coordinated trends may limit effective early B cell priming and thereby constrain downstream transcriptional and cellular responses following vaccination. In contrast, RS individuals displayed moderate early immune activation, as evidenced by partial enrichment of B cell-related and innate immune-related pathways from day 0 to day 7. These transcriptional signatures were reflected in immune cell dynamics, with RS individuals showing inferred increases in naive B cells and decreases dendritic cells. The observed increase in naive B cells is notable, as higher baseline levels of this population have been associated with stronger antibody responses following vaccination⁵⁴. Dendritic cells, as key antigen-presenting cells, contribute to the regulation of adaptive immunity by shaping helper T cell activation and promoting B cell responses⁵⁵. This decrease in inferred dendritic cell proportions may reflect their migration from peripheral blood to secondary lymphoid tissues or their activation status during early immune responses. Together with the increase in naive B cells, these changes are likely involved in rapid B cell–mediated responses that could facilitate the initiation of antibody production. Similarly, CI individuals demonstrated an early divergence from both LR and RS groups by day 7, despite minimal changes in gene expression. This transcriptional divergence was associated with enriched activity in antigen processing, type 1 helper T cell (Th1), and type 2 helper T cell (Th2) pathways. SsGSEA further revealed significantly higher enrichment of the “positive regulation of type 2 immune response” in CI, consistent with the Th2-biased transcriptional signature observed in GSEA. Th2-biased responses have previously been linked to robust humoral immunity following SARS‑CoV‑2 vaccination⁵⁶, suggesting that early Th2 activation may be favorable for higher antibody production in this group. These transcriptional signatures were reflected in immune cell dynamics. Specifically, CI individuals showed increased naive B cells, partially validated by flow cytometry, along with inferred increases in naive CD4 T cells, M2 macrophages, and regulatory T cells. In line with the Th1-associated transcriptional signatures identified by GSEA at day 7, CI individuals also tended to exhibit higher IFN-γ⁺, IL-2⁺, and dual IFN-γ⁺IL-2⁺ T cell responses at day 30. Such vaccine-induced IFN-γ⁺ and IL-2⁺ T cell responses have been shown to correlate with stronger NAb responses⁵⁷. Overall, coordinated early immune activation is likely associated with the stronger NAb responses observed in CI individuals.

Building on the immune signatures identified earlier, we next explored post-transcriptional regulation as a key mechanism driving divergent NAb responses. Our analysis revealed that AS in HLA class I genes may contribute to the heterogeneity of NAb responses. These molecules are expressed on almost all nucleated cells and mediate immune recognition by presenting endogenous peptides to CD8⁺ T cells and engaging NK cell receptors⁵⁸. Among them, HLA-F functions primarily as an immunomodulatory molecule through interactions with activating and inhibitory NK cell receptors⁵⁹, whereas the HLA‑H signal peptide can mobilize HLA‑E to the cell surface on various immune cells⁶⁰. AS can modulate HLA expression by altering mRNA stability, translational efficiency, and the production of distinct protein isoforms⁶¹. We observed significant AS events in HLA-F and HLA-H between RS and CI, which may influence the regulation of these genes. The enrichment of the “regulation of antigen receptor mediated signaling pathway” in CI but not in RS is consistent with potential functional consequences of these post-transcriptional differences. Together, these findings suggest that post-transcriptional regulation of HLA-F and HLA-H may contribute to the observed heterogeneity in NAb responses.

In addition to AS, we observed a general tendency toward 3’UTR shortening in CI individuals compared with both LR and RS groups. In the context of immune activation, shortened 3’UTRs have been shown to be depleted of immune-induced miRNA target sites⁶². In addition, APA-mediated 3’UTR shortening can enhance transcript stability and protein synthesis⁶³. These findings suggest that 3’UTR shortening may enable immune genes to evade post-transcriptional repression and sustain higher expression levels. Within this framework, 3’UTR shortening was especially prominent in immune-relevant genes, with IKBKE shortened in CI compared with LR and HRAS shortened in CI compared with RS. IKBKE encodes a non-canonical IκB kinase that promotes type I interferon production and NF-κB activation downstream of pattern-recognition receptors⁶⁴. HRAS, a Ras GTPase, activates the MAPK cascade downstream of antigen receptor signaling to regulate immune cell activation and proliferation⁶⁵. In this context, preferential 3’UTR shortening of IKBKE and HRAS in CI individuals may contribute to enhanced immune activation by potentially favoring higher or more sustained expression of these signaling molecules.

