A Distinct Effector B Cell Population Drives Autoantibody Production in SARS-CoV-2 Infection

SUMMARY

Autoantibodies (autoAbs) are linked to mortality and Long COVID, yet their cellular origins remain unclear. We analyzed the INCOV cohort and identified 12 age- and sex-matched participants with varying autoAb abundance and integrated single-cell RNA-seq and ATAC-seq data from B cells, plasma proteomics, proteome-wide autoAb profiling, clinical data, and in vitro assays. AutoAb abundance inversely correlated with neutralizing IgG and declined as infection resolved, paralleling the contraction of atypical memory B cells (AtMs). In vitro, AtMs preferentially differentiated into autoAb-producing antibody-secreting cells upon TLR7/8 stimulation. CD11c⁺ AtMs (double-negative 2, DN2s) in autoAb-high individuals exhibited increased TLR7 signaling, oxidative stress, and isotype switching, regulated by transcription factors T-bet and XBP1. Integrated genetic and genomic analyses showed that DN2s had the strongest enrichment for autoimmune trait heritability and inferred regulatory effects of autoimmune risk variants among B cell subsets. These findings identify DN2s as key precursors of autoAb-producing cells during SARS-CoV-2 infection.

eTOC

Autoantibodies (autoAbs) are linked to mortality and Long COVID, yet their cellular origins remain unclear. Yuan et al. identify a distinct B cell population (DN2 cells) as a key source of autoAbs during SARS-CoV-2 infection. These cells mirror lupus DN2s and carry a strong autoimmune genetic signature, linking infection-triggered autoAbs to autoimmune predisposition and long COVID.

Graphical Abstract

INTRODUCTION

Autoantibodies (autoAbs) are implicated in severe outcomes of acute Coronavirus Disease 2019 (COVID-19)^1–6 and are associated with elevated risk for post-acute sequelae of COVID (PASC, or Long COVID)^7–11. Acute SARS-CoV-2 infection also increases the risk of new-onset autoimmune conditions^12–14, including systemic lupus erythematosus (SLE). These observations underscore the need to move beyond these associations and define the cellular and molecular mechanisms of autoAb production in COVID-19.

In SLE, autoAbs are primarily produced by antibody-secreting cells (ASCs) differentiated from double negative 2 cells (DN2s)¹⁵. DN2s are a major subset of atypical memory B cells (AtMs), lack canonical naïve (IgD) and memory (CD27) markers, express elevated CD11c, and can constitute up to 80% of circulating B cells¹⁵. DN2s are driven by non-germinal-center (GC) activation and are highly responsive to innate stimuli via Toll-like receptor 7 (TLR7) signaling¹⁵. In COVID-19, we and others have identified an expansion of AtMs with low somatic hypermutation (SHM) rates^11,16–18, suggesting a non-GC origin similar to SLE. However, not all COVID-19 patients with elevated AtMs exhibit high autoAb abundance¹¹, indicating variability in their pathogenic role. Beyond autoimmunity^19–21, AtMs are expanded in aging^22–24 and infectious diseases like HIV and malaria^25–27, where they are associated with GC-independent responses²⁸. AtMs are also found in healthy immune responses to vaccination and capable of rapid and robust recall responses²⁹.

We hypothesize that CD11c⁺ AtMs, or DN2s, represent a distinct, hyperactive B cell lineage in SARS-CoV-2 infection that is associated with autoAb production. We sought to go beyond established epidemiologic associations between autoAbs and adverse post-COVID outcomes by defining the molecular features of the B cell response associated with autoAb production. To this end, we mined a longitudinal, multi-omic data set collected from 209 participants with COVID-19 to select a subcohort of sex- and age- matched female patients with similar AtM compositions but varying plasma autoAb abundance. Using circulating B cells and plasma, we integrated multi-omic single-cell transcriptomics (scRNA-seq) and transposase-accessible chromatin sequencing (scATAC-seq), plasma proteomics and proteome-wide autoAb profiling. This integrated analysis was used to define B cell types and lineages, identify key transcription factors (TFs), cis-regulatory elements, and differentially expressed genes (DEGs) associated with autoAb production, and to assess their functional roles in vitro. We explore the genetic basis of susceptibility to these responses by evaluating cell type-specific heritability and regulatory effects of autoimmune genetic variants. These analyses provide insights into the molecular and genetic basis of infection-induced autoimmunity in COVID-19 and its downstream outcomes, including Long COVID.

RESULTS

AutoAb-high and autoAb-low individuals are matched in clinical features but differ in autoAb abundance

We selected a nested cohort from the 209-person INCOV cohort, which is a well-characterized, diverse cohort for acute and post-acute COVID-19^11,18 (STAR Methods). Because autoAb prevalence is higher in women and older adults³⁰, we selected patients of the same sex and similar age with markedly different plasma autoAb abundance. After applying these criteria, 12 patients remained eligible. These were divided into autoAb-high (75±7 years, n=4; at least one ELISA autoAb > 3.5 standard deviations [SD] above pre-COVID healthy controls) and autoAb-low (74±9 years, n=8; all < 2 SD) (Figures 1A; STAR Methods). The two groups showed similar AtM compositions, and had no significant differences in sample collection time, disease severity, or comorbidities (Tables S1.1–S1.4). Additionally, we selected a 101-patient validation cohort using the same autoAb criteria but without age or sex restrictions. All samples were collected before April 2021 (pre-Omicron, pre-vaccination). Symptoms associated with long COVID were reported by most participants (Tables S1.5–1.6) and did not differ significantly between the two autoAb groups (Tables S1.3–1.4), and it was not used as a selection criterion.

Figure 1. AutoAb-high COVID-19 patients show proteome-wide autoreactivity, attenuated neutralizing antibodies, and a convalescent type III IFN response.

(A) Overview of cohort selection, sample types analyzed, and assays performed. EHR, electronic health records; WGS, whole genome sequencing; autoAb-high, patients with any autoantibody (autoAb) > 3.5 standard deviations (SD) above the pre-COVID healthy control mean; autoAb-low, patients with all autoAb < 2 SD above the pre-COVID healthy control mean.

(B) Box plots showing autoAb optical density at 450nm (OD450) z-scores (relative to pre-COVID healthy controls) for anti-nuclear antibodies (ANAs) (left) and autoAbs against IFN-α2 (right) measured by ELISA in autoAb-high and autoAb-low patients.

(C) Box plots showing quantile-normalized median fluorescence intensity (MFI) for four major protein categories measured using HuProt arrays in autoAb-high and autoAb-low patients.

(D) Scatter plot showing the correlation between ANA OD450 z-scores measured during acute infection and neutralizing antibody titers against SARS-CoV-2 at 3 months. AutoAb-high (n = 4) and autoAb-low (n = 7) samples were fitted with separate linear regression lines with 95% confidence intervals. Pearson correlation coefficients and p-values are shown.

(E) Same as (D), but for the 209-patient cohort, showing only samples with ANA > 2 SD above the healthy control mean.

(F) Box plots showing autoAb OD450 z-scores during acute infection (around diagnosis) and at 3, 6, 12, 18, and 24 months post-diagnosis. Samples from the same patients are connected by lines.

(G) Heatmap of selected plasma proteins significantly elevated in autoAb-high patients compared with autoAb-low patients at 3 months. Samples (rows) are hierarchically clustered.

(H) Top enriched Gene Ontology Biological Process (GO BP) terms of plasma proteins significantly elevated in autoAb-high patients compared with autoAb-low patients at 3 months. p-values are Benjamini-Hochberg adjusted. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001.

See also Figure S1 and Tables S1–2.

We used ELISA to quantify 8 autoAbs associated with COVID-19 and/or systemic rheumatic diseases^4,31,32: anti-interferon α2 (IFN-α2) and anti-nuclear autoantibodies (ANAs) against Jo-1, La/SS-B, ribosomal phosphoprotein P1 (P1), Ro/SS-A, U1-snRNP, CENP-B, SmD, and Scl-70 (Figures 1B, S1A, and S1B). To examine autoreactivity beyond anti-IFN-α2 and ANAs, we performed a HuProt proteome autoAb assay (Figure S1C), which corroborated the ELISA findings for IFN-α2 and Ro/SS-A (TROVE2) and revealed increased reactivity against multiple IFN-α subtypes (Figure S1D). Grouping antibodies into four protein classes (nuclear, cytosolic, membrane, and secreted), the autoAb-high group showed significantly higher reactivity against all classes in both cohorts (Figures 1C and S1E), indicating a proteome-wide increase in autoreactivity in the autoAb-high group, particularly against type I IFNs and nuclear antigens.

The autoAb-high group exhibited lower SARS-CoV-2 receptor-binding domain IgG and neutralizing antibodies at convalescence (3 months post-diagnosis) (Figure S1F). Within this group, convalescent neutralizing antibodies negatively correlated with acute ANA abundance (Figure 1D). This anti-correlation was highly significant in patients with ANA z-score > 2 SD in the full INCOV cohort (Figure 1E), consistent with our previous report¹¹ and suggesting attenuated viral-specific antibody responses when autoAbs are high. AutoAb production appeared infection-triggered in most individuals, declining to baseline upon recovery (Figure 1F).

These analyses define two sex- and age-matched groups that differ mainly in autoAb abundance, with autoAb-high individuals showing proteome-wide autoreactivity and attenuated neutralizing antibody responses.

Type III IFN and B cell responses characterize autoAb-high individuals at convalescence

Plasma proteomic profiling revealed that, during acute infection, inflammation did not differ significantly between the two autoAb groups, as assessed by an inflammation panel (Figure S1G). At convalescence, however, the autoAb-high group showed significant elevation of IFNL1 (the type III IFN) and its receptors IL10RB and IFNLR1 (Figure 1G), with the type III IFN response the most enriched Gene Ontology (GO) pathway (Figure 1H; Table S2). Proteins associated with B cell activation and B cell proliferation were also elevated in the autoAb-high group, as was ACE2, which rose during acute infection and remained elevated into convalescence (Figure S1H), suggesting persistent shedding. These findings were recapitulated in the 101-patient cohort (Figures S1H–J). Thus, autoAb-high individuals are distinguished not by acute inflammation, but by a convalescent signature of elevated type III IFN, B cell activation and proliferation, and persistent ACE2.

In vitro stimulation of patient-derived B cells identifies AtMs as autoAb-producing precursors

Non-GC activation of AtMs through dysregulated TLR7 signaling, leading to pathogenic autoAb production, is a hallmark of SLE pathogenesis¹⁵. To investigate whether a similar pathway operates in SARS-CoV-2 infection, we stimulated naïve B cells, AtMs, switched memory B cells (SWMs), and ASCs sorted from PBMCs with TLR7/8 agonists (Figures 2A and S2A), quantifying total IgG, anti-IFN-α2 IgG, ANAs, and SARS-CoV-2 spike-specific IgG in culture supernatants over time (Figure S2B). Additionally, we measured cell counts and ASC frequencies to assess cell expansion and differentiation, respectively.

