Tumor-Infiltrating Plasma Cell Profiling after PD-1 Blockade Reveals Tumor-Specific Antibodies

Summary

The role of tumor-infiltrating B cells (TIL-Bs) in shaping anti-tumor responses in the context of immune checkpoint blockade remains incompletely understood. Here, we interrogate the humoral response in resected lung tumors from patients with non-small cell lung cancer (NSCLC) treated with neoadjuvant PD-1 blockade. We find that tumors orchestrate tertiary lymphoid structures with CD138⁺ plasma cells, from which we clone recombinant monoclonal antibodies (mAbs) using B cell receptors (BCR) exhibiting somatic hypermutation and class-switching. Several mAbs bind cell surface citrullinated proteins, characteristic of cancer cells. CAR T redirected with the scFv of our lead candidate antibody (PC-1), specifically target tumor cells and tumor-promoting myeloid cells in vivo without off-target activity. Moreover, ablation of the citrullination enzyme PADI2 in tumor-bearing mice eliminates reactivity to PC-1 and cytotoxic killing by the CAR. Our results implicate a therapeutic potential for tumor-infiltrating plasma cells that may be harnessed for cancer treatment.

eTOC Blurb

Meyerhoff et al. identify tumor-infiltrating plasma cells as a source of antibodies with therapeutic potential. Their lead antibody, PC-1, recognizes citrullinated antigens on cancer and tumor-promoting myeloid cells, and can be repurposed to guide CAR T-cell therapy.

Graphical Abstract

Introduction

There is growing evidence demonstrating the importance of tumor infiltrating B (TIL-B) cells in both anti-tumor immune responses and hindrance of tumor progression. To do so, TIL-Bs serve diverse roles in anti-tumor immunity including presenting antigen to T cells ^1,2, regulation of T cell function ^2–4, and secretion of anti-tumor antibodies ⁵. Moreover, TIL-Bs are associated with improved survival in many cancers including soft tissue sarcoma, pancreatic ductal adenocarcinoma, and non-small cell lung cancer (NSCLC) ^6–12. Whereas the canonical mechanism of action for blocking the programmed cell death-1 (PD-1) immune checkpoint is facilitation of anti-tumor T cell response ^13–15, there is emerging evidence that programmed cell death ligand-1 (PD-L1) blockade influences B cell responses as well ^9,16,17. Indeed, trials evaluating neoadjuvant PD-1 blockade report tertiary lymphoid structure (TLS) maturation and abundance in resection specimens from many patients with non-small cell lung cancer (NSCLC) treated with neoadjuvant chemotherapy and PD-1 blockade compared to untreated or those treated with chemotherapy alone ^18,19. In these studies, increased TLS abundance correlated with major pathologic response and disease-free survival ^18,19. Furthermore, increased frequency of intratumoral CD138⁺ plasma cells has been shown to correlate with improved clinical outcomes after PD-1 blockade in NSCLC ¹⁰. Similarly, our group has also observed plasma cell infiltrates in resection specimens from patients treated with neo-adjuvant PD-1 blockade ²⁰. However, the anti-tumor roles of infiltrating plasma cells in NSCLC patients are largely unknown.

Advancements in single cell RNA sequencing (scRNA-seq) technologies have permitted high-resolution of B cell populations in the tumor microenvironment (TME). It is now known that TIL-Bs span several major lineages and functional states including naïve B, activated B, germinal center B, memory B, atypical memory B, and antibody secreting plasma B cells (ASCs) ^2,4,21. As such, the diversity and functions of TIL-Bs are increasingly understood. However, despite recent efforts to generate atlases of intratumoral B cells, tumor antigen specificities of B cell receptors (BCRs) across these TIL-B populations are less well investigated ^2,4,21.

Chimeric antigen receptor (CAR) T-cell therapy is being actively investigated as a strategy for the treatment of solid tumors, including non-small cell lung cancer (NSCLC), as durable responses to immune checkpoint blockade are limited to a subset of patients. By redirecting autologous T cells against tumor-associated surface antigens such as EGFR or mesothelin, CAR T-cell platforms aim to bypass endogenous antigen presentation and mediate direct cytotoxicity against malignant cells ²². Multiple biological and clinical constraints, however, remain unresolved. A key challenge is the identification of antigens with sufficient tumor restriction to minimize on-target, off-tumor toxicity. In addition, immunosuppression driven by myeloid cells and fibroblasts presents a substantial barrier to effective intratumoral T-cell accumulation, effector activity and long-term persistence.

The observation that anti-PD-1 treated tumors have an induced accumulation of plasma cells suggests that these may be enriched sources of tumor cell-specific antibodies that can be mined for development of tumor therapeutics. In this study we utilized tumor samples from patients with NSCLC treated with neoadjuvant PD-1 blockade ^20,23. To define TIL-B BCR reactivity for tumor antigens, we generated recombinant monoclonal antibodies (mAbs) based on sequenced TIL-B BCR genes that were identified among tumor-infiltrating plasma cells that we studied using a high-throughput technology that enabled transcriptome sequencing and full-length variable region of paired heavy and light chain genes for the BCR per individual cell. Here, we were able to use the transcriptome data to define TIL-B populations, including plasma cells, and select BCR sequences from these different TIL-B populations for further functional characterization of tumor reactivity and anti-tumor therapeutic potential. We show that these antibodies can recognize tumor antigens that are uniquely expressed on the cell surface following post-translational modification. Furthermore, we develop a CAR T construct bearing the BCR of our lead candidate antibody (PC-1) and show that this construct killed tumor cells and tumor-promoting myeloid cells without deleterious off-target effects, thus providing robust proof-of-concept data supporting a role for mining tumor antibody repertoires for development of novel tumor therapeutics.