We further examined the contribution of distinct BCR repertoire features to the heterogeneity in NAb responses across groups. CI individuals preferentially utilized clonotypes such as IGHV3-23-IGHJ4, which has been commonly reported in vaccine recipients⁶⁶ and linked to monoclonal expansions targeting the spike protein⁶⁷. Additionally, CI individuals frequently used IGHV4-59-IGHJ4, corresponding to the public antibody M15, a clonotype observed after both SARS-CoV-2 infection and vaccination⁶⁸. The preferential use of these clonotypes may contribute to the robust and sustained NAb response observed in CI. In contrast, the RS group showed enrichment for IGHV3-74-IGHJ4 and IGHV3-53-IGHJ4, which have been implicated in B cell responses to SARS-CoV-2 antigens⁶⁹ and RBD-specific neutralizing antibodies⁷⁰, respectively. These clonotypes may facilitate an early antibody response, although they appear to be less associated with sustained NAb responses compared with CI individuals. Interestingly, LR individuals predominantly utilized IGHV1-69-IGHJ2, a clonotype linked to SARS-CoV-2 reactivity⁷¹, despite exhibiting weaker NAb production in this group. Notably, IGHV3-30-IGHJ4, a public clonotype reported across multiple individuals following SARS-CoV-2 exposure⁷², was preferentially utilized by all three groups. Although shared across groups, differences in clonal expansion may contribute to variability in NAb quality and magnitude.

To identify individuals likely to exhibit low NAb responsiveness, we developed predictive models based on pre-vaccination gene expression profiles. These models successfully identified LR individuals, demonstrating high accuracy in predicting suboptimal vaccine responses. The ability to predict LR individuals prior to vaccination provides an opportunity for tailored vaccination strategies, such as adjusted dosing schedules or enhanced formulations. By incorporating baseline immune features, this model offers a promising approach to optimize vaccine efficacy in susceptible individuals, potentially enabling more personalized and effective vaccination strategies.

In summary, this study provides valuable insights into the heterogeneity of NAb responses following vaccination, identifying distinct immune phenotypes and key molecular features that drive divergent immune outcomes. Our findings emphasize the complexity of vaccine-induced immune responses. They move beyond traditional binary classifications and offer a more refined understanding of individual variability. Through multi-layered transcriptomic analysis, we have uncovered novel molecular signatures that could inform the development of personalized vaccination strategies. However, several limitations should be considered. The small sample size may limit the generalizability of these findings. Additionally, our analysis focused on short-term immune responses, and the long-term durability of NAb responses remains to be fully assessed. Moreover, PBMC isolated by Ficoll density gradient centrifugation may result in residual RBC carryover. Consistent with this, xCell analysis indicated uniformly low RBC-associated transcriptional signals across samples. Finally, the absence of single-cell transcriptomic data limits our ability to resolve immune cell subsets and their functional states, which could partially contribute to heterogeneity in neutralizing antibody responses. Future studies should expand cohort sizes and include more diverse populations, extend follow-up beyond 30 days to assess the durability of immune responses, and apply single-cell immune profiling to better characterize immune cell heterogeneity.