Figure 2. B cell in vitro stimulation assays identify AtMs as autoAb-producing precursors.

(A) Top, graphical illustration of the experimental workflow. Bottom, graphical illustration of quantitative assays performed.

(B) Scatter plots fitted with linear regression lines and 95% confidence intervals (grey shaded areas) showing the correlation of atypical memory B cell (AtM) (top) and switched memory B cell (SWM) (bottom) expansion with supernatant IFN-α2 IgG optical density at 450nm (OD450) (left), anti-nuclear antibody (ANA) IgG OD450 (middle) and SARS-CoV-2-specific IgG OD450 (right). Pearson correlation coefficients and p-values are shown.

(C) Box plots showing percentages of antibody-secreting cells (ASCs) at the end of the culture (day 16) of AtM and SWM cultures from autoAb-high and autoAb-low patients. **p < 0.01.

(D) Box plots, from left to right, showing peak total IgG, anti-IFN-α2 IgG, ANA IgG, and SARS-CoV-2-specific IgG OD450 from AtM and SWM cultures from autoAb-high and autoAb-low patients. *p < 0.05, ns, not significant (p > 0.05).

See also Figure S2.

With TLR7/8 agonist R848, AtMs and SWMs cultures expanded substantially (20–30 fold) and differentiated into IgG-producing ASCs, whereas naïve B cells and ASCs exhibited minimal expansion (1–2 fold), potentially explaining their low antibody signals (Figure S2C). Notably, no cultures survived without R848 (Figure S2D). AtM expansion correlated with anti-IFN-α2 and ANA production (Figures 2B top and S2E), whereas SWM expansion correlated with all measured IgG types, including autoAbs and, most strongly, SARS-CoV-2-specific IgG (Figures 2B bottom and S2E). Anti-IFN-α2 and ANA production were highly correlated, especially in AtMs (Figure S2E). Thus both AtMs and SWMs can generate autoAbs in vitro, while SWMs additionally produce SARS-CoV-2-specific IgG.

AtMs from the autoAb-high group differentiated into ASCs more efficiently (Figure 2C) and produced more total IgG and autoAb IgG (Figure 2D) than those from the autoAb-low group. This capacity was not restricted to acute infection. AtMs from a convalescent autoAb-high individual still differentiated into autoAb-producing cells upon TLR7/8 stimulation (Figure S2D), though this observation requires confirmation in a larger cohort.

Together, these assays identify AtMs as non-GC TLR7/8-driven precursors of autoAb-producing cells, with autoAb-high AtMs preferentially skewed toward autoAb production, as in SLE.

Single-cell multi-omics identifies a distinct DN2 population in acute SARS-CoV-2 infection

To probe for potential differences in the B cell biology between autoAb-high and autoAb-low patients, we performed scRNA-seq and scATAC-seq on nuclei isolated during acute infection from the 12-patient cohort. From over 150,000 B cells sequenced, 96,551 passed quality control (QC; STAR Methods). We used MIRA³³ to integrate scRNA-seq and scATAC-seq data to identify cell types and gene regulation along pseudotime trajectories, resolving seven distinct B cell types: naïve (54.4%), AtM (DN2 [6.6%] and DN3 [2.1%]), intermediate Memory (int. Mem, 4.5%), unswitched memory (USM, 10.3%), SWM (9.7%), and ASC (12.5%) (mean proportions across samples; Figures 3A). scATAC-seq revealed no additional populations beyond those defined by scRNA-seq alone. Nearly all populations were represented by multiple donors, indicating that the data were not dominated by any single individual (Figure S3A).

Figure 3. Single-cell multi-omics identifies DN2s as a distinct B cell population in acute SARS-CoV-2 infections.

(A) Joint scRNA-seq and scATAC-seq UMAP of 96,551 B cells from the 12-patient cohort.

(B) UMAP of B cells color-coded by expression of marker genes for DN2s.

(C) Dot plot showing normalized expression of marker genes across B cell types. Dot size and color represent the fraction of cells expressing the gene and the column-scaled mean of expression among expressed cells within each cell type, respectively.

(D) Normalized pseudo-bulk chromatin accessibility profiles for each B cell type at the DN2 marker gene ITGAX. Maximum accessibility was clipped at 50% of the maximum value. Promoter regions are highlighted in gray and the gene model is shown at the bottom.

(E) Heatmap showing column-scaled normalized accessibility of differentially accessible peaks in each B cell type. Peak annotations are shown in the top bar.

(F) Box plot showing percentages of AtMs at Acute-1 (around diagnosis), Acute-2 (approximately within week after Acute-1), and at 3 months. **p < 0.01.

(G) B cell lineage inferred from the scRNA-seq and scATAC-seq data.

(H) Bar plot showing somatic hypermutation rates, defined as the percentage of gaps and mismatches in the CDR1–3 regions of the heavy-chain sequence relative to the top germline sequence hit. Data are represented as mean ± SEM. **FDR < 0.01, ****FDR < 0.0001, ns, not significant (FDR > 0.05).

See also Figure S3.

AtMs were classified by CD27, CD21, and TBX21 expression, and DN2s were distinguished from DN3s by ITGAX (Figures S3B–D). DN2s were characterized by expression of TBX21 (encodes T-bet), ITGAX (encodes CD11c), SOX5, ZEB2, FCRL3/5, MPP6, and DAPP1 (Figures 3B and 3C). Only a small fraction of cells expressed T-bet, consistent with prior studies^34,35, but where expressed it was strong and highly specific to DN2s (Figures 3B and 3C). T-bet and CD11c are well-established AtM markers across malaria, HIV, and autoimmunity^27,36–39, and T-bet is highly expressed in COVID-19 DN2s¹⁷. Chromatin accessibility around the ITGAX promoter is the highest in DN2s (Figure 3D). These markers have established roles in B cells: SOX5 decreases B cell proliferation but promotes plasmablast differentiation⁴⁰, ZEB2 drives CD11c⁺ age-associated B cell development⁴¹, and FCRL5 can disrupt B cell anergy and facilitate TLR signaling, promoting autoimmunity⁴². Their elevation in SLE and healthy DN2s¹⁵ suggests a shared marker expression profile with COVID-19 DN2s.

ASCs and DN2s had the most differentially accessible regions (DARs), whereas naïve and memory B cell types had few (Figure 3E), suggesting high transcriptional activity in these effector populations. These cell type-specific DARs were also enriched for promoter (13.3%) and exon (3.1%) regions versus all peaks (7.6% and 2.5%). Circulating AtMs decreased significantly from acute COVID-19 to convalescence (Figure 3F), mirroring the autoAb decline (Figure 1F), whereas other B cell types showed no significant changes over time (Figure S3E).

Together, single-cell multi-omics resolves seven B cell types and identifies DN2s as a transcriptionally active, T-bet⁺CD11c⁺ population whose marker profile mirrors SLE DN2s.

Lineage analysis identifies key regulators of AtM differentiation

MIRA lineage inference showed that DN2 and DN3 AtMs resembled naïve B cells transcriptomically and epigenetically, forming an alternative lineage to SWMs (Figure 3G; STAR Methods). AtMs had significantly lower SHM rates than SWMs, which showed the highest among all subsets (INCOV CITE-seq data¹¹; Figure 3H), consistent with GC-independent activation.

To identify key regulators driving AtM and SWM differentiation, we analyzed RNA and ATAC MIRA latent variables along the pseudotime trajectory (Figure S3F; STAR Methods). Naïve B cells were enriched for SPI1 and SPIB (environmental response and GC formation for high-affinity antibody production⁴³) and CTCF (proliferation over differentiation via PRDM1 repression⁴⁴). The AtM branch was distinguished by ATAC variable a1 (T-box and AP-1 motifs [JUN, JUNB, JUND, FOS, FOSL1/2, FOSB, BATF3]; B cell activation, class switching, and plasma cell differentiation^45–47), with terminal AtM formation marked by a9 (OCT2; BCR signaling, T-B interactions, and T follicular helper cell differentiation⁴⁵) and RNA variable e2 (MPP6, FCRL5, IFNLR1, and FGFR1; antigen-receptor and type III IFN signaling). The SWM branch was defined early by a0 (NFKB1/2 and TCF3/4; B cell activation and GC development^45,48), and late by a10 (EGR2, EGR3; inhibition of excessive cytokine secretion and age-associated B cell accumulation^49,50).

Thus AtMs likely arise from naïve B cells and adopt an activated phenotype, whereas SWMs follow a restrained pathway towards GC development.

Transcriptomes of autoAb-high DN2s are enriched for TLR7 signaling

To elucidate cell type-specific differences between autoAb-high and autoAb-low individuals, we first examined the B cell compositions. Unlike active SLE, where DN2s are markedly expanded^15,39, no subset was significantly overrepresented in either group (Figure S4A). Transcriptomic analysis of B cell subsets (Table S3) revealed that DN2s exhibited the most pronounced differences between groups, with the “Trafficking and Processing of Endosomal TLR” pathway as the most enriched in DEGs in autoAb-high DN2s (combined score 122; Figures 4A and 4B). This included upregulated TLR7 (an ssRNA sensor) and HSP90B1, and downregulated UNC93B1 (Figure 4A). HSP90B1, a master TLR chaperone, is required for antibody production under TLR stimulation⁵¹, whereas the TLR chaperone UNC93B1 prevents TLR localization to the plasma membrane where self-nucleic acids may be present⁵². UNC93B1 D34A mutant mice develop severe autoimmune responses and lethal inflammation⁵³. STAT4 was upregulated in autoAb-high DN2s (Figure 4A), potentially skewing B cells toward non-GC activation and differentiation into plasmablasts^54,55. CD79B, a subunit of the BCR complex, was slightly but significantly downregulated in DN2s from autoAb-high patients (Figure 4A; Table S3). MHC class II and associated antigen presentation proteins were unchanged (Table S3). Together, these findings support enhanced endosomal TLR7 signaling in autoAb-high DN2s.

Figure 4. AutoAb-high DN2s show enhanced endosomal TLR7 signaling and distinct T-box and XBP1 cis-regulation.

(A) Heatmap of selected DEGs in DN2s that distinguish autoAb-high and autoAb-low patients. Samples (columns) are ordered by ANA OD450 z-score shown in the top bar plots. The log₂ fold change between autoAb-high and autoAb-low patients is shown on the right. *p < 0.05, **p < 0.01, ****p < 0.0001.

(B) Top 10 enriched Reactome 2024 pathways of DEGs (n=330) in autoAb-high DN2s compared to autoAb-low DN2s.

(C) Box plots of expression of FOS (left) and FOSB (right) across all B cell types and B cells jointly (All B) in autoAb-high and autoAb-low patients.

(D) Heatmap showing the top 10 enriched motifs per cell type and autoAb group, with rows and columns hierarchically clustered and TF families annotated at the bottom. Exact motifs and families are provided in Table S4.