Results

Plasma cell accumulation in NSCLC tumors following PD-1 blockade

The previously reported TOP1501 clinical trial (NCT02818920) was a single-arm phase II study evaluating pembrolizumab in patients with clinical stage IB, II, or IIIA NSCLC prior to surgical resection, demonstrating not only safety and feasibility, but promising major pathologic response, progression free survival, and overall survival as previously reported ^20,23. From surgically resected lung tumors of patients in this study which had sufficient banked sample for evaluation, we identified five post-resection tumor specimens in which TIL-Bs were present on flow cytometric analyses ²⁴ and had sufficient banked samples for study (either paraffin-embedded tumor (patients 001, 004, and 006) or disaggregated tumors (patients 004, 201, and 301)). The clinical characteristics and treatment outcomes for the 5 patients whose tumor specimens were utilized are summarized in Table S1.

TIL-Bs have been associated with improved survival in many cancers, including non-small cell lung cancer (NSCLC) ^6–12, and because PD-1 blockade influences B cell responses ^9,16,17, we first sought to determine how PD-1 blockade might be augmenting the intratumoral humoral immune response in NSCLC.

TLS formation in NSCLC tumors following PD-1 blockade

Using CO-Detection by indEXing (CODEX) multiplex immunofluorescence assay, we investigated TLS density and immune cell composition in post-treatment NSCLC resection specimens of three patients with paraffin-embedded tissue available for study (001, 004, and 006) including a pre-treatment core needle biopsy specimen from patient 004. A list of cell-surface markers is shown for reference in Table S2. Compared to patient 004 pre-treatment biopsy, the resected tumor post treatment for each of the three patients (001, 004 and 006) contains numerous TLSs, including mature TLS containing germinal center CD4⁺ T cells and Ki67⁺CD20⁺ B cells surrounded by CD138⁺CD38⁺ ASCs (Figures 1A–E), which suggests cross-talk between T and B cells in intratumoral TLS and subsequent TIL-B diversification and ASC generation. We additionally observed prevalent accumulation of CD138⁺CD38⁺ ASCs throughout the tumor stroma and parenchyma (Figures 1A–D), thus implicating a role for these cells in the TME that is understudied.

Figure 1: PD-1 blockade is associated with mature TLS and ASC accumulation.

(A–D) Representative CODEX images of NSCLC tumors after neoadjuvant PD-1 blockade. Areas of plasma cell infiltration are marked by white arrows and distinct TLSs or lymphoid aggregates are marked by white asterisks. CK5 tumor marker is shown in cyan. (E) Representative mature TLS containing CD4⁺ T cells, Ki67⁺CD20⁺ B cells, and CD138⁺CD38⁺ ASCs. (F) Ki67 expression across immature, quiescent mature, and proliferating mature TLSs. Nuclei are stained with DAPI (blue). (G) Distribution of TLS maturation states across tumor specimens. See Tables S2, S3 and Figure S1.

Thus, we next sought to further interrogate the cell compositions and maturation states of TLSs throughout the resected tumor specimens. Following cell-type neighborhood analysis, cells within TLSs were grouped into spatially coherent patches and TLS maturity defined by B cell enrichment and Ki67 expression as a marker for proliferative activity, thus segregating TLS into three separate groups (immature, quiescent mature, and proliferating mature TLS) (Figure 1 F–G, Figures S1A–H). As expected, increased Ki67 expression was seen in proliferating mature TLSs as compared to quiescent mature and immature TLSs (Figure 1F, Table S3). There was little difference between TLS maturation states between specimens studied (Figure 1G) nor was there association with pathologic response rates (Figure SI–J), possibly related to sampling variability of tissue available for study from surgical resection specimens. Taken together these findings suggest that the TME in the context of PD-1 blockade may promote an environment that supports tumor B cell development.

Single cell profiling of total TIL-B cell populations in NSCLC

As we found evidence of mature TLSs in resected NSCLCs, presumably a requisite of antigen-driven positive selection and expansion of affinity-matured B cell clones, we next sought to define the phenotypic differentiation of TIL-Bs in the TME. Using available disaggregated tumor samples from the TOP1501 study, we performed single cell transcriptome sequencing from CD138- (patient 004 and 301) and CD19- enriched (patient 201) NSCLC TIL-Bs and combined these datasets for analysis (Figure 2A). We studied 7,150 quality-filtered TIL-Bs across 25 clusters that were annotated using the Azimuth tonsil reference dataset ²⁵ (Figures S2A–2C). To investigate the presence of atypical memory B cells in our dataset, we assessed for ITGAX (CD11c), a marker found within but not exclusive to this population ^26,27. We identified seven major B cell clusters (naïve-like B, activated naïve B, cycling B, germinal center (GC) B, memory B (MBC), atypical memory (AtM) B, and ASCs) that also showed corresponding gene expression profiles for these populations in previously published studies ^2,4,21 (Figure S2D and S2E; see methods), including enrichment for TCL1A, YBX3, FCER, and IGHD expression in naïve-like B cells, CD83, MIR155HG, NFKB1, CD69, and CD83 for activated naïve B cells, MKI67, STMN1, and HMGB2 for cycling B cells, BCL6 and AICDA for GC B cells, CD27 and TNFRSF13B for MBCs, DUSP4, FCLR5, ZEB2 and ITGAX (CD11c) for AtM B cells, and MYDGF, JCHAIN, XBP1, and MZB1 enriched in ASCs (Figure S2D and SE). Next, we used gene set-based analyses to further validate molecular phenotyping of our TIL-B annotations, demonstrating high expression of genes associated with canonical B cell populations such as naïve-like B, GC B, atypical memory B, and plasma cell/ASCs (Figure 2C) ². A full list of genes used in the gene set-based analysis is shown in Table S4. Given the smaller number of TIL-Bs in our dataset, we did not attempt to further subclassify these TIL-B populations (e.g. sub-populations of activated naïve, GC, MBC, and ASCs) as has been done in previous studies using larger datasets ^2,4,21

Figure 2: Single-cell profiling identifies diverse TIL-B populations in NSCLC.