Methods

Human participants

This study was conducted at Yunnan University Affiliated Hospital between June and August 2021, a period during which SARS-CoV-2 community transmission in China was effectively suppressed under strict public health policies^73,74. A total of 73 participants aged 18–50 years were enrolled. Participants were eligible if they were aged 18–65 years, had no history of SARS-CoV-2 vaccination, had no diagnosed autoimmune diseases or ongoing immunosuppressive therapy, and had available blood samples collected at day 0 (prior to the first dose), day 7 (7 days after the second dose), and day 30 (30 days after the second dose). To ensure that participants were SARS-CoV-2 naive, multiple screening measures were applied: participants underwent repeated community-based SARS-CoV-2 RT-PCR testing to rule out historical infections⁷⁵, reported no history of COVID-19–related symptoms or known exposure, and tested negative for SARS-CoV-2 immediately prior to enrollment, consistent with previous vaccine trials⁷⁶. Baseline serological measurements showed uniformly low levels of anti-RBD IgG and IgM. Together, these findings demonstrate that participants had no evidence of recent or prior SARS-CoV-2 infection. All participants received a two-dose regimen of inactivated SARS-CoV-2 vaccine (BBIBP-CorV, CoronaVac, or mixed doses). Mixed doses refer to one dose of BBIBP-CorV and one dose of CoronaVac, in any order. BBIBP-CorV (HB02 strain, wild-type SARS-CoV-2) and CoronaVac (CN02 strain, wild-type SARS-CoV-2) are β-propiolactone-inactivated whole-virion vaccines containing aluminum hydroxide as an adjuvant^77,78. Peripheral blood samples were collected at day 0, day 7, and day 30. This study was approved by the Human Research Ethics Committee of Yunnan University (approval number: CHSRE2021020). Written informed consent was obtained from all participants.

Serum and PBMCs isolation

Blood samples were centrifuged at 2000 g for 10 min to obtain serum, which was aliquoted and stored at −80 °C until neutralizing antibody detection. PBMCs were isolated using Ficoll density gradient centrifugation, as described in our previous work¹⁸. Briefly, peripheral blood samples were first mixed 1:1 with RPMI-1640 medium (VIVACELL, China). The mixture was then gently layered onto Ficoll-Paque (STEMCELL, Canada) and centrifuged at 1455 × g for 30 min at room temperature without brake. Next, the PBMC-containing buffy coat was collected, washed with RPMI-1640, and centrifuged at 524 × g for 10 min. The resulting cell pellet was resuspended in cold cryopreservation medium consisting of 90% fetal bovine serum (VIVACELL, China) and 10% Dimethyl sulfoxide (DMSO) (Solarbio, China). Afterwards, the suspension was aliquoted into cryovials and frozen at –80 °C overnight using a Mr. Frosty freezing container (Thermo Scientific). Finally, the samples were transferred to liquid nitrogen for long-term storage until RNA extraction.

Detection of wild-type SARS-CoV-2 RBD-specific neutralizing antibodies

Surrogate NAb responses against SARS-CoV-2 were measured using a commercially available magnetic particle chemiluminescence immunoassay (MCLIA) with the SARS-CoV-2 Neutralizing Antibody Detection Kit (Bioscience Co., Chongqing, China), following the manufacturer’s instructions, as described previously¹⁸. This assay specifically detects NAbs targeting the receptor-binding domain (RBD) of the spike protein from wild-type SARS-CoV-2 strains included in BBIBP-CorV and CoronaVac. It has been shown to correlate strongly with live-virus neutralization¹⁸, supporting its reliability. In the assay, 300 µL of serum was incubated with RBD-coated magnetic beads and ACE2 protein labeled with alkaline phosphatase (ALP). NAbs in the serum compete with labeled ACE2 for binding to RBD. After washing away unbound components, a chemiluminescent substrate was added, and emitted relative luminescence units (RLU) were measured using an automated chemiluminescence analyze. NAb levels were expressed as the chemiluminescence signal relative to the cutoff (S/CO). The cutoff value was defined by the receiver operating characteristic curves. All samples were assayed in duplicate with the kit’s internal quality controls to ensure reliable results.