(E) UMAP of B cells colored by motif enrichment scores for T-bet.

(F) Bar plot showing the number of genes with varying numbers of gl-cCREs identified. The x-axis indicates the number of gl-cCREs linked to a gene (1 to 14), and the y-axis represents the number of genes. For example, approximately 1,500 genes had only one cCRE linked to them in autoAb-high patients.

(G) Heatmap showing column-scaled pseudobulk normalized chromatin accessibility of gl-cCREs across autoAb groups for each B cell type. AutoAb- and cell type-specific gl-cCRE modules were identified using non-negative matrix factorization, and columns are ordered according to these modules.

(H) Scatter plot showing enriched TF motifs in autoAb-high- and DN2-specific gl-cCREs (left), autoAb-high- and ASC-specific gl-cCREs (middle), and autoAb-low- and ASC-specific gl-cCREs (right). Motifs with a fold enrichment > 1.6, and adjusted p-values (Benjamini-Hochberg) < 0.01 are shown with larger dot sizes and labeled with motif names.

See also Figure S4 and Tables S3–5.

AutoAb-high DN2s exhibited a metabolically activated phenotype characterized by excessive oxidative stress, with downregulation of the reactive oxygen species (ROS)-inhibiting gene TXN⁵⁶ and upregulation of ROS-stimulated genes CDK6, TNIK and FOS (Figure 4A). FOS, an AP-1 component, is critical in B cell proliferation and ASC differentiation⁵⁷. FOS and FOSB were upregulated across all B cell subsets in autoAb-high individuals (Figure 4C). AutoAb-high DN2s also presented an isotype switched phenotype, with upregulated IGHG3–4 and IGHA2 and downregulated IGHM (Figure 4A).

B cells expressed the highest IFNLR1 among circulating immune cells (Figure S4B), and DN2s the highest among B cell subsets (Figure 3C). AutoAb-high DN2s upregulated IFNLR1 even further (Figure 4A), suggesting those DN2s are uniquely poised to respond to elevated plasma type III IFN abundance (Figure 1G).

The source of TLR7 stimulation, whether self-nucleic acids from damaged cells or SARS-CoV-2 ssRNA, remains an open question. All enrolled patients had confirmed COVID-19 diagnoses via positive nasal swabs. SARS-CoV-2 viremia showed no significant difference between groups in either the 12- or the 101-patient cohort (Figure S4C), though tissue viral loads remain unknown. To assess ssRNA from damaged cells, we examined inflammation and tissue-damage markers using clinical labs and plasma proteomics. Inflammation was not elevated in the autoAb-high group, as assessed by an inflammation panel (Figure S1G) and pro-inflammatory markers, including C-reactive protein, IL-6, IL-1A, TNF, and IFNG (Figures S4D and S4E). Canonical tissue-damage markers, such as KRT19 and AGER, showed no differences between the two groups (Table S2).

Together, autoAb-high DN2s display a distinct transcriptomic program of enhanced endosomal TLR7 signaling, oxidative stress, isotype switching, and potential heightened type III IFN responsiveness, without evidence of greater viral load, systemic inflammation, or tissue injury.

T-bet activity characterizes DN2s but is independent of autoAb status

To investigate TF activity, we calculated single-cell motif scores using ChromVAR⁵⁸ (STAR Methods). TF motif enrichment patterns were associated with B cell types, not autoAb status (Figure 4D; Table S4). ASCs and DN2s exhibited the highest motif activity, consistent with their greater DARs compared to other B cell types (Figure 3E), suggesting a more transcriptionally active state in these cells.

DN2s showed a unique motif enrichment profile, sharing ETS motifs with non-ASCs (B cell survival and identity, restraining differentiation) and AP-1 motifs with ASCs. ASCs showed unique enrichment for IRF and STAT motifs (antibody production), whereas DN2s were uniquely enriched for T-box motifs, with T-bet identified as the top enriched motif (Figures 4D and 4E). Thus, transcription factor motif enrichment in open chromatin regions distinguishes B cell types rather than autoAb groups, with T-box/T-bet motifs specifically marking DN2s.

T-box and XBP1 programs distinguish the cis-regulatory landscape of autoAb-high DN2s

To elucidate active cis-regulatory mechanisms specific to each B cell type and autoAb group, we identified candidate cis-regulatory elements (cCREs) linked to their target genes (gene-linked cCREs, or gl-cCREs). Links were supported by gl-cCRE-promoter co-accessibility and gl-cCRE-gene expression correlation (STAR Methods)⁵⁹. AutoAb-high individuals had fewer gl-cCREs per gene (Figure 4F), suggesting looser cis-regulation and potential loss-of-function pathogenesis. As expected, gl-cCREs were predominantly located in non-coding genomic regions (Figure S4F).

We clustered gl-cCREs based on chromatin accessibility using non-negative matrix factorization (NMF) to identify cell type- and autoAb-specific cis-regulatory patterns (STAR Methods; Table S5). While both SWMs and DN2s can yield autoAb-producing ASCs, autoAb-associated cis-regulatory differences were most pronounced in ASCs and DN2s, with DN2s showing the clearest differences among B cell populations (Figure 4G). Within autoAb-high DN2-specific gl-cCREs, motifs for T-box and those for regulation of antibody secretion (XBP1, CREB3, and ATF1/6) were significantly enriched (Figure 4H left). These results suggest that, despite broad enrichment of ETS, AP-1, and T-box motifs in DN2s (Figure 4D), T-box and XBP1 programs are the most distinguishing cis-regulatory features of autoAb-high DN2s.

IRF4, IRF5, and IRF8 motifs^60,61 were enriched in both autoAb-high and autoAb-low ASC gl-cCREs (Figure 4H middle and right). IRF9 motifs were enriched in autoAb-high ASC gl-cCREs (FDR < 1 × 10⁻⁶⁰). As IRF9 is crucial for TLR7-mediated B cell activation by regulating TLR7 expression and is required for IgG autoAb production⁶², this may reflect enhanced TLR7 signaling. Other gl-cCRE modules showed minimal motif enrichment.

Together, these analyses show that T-box and XBP1 programs distinguish the cis-regulatory landscape of autoAb-high DN2s, while IRF9 enrichment marks autoAb-high ASCs.

T-bet and AP-1 networks establish DN2 identity, while T-bet and XBP1 exert stronger regulatory activity in autoAb-high DN2s

The roles of T-bet, AP-1, and XBP1, which stood out in the lineage, motif enrichment and gl-cCRE analyses of DN2s (Figures S3F, 4D and 4H), were further explored by performing TF footprinting and regulatory network analysis. Footprinting detects reduced Tn5 cleavage (“footprints”) at bound motifs, indicating TF occupancy. T-bet exhibited the deepest footprint in DN2s, suggesting the strongest TF occupancy in this population compared with other B cell types (Figure 5A).

Figure 5. T-bet and AP-1 networks establish DN2 identity, with elevated T-bet and XBP1 target gene expression in autoAb-high DN2s.

(A) Tn5 insertion bias-corrected footprinting of T-bet across B cell types. Binding motif shown at top.

(B) Scatter plot of T-bet motif enrichment versus the background-corrected average expression of T-bet target genes through promoters (left) and gene-linked candidate cis-regulatory elements (right) for each cell type.

(C) Violin and box plots showing T-bet motif enrichment in autoAb-high and autoAb-low DN2s.

(D) Tn5 insertion bias-corrected footprinting of T-bet in autoAb-high and autoAb-low DN2s. Binding motif shown at top.

(E) Violin and box plots of T-bet target gene scores through promoters in autoAb-high and autoAb-low DN2s.

(F) UMAP of B cells colored by motif enrichment scores for XBP1.

(G) Violin and box plots showing XBP1 motif enrichment in autoAb-high and autoAb-low DN2s.

(H) Violin and box plots of XBP1 target gene scores through promoters in autoAb-high and autoAb-low DN2s.

(I) T-bet regulatory network showing selected target genes of T-bet in DN2s.

(J) Genomic tracks of a peak containing the promoter region of an example T-bet target gene (FCRL3, a DN2 marker that is upregulated in autoAb-high DN2s), showing corrected cut-site signal tracks (top) and TOBIAS footprint scores (bottom). Detected motifs are highlighted in grey, and gene locations are indicated by black boxes at the top, with arrows indicating gene strand. **p < 0.01, ****p < 0.0001.

See also Figure S5.

To quantify TF-mediated gene regulation, we computed per-TF target gene scores, defined as the background-corrected mean expression of genes linked through promoters or gl-cCREs (STAR Methods). B cell types with higher T-bet motif enrichment, as measured by ChromVAR deviation scores, also displayed elevated expression of target genes (Figure 5B), supporting the role of T-bet as a transcriptional activator.

We next examined whether T-bet regulation differed by autoAb status within DN2s. T-bet motif enrichment, assessed by ChromVAR, was only slightly elevated in autoAb-high relative to autoAb-low DN2s (Figure 5C), consistent with TF footprinting results showing similar occupancy between the two groups (Figure 5D). However, T-bet promoter-linked target gene expression was significantly higher in autoAb-high DN2s (Figure 5E), indicating stronger downstream transcriptional activity through promoter-mediated regulation despite similar motif accessibility and TF occupancy.

XBP1 motifs were enriched specifically in DN2 and ASC (Figure 5F). Motif enrichment is comparable between autoAb-high and autoAb-low DN2s (Figure 5G). However, similar to T-bet, XBP1 target gene scores were significantly higher in autoAb-high DN2s (Figure 5H). This suggests greater XBP1 regulatory activity in autoAb-high DN2s and is consistent with their greater protein secretion capacity observed in the in vitro stimulation assays (Figures 2C and 2D).

For the AP-1 TF family, which ranked second in motif enrichment in DN2s (Figure 4D), we focused on FOS, the top-enriched AP-1 TF. FOS motifs were enriched in both DN2s and ASCs (Figure S5A), with the strongest footprinting in DN2s (Figure S5B), indicating robust occupancy. Higher FOS motif enrichment was associated with increased expression of both promoter- and gl-cCRE- linked target genes (Figure S5C), suggesting that FOS also acted as a transcriptional activator. However, neither FOS motif enrichment nor expression of FOS target genes differed significantly by autoAb status (Figures S5D–F), suggesting that FOS is more tied to DN2 cell identity and activation state than to autoAb-associated differences.

We constructed regulatory networks for T-bet and FOS (STAR Methods), showing multiple DN2 marker genes and DEGs distinguishing autoAb-high and autoAb-low DN2s as the target genes (Figures 5I and S5G). T-bet regulated the DN2 markers SOX5, FCRL3 and MPP6 (Figure 5I), while FOS regulated SOX5, ZEB2, and ITGAX (Figure S5G). T-bet also regulated FOS directly (Figure 5I). Notably, the two most highly expressed DN2 marker TFs SOX5 and ZEB2 were regulated through gl-cCREs, whereas other marker genes and DEGs (ITGAX, MPP6, FCRL3) were regulated through promoters. This indicates that T-bet and FOS control DN2 identity through enhancer-mediated TF regulation, whereas promoter-mediated regulation drives the expression of effector genes critical for DN2 function.