(A) Workflow for single-cell transcriptome and paired BCR profiling of NSCLC TIL-Bs. (B) UMAP of annotated TIL-B populations. (C) Heatmap for gene set-based validation of major B cell states. (D) Immunoglobulin isotype distribution across annotated NSCLC TIL-B populations. (E) Somatic hypermutation frequency by annotated NSCLC TIL-B population. (F) Variable heavy and light chain gene usage across annotated NSCLC TIL-B populations. (G) Pictorial summarization of produced mAbs which contain shared clonal lineages across TIL-B populations. The top bold number indicates the number of sequences present in that population and the bottom number indicates average percent VH mutation frequency for those sequences. The asterisk denotes the PC-1 lineage from which the PC-1 mAb was derived and developed into a CAR T in subsequent figures See also Table S4 and Figure S2.

We next combined computationally reconstructed BCR sequences with the annotated scRNAseq data yielding 5,281 functional heavy and light chain pairs for analysis. The naïve-like B cell population was dominated by IgM and IgD isotypes whereas GC, MBC, AtM, and ASC populations exhibited a variety of isotypes indicative of class-switching in these populations (Figure 2D). Somatic hypermutation (SHM) was determined for each TIL-B population, demonstrating decreased SHM frequency (average) in the naïve-like B cell population and higher rates of SHM in the GC, MBC, and ASC populations (Figure 2E). There was no major difference in variable heavy, kappa, or lambda gene family usage amongst TIL-B populations (Figure S2F). Similarly, no significant differences were observed in variable heavy and light chain use across TIL-B populations (Figure S2F). We next assessed for evidence of shared clonal lineages across TIL-B populations identifying 39 shared clonal lineages (12 of which with BCRs expressed as recombinant monoclonal antibodies), thus highlighting evidence of TIL-B differentiation and clonal expansion within NSCLC tumors (Figure 2G).

Tumor reactivity of antibodies derived from different TIL-B cell populations

Given previous observations for expansion of intratumoral ASCs following PD-1 blockade in NSCLC ²⁸, we sought to investigate the immunogenetic hallmarks of antigen-experience tumor-reactive ASCs. To do so, disaggregated tumors were enriched for CD138⁺ B cells for subsequent BCR sequencing of paired heavy (V_H) and light (V_κ/V_λ) chain genes as defined in the methods. We obtained 1,353 unique BCR clonotypes for ASC plasma cells from patient 004, and 1,671 from patient 301. In both patients we observed clonal expansion of TIL-Bs as evidenced by the presence of more than one member per clonotype (Figure 3A, Figure S2G). Variable heavy (V_H3), variable kappa (V_κ1), and variable lambda (V_λ3) genes families were the most common amongst TIL-B ASCs (Figure S2H), and tumor-derived V_H, V_κ, and V_λ genes were mutated (Figure S2I), suggestive of antigen-driven maturation of these ASCs. Additionally, tumor infiltrating ASCs were primarily class switched; IgG1 was the most frequent isotype (60.5% and 43.3%) followed by IgA1 (23.8% and 13.5%) for patient 004 and 301, respectively (Figure S2J).

Figure 3: Tumor-reactive antibodies arise from multiple antigen-experienced TIL-B states.

(A) Expanded ASC clonotypes from CellRanger analysis were selected for recombinant antibody production. Each colored slice represents a single clonal family. White slices represent all unique (n=1) clonotypes. (B) Heavy and light chain mutation frequencies and (C) HCDR3 lengths of produced ASC-derived antibodies. (D) Isotype distribution of produced ASC-derived antibodies. (E) Heavy and light chain usage of produced ASC-derived antibodies. (F) Cell-surface binding of produced ASC-derived antibodies to Calu-6 cells. Cell surface antibody binding detected with AF488 anti-human IgG (green). Staining for cell membrane (Cell Mask) is in red and nuclei (DAPI) in blue. (G) Flowchart describing workflow for selecting TIL-B populations for additional recombinant antibody production. (H) Mutation frequencies, (I) HCDR3 lengths, and (J) isotype distribution of antibodies from produced antigen-experienced TIL-B populations. (K) Heavy and light chain gene usage of produced antibodies from antigen-experienced TIL-B populations. (L) Cell-surface binding of produced recombinant antibodies to A549 cells. Images for cell surface binding are representative of at least duplicate experiments. See Figures S3–4.

We next sought to define tumor reactivity of antibodies made by TIL-Bs ASCs. To do so, we produced 15 recombinant mAbs from patient 004 and 301 that were representative of the most expanded TIL-B ASC clonotypes (Figure 3A, Figure S3A). Mean percent V_H and V_κ/_λ mutation frequencies, and HCDR3 length was 5.3%, 3.5%, and 15 amino acids, respectively (Figures 3B–C). These B cells were all class-switched (IgG1 the most common) (Figure 3D) with varying heavy and light chain gene usage, summarized in Figure S3A. Each antibody was assessed for the ability to bind to cell surface-expressed antigens using immunofluorescent staining of the immortalized Calu-6 NSCLC cell line (Figure 3E, Figure S3B). In total, 11/15 (73%) of these mAbs demonstrated cell surface antigen reactivity of varying staining intensity.

By studying antibodies from only the most expanded TIL-B ASC clonotypes, we reasoned that we may miss potent anti-tumor antibodies from other TIL-B populations, including those that have not yet expanded. To assess tumor antigen reactivity present in other TIL-B populations identified in our transcriptome analysis, we next generated 17 recombinant mAbs derived from GC, MBC, and AtM TIL-B populations in addition to another 12 recombinant mAbs from less clonally expanded but either extensively mutated or highly diversified ASC clonal lineages, as summarized in Figure 3F. Immunogenetics for the 29 total mAbs expressed across TIL-B populations are summarized in Figure S4A. Mean percent V_H and V_κ/_λ mutation frequencies, and HCDR3 length for the produced mAbs were 8%, 5%, and 16 amino acids, respectively (Figures 3G–H). These B cells were all class-switched (IgG1 the most common) (Figure 3H) with varying heavy and light chain gene usage, summarized in Figure 3I. We next tested each antibody for cell surface-antigen reactivity using immunofluorescent staining of the immortalized A549 NSCLC cell line (Figure 3J, Figure S4B). In total 15/29 mAbs (52%) demonstrated cell surface antigen reactivity, thus indicating that tumor antigen reactivity is found across antigen-experienced TIL-B populations including germinal center, memory, and atypical memory B cell populations, in addition to ASCs.