Detection of IgG and IgM antibodies specific to the wild-type SARS-CoV-2 RBD

IgG and IgM antibodies against wild-type SARS-CoV-2 were detected in serum samples using commercially available MCLIA kits specifically designed for IgG or IgM detection, respectively (Bioscience Co., Chongqing, China), according to the manufacturer’s instructions, as described previously¹⁸. This assay specifically detects IgG or IgM targeting the RBD of the spike protein from wild-type SARS-CoV-2 strains. In this indirect assay, antibodies in the serum bind to recombinant RBD antigens coated on magnetic beads. An ALP-labeled secondary antibody binds to the captured antibodies, forming a complex. After removing unbound components by washing, a chemiluminescent substrate was added, and the emitted RLU was measured using an automated analyzer. Antibody levels were expressed as the chemiluminescence signal relative to the cutoff (S/CO). The cutoff was defined using receiver operating characteristic curves. All samples were assayed in duplicate with the kit’s internal quality controls to ensure reliable results.

Detection of wild-type SARS-CoV-2-specific T cell responses

SARS-CoV-2 specific IFN-γ⁺, IL-2⁺, IFN-γ⁺IL-2⁺ T cell responses were evaluated using the FluoroSpot Path: SARS‑CoV‑2 (S1scan+SNMO) Human IFN‑γ/IL‑2 (Mabtech, Sweden) following the manufacturer’s instructions and as described in our previous work¹⁸. This assay detects T cell responses targeting multiple proteins from wild-type SARS-CoV-2, including spike (S), nucleocapsid (N), membrane (M), and open reading frame (ORF)-derived epitopes. Briefly, cryopreserved PBMCs were thawed, rested, and seeded at 2.5 × 10⁵ cells per well. Cells were stimulated with SARS-CoV-2 peptide pools (S1 scanning pool: 166 overlapping 15-mer peptides covering the S1 domain of spike; SNMO defined pool: 47 peptides from S, N, M, ORF-3a, and ORF-7a) in the presence of co-stimulatory anti-CD28 mAb and incubated at 37 °C in a humidified incubator with 5% CO₂ for 32 h. Unstimulated wells containing equivalent concentrations of DMSO were included as negative controls, and wells stimulated with anti-CD3 mAb served as positive controls. Spot-forming cells were detected using the Mabtech IRIS reader. The results were expressed as the number of antigen-specific cells per 10⁶ PBMCs after subtraction of background counts from negative control wells.

Flow Cytometry of PBMCs

Frozen PBMCs were thawed in a 37 °C water bath with gentle agitation. Thawed cells were immediately diluted in 5 mL of pre-warmed RPMI 1640 medium supplemented with 10% fetal bovine serum (FBS) and centrifuged at 500 × g for 5 minutes. After the supernatant was carefully aspirated, cells were resuspended in 1 mL of culture medium for cell counting. The cell suspension was then rested at 37 °C in a humidified incubator with 5% CO₂ for 1 h before staining. Following incubation, cells were washed with phosphate-buffered saline (PBS) and resuspended in 100 μL of PBS. Dead cells were excluded using Zombie UV viability dye (BioLegend) by incubating samples at room temperature in the dark for 10 min. For surface staining, cells were incubated with a panel of fluorochrome-conjugated antibodies (listed in Supplementary Data 5) at room temperature in the dark for 30 min. Stained cells were washed with PBS and resuspended in 200 μL of PBS for flow cytometric acquisition. Data were acquired on a NovoCyte Opteon spectral flow cytometer (Agilent) and analyzed using FlowJo software (Tree Star) with manual gating strategies.

Unsupervised analysis of neutralizing antibody dynamics

Two unsupervised approaches were applied to the pairwise changes in neutralizing antibody titers (day 7- day 0, day 30 - day 0, and day 30 - day 7). NAb response trajectories were projected into a two-dimensional space using the R package Rtsne (https://github.com/jkrijthe/Rtsne) with parameters perplexity = 10 and max_iter = 1000, and visualized using ggplot2⁷⁹. Samples were clustered based on the Euclidean distances of titer changes using complete linkage to maximize inter-cluster dissimilarity. The optimal number of clusters (K = 3) was determined by gap statistic analysis with the R package cluster⁸⁰. The resulting dendrogram was segmented into three distinct clusters using the R package factoextra (https://cran.r-project.org/web/packages/factoextra/index.html).