To provide locus-level evidence of T-bet occupancy beyond the aggregated footprints, we applied TOBIAS⁶³ for genome-wide footprinting. The promoter of the T-bet target gene FCRL3 (a DN2 marker upregulated in autoAb-high DN2s) contained two T-bet motifs with high footprinting (Figure 5J), confirming locus-level occupancy.

Together, the findings illustrated in Figures 4 and 5 reinforce the pivotal roles of T-bet and AP-1 in establishing DN2 identity and point to the stronger T-bet and XBP1 regulatory activities in autoAb-high cells. Notably, T-bet and AP-1 are also most enriched in DN2 accessibility in SLE epigenetic studies⁶⁴. In contrast, the lack of distinct TF motifs associated with SWMs in the autoAb-high group (Figures 4D and 4G) suggests that autoAbs originating from SWMs may arise from a recall response rather than a de novo activation driven by infection-associated epigenetic changes.

DN2s are the most enriched for autoimmune heritability among B cell types

Given the strong genetic basis of SLE susceptibility^65,66, we asked whether genetic risk for common diseases might converge on the DN2s identified here, linking the regulatory programs associated with autoAb production in COVID-19 to broader health conditions. Using Linkage Disequilibrium Score Regression (LDSC)⁶⁷, we partitioned the heritability of 28 genome-wide association study (GWAS) traits, including social factors, anthropometric traits, and autoimmune diseases, across cell type-specific ATAC peaks (Figure 3E; STAR Methods). As expected, B cells were significantly enriched for autoimmune trait heritability (Figure 6A). Among all B cell subsets, DN2s exhibited the strongest enrichment for several autoimmune traits, including multiple sclerosis (FDR=6×10⁻⁶), primary biliary cirrhosis (FDR=2×10⁻⁴), type 1 diabetes (FDR=6×10⁻⁴), inflammatory bowel disease (FDR=0.001), Crohn’s disease (FDR=0.005), rheumatoid arthritis (FDR=0.01), and SLE (FDR=0.001). Consistently, SLE age-associated B cells (CD19⁺CD27⁻CD21⁻CD10⁻) have been reported to express the highest expression of SLE risk genes among immune cell types⁶⁵. Thus, among B cell subsets, DN2 chromatin is preferentially enriched for the heritability of autoimmune traits, including SLE.

Figure 6. DN2s are enriched for autoimmune heritability and risk-variant effects.

(A) Heatmap showing the cell type-specific linkage disequilibrium score regression (LDSC) enrichment of GWAS traits across B cell types. p-values are Benjamini-Hochberg adjusted. Adjusted p-values (FDR) are overlaid on the heatmap when < 0.05. *FDR < 0.05, **FDR < 0.01, ***FDR < 0.001, and ****FDR < 0.0001.

(B) Scatter plot of the difference in motif binding strength (risk allele minus reference allele) versus the contribution score of the reference allele in DN2s, color-coded by the difference in contribution scores (risk allele minus reference allele) in DN2s. Each dot represents a variant overlapping a TF motif (FIMO p-value < 1 × 10⁻⁴). The percentage of variants with a reference allele contribution score greater than 0.1 is shown in the top right corner. Histograms of the x-axis values (reference allele contribution scores) and y-axis values (motif binding strength differences) are shown on the top and left, respectively.

(C) Single-base contribution scores for the 50-base pair sequence surrounding the variant rs11553760 in DN2s (top) and ASCs (bottom). The variant position is marked by two vertical red lines. T-bet Binding motif shown at top.

(D) Odds ratio for developing anti-nuclear antibody (ANA) (left) and post-acute sequelae of SARS-CoV-2 infection (PASC) (right) for the following independent factors: common control factors including sex (female), age per decade, BMI (kg/m²), confirmed diagnosis of autoimmune disease, and autoimmune polygenic risk score (PGS).

See also Figure S6 and Table S6.

DNA sequence models infer preferential SLE variant effects in DN2s

Because LDSC analysis suggested that autoimmune risk loci are more likely to exert regulatory effects in DN2s compared to other B cell types, we investigated DN2-specific regulatory consequences of these variants. Here we use SLE as a representative case study, because of the depth of associated literature. We quantified changes in motif binding strength between reference and risk alleles using negative log10-transformed FIMO p-values (STAR Methods). Most SLE risk variants reduced motif binding strength relative to the reference variants (Figure 6B y-axis histogram), suggesting that these variants may disrupt motif binding sites, contributing to loss-of-function effects.

To evaluate the magnitude and specificity of these motif binding alterations, we trained a deep learning DNA sequence model to learn regulatory grammars from chromatin accessibility data^68–70 and to infer cell type-specific single-base contribution scores, each nucleotide’s predicted contribution to chromatin accessibility⁷¹. Correlations between the ground truths and predictions (Figure S6A) captured expected similarities and differences between cell types (Figure S6B).

Our analysis revealed that 16.2% of SLE motif-binding reference variants had contribution scores greater than 0.1 in DN2s (Figure 6B), compared with 12.1% in ASCs and far fewer in other cell types (Figure S6C). This pattern is consistent with the LDSC results and suggests that DN2s are the B cell population in which SLE risk variants are most likely to exert effects through altered motif binding. Notably, higher reference allele scores were associated with greater reductions upon substitution with the risk allele (Figure 6B), indicating that variants in highly accessible regions are particularly sensitive to allelic changes, potentially disrupting regulatory function.

This framework provides a powerful tool for deciphering how individual variants act in a cell type-specific manner, beyond GWAS associations. For example, the synonymous variant rs11553760 in the gene CSK lies within a T-bet motif in the reference allele (FIMO p-value = 5.9 × 10⁻⁵), and this T-bet motif is predicted to be highly accessible in DN2s (Figure 6C top). Substituting the reference allele (C) with the risk allele (T) enhanced T-bet binding (FIMO p-value = 2.9 × 10⁻⁶) and created twelve additional T-box motif matches. Consistently, the risk allele was predicted to increase sequence accessibility at this locus in DN2s but not ASCs (Figure 6C bottom), highlighting cell type specificity. This analysis suggests that rs11553760, despite being synonymous, may influence CSK gene regulation in DN2s through a gain-of-function regulatory mechanism involving enhanced T-bet activity. Details for all SLE GWAS variants overlapping TF motifs, their binding strengths and cell type-specific contribution scores, are in Table S6. Together, these models indicate that SLE risk variants preferentially reshape regulatory activity in DN2s, exemplified by rs11553760, predicted to enhance T-bet binding at CSK.

Autoimmune polygenic risk score independently associates with both ANA and PASC

The enrichment of autoimmune heritability in DN2 open chromatin suggested a genetic connection between the DN2 program and autoimmune disease risk. We therefore asked whether broader autoimmune genetic predisposition was associated with autoAb positivity and PASC in an independent population-scale dataset. We analyzed the All of Us (AoU) cohort⁷², which includes over 250,000 participants with whole-genome sequencing (WGS) and linked health records, testing whether autoimmune polygenic risk score (PGS)⁷³ is independently associated with ANA positivity (n=14,972, 3,416 ANA-positive), accounting for autoimmune disease status and common confounders (STAR Methods). The presence of autoimmune disease and female sex were strong predictors of ANA positivity, with odds ratios (ORs) of 2.59 and 1.57, respectively (Figure 6D left). The autoimmune PGS showed a dose-dependent trend, with individuals ≥ 1 SD above the mean PGS exhibiting increased risk of ANA positivity. Notably, autoimmune disease and autoimmune PGS were independent predictors of ANA positivity.

We repeated these analyses with long COVID as the outcome. Patients with newly emergent symptoms post-COVID (n = 939) were identified from 18,165 individuals with a recorded COVID-19 diagnosis or positive SARS-CoV-2 test. Again, the presence of autoimmune disease was a strong predictor of long COVID (OR=1.74), and the autoimmune PGS was a significant independent predictor, showing a similar dose-dependent trend as observed for ANA (Figure 6D right). These findings suggest that a shared genetic predisposition may underlie both ANA positivity and increased long COVID susceptibility. Together with prior evidence that autoAbs arising during acute COVID-19 are associated with Long COVID¹¹, these results highlight that some Long COVID cases may have an autoimmune component, potentially driven by genetic variants that promote autoAb production and chronic inflammation.

DISCUSSION

We previously reported that AtMs, identified as precursors of autoAb-producing cells in SLE¹⁵, emerged in acute SARS-CoV-2 infections with low SHM rates¹⁸. These observations raised the possibility that rapid, non-GC maturation of AtMs contributes to autoAb production in COVID-19. We therefore investigated B cells from a cohort of 12 age- and sex-matched female COVID-19 patients with similar frequencies of AtMs but significant differences in plasma autoAb abundance.

Expansion of AtMs correlated with the abundance of secreted autoAbs, but not with SARS-CoV-2-specific antibodies, suggesting that the dominant output of this population is more closely linked to autoreactivity than to virus-specific humoral responses. Our results are consistent with previous reports that naïve-derived, minimally mutated B cells within the DN2-associated pathway can produce antibodies that are dual-reactive for SARS-CoV-2 and self-antigens⁵, as a subset of the AtM samples in our study produced both autoAbs and SARS-CoV-2-specific antibodies. This suggests that the AtM pathway is functionally heterogeneous but, in autoAb-high patients, is preferentially skewed toward autoAb production.

A clear difference between autoAb-high and autoAb-low individuals was the enrichment of innate sensing pathways, particularly in DN2s. The pronounced enrichment of TLR7 signaling in autoAb-high DN2s mirrors pathways previously described in SLE¹⁵. Upregulation of TLR7 and HSP90B1, coupled with downregulation of UNC93B1, suggests enhanced endosomal TLR responsiveness in autoAb-high DN2s, potentially triggered by viral RNA or self-nucleic acids released from damaged cells. Consistent with this model, autoAb-high AtMs more efficiently differentiated into ASCs and produced significantly more autoAbs, but not SARS-CoV-2-specific Abs, in response to R848-dependent TLR7/8 stimulation in vitro. Together, these findings support a model in which heightened endosomal TLR responsiveness is a key step toward autoAb production in COVID-19.

Whether this reflects greater extrinsic stimulation or an intrinsically heightened cellular responsiveness is an important question, but several observations argue against a model driven solely by acute inflammation. The lack of significant differences in disease severity, SARS-CoV-2 viremia, systemic inflammatory markers, or tissue injury markers argue that environmental cues alone are unlikely to explain the stronger TLR-responsive program. In addition, AtM and SWM obtained at convalescence from an autoAb-high patient retained the capacity to produce autoAbs upon R848 stimulation after the infection had resolved, indicating that this autoreactive bias is not restricted to the acute phase. Together, these findings support the possibility of a durable predisposition that favors autoAb production.