Of the 29 mAbs studied for cell surface antigen-reactivity, PC-1 demonstrated the most intense cell surface reactivity (Figure 3F). Moreover, PC-1 was one of the most clonally expanded lineages in the ASC repertoire bearing 107 recovered members. These data suggest remarkable maturation of an ASC lineage to tumor antigens in this representative antibody lineage; thus we chose PC-1 for further workup as the lead tumor-reactive antibody in our study.

Citrullinated vimentin is a common target of tumor antibodies

Since PC-1 mAb demonstrated robust cell surface binding to the Calu-6 NSCLC cell line, we decided to investigate this antibody further for reactivity to cell surface antigens expressed by other cancer cell lines of diverse anatomic origin. We found the PC-1 bound to antigen or antigens present on the cell surface of tumor cells of diverse anatomic origin, lung, mesothelium, ovary, and uterus (Figure 4A), suggesting that PC-1 recognizes a tumor antigen shared amongst tumor cells. To determine the antigen(s) bound by PC-1, we performed immunoprecipitation (IP) of cell line lysates followed by mass spectrometry to identify a predominant 47 kDa band seen on SDS-PAGE/Western blot among, identifying 1,885 candidate proteins (Table S5). Following down-selection of candidate proteins bound by PC-1 (see methods and Figure 4B), we identified vimentin as a top candidate antigen bound by PC-1 mAb. Although vimentin is a cytoplasmic structural protein, it is known to be expressed on the cell surface in its citrullinated form, a post-translational modification known to provide a source of tumor antigens ^29–31. As such, we demonstrated binding of PC-1 mAb to recombinant citrullinated vimentin but not unmodified vimentin (Figure 4C), in addition to other candidate targets identified in the mass spectrometry data, such as citrullinated and unmodified Hsp90, another known tumor antigen with roles in tumorigenesis and metastatic potential ^32,33, citrullinated and unmodified-calreticulin that is known to be expressed on the surface of tumor cells following T cell injury and immunogenic cell death (ICD) ^34–37, and the tumor suppressor gene p53. Curiously, several tumor mAbs tested demonstrated varying levels of reactivity to citrullinated calreticulin, suggesting this may be a common tumor antigen (Figure 4C). PC-1 bound with nanomolar affinity to citrullinated vimentin (K_D 1.93 × 10⁻⁸ M), calreticulin (K_D 1.89 × 10⁻⁸ M), and p53 (K_D 1.36 × 10⁻⁸ M), thus confirming high affinity of this antibody for these antigens (Figure 4D).

Figure 4: PC-1 binds citrullinated vimentin, a ubiquitous cell surface-expressed tumor antigen.

(A) PC-1 binding by FACS to the cell surface of immortalized tumor cells of diverse origins. (B) Antigen discovery workflow using immunoprecipitation, mass spectrometry, and criteria for candidate prioritization. (C) Heatmap for ELISA binding of PC-1 and other tumor-derived antibodies to candidate antigens. (D) Surface plasmon resonance analysis of PC-1 binding to candidate antigens. Table listing kinetics and affinity of PC-1 binding to antigens shown below. (E) Binding of PC-1 and additional tumor-derived antibodies to cyclic citrullinated peptides in ELISA. (F) PC-1 autoreactivity testing against clinically relevant nuclear autoantigens in a multiplex immunofluorescence assay. 4E10, a known poly/autoreactive anti-HIV-1 mAb was included as a positive control. CH65, an anti-influenza hemagglutinin mAb was included as a negative control. (G) Assessment of PC-1 binding polyspecificity against a panel of 23,004 human proteins in the HuProt microarray (CDI labs). (H) Alignment of PC-1 germline-reverted and consensus “mature”. (I) Binding of germline-reverted PC-1 and “mature” PC-1 to citrullinated tumor antigens in ELISA. Binding data are representative of at least duplicate experiments. See Tables S5–6, and Figures S5.

We next sought to determine if binding to citrullinated antigens is prevalent in serum antibodies from all the patients from the TOP1501 clinical trial and from whom we obtained samples for our study outlined above. We found no significant difference in plasma antibody binding to these antigens compared to normal donor controls, nor boost of binding titers to these antigens following administration of systemic pembrolizumab (Figure S5A), suggesting that these responses may be localized to the TME. Lastly, there was also no correlation of pathologic response with plasma antibody binding to these antigens (Figure S5B).

Ontogeny of tumor-reactive antibodies

Given prior reports of autoreactive antibodies binding citrullinated peptides ^31,38, we tested PC-1 plus nine additional antibodies isolated from patient 004 for binding to cyclic citrullinated peptides. None of these antibodies bound cyclic citrullinated peptides, suggesting that PC-1 likely represents a different class of antibody than what is associated with rheumatoid arthritis (Figure 4E). To further exclude pathologic autoreactivity, we tested PC-1 mAb against a panel of nuclear autoantigens associated with various autoimmune diseases including Systemic lupus erythematosus, Sjogren’s syndrome, and Scleroderma amongst others in a clinically-validated AtheNA-Multi-Lyte assay (Athena Diagnostics) which has been used previously to define monoclonal antibody autoreactivity ^39–41. PC-1 did not demonstrate autoreactivity to these antigens, thus further suggesting lack of pathologic autoreactivity by this antibody (Figure 4F). Additionally, we investigated chart records of the TOP1501 subjects included in this study for pre-existing of or new manifestations of autoimmune diseases (Table S1). Specifically, subject 004 from which the PC-1 mAb was isolated from, did not have any clinical manifestations of rheumatoid arthritis either before or after PD-1 blockade, however, did develop clinical symptoms suggestive of pneumonitis/bronchitis after the third dose of pembrolizumab which resolved following treatment with corticosteroids (Table S1). Polyreactivity of PC-1 was further assessed in a microarray comprised of anatomically diverse normal tissues from 49 human donors demonstrating intense cell-surface binding to thyroid epithelial cells and variability of binding to cytosolic antigens observed for other tissue types (Figure S5C–D). To address the issue for potential broad polyreactivity, PC-1 was additionally tested at a concentration of 1 μg/mL against a panel of 23,004 proteins and isoform variants in the HuProt microarray (CDI Labs), demonstrating binding above cut-off (Z score > 3) in only 0.6% of proteins (Figure 4G, Table S6). Additionally, this assay also identified binding to p53 and Hsp90a proteins as discovered from IP-mass spec screenings described earlier (Figure 4G).