RNA-seq library preparation and sequencing

Total RNA was extracted from PBMCs using TRIzol reagent (Invitrogen) following the manufacturer’s instructions. mRNA was subsequently fragmented into small pieces and used for library preparation with the SMARTer Stranded Total RNA-Seq Kit v3 - Pico Input Mammalian (Takara, Japan), according to the manufacturer’s protocol. The libraries were sequenced on the NovaSeq 6000 platform (150 bp paired-end reads).

Quantification of expressed genes, identification and quantification of alternative splicing events and alternative polyadenylation events

Adapter and quality trimming were performed using Cutadapt⁸¹, followed by quality control and read filtering using Fastp⁸². Trimmed reads were aligned to the human GRCh38 reference genome using STAR⁸³. PCR duplicates were removed using UMI-tools based on unique molecular identifiers⁸⁴. Unique mapping reads were extracted by Bamtools⁸⁵. Gene-level expression counts were generated for each sample using the featureCounts function from the Subread package⁸⁶. AS analysis was performed as follows: intron junctions were extracted from BAM files using RegTools⁸⁷, and AS events were identified and quantified using LeafCutter⁸⁸. For each intron excision (splice junction) within a cluster, the PSI value represents the proportion of reads supporting that junction relative to the total reads in the cluster. A PSI of 1 indicates that the junction is predominantly used, whereas a PSI close to 0 indicates rare usage of that junction. For APA analysis, BAM files were converted to bedGraph format using BEDTools⁸⁹. APA events were identified and distal polyadenylation site usage quantified using DaPars2⁹⁰. The percentage of PDUI was defined as the proportion of transcripts using the distal polyA site: PDUI = 1 indicates exclusive usage of the distal site, and PDUI = 0 indicates exclusive usage of the proximal site.

Differential expression, AS events and APA events analysis

Differential gene expression analysis was performed using DESeq2⁹¹. Day 0 group comparisons used the standard DESeq2 design; day 7 group comparisons included vaccine product combination as a covariate; and day 0 versus day 7 changes were tested using a paired design with subject ID as a blocking factor. DEGs were defined as those with a false discovery rate (FDR) < 0.05 and |log₂FC | > 0.5. Differential splicing analysis was performed using the LeafCutter⁸⁸ leafcutter_ds.R script, with vaccine product combination included as a covariate to control for potential confounding by vaccine product. DASEs were defined as those with FDR < 0.05 and |ΔPSI | > 0.1. Differential APA analysis was performed using a quasibinomial generalized linear model that included vaccine product combination as a covariate. Pairwise contrasts between groups were obtained from the fitted model using R packages emmeans (https://rvlenth.github.io/emmeans/). DAPAEs were defined as those with FDR < 0.05 and |ΔPDUI | > 0.1. The FDR was calculated by adjusting P values for multiple testing using the Benjamini-Hochberg (BH) procedure.

Assessment of RBC contamination

To evaluate potential residual RBC contamination in PBMC samples, erythroid cell enrichment scores were calculated using xCell³¹. Scores were computed for each sample and visualized for all three responder groups.

Functional enrichment analyses

GO or KEGG pathway enrichment analysis was conducted using the R package clusterProfiler⁹² to explore the biological functions of candidate genes. GO terms or KEGG pathways with an FDR < 0.05 were considered significantly enriched. GSEA were obtained from the GO Biological Process category in the Molecular Signatures Database (MSigDB) (http://gsea-msigdb.org). GSEA was performed using the R package fgsea⁹³ with log₂FC values for all detected genes. To identify immune-related gene sets, we filtered terms containing keywords such as “_IMMUNE”, “_T_CELL”, “_B_CELL”, “_ANTIGENIC”, “_ANTIGEN”, “_MHC”, “_DENDRITIC_CELL”. GO terms or pathways with an FDR < 0.05 and an absolute normalized enrichment score (NES) > 1 were considered significant. P-values were adjusted using the BH procedure. All enrichment results were visualized using the R package ggplot2⁷⁹.