For TF regulation, multiple lines of evidence support a T-bet-associated regulatory program in autoAb-high DN2s: specific T-bet expression; its motif as the most accessible in DN2s; greater motif enrichment in autoAb-high cis-regulatory elements; higher target gene expression in autoAb-high DN2s; and footprinting consistent with T-bet occupancy. T-bet regulation also characterizes AtMs in SLE, vaccinations, and other infections^{15,19,21,25–29,36,37}. Prior perturbation studies in mouse models have shown that T-bet knock-down in B cells reduces autoAb production and mitigates SLE pathology^38,74, though its causal role has not been tested in human B cells. By contrast, XBP1 motifs were enriched not broadly in DN2s but specifically in cis-regulatory elements in autoAb-high DN2s. Given XBP1’s role in protein secretion⁷⁵, this aligns with the enhanced ASC differentiation and antibody production in vitro. With fewer gl-cCREs per gene, these findings indicate autoAb-high DN2s are poised for rapid differentiation.

AutoAbs were generated from both AtMs and SWMs in vitro, and autoAb-high SWMs produced more ANA and SARS-CoV-2 Spike-specific IgGs following R848 stimulation. However, unlike DN2s, single-cell analyses did not identify clear transcriptomic or regulatory differences between autoAb-high and autoAb-low SWMs. These SWMs may represent memory B cell recall responses formed from previously activated AtMs²⁹, though we did not test this directly. In contrast, the pronounced transcriptomic and regulatory differences observed in DN2s between the two autoAb groups suggest that DN2s exhibit the clearest molecular association with the autoimmune component of SARS-CoV-2 infection and Long COVID, consistent with their enrichment for autoimmune trait heritability.

These mechanistic differences may also have consequences for the quality of antiviral humoral immunity and the trajectory of recovery. The negative correlation between autoAb and SARS-CoV-2 neutralizing IgG production suggests that elevated autoAbs, even if non-functional, are associated with impaired protective B cell immunity, and may contribute to more severe disease and delayed recovery during COVID-19. Consistent with this possibility, IFN-λ was significantly elevated in autoAb-high plasma at convalescence. Although early IFN-λ can have anti-viral protective effects, persistent or late IFN-λ during recovery has been reported to impair lung epithelial proliferation following respiratory viral infections⁷⁶. ACE2 shedding into blood has been observed during SARS-CoV-2 infection and has been associated with endothelial dysfunction⁷⁷. Elevated ACE2 into convalescence may reflect ongoing epithelial or endothelial perturbation, though it can not distinguish delayed tissue repair from persistent viral reservoirs. Together, these observations suggest that the heightened responses may be associated with impaired recovery and could contribute to long COVID pathogenesis.

Notably, DN2s expressed the highest IFNLR1 among circulating immune populations, with further upregulation in autoAb-high patients. Whether type III IFN signaling drives their differentiation and autoAb production remains to be determined.

The strong enrichment of autoimmune trait heritability in DN2s further supports the relevance of this population to autoimmune pathogenesis. Notably, 16.2% of SLE motif-binding variants in DN2s were accessible (contribution scores > 0.1), and most were predicted to disrupt motif binding for the risk allele, suggesting that autoimmune risk variants may frequently perturb transcriptional regulation in this population. Our analysis of rs11553760 in CSK, predicted to enhance T-bet binding in DN2s, illustrates how the sequence model resolves how individual risk variants reshape regulatory circuits in specific B cell populations. These analyses connect the DN2 program to established autoimmune risk and support an intrinsic predisposition, not only acute environmental triggers.

The transcriptomic, regulatory, and genetic features of COVID-19 DN2s closely resemble SLE DN2s: both express high T-bet, CD11c, SOX5, ZEB2, FCRL5^15,39 and share non-GC activation, autoreactivity, innate sensing, T-bet regulation, and autoimmune risk variants^15,39,64,65. T-bet/TLR7 involvement in DN2 biology is well documented in SLE. Our study extends this to acute SARS-CoV-2 infection, identifying an autoAb-associated DN2 program that overlaps with autoimmune DN2s in transcriptomics, gene regulation, and genetics. This connection links infection-induced autoAbs to established autoimmune B cell pathogenesis. Two differences nonetheless emerge. First, DN2 frequencies were similar between the autoAb groups and returned to baseline upon recovery, unlike SLE, where DN2s expand substantially and may persist despite treatment^15,39. Second, while elevated serum type I IFN and IFN-stimulated gene expression are prominent in SLE^39,78, we observed neither in acute autoAb-high DN2s, but instead elevated type III IFN signaling and persistently elevated ACE2 at convalescence, possibly reflecting prolonged tissue repair rather than a lupus-like IFN program. Thus COVID-19 and SLE DN2s may share core mechanisms of B cell tolerance breakdown and endosomal TLR signaling. In acute SARS-CoV-2 infection, this program may rapidly generate broadly reactive antibodies that complement antiviral defense, but may also come at the expense of virus-specific humoral immunity. If repeatedly driven by nucleic acid stimulation and unresolved, such responses could enter GC reactions and form durable memory, consistent with the enhanced differentiation and antibody production of autoAb-high SWMs in vitro, one mechanism by which SARS-CoV-2 infection may increase the risk of subsequent autoimmune diseases^12,14.

Following Eisenbarth et al.⁷⁹, we propose renaming AtM cells as primary-switched non-GC B cells (PSNs): primary (transcriptomically proximal to naïve B cells with low somatic hypermutation [7% vs. 15% in SWMs] indicating limited GC participation), switched (IgD⁻ phenotype), and non-GC (anatomical location not directly determined, favoring a process- over location-based definition). High T-bet expression and strict TLR7/8-dependent differentiation mark them as recently activated GC-independent effectors rather than classical memory. We note that a GC-independent origin is not unique to this population, as SWMs can also arise independently of the GC. We further designate the DN2 subset effector PSNs, given their poised differentiation into ASCs and CD11c expression⁷⁹.

The link between autoAb production, effector PSN cell activation, and autoimmune disease genetics suggests that a subset of long COVID cases may have an autoimmune component driven by dysregulated effector PSNs. Our findings nominate endosomal TLR signaling and the T-bet-associated regulatory pathways as candidate therapeutic axes, although direct perturbation studies will be required to establish causality. Together, the similar peripheral inflammatory and virology profiles between the autoAb groups, the persistence of autoAb-producing capacity into convalescence, and the enrichment of autoimmune heritability and predicted variant effects in PSNs support a response potentially encoded by stable epigenetic states and/or inherited genetic predispositions. Integrating genomic information with immune profiling may therefore help identify individuals at heightened risk for persistent autoimmune sequelae after SARS-CoV-2 infection.

Limitations of the Study

The relatively small sample size (n=12 and n=101 for the validatin cohort) of this nested case-control study may limit the generalizability of our findings. Additionally, tissue-specific analyses could further clarify the source of TLR7 stimulation and distinguish contributions from viral versus self-derived nucleic acids. Finally, this study did not sequence BCRs or label cells with viral or autoAb-specific tags, precluding analysis of the clonal evolution of antigen-specific B cells.

RESOURCE AVAILABILITY

Lead Contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Dr. James R. Heath (jim.heath@isbscience.org).

Materials availability

This study did not generate new unique reagents.

Data and Code Availability

• All multiome scRNA-seq and scATAC-seq data used in this study can be accessed by ArrayExpress. The accession number is listed in the Key resources table, and data are publicly available from the date of publication.

• This paper does not report original code. Custom scripts are available upon request.

• Any additional information required to reanalyze the data reported in this work is available from the Lead Contact upon request.

Key resources table.