Additionally, it is known that tumor antibodies can arise from auto and/or polyreactive precursors ⁵. To determine if the demonstrated reactivity by PC-1 was germline-encoded, we generated a germline-reverted PC-1 and tested for binding to unmodified and citrullinated vimentin, p53, calreticulin, and Hsp90a (Figures 4H). We found that germline PC-1 mAb had low level binding to the citrullinated antigens tested in contrast with the mature PC-1 mAb in agreement with the mature antibody having more somatic hypermutations (Figure 4I) that is associated with increased affinity. These data also suggested that PC-1 Ab lineage affinity matured for tumor reactivity in the TME.

Determination of the epitope and structure of PC-1

We next sought to define the epitope of the PC-1 mAb on calreticulin and p53, both multidomain proteins (Figure 5A) ^42,43 bound by PC-1 with high affinity. PC-1 bound the p53 DNA-binding core domain, although with substantially lower affinity than full-length p53 (0.625–10 μM affinity versus 1–16 nM) (Figure 5B), suggesting that the core domain may be a key component of the PC-1 epitope, or that there could be conformational variability in the core domain expressed on its own as opposed to the full-length p53 protein. CALR consists of three primary domains: N, P, and C, with each domain roughly equal in size ⁴². PC-1 bound the NP construct but not the NC construct (Figure 5B). PC-1 bound the CALR-NP domain with similar affinity to full-length CALR (108–863 nM versus 4–64 nM), suggesting that P domain was the primary target of the PC-1 antibody. Binding to the P domain was observed, supporting the P domain as the primary site of recognition, however solubility and aggregation issues prevented binding kinetics analyses. Quality control data of proteins produced for kinetic studies are shown in Figure S6A.

Figure 5. Structure of the PC1 Fab and the location of its epitope on p53 and Calreticulin.

(A) Domain architecture of p53 and calreticulin constructs used for binding studies. (B) Surface plasmon resonance analysis of PC-1 binding to p53 and calreticulin domains. Table listing kinetics and affinity of PC-1 binding to antigens shown below. (C) Top: Cartoon representation of the PC-1 fab domain, with heavy and light chains colored green and blue respectively. CDR loops are colored pink, blue, and red for loops one, two, and three, respectively. Bottom: Cartoon representation of the paratope region of PC1 demonstrating presence of multiple aromatic and polar residues line the putative paratope region. See Table S7 and Figure S6.

We further solved the ligand free or apo PC-1 Fab crystal structure to 2.2 Å resolution (Table S7). The PC1 HCDR2 and HCDR3 are 17 and 15 amino acids in length, respectively, and the LCDR1 loop is 17 amino acids in length. That the CDRs are extended in length and incorporate bulky hydrophobic and polar residues in the light chain (Y31, Y98, H98, Y102) and the heavy chain (W33, W110) (Figure 5C), are consistent with a binding surface capable of accommodating structurally diverse epitopes. We attempted to co-crystallize the Fab domain of PC-1 with the epitope domains determined from the binding data, but unfortunately, this proved difficult. We then turned to, computational models using AbEpiTope-1.0 ⁴⁴ of PC-1 bound to the calreticulin P domain or p53 core domain predicted antigen-Fab contacts and potential shared paratope features involved in antigen recognition (Figure S6B–D). Specifically, the P-domain prediction shows extensive hydrogen bonding interactions between the Fab and the P-domain. Of note is a bidentate interaction between D55 and D57 of the heavy chain CDR2 and K248 of the P-domain. The p53 core and the Fab also share extensive interactions. There are two residues: R59 of the heavy chain, and Y98 of the light chain that, in our models, show binding interactions with both the P-domain and the p53 core domain. Although antigen-bound structures will be required for definitive epitope assignment, our data raised the hypothesis that PC-1 recognizes conformational features shared across distinct tumor-associated antigens.

Increased expression of citrullinated vimentin in human tumor cells and tumor-promoting cells within the tumor microenvironment

Since PC-1 has high affinity for citrullinated vimentin and other citrullinated proteins that are known to be expressed on the surface of tumor cells ^29–31,45, we next tested whether PC-1 would bind the surface of patient-derived tumor cells. Indeed, PC-1 bound to freshly disaggregated patient-derived tumor cells from patients with high grade serous ovarian carcinoma (HGSOC) but not paired normal tissue (Figure 6A, Figure S7A). Additionally, PC-1 was found to bind to tumor promoting macrophages, myeloid-derived suppressor cells (MDSCs) and exhausted CD4⁺ T cells present in HGSOC tumors and similarly, tumor promoting cells (including cancer-associated fibroblasts (CAF)) in NSCLC malignant effusions (Figure 6A–B, Figure S7A–B). Of note, PC-1 did not react with immune cells from healthy donor PBMCs (Figure 6C, Figure S7A). In summary, these data demonstrate the ability of PC-1 to recognize patient-derived tumor cells and tumor-promoting myeloid cells present in the TME, thus highlighting the potential of this antibody specificity to be exploited for therapeutic development.

Figure 6. PC-1 binds patient-derived tumor cells and tumor-promoting myeloid cells.