Gene set variation analysis (GSVA)

To quantify pathway activity at the single-sample level, enrichment scores were computed using the ssGSEA algorithm from the R package GSVA⁹⁴, based on gene expression profiles and gene sets from the GO Biological Process ontology. TF activity was inferred based on the expression patterns of their target genes. The ssGSEA algorithm was used to estimate TF activity scores. For each regulatory module, overall activity was defined as the mean of individual TF activities. For paired day 0 versus day 7 comparisons, statistical significance was assessed using a two-tailed paired Wilcoxon signed-rank test. For between-group comparisons at day 7, a linear model including vaccine product combination as a covariate was fitted, and pairwise contrasts between groups were obtained using the R package emmeans, with multiple testing correction via the BH method. An FDR < 0.05 was considered significant. The results were visualized using the R package ggplot2⁷⁹.

Immune cell fraction deconvolution

Using LM22 as a reference signature, the relative proportions of 22 immune cell types were estimated for each sample based on the normalized gene expression matrix, using the R package CIBERSORT⁹⁵. The estimated proportions for each sample summed to 1. For paired day 0 and day 7 comparisons, statistical significance was assessed using a two-tailed paired Wilcoxon signed-rank test. For between-group comparisons at day 7, cell-type proportions were analyzed using a beta regression model from R package betareg⁹⁶ that included vaccine product combination as a covariate, followed by pairwise contrasts with the R package emmeans. For all analyses, multiple testing correction was performed using the BH method, and an FDR < 0.05 was considered significant. Results were visualized as bar plots using the R package ggplot2⁷⁹.

Weighted gene co-expression network analysis

The gene expression matrix was filtered to retain genes with variance > 0. A weighted gene co-expression network was constructed using the package WGCNA⁹⁷. The optimal soft-thresholding power was determined using the pickSoftThreshold function to approximate scale-free topology. A soft-thresholding power of 12 was selected when the scale-free topology fit index exceeded 0.85. The resulting adjacency matrix was transformed into a topological overlap matrix (TOM), and genes were hierarchically clustered based on TOM-based dissimilarity. Modules were identified using a dynamic tree cut algorithm with a minimum module size of 50 genes and a module merging threshold of 0.25.

Key module selection related to change in neutralization antibody

To identify modules associated with changes in neutralizing antibody levels, Pearson correlation analysis was performed between module eigengenes and neutralizing antibody changes (day 7–day 0, day 30–day 0, day 30–day 7). Modules with |Pearson’s r | > 0.3 and FDR < 0.05 were considered significantly correlated. Within each key module, scatter plots were generated to illustrate the linear relationship between gene significance (GS, correlation between individual gene expression and phenotype) and module membership (MM, correlation between gene expression and module eigengene). Modules were further filtered by requiring |Pearson’s r | > 0.3 and P < 0.05 for GS-MM correlation. Hub genes were identified within key modules using the networkScreening function, based on GS and MM. Genes with |cor.Weighted | > 0.3 and q-weighted < 0.001 were defined as hub genes. FDR was calculated using the g using the BH method.

Identifying core TFs of hub genes

To identify core TFs regulating hub genes, cis-regulatory motif enrichment analysis was performed using RcisTarget⁹⁸. The analysis used the motif ranking database hg38__refseq-r80__10kb_up_and_down_tss.mc9nr downloaded from the cisTarget resources website (https://resources.aertslab.org/cistarget/). TFs with an absolute NES > 3.0 were considered significantly enriched. To further identify genes co-expressed with these enriched TFs, gene regulatory networks were inferred using the GENIE3⁹⁹ with a weight threshold > 0. TF–target gene pairs that appeared in both the RcisTarget and GENIE3 results were retained. TFs present in this intersection were defined as core regulators of the hub genes.