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Anti-human CD66b-biotin	Miltenyi Biotech	Cat#130-118-983
BV421 anti-human CD19	Biolegend	Cat#302233
FITC anti-human CD3	Biolegend	Cat#317305
FITC anti-human CD14	Biolegend	Cat#325603
FITC anti-human CD16	Biolegend	Cat#302005
Alexa Fluor 488 anti-human CD56	Biolegend	Cat#318312
PerCP/Cy5.5 anti-human IgD	Biolegend	Cat#348208
APC anti-human CD38	Biolegend	Cat#356605
APC/Cy7 anti-human CD27	Biolegend	Cat#302815
Goat anti-human IgG for B cell stimulation	Jackson ImmunoResearch	Cat#109-066-097
Rabbit anti-IFN-α2	Sinobiological	Cat#13833-R001
Rabbit anti-SNRP70	Novus Biologicals	Cat#NBP1-57487
Anti-SARS-CoV-2 Spike S1 Antibody	Abcam	Cat#ab273073
Anti-Human IgG-Peroxidase Antibody	Sigma-Aldrich	Cat#A6029
Alexa Fluor 546 anti-Human IgG for HuProt	ThermoFisher	Cat#A-21089
Chemicals, peptides, and recombinant proteins
RPMI 1640 Medium	Gibco	Cat#11875093
RBC lysis buffer	Miltenyi Biotech	Cat#130-094-183
Pan B isolation kit	Miltenyi Biotech	Cat#130-101-638
Trizma Hydrochloride Solution pH 7.4	Sigma-Aldrich	Cat#T2194
Sodium Chloride Solution	Sigma-Aldrich	Cat#59222C
Magnesium Chloride Solution	Sigma-Aldrich	Cat#M1028
Tween-20 for nuclei lysis	BioRad	Cat#1662404
Nonidet P40 Substitute	Sigma-Aldrich	Cat#74385
10% BSA	Miltenyi Biotech	Cat#130-091-376
DTT	Sigma-Aldrich	Cat#646563
RNase Inhibitor	Sigma-Aldrich	Cat#3335402001
Digitonin	ThermoFisher	Cat#BN2006
Nuclease-free water	ThermoFisher	Cat#AM9937
Tween-20 for ELISA	ThermoFisher	Cat#28320
IFN-α2	Miltenyi Biotech	Cat#130-093-874
U1-snRNP	Diarect	Cat#A13000
Ribosomal Phosphoprotein P1	Diarect	Cat#A14200
Ro/SS-A	Diarect	Cat#A12700
La/SS-B	Diarect	Cat#A12800
Jo-1	Diarect	Cat#A12900
Centromere protein B	Diarect	Cat#A12500
SmD	Diarect	Cat#A11700
Scl-70	Diarect	Cat#A14500
Fc Receptor Blocking Solution	Biolegend	Cat#422302
Monocyte Blocker	Biolegend	Cat#426102
FBS for B cell culture	Cytiva	Cat#SH30070.03
β-mercaptoethanol	Invitrogen	Cat#21985023
HEPES	Invitrogen	Cat#15630080
Sodium pyruvate	Invitrogen	Cat#11360070
Penicillin-Streptomycin	Invitrogen	Cat#15070063
MEM non-essential amino acids	Invitrogen	Cat#11140050
GlutaMAX	Invitrogen	Cat#35050061
IL-2	PeproTech	Cat#200-02
IL-21	PeproTech	Cat#200-21
BAFF	PeproTech	Cat#310-13
R848	invivoGen	Cat#tlrl-r848
IFN-γ	Biolegend	Cat#570202
SARS-CoV-2 Spike Protein S1/S2	Invitrogen	Cat#RP-87680
TMB Peroxidase Substrate Kit	Seracare	Cat#5120-0047
PBS, 1X	Fisher Scientific	Cat#21-040-CV
Critical commercial assays
Chromium Next GEM Single Cell Multiome ATAC + Gene Expression Reagent Bundle	10X Genomics	1000285
Chromium Next GEM Chip J Single Cell Kit	10X Genomics	1000234
IgG Human Uncoated ELISA Kit	Invitrogen	39-50550-65
Deposited data
Processed scRNA-seq and scATAC-seq data	This paper	E-MTAB-13745
Supplemental Tables S1–S6	This paper; Mendeley Data	https://data.mendeley.com/drafts/866gsnhvx5
Software and algorithms
Cell Ranger ARC v1.0.1	10x Genomics	RRID:SCR_023897
FlowJo	FlowJo	RRID:SCR_008520
MIRA v1.0.4	Lynch et al.33	https://github.com/cistrome/MIRA
chromVAR v1.20.2	Schep et al.58	https://github.com/GreenleafLab/chromVAR; RRID:SCR_026570
TOBIAS v0.14.0	Bentsen et al.63	https://github.com/loosolab/TOBIAS
LDSC v1.0.1	Finucane et al.67	https://github.com/bulik/ldsc; RRID:SCR_022801
CREsted v1.5.0	Kempynck et al.71	https://github.com/aertslab/crested
MACS v2.2.6	Zhang et al.81	RRID:SCR_013291
Signac v1.10.0	Stuart et al.82	https://stuartlab.org/signac/; RRID:SCR_021158
GenomicRanges v1.50.2	Lawrence et al.83	https://github.com/Bioconductor/GenomicRanges; RRID:SCR_000025
SoupX v1.6.2	Young et al.84	https://github.com/constantAmateur/SoupX; RRID:SCR_019193
scDblFinder v1.12.0	Germain et al.85	https://github.com/plger/scDblFinder; RRID:SCR_022700
Amulet v1.1	Thibodeau et al.86	https://github.com/UcarLab/AMULET
GATK v4.1.9.0	Poplin et al.88	https://gatk.broadinstitute.org/hc/en-us; RRID:SCR_001876
vcftools v0.1.16	Danecek et al.89	https://vcftools.github.io/index.html; RRID:SCR_001235
bcftools v1.8	Danecek et al.90	https://samtools.github.io/bcftools/; RRID:SCR_005227
CrossMap v0.5.2	Zhao et al.91	https://github.com/liguowang/CrossMap; RRID:SCR_001173
VarTrix v1.1.22	10x Genomics	https://github.com/10XGenomics/vartrix
Vireo v0.5.0	Huang et al.92	https://github.com/single-cell-genetics/vireo; RRID:SCR_027463
Harmony v0.1.1	Korsunsky et al.95	https://github.com/immunogenomics/harmony; RRID:SCR_022206
leidenalg v0.9.0	Traag et al.96	https://github.com/vtraag/leidenalg; RRID:SCR_028421
UMAP v0.5.3	McInnes et al.97	RRID:SCR_018217
IgBLAST v1.16.0	Ye et al.99	https://ncbi.github.io/igblast/; RRID:SCR_002873
Seurat v4.3.0	Hao et al.100	https://satijalab.org/seurat/; RRID:SCR_016341
edgeR v3.40.2	Chen et al.101	https://bioconductor.org/packages/release/bioc/html/edgeR.html; RRID:SCR_012802
HOMER v4.11	Heinz et al103	http://homer.ucsd.edu/homer/; RRID:SCR_010881
Cicero v1.3.9	Pliner et al.104	https://github.com/cole-trapnell-lab/cicero-release
Igraph v1.5.0	Csárdi et al.105	https://igraph.org/; RRID:SCR_019225
SnapATAC2 v2.9.0.dev0	Zhang et al.106	https://scverse.org/SnapATAC2/; RRID:SCR_026622
FIMO MEME Suite 5.5.8	Grant et al.107	https://meme-suite.org/meme/doc/fimo.html; RRID:SCR_001783
Other
384-well plates	Thermo Fisher	Cat#464718
LS MACS columns	Miltenyi Biotech	Cat#130-042-401

STAR METHODS

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

Study participants

The study cohort is a subset of the INCOV cohort published previously^11,18. Procedures for the INCOV study were approved by the Institutional Review Board (IRB) at Providence St. Joseph Health with IRB study number STUDY2020000175 and the Western Institutional Review Board with IRB study number 20170658. In this study, PBMCs from participants were collected at enrollment close to diagnosis (Acute-1), 1 week (Acute-2), and 3 months following enrollment. Potential participants were identified at five hospitals of Swedish Medical Center and affiliated clinics located in the Puget Sound region near Seattle, WA. All enrolled patients provided written in-person informed consent. Individuals from the INCOV cohort were selected for study based on the following criteria: 1. High abundance of at least one autoAb (autoAb-high, z-score≥3.5) or low abundance of all autoAbs (autoAb-low, z-score<2) at initial diagnosis (Acute-1); 2. We limited to a single sex, all female or all male; 3. At similar ages. AutoAb-high individuals were selected with ages differing by less than 15 years. AutoAb-low individuals were selected to match the age of autoAb-high individuals. After applying the above selection criteria, we obtained a list of 12 patients who are all females and of similar ages (mean±SD: 75±7 and 74±9 years old, respectively). To extend the analysis to the full INCOV cohort when data is available, we applied the same autoAb filtering criteria but excluded the age and sex filters, identifying a cohort of 101 patients of autoAb-high (n=35, 63% females; mean±SD: 56±18 years old) and autoAb-low (n=66, 48% females; mean±SD: 55±17 years old) individuals. See Table S1 for patient metadata.

METHOD DETAILS

Single-cell multiome assay

PBMCs were isolated from whole blood and stored until use in liquid nitrogen as previously described^11,18. Chromium Single Cell Multiome ATAC and Gene Expression kits (10X Genomics) were utilized to analyze both the transcriptome and genome-wide chromatin accessibility from the same single cells. Briefly, cryopreserved PBMCs were thawed in RPMI with 10% FBS and resuspended in 1X red blood cell lysis solution (Miltenyi Biotech) to lyse residual red blood cells. Neutrophils were depleted and B cells were enriched from PBMC samples using CD66b-biotin antibodies and the B cell enrichment kit (Miltenyi Biotech). Enriched B cells were lysed in Lysis Buffer (10 mM Tris-HCl, 10 mM NaCl, 3 mM MgCl₂, 0.1% Tween-20, 0.1% NP-40, 0.01% digitonin, 1% BSA, 1 mM DTT and 1 U/μl RNase inhibitor in nuclease-free water) for three minutes on ice. Lysed nuclei were washed three times in Wash Buffer (10 mM Tris-HCl, 10 mM NaCl, 3 mM MgCl₂, 1% BSA, 0.1% Tween-20, 1 mM DTT and 1 U/μl RNase inhibitor in nuclease-free water) before being loaded onto the Chromium Next GEM Chip J. Gene expression and ATAC libraries were prepared according to manufacturer’s guidelines and sequenced on Nextseq 2000 (Illumina).

Human proteome-wise autoAb measurements

AutoAbs against human proteome were measured using the HuProt human proteome microarray (CDI Labs). Briefly, the microarrays were blocked using 5% BSA in 1X TBST buffer. Serum / plasma samples from the participants were diluted 1:1000 in blocking buffer and probed on the microarrays. After 1h incubation at RT with gentle shaking, microarrays were washed with 1X TBST three times for 10 min each. Anti-Human IgG was diluted 1:2000 in blocking buffer and was added to the microarrays. The microarrays were incubated for 1 hr with gentle shaking, and then washed with 1X TBST and dried in a centrifuge. The signal intensity of the protein spots was determined using the GenePix 4000B scanner.

Plasma proteomics

Plasma proteome was assayed using Olink’s Explore platform that uses high-sensitivity Proximity Extension Assay dual antibody technology with next-generation sequencing read-out to evaluate plasma abundance of K=3,072 proteins⁸⁰. Library preparation from 50 μl of plasma, sequencing (at ISB using Illumina’s NextSeq 2000, P2 flow cell), and protein abundance estimation were performed per manufacturer’s protocol. Samples underwent internal QC and quality assurance as part of the standard pipeline implemented in NPX Manager cloud software that utilizes a multi-indicator strategy. Six to eight pooled internal control samples were included on each plate to enable bridge normalization. Longitudinal samples were always plated together. Olink Explore provides estimates of normalized relative protein abundance on a log2 scale.

Whole genome sequencing

Whole genome sequencing (WGS) library preparation was performed on extracted genomic DNA using Illumina DNA PCR-Free Prep Tagmentation kits according to the manufacturer’s instructions. The resultant libraries were quantified by fluorometer and were pooled at equimolar concentration. Pooled libraries were loaded onto NovaSeq S4 (Illumina) flow cells and sequenced at 30x or higher coverage on a NovaSeq 6000 instrument. Following sequencing, data were transferred onto a DRAGEN Bio-IT platform (Illumina) for FASTQ generation, alignment, duplicate marking and variant calling using default flags. The resulting bam and gvcf files were used in subsequent analyses.

Plasma autoAb measurements

Plasma autoAb measurements were adopted from the protocol described previously¹¹. In brief, 384-well plates were coated with 2 μg/mL of recombinant IFN-α2 (Miltenyi Biotech), U1-snRNP (Diarect), P1 (Diarect), Ro/SS-A (Diarect), La/SS-B (Diarect), or Jo-1 (Diarect), followed by incubating with 1:50 dilutions of plasma samples in duplicates. In the ANA measurement, 384-well plates were coated with a mix containing 2 μg/mL of U1-snRNP, P1, Ro/SS-A, La/SS-B, Jo-1, CENP-B, SmD, and Scl-70. End-point OD at 450nm was measured and recorded.