(A) Flow cytometric analysis of PC-1 binding to tumor cells and tumor associated T cells from freshly disaggregated HGSOC tumors from four patients. Representative FACS plots shown. (B) Flow cytometric analysis of PC-1 binding to tumor-associated CD4⁺ or CD8⁺ T cells. CD4+ T cells were assessed for PD-1 and TIM-3 expression as markers of terminal exhaustion. Representative FACS plots shown. (C) Flow cytometric analysis of PC-1 binding to PBMCs from healthy donors or individuals with HGSOC (n=4 each). (D) Representative FACS histograms comparing PC-1 binding to tumor-associated macrophages (CD45⁺ CD14⁺ DR⁺), M-MDSCs (CD45⁺ CD14⁺ DR⁻) and lymphocytes (CD3⁺ CD45⁺) found in CTCL, NSCLC, and HGSOC tumors. Healthy donor PBMCs were included as a control. See Figure S7.

PC-1 CAR T cells specifically targeted tumor cells and tumor-promoting myeloid cells in vivo without off-target activity

Chimeric antigen receptor (CAR) T-cell therapy is being actively investigated as a strategy for the treatment of solid tumors as durable responses to immune checkpoint blockade are limited to a subset of patients. A major challenge, however, is the identification of antigens with sufficient tumor restriction to minimize on-target, off-tumor toxicity. Given the increased expression of citrullinated vimentin in human tumor cells and tumor-promoting cells within the tumor microenvironment, we sought to demonstrate that citrullinated vimentin is an effective tumor-selective antigen. Thus, we engineered T cells to express a CAR incorporating the PC-1 mAb derived scFv as the recognition element. We selected a CAR T platform for proof-of-concept studies because recent studies have shown that CAR-redirected T cells have exhibited more robust activity against at least osimertinib-resistant tumors than antibody-drug conjugates built from the same parental antibody ⁴⁶. Using the soluble chain fragment variable (scFv) region cloned from PC-1 mAb, we generated CD28 and 4–1BB co-stimulated T cells using activated αβ T cells from healthy donors (see methods); a widely established approach for development and assessment of CAR T cells bearing mAb-derived scFv ^47–49.

We found that CD28- and 4–1BB-costimulated PC-1 CAR T cells killed Calu-6 and A427 NSCLC cells in vitro compared with mock-transduced controls, reducing Calu-6 cell index by 80% and 92% and A427 cell index by 64% and 60%, respectively as measured in the Xcelligence in vitro cytotoxicity assay (see methods) (Figures 7A–B). In NSG xenograft models, CD28-costimulated PC-1 CAR T cells controlled established Calu-6 and A427 tumor growth compared with mock-transduced T cells (Figures 7C–E). PC-1 also bound mouse tumor-associated macrophages and splenic macrophages in A427-bearing Rag1−/− mice, supporting use of immunocompetent models to evaluate activity and potential toxicity (Figure 7F).

Figure 7. CAR T cells against citrullinated vimentin effectively mediate antitumor activity without detectable off-tumor toxicity.

(A, B) In vitro killing of Calu-6 (A) and A427 (B) cells by human PC-1 CAR T cells bearing CD28 or 4–1BB costimulatory domains compared with mock-transduced T cells. (C) Schematic of NSG mouse xenograft treatment model. (D,E) Tumor growth in Calu-6 (D) and A427 (E) xenograft-bearing NSG mice treated with CD28-costimulated PC-1 CAR T cells or mock-transduced T cells (n=9, pooled from 2 independent experiments). (F) PC-1 binding to tumor cells (EPCAM⁺), tumor-associated macrophages (CD45⁺ F4/80⁺ MHCII⁺), and splenic macrophages (CD45⁺ F4/80⁺ MHCII⁺) in A427-bearing Rag2^−/− mice. Quantification of cells expressing citrullinated vimentin is shown by percentage at right (n=9 mice). (G) PC-1 binding to parental and PAD2-deficient A427 cells. Representative histograms (left) and scatter plots (right) are shown. (H) In vitro killing of PAD2-deficient A427 cells by PC-1 CAR T cells compared with mock-transduced T cells. (I, J) In vitro killing of murine KPM (I) and UPK10 (J) tumor cells by murinized PC-1 CAR T cells. (K) Schematic of syngeneic UPK10 intraperitoneal tumor model and CAR T treatment. (L) Survival of UPK10-bearing B6 mice treated with murinized PC-1 CAR T cells or mock-transduced T cells. (M-O) Histologic and serum chemistry assessment of off-tumor toxicity after PC-1 CAR T treatment. Scale bars denote 400 μm. For cytotoxicity data (panels A-B, I-J) significance was determined by two-way repeated measures ANOVA. For tumor growth curves (panels D-E), significance was determined by two-way repeated measures ANOVA, followed by a post-hoc Šídák’s multiple comparisons test. For survival data (panel L), significance was determined by Log-Rank test. A p value < 0.05 was considered statistically significant and is denoted as follows: *p<0.05, **p<0.01. Error bars reflect standard deviation.

To test whether PC-1 CAR T activity depended on PAD2-mediated citrullination, we generated PAD2-deficient A427 cells. PAD2 loss reduced PC-1 surface binding from 82.6% in parental A427 cells to 6.4% in PAD2-deficient cells and eliminated killing by CD28- or 4–1BB-costimulated PC-1 CAR T cells (Figures 7G–H). These data support PAD2-dependent citrullinated cell-surface antigens as functionally relevant targets of PC-1 CAR T cells.

PC-1 CAR T cells demonstrated anti-tumor activity and improved survival in immunocompetent mouse models of cancer

To better characterize the effects of CAR T cells directed against PC-1 in vivo, we sought to use a syngeneic, immunogenic mouse model previously developed in our lab ⁵⁰. Human and mouse vimentin share >97% identity, including conservation of Arginine residues. To test whether mouse cancers can translocate citrullinated vimentin to the cell surface of tumor associated myeloid cells, we challenged B6 mice with orthotopic ID8-Defb29/Vegfa tumors, an ascitogenic model generated by our team ⁵¹.