BCR repertoire reconstruction

Bulk RNA‑seq of PBMCs captures a subset of V(D)J-containing transcripts derived from these lymphocytes¹⁰⁰. These transcripts can be leveraged to reconstruct BCR repertoires using computational tools such as TRUST4³². Specifically, raw sequencing files were processed using TRUST4 to extract BCR reads, assemble full-length IGH sequences, and annotate V, D, J, and constant (C) gene usage as well as CDR3 sequences using default parameters³². For downstream analysis, IGH sequences were selected based on the following criteria: (1) the chain must be a productive IGH; and (2) only samples with more than 100 productive IGH sequences were retained³³.

Clonotype definition and clone diversity

Clonotypes were defined for IGH based on identical V and J germline gene assignments and ≥85% amino acid sequence identity in the CDR3 region using Change-O¹⁰¹. The abundance of each clonotype was estimated based on the number of supporting IGH contigs, reflecting the relative transcriptional abundance of each clonotype. Additionally, the contribution of each clonotype was defined as the proportion of total IGH contigs in the sample represented by that clonotype. Repertoire diversity was quantified using the Shannon index, which was calculated based on the observed clonotype distribution using the following function:

$$H=-\sum _{i=1}^{N}p\left(i\right)\log 2p\left(i\right)$$

where N was the total number of unique CDR3, and p(i) was the frequency of a single CDR3.

Multiple machine learning algorithms and candidate hub genes screening

To identify key biomarkers associated with LR individuals, we observed limited significant DEGs were detected between LR and CI groups at day 0. Therefore, genes with |log₂FC | > 0.5 at day 0 were selected as candidate features for predictive modeling. Two machine learning algorithms, Random Forest (RF) and LASSO regression, were applied to these DEGs. Machine learning analyses were conducted using the R packages rfPermute (https://cran.r-project.org/web/packages/rfPermute) and glmnet¹⁰². Genes identified by both RF and LASSO were considered robust candidates for predicting low responders.

Prediction model construction and receiver operating characteristics (ROC) curve

Expression profiles of candidate genes were used to develop a predictive model for identifying LR individuals. The model was constructed using the R package caret framework (https://github.com/topepo/caret/) with the xgbTree method, which implements the XGBoost algorithm. Samples were randomly divided into a training set (n = 26) and a testing set (n = 18) at a 6:4 ratio. The model was trained using a small predefined hyperparameter grid (varying the learning rate) with 5-fold cross-validation, and class imbalance was addressed by up-sampling. Model performance was evaluated on the independent test set by plotting ROC curves and calculating the area under the curve (AUC) using the R package pROC¹⁰³.

Statistics

Statistical analyses were performed in R and are specified in each corresponding figure caption. Tests for statistically significant differences in continuous variables among groups at day 0 were performed using the Kruskal–Wallis test. For paired samples within the same individuals (pre- vs post-vaccination), the paired Wilcoxon signed-rank test was applied. Between-group comparisons at day 7 were assessed using linear models adjusted for vaccine product combination, with Type III ANOVA used to evaluate overall group effects. Pairwise contrasts between groups were obtained using the R package emmeans. Differences in categorical variables were evaluated using Fisher’s exact test. P values were adjusted for multiple testing using the BH correction. Pairwise correlations were assessed using Pearson correlation.

Author contributions

R.C., X.P. and Z.S.Z. conceived and designed the study. M.Y. performed the preparation of RNA-seq libraries. Q.Q.W., H.J.H. and L.Y.Q. analyzed the RNA-seq data and generated the figures. X.P.M. performed Flow cytometry. B.H.C., X.L. and L.X. gave recommendations for analysis. Q.Q.W. wrote the manuscript. R.C., Y.L.C., C.M.L. and Z.S.Z. reviewed and edited the manuscript.

Data availability

The raw sequence data reported in this paper have been deposited in the Genome Sequence Archive¹⁰⁴ in National Genomics Data Center¹⁰⁵, China National Center for Bioinformation/Beijing Institute of Genomics, Chinese Academy of Sciences (GSA-Human: HRA012153) that are publicly accessible at https://ngdc.cncb.ac.cn/gsa-human.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41541-026-01386-z.