In vitro stimulation and culture of B cells

PBMCs from acute infections were stained with the following antibody-dye conjugates and blockers before being sorted into 96-well cell culture plates using the BD FACSAria Fusion sorter at the Benaroya Research Institute: BV421-CD19, FITC-CD3, FITC-CD14, FITC-CD16, Alexa488-CD56, PerCP/Cy5.5-IgD, APC-CD38, APC/Cy7-CD27, Fc blockers, Monocyte blockers (all from Biolegend). Four B cell types were sorted, including ASCs (CD3⁻CD14⁻CD16⁻CD56⁻CD19⁺CD27⁺CD38⁺), Naïve (CD3⁻CD14⁻CD16⁻CD56⁻CD19⁺CD27⁻IgD⁺), SWM (CD3⁻CD14⁻CD16⁻CD56⁻CD19⁺CD27⁺IgD⁻), and AtMs (CD3⁻CD14⁻CD16⁻CD56⁻CD19⁺CD27⁻IgD⁻). Sorted cells were split into 2,000–3,000 cells per well and cultured in RPMI1640 with the following supplements (all from Invitrogen except noted separately): 10% FBS (Cytiva), 55 μM β-mercaptoethanol, 10 mM HEPES, 1 mM sodium pyruvate, 100U/ml penicillin/streptomycin, 1X MEM non-essential amino acids, and 1X GlutaMax. The following cytokines were added fresh into the medium: 50 ng/ml IL-2 (PeproTech), 10 ng/ml IL-21 (PeproTech), 10 ng/ml BAFF (PeproTech), 1 μg/ml R848 (invivoGen), 20 ng/ml IFN-γ (Biolegend), and 10 μg/ml anti-human IgG (Jackson ImmunoResearch, only for the first four days). Half of the supernatant was carefully removed and saved, and the same amount of fresh complete medium with cytokines was added to the cultures every 4 days. After 16 days of culture, supernatants were saved and cells were analyzed on the Attune flow cytometer (Thermo Fisher) to quantify the percentage of ASCs using FlowJo. A cell expansion score was calculated by dividing the cell number at the end of the culture by the starting cell number. ELISA for measuring total IgGs and IgGs against IFN-α2, ANAs and SARS-CoV-2 in the supernatants were performed. Detailed ELISA procedure is discussed below.

ELISA measurements of IgG in culture supernatants

Total IgGs from the culture were measured using the IgG Human Uncoated ELISA Kit (invitrogen) on a 384-well plate according to manufacturer’s instructions. Briefly, 384-well plates (ThermoFisher) were coated with 30 μL of capture antibody in Coating Buffer overnight at 4°C. Plates were washed twice with 120 μL of Wash Buffer (phosphate buffered saline (PBS) containing 0.05% Tween-20) and blocked with 75 μL Blocking Buffer for 2 hours at room temperature (RT). After blocking, wells were washed twice and incubated with 20 μL samples for 2 hours at RT. Human IgG standards at eleven serial three-fold dilutions starting from 1 μg/mL were used as positive controls. After incubation, wells were washed four times and incubated with 30 μL of Detection Antibody for 1 hour. Finally, wells were washed four times and incubated with 30 μL of Substrate Solution for 20 minutes at RT before quenching with 30 μL of 1M H₂SO₄. End-point OD at 450nm was measured on a Spectramax Plate Reader.

The same procedure was followed for measuring IgGs against IFN-α2, nuclear proteins and SARS-CoV-2 spike protein with the following modifications. For measuring IgGs against IFN-α2 and SARS-CoV-2, 384-well plates were coated with 30 μL of 10 μg/mL IFN-α2 (miltenyi) or SARS-CoV-2 Spike protein (invitrogen). For measuring IgGs against nuclear proteins, 384-well plates were coated with a mix containing 2 μg/mL of U1-snRNP, P1, Ro/SS-A, La/SS-B, Jo-1, CENP-B, SmD, and Scl-70 (all from Diarect). Wells were blocked with 5% BSA in Wash Buffer. Antibodies against IFN-α2 (Sinobiological), U1-snRNP (Novus Biologicals), and SARS-CoV-2 S1 (Abcam) at eleven serial three-fold dilutions starting from 2 μg/mL were used as positive controls. Peroxidase-conjugated goat anti-human IgGs (Sigma Aldrich, 1:1,000 dilution) were used as detection antibodies. 3,3’,5,5’-tetramethylbenzidine (TMB) (Seracare) was used as substrate solution.

QUANTIFICATION AND STATISTICAL ANALYSIS

Plasma proteomics analysis

The Wilcoxon rank-sum test was used to compute p-values comparing the autoAb-high and autoAb-low groups for each of the 2,941 measured plasma proteins. Proteins with a p-value < 0.05, a mean > 0 in the autoAb-high group, and a higher mean in the autoAb-high group compared to the autoAb-low group were used as input for GO enrichment analysis of Biological Process 2025. An additional filter, requiring the autoAb-high group to have a mean at least 0.5 greater than the autoAb-low group, was applied to the 12-cohort analysis. A total of 141 proteins meeting these criteria were identified in the 12-patient cohort, and 69 in the 101-patient cohort. All elevated proteins and their enriched pathways are listed in Table S2.

Human proteome-wise autoAb analysis

Raw signal data were extracted from .gpr files, and paired spots for each antigen were averaged. Signal intensities were quantile-normalized to ensure a consistent distribution across all samples and to minimize technical variation between arrays. The normalized data were used for downstream analyses, including unsupervised clustering of samples and antigens and comparison of autoAbs between the autoAb-high and autoAb-low groups.

Single-cell multiome data preprocessing

Multi-omic sequencing data were mapped and referenced against the GRCh38 human reference genome using the Cell Ranger ARC Software (10X Genomics). Cells from the same library were demultiplexed using WGS data as described in the next section. 151,786 cells were sequenced. Low-quality cells were removed by filtering out cells with low RNA counts (≤600) and low ATAC counts (≤800); high RNA counts (≥25,000) and high ATAC counts (≥100,000); high percentage of reads mapped to mitochondrial genome (≥20%); nucleosome signal≥2 and transcription start site enrichment ≤1.5 and ≥50. After the aforementioned QC filtering, ATAC peaks were called from each batch using MACS2⁸¹ with the CallPeaks function from Signac⁸². Called peaks were merged into a consensus peak list using the reduce function from GenomicRanges⁸³. RNA raw cell counts were corrected using SoupX to remove ambient RNA contamination⁸⁴. Cell doublets from each batch were identified through three methods: WGS-informed single nucleotide polymorphisms (SNP) doublets, scDblFinder for RNA doublets⁸⁵ and a coverage-based scoring for ATAC doublets⁸⁶. Doublets were visualized in 2-D UMAP and clusters with high average doublet scores calculated from any of the methods were removed accordingly. Non-B cell clusters were identified based on canonical B, T, NK, and myeloid cell marker expression and removed accordingly. After removing low-quality cells, residual non-B cells, and doublets, 96,551 B cells were retained for downstream analysis.

Single cell demultiplexing using genetic variants

We wrote a Snakemake⁸⁷ workflow (https://github.com/racng/snakemake-merge-wgs) for processing GVCF files from WGS to generate a multi-sample VCF file of exon variants. GVCFs are combined for specified genomic region (autosomes and sex chromosomes) using GATK GenomicsDBImport to generate a GenomicsDB datastore, which is then used by GenotypeGVCFs for joint calling of variants⁸⁸. After removing indels using vcftools⁸⁹ and excluding intron variants via bcftools⁹⁰, the remaining exon variants were lifted over to GRCh38 (hg38) using CrossMap⁹¹ and filtered again to remove KI27 contigs and duplicated variants.

By default, Cell Ranger ARC cell calling is capped at 20,000 cells per library. To increase the number of cells called, we defined thresholds for GEX UMIs and ATAC fragments based on their knee plots, and kept cells that are below either threshold. For each 10x library, GEX and ATAC BAM alignment files from Cell Ranger ARC were combined. Using the processed VCF file and modified cell calling, single cell variants were extracted from the combined BAM files via 10x Genomics VarTrix with the coverage scoring method. We then used vireo⁹² to assign donor identity and doublets based on the processed single-cell and WGS variants.

Single-cell multiome B cell type identification

Raw RNA and ATAC count matrices were used as input to train two independent MIRA topic models, yielding two separate latent spaces for each single cell³³. MIRA employs a deep variational autoencoder-based topic modeling framework that has been successfully applied to the integration and analysis of multimodal single-cell datasets owing to its advantages in nonlinear dimensionality reduction, denoising of sparse count data, and correction for covariate effects^33,93,94. The resulting MIRA latent spaces support a range of downstream applications, including inference of regulatory programs, pseudotime trajectory projection, and joint multi-omic visualization. The RNA and ATAC latent spaces were concatenated and corrected for batch effects using Harmony⁹⁵. A k-nearest neighbors (KNN) graph was constructed from the corrected joint latent spaces and was used for clustering using the Leiden algorithm⁹⁶ and further dimensionality reduction using UMAP⁹⁷. Cell clusters identified from the joint latent space were manually annotated into B cell types based on the top expressed genes in each cluster and canonical B cell markers. Memory B cells were further reclustered to define DN2 and DN3 populations based on the expression of CD27, CD21, TBX21, and ITGAX. The markers used for cell type annotation are shown in Figure 3C.

B cell lineage inference

B cell lineage was inferred using MIRA³³. Briefly, the KNN graph built from the joint MIRA latent space was used to construct a Markov chain model using the Palantir algorithm⁹⁸. Then, each cell was assigned probabilities of reaching all the possible terminal states. B cell lineages were then inferred using cell state tree structures resolved from these terminal state probabilities.

Somatic hypermutation analysis

Somatic hypermutation rates were calculated as the percentage of gaps and mismatches in the CDR1–3 regions of the heavy-chain sequence compared to the top germline sequence hit identified through IgBLAST⁹⁹.

Differential gene expression and enrichment analysis

Differentially expressed genes (DEGs) for B cell types were identified using the FindMarkers function from the Seurat package¹⁰⁰, with default settings except for the following parameters: logfc.threshold=1 and min.pct = 0.2.

DEG analysis between the autoAb-high and autoAb-low groups was performed for each B cell type using the edgeR package with a quasi likelihood F-test¹⁰¹. Pseudobulk gene expression profiles were generated by summing raw counts from all cells of each sample within a given cell type. These pseudobulk counts were fitted to a generalized linear model in edgeR using the glmQLFit function. The glmQLFTest function was then used to identify DEGs between the autoAb-high and autoAb-low groups for each cell type. DEGs with an absolute log2 fold change greater than 0.5 and a p-value less than 0.2 were retained for enrichment analysis. To identify enriched pathways in the DEGs between the autoAb groups in each cell type, the identified DEGs from both directions (up- and down- regulated) were used as input for the enrichR package¹⁰² to identify enriched REACTOME 2024 pathways. The list of DEGs and the enriched pathways are provided in Table S3.