For evaluating PC-1 CAR T activity in an immunocompetent setting, we generated murinized CD28-costimulated CAR T cells using the PC-1 VH/VL CDR sequences integrated into a highly similar IGHV1–80*01/IGHJ1*01; IGKV8–30*01/ IGKJ2*03 IgG1 backbone to prevent in vivo neutralization. We used the murinized scFv of antibody PC-1 to generate CD28 co-stimulated mouse CAR T cells, with the murine counterparts of the signaling motif used in conventional human CAR T cells ⁵². Murinized PC-1 CAR T cells killed murine KPM and UPK10 tumor cells in vitro, reducing cell index by 65% and 44%, respectively, compared with mock-transduced controls (Figures 7I–J). In an orthotopic syngeneic UPK10 model, treatment of established tumors with murinized PC-1 CAR T cells improved survival compared with mock-transduced T cells, despite the aggressive progression of this orthotopic model ⁵³ and that treatment was without lymphodepletion (Figures 7K–L).

We next assessed off-tumor toxicity in UPK10-bearing mice treated with PC-1 CAR T cells or mock controls. Histologic evaluation of lung, heart, liver, kidney, and brain showed no evidence of tissue injury, and serum markers of renal, pancreatic, and hepatic function were not significantly different between groups (Figures 7M–N). In contrast, tumors from PC-1 CAR T-treated mice showed widespread fat necrosis compared with mock controls (Figure 7O). Together, these findings demonstrate antitumor activity of PC-1 CAR T cells without detectable off-tumor toxicity in an immunocompetent HGSOC model.

Discussion

TIL-Bs bridge innate and antigen-driven immunity through antigen presentation and antibody secretion ¹¹, yet the antigen specificities and therapeutic potential of tumor-infiltrating B cell responses remain incompletely defined. Here, we performed paired single cell transcriptomics and BCR profiling of TIL-Bs, thus allowing for in-depth analysis of BCR specificity amongst TIL-Bs present in NSCLC treated with PD-1 blockade. We demonstrated that antibodies isolated from various TIL-B populations react with intracellular proteins ectopically expressed on the surface of cancer cells but not normal cells. Moreover, these antigens were not patient-specific and raised the possibility that these patient-derived tumor antibodies can be repurposed into novel cancer therapeutics such as CAR T cells. In fact, we showed that CAR T cells designed with our lead candidate tumor-reactive mAb, PC-1 specifically killed tumor cells and tumor-promoting myeloid cells without off-target activity. Thus, we developed a high-throughput pipeline with broad applicability in assessing tumor-reactive antibodies from TIL-B populations and showed the therapeutic potential of tumor-infiltrating plasma cells in NSCLC patients following anti-PD1 antibody treatment.

Multiple TIL-B states have been associated with favorable outcomes across solid tumors, including GC B cells ¹², MBCs ^54–56, ASCs ^10,28 and atypical MBCs ^2,57,58, yet their antigen specificities remain incompletely defined. Here, we identify tumor-reactive, class-switched antibody lineages across antigen-experienced TIL-B populations, supporting T-dependent diversification and maturation of humoral responses within the TME, and raised the possibility that endogenous TIL-B responses may contribute to antitumor immunity through direct antigen recognition and Fc-dependent effector mechanisms as previously demonstrated ⁵. Previous single-cell studies have defined the phenotypic diversity and functional maturation of TIL-B populations across tumor types ^2,4,21, but antigen specificity remains less well characterized. Previously, Mazor et al. utilized single cell plasma cell sorting followed by PCR and production of recombinant mAbs, defining tumor antibody specificity in patients with HGSOC ⁵. By integrating single-cell transcriptomics with paired BCR sequencing and recombinant antibody expression, our approach links TIL-B cell state to antibody specificity and enables functional interrogation of tumor-reactive humoral responses across multiple TIL-B states, which may give insight into the importance of the TIL-B responses as antibodies to extracellular and secreted proteins contribute to the effectiveness of immune checkpoint inhibitors ^28,59.

In our analysis of antibody specificity, PC-1 mAb representing one of the most clonally expanded plasma cell lineages, recognized citrullinated proteins including vimentin. Citrullination, associated with states of cell stress and inflammation can result in ectopic expression of post-translationally modified intracellular self-proteins on the cell surface and as such, there is growing interest in citrullination as a source of immunogenic tumor antigens ^{29–31,33,60,61}. Of note, citrullinated vimentin and Hsp90 are intracellular proteins known to be ectopically expressed on the cell surface on tumor cells ^32,33,60,62 and citrullinated calreticulin is known to be expressed on tumor cell surface following T cell induced injury and immunogenic cell death (ICD) ³⁷; all of which were citrullinated proteins identified in our study. With regards to immunogenicity for these types of citrullinated proteins, autophagy of tumor cells has been shown to facilitate processing and presentation of intracellular citrullinated peptides on MHC class II molecules to CD4⁺ T cells, which may permit citrullinated proteins to contribute to the development of TLS and TLS-associated GC reactions in the TME ^63–65. As the PC-1 lineage was recovered from CD138-enriched TIL-Bs, all but one of the PC-1 lineage members were in the PC/ASC B cell pool, unfortunately limiting our ability to map progression of this clonal lineage through a GC reaction. We do hypothesize a role for T-cell mediated positive selection in the GC for this lineage however, as acquisition of somatic mutations in the mature sequence corresponded with nanomolar affinity of the PC-1 mAb for citrullinated vimentin.

Despite reactivity with citrullinated antigens, PC-1 did not display features similar with antibodies associated with autoimmune rheumatoid arthritis ⁶⁶ and the patient from which PC-1 was derived did not develop rheumatoid arthritis after immune checkpoint blockade. Thus, PC-1 reactivity likely reflects a tumor- and damage-associated state in which PAD-mediated citrullination coincides with aberrant cell-surface exposure of intracellular proteins. Citrullination can occur during inflammatory and tissue injury states, including settings of calcium flux, oxidative stress, neutrophil activation/NETosis, macrophage activation, and immunogenic cell death ^{30,31,33,60,62,67}. However, in normal tissues these processes are typically constrained and transient, whereas the TME provides a chronic inflammatory niche characterized by persistent cellular stress, cell death, hypoxia, myeloid activation, and tissue remodeling. This may explain why PC-1 recognized cancer cells and cancer-associated immune/stromal cells but showed limited binding to paired normal tissues or healthy donor PBMCs.