Differential accessibility analysis

Term frequency-inverse document frequency (TF-IDF) normalization was performed in Signac on the ATAC count data that accounts for sequencing depth and gives high values to rare peaks⁸². Normalized ATAC counts were used as input to the Seurat function FindMarkers to identify cell type-specific differentially accessible regions with the following parameters: min.pct=0.01, logfc.threshold=0.25, and test.use = “wilcox”. Peaks with p-values <0.01 were kept for further analysis. Peaks specific to fewer than three groups were retained as cell type-specific peaks. The number of cell type-specific peaks identified for each cell type were as follows: 38,983 for ASC; 15,132 for DN2; 1,173 for DN3; 3,031 for int. Mem; 4,872 for naïve; 1676 for SWM; and 2,079 for USM. Peak annotation was performed using the program annotatePeaks.pl in HOMER¹⁰³.

TF motif enrichment and footprinting

Single-cell TF motif activity scores corrected for confounding factors such as GC content and Tn5 insertion were calculated for 879 TFs from the JASPAR 2024 database filtered for CORE collection and vertebrates using chromVAR⁵⁸ from Signac. ChromVAR evaluates enrichment of TF motifs in accessible chromatin regions in each cell, enabling the assessment of TF activity potential.

In Figure 4D, motif enrichment for each autoAb group and cell type in was calculated by averaging motif scores from up to 500 randomly sampled single cells per sample to account for cell frequency biases. The top 10 enriched TF motifs for each cell type and autoAb group were selected as columns.

The Footprint function in Signac was used to perform the footprinting analysis of TFs in Figures 5A, 5D, S5B, and S5E.

TOBIAS was used to investigate genome-wide TF binding. Pseudo-bulk ATAC BAM files were generated per B cell type using SAMtools⁹⁰, which were then corrected for Tn5 insertion bias using TOBIAS ATACorrect. The corrected single-basepair cut-site tracks were divided by the total genome-wide signal sum and multiplied by 10⁶ (CPM-style normalization). TOBIAS FootprintScores was applied to the normalized cut-site signals to calculate a continuous footprinting score across regions that represents protein binding. TOBIAS BINDetect was used to detect bound sites for T-bet across the genome for each B cell type. Finally, the TOBIAS PlotTracks command was used to plot normalized corrected genome tracks and footprinting scores at specific regions shown in Figure 5J.

Identification and clustering of gene-linked candidate cis-regulatory elements

Cell type-specific co-accessibility of peaks were calculated using Cicero¹⁰⁴. To identify cis-regulatory elements (cCREs) and their target genes, co-accessible peaks and peak-gene correlations were analyzed for each B cell type and autoAb group. First, co-accessible peaks were identified using Cicero, retaining those with co-accessibility values greater than 0.01. Pairs of co-accessible peaks where one peak was located in the promoter region of a gene (the target gene) were retained for further analysis. For all co-accessible peaks identified in each B cell type, sample-wise Pearson correlations were calculated between the expression of the target gene and the chromatin accessibility of the paired co-accessible peak. Co-accessible peaks with correlations in the 95^th percentile, a p-value < 0.01, and a distance greater than 10,000 base pairs were defined as cCRE-target gene pairs, focusing on distant interactions such as enhancer-promoter pairs. Accordingly, the identified cCREs were classified as gene-linked-cCREs (gl-cCREs).

To identify autoAb group- and cell type- specific gl-cCRE modules, TF-IDF-normalized scATAC-seq data were subset to the identified 12,089 gl-cCREs and pseudobulk accessibility was calculated by averaging the normalized accessibility for each B cell type and autoAb group combination. Non-negative matrix factorization was performed using the NMF R package to cluster gl-cCREs into 14 modules. The basis matrix (W; 12,089 gl-cCREs × 14 modules) represented gl-cCRE contributions to each module, and the coefficient matrix (H; 14 modules × 14 groups) indicated module activity per group. Each gl-cCRE was assigned to the module with the highest contribution score, and modules were mapped to specific B cell type and autoAb group combinations by identifying the group with the maximum coefficient in H. Results were visualized in a heatmap of z-score-normalized accessibility of module-assigned gl-cCREs, with rows ordered by group and columns by module in Figure 4G.

TF target score and regulatory network analysis

We identified target genes of TFs in each cell type. Promoter-linked target genes were identified as those whose promoters are accessible and contain the TF’s binding motifs. cCRE-linked target genes are identified as those whose promoter is co-accessible with that cCRE and the expression of that target gene is correlated with the accessibility of that cCRE. Single-cell target gene scores were calculated for each TF by averaging gene expression of the promoter-linked target genes subtracted by the expression of randomly sampled cells using the function AddModuleScore in Seurat. For constructing the TF regulatory network, the R package igraph was used for visualization¹⁰⁵. Only target genes that are TFs, cell type marker genes, or upregulated genes in the autoAb-high group compared to the autoAb-low group were shown in the network.

Cell type-specific linkage disequilibrium score

Each cell type-specific peak was extended by 2,000 base pairs (bp) in both directions to capture potential regulatory regions. These peak coordinates were saved as BED files and converted from hg38 to hg19 using the UCSC liftOver tool.

To run cell type-specific linkage disequilibrium score (LDSC) analysis, annotation files for each peak sets were generated using the make_annot.py with the 1000 Genomes European Phase 3 reference and a window size of 100k bp. LD scores were computed using ldsc.py with the --l2 flag, the same 1000 Genomes BIM file, a 1 cM LD window, and the --thin-annot flag to focus on the custom annotation. GWAS summary statistics for common traits were downloaded from https://alkesgroup.broadinstitute.org/sumstats_formatted/. Partitioned heritability for each GWAS trait was estimated using the ldsc.py script with the --h2-cts flag, baseline v1.2 annotations, a 1 cM LD window, and the --thin-annot flag to focus on the custom annotation. p-values were adjusted for multiple correction using Benjamini-Hochberg and reported in Figure 6A.

DNA sequence modeling and sequence scanning for motif matches

The peak regression model was trained using CRESted⁷¹. Peaks was called within each cell type using MACS3 and merged into uniform 500-bp regions using SnapATAC2¹⁰⁶. Cut-site Bigwig files for each cell type were exported using the snapatac2.export.export_coverage function with the following parameters: bin_size=10, normalization=“CPM”, and counting_strategy=“insertion”. The dataset was randomly split into 80% training, 10% testing, and 10% validation subsets. The model was initially trained on all scATAC-seq peaks, which were extended to 1500 bp in length, and subsequently fine-tuned using cell type-specific peaks with a Gini index greater than 1. Contribution scores for a given sequence were plotted using the function crested.pl.patterns.contribution_scores in Figure 6C.

GWAS summary statistics for systemic lupus erythematosus (SLE) were obtained from EMBL-EBI (accessed February 3, 2025). Contribution scores for SLE genetic variants, reflecting the impact of reference and alternative alleles on the chromatin accessibility of the surrounding 1,500-bp sequence, were calculated for each B cell type using the calculate_contribution_scores_sequence function in CRESted.

To identify TF motifs overlapping each variant and quantify motif binding strength, 15-bp sequences centered on the variant position for both the reference and alternative alleles were analyzed using the FIMO tool from the MEME Suite¹⁰⁷. Motifs were searched against the JASPAR 2024 Core Non-redundant Vertebrates database (879 motifs). A FIMO p-value less than 1 × 10⁻⁴ was applied to identify significant motif matches. For each matched motif for the reference allele, FIMO was rerun on the alternative allele sequence to calculate the change in motif binding strength, defined as the difference in negative log10-transformed p-values between the reference and alternative alleles. Motif matches, p-values, binding strength changes for each variant and each motif match, and new binding motifs for the alternative allele are reported in Table S6.

Polygenic risk score

We assessed whether a polygenic risk score (PGS)⁷³ for autoimmune disease was associated with both post-acute sequelae of SARS-CoV-2 infection (PASC) and elevated autoAb production. We leveraged the NIH All of Us (AoU) Research Program cohort, comprising over 250,000 participants with available WGS data and linked health records. We identified individuals within AoU who either had a recorded COVID-19 diagnosis or a positive SARS-CoV-2 test (n = 18,165). PASC phenotypes were derived by mapping ICD-10 codes from AoU medical records to the symptom-based PASC definition established by Thaweethai et al.¹⁰⁸. For each individual, we aggregated PASC-associated diagnoses recorded during the year preceding the pandemic and the year following COVID-19 infection. Individuals with newly emergent PASC symptoms post-COVID (n = 939) were included in the primary analysis. Autoimmune PGS values were computed with a published polygenic score for general autoimmune disease⁷³. Logistic regression models were constructed with PASC status as the dependent variable and the following independent variables: age, sex, body mass index (BMI), genetic ancestry principal components, clinical diagnosis of autoimmune disease (yes/no), and autoimmune PGS (standardized). To evaluate associations with autoAb positivity, we performed a parallel analysis using ANA test results available in AoU (n = 14,972 tested individuals, of whom 3,416 were ANA-positive, regardless of COVID-19 diagnosis). Logistic regression models for ANA positivity used the same covariates as in the PASC analysis. Odds ratios (OR) with 95% confidence intervals (CI) were calculated for both autoimmune diagnosis and PGS categories. Autoimmune disease and autoimmune PGS were treated as independent predictors in all models.

Supplementary Material

2

Table S1. Patient metadata. Related to Figure 1.

Detailed metadata including the timing of sample collection, demographics, Who Ordinal Scale (WHO) for disease severity, hospitalization/ICU status, COVID-19 treatments, comorbidities, and PASC symptoms. Statistical comparisons between the two autoAb groups for both the 12-patient and 101-patient cohorts across the aforementioned features are also reported.

3

Table S2. Plasma protein enrichment analysis. Related to Figure 1.

Differentially elevated proteins and their enrichment of Gene Ontology Biological Process 2025 terms, comparing the autoAb-high and autoAb-low groups in the 12-patient cohort and 101-patient cohorts.

4

Table S3. Differentially expressed gene analysis. Related to Figure 4.

Differentially expressed genes and their enrichment of Reactome 2024 pathways, comparing the autoAb-high and autoAb-low groups in the 12-patient cohort.

5

Table S4. TF motif enrichment analysis. Related to Figure 4.

TF motifs in Figure 4D (columns, left to right) and their corresponding TF family names.

6

Table S5. Gene-linked candidate cis-regulatory elements identified through integrated scATAC-seq and scRNA-seq analysis. Related to Figure 4.

Details of each gene-linked candidate cis-regulatory element identified, including its genomic location, the promoter it is co-accessible with, the gene it is linked to, and its NMF module assignment.

7

Table S6. SLE GWAS variants with overlapping TF binding motifs. Related to Figure 6.

Details for all SLE GWAS variants with overlapping TF binding motifs, including binding strengths for reference and risk alleles, newly created motif binding sites upon substituting with risk alleles, and cell type-specific contribution scores.

Highlights.

• AutoAb levels track atypical memory B cell (AtM) expansion and decline upon recovery

• AtMs are precursors of autoAb-producing cells in COVID-19

• DN2 B cells drive autoAb production via TLR7/8 signaling, regulated by T-bet and XBP1

• COVID-19 DN2s resemble lupus DN2s and carry the top autoimmune risk variant signal