The mechanism underlying PC-1 cross-reactivity with citrullinated vimentin, calreticulin, p53, and Hsp90 remains unresolved, but this phenotype is consistent with prior reports of polyreactive tumor antigen-reactive antibodies in ovarian carcinoma and melanoma ^5,68. Such polyreactivity may be advantageous in overcoming antigenic diversity within and across tumors. Because sequence alignment did not reveal obvious shared linear epitopes, and PC-1 did not bind cyclic citrullinated peptides, we suspect that citrullination exposes or stabilizes a shared conformational feature rather than serving as the sole binding determinant. Interestingly, despite increased affinity of the antibody to citrullinated vimentin compared to the unmodified protein, the PC-1 mAb did not react with cyclic citrullinated peptides, suggestive that the presence of citrulline alone is not a requirement of binding by this lineage. Defining the PC-1 epitope on citrullinated vimentin remains an area of active investigation.

Durable responses to immune checkpoint blockade are limited to a subset of patients, and as such, adjuvant therapies, such as CAR T-cell therapy is being investigated as an attractive strategy for the treatment of solid tumors, including NSCLC, as they can redirect autologous T cells against tumor-associated surface antigens to bypass endogenous antigen presentation and mediate direct cytotoxicity against malignant cells ²². A key challenge, however, for CAR T development is identification of antigens with sufficient tumor specificity to minimize toxicity. In addition, immunosuppression driven by myeloid cells and fibroblasts presents a substantial barrier to effective intratumoral T-cell accumulation, effector activity and long-term persistence. Remarkably, we found that citrullinated vimentin, a target of PC-1 was expressed not only on the surface of tumor cells, but also on tumor-promoting myeloid cells. When developed into a CAR-T construct bearing the PC-1 mAb specificity, we showed that this construct also killed both tumor and tumor-promoting myeloid cells (and did so without deleterious off-target effects), thus providing robust proof-of-concept data supporting a role for mining tumor antibody repertoires for development of novel tumor therapeutics.

In summary, our study provides a platform for computational analysis of TIL-B responses enabling rapid and high-throughput for antigen-agnostic isolation of tumor-reactive antibodies, demonstrating tumor antigen reactivity amongst heterogeneous TIL-B populations that develop as part of the anti-tumor humoral response. More importantly, identification of patient-derived mAbs and their cognate antigens may aid in developing new tumor therapies such as antibody-drug conjugates and CAR T cells and serve as candidate tumor antigens to be incorporated into future cancer vaccine platforms.

Limitations of the Study

This study is limited by the small number of clinical specimens with sufficient residual tissue and plasma cell infiltrates after neoadjuvant checkpoint blockade. Limited tissue availability also precluded systematic analysis of paired adjacent normal lung tissue or draining lymph nodes as potential sources of precursor B cells that populate the TME and produce tumor-reactive antibodies. In addition, because we produced antibodies from selected clonally expanded, highly mutated, or diversified BCR lineages, our approach likely under sampled the full diversity of tumor-reactive TIL-B specificities. We also cannot determine whether the antibody-secreting cells studied arose within intratumoral TLS or emigrated into the TME from systemic circulation. Finally, because our therapeutic studies focused on Fc-independent applications of tumor-derived antibodies, we did not evaluate Fc-dependent effector functions such as ADCC or ADCP. Despite extensive efforts, we were also unable to resolve the structure of PC-1 bound to citrullinated vimentin, likely related to the conformational flexibility of the antigen. Despite these limitations, our findings support the intratumoral BCR repertoire as a rich source of antibodies with therapeutic potential.

Resource Availability

Lead Contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Ryan Meyerhoff (ryan.meyerhoff@duke.edu).

Materials Availability

All unique/stable reagents generated in this study are available from the lead contact with a completed materials transfer agreement.

Data and Code Availability

All single cell sequencing data reported in this manuscript has been submitted to the SRA database and are accessible via BioProject accession number PRJNA1292770. Crystal structure of the PC-1 fab was submitted to the Protein Data Bank (PDB) and is accessible via ID number 9ZQF. All mass spectrometry raw data, metadata, and database search results have been submitted to the ProteomeXchange consortium (PXD079756) via the MassIVE repository (MSV000102166) and can be accessed at ftp://massive-ftp.ucsd.edu/v13/MSV000102166/. All raw and processed data from testing PC-1 mAb binding in the HuProt microarray have been submitted to Zenodo and can be accessed at https://doi.org/10.5281/zenodo.21015949.

STAR Methods

EXPERIMENTAL MODEL AND STUDY PARTICIPANTS.

Human samples.

Human resected lung tumor specimens were obtained from the TOP1501 clinical trial (NCT02818920) ²⁰. Briefly, this was a previously reported single-arm, triple-center phase II trial of neoadjuvant single-agent pembrolizumab in 35 patients with clinical stage IB, II, or IIIA NSCLC as assessed according to the American Joint Committee on Cancer staging system, seventh edition ²⁰. Patients received two cycles of 200 mg intravenous pembrolizumab prior to surgery. Patient characteristics and study design are further outlined in Wood C. et al. ²⁰. The study protocol was approved by the Institutional Review Boards of each participating center (IRB# Pro00071629) and informed consent was obtained from each patient prior to enrollment in the trial in accordance with federal and institutional guidelines. Clinical characteristics for each patient included for analysis in this study are shown in Supplemental Table S1.

A normal tissue microarray (CHTN_Norm3) was purchased from the Mid-Atlantic Cooperative Tissue Network division at the University of Virginia. Non-neoplastic adult tissue obtained from surgical resection specimens were sampled three times with a needle biopsy for the array. In total, 49 different tissue types, with three core samples each, were obtained from 49 patients (Summarized in Figure S5D).

Mouse models.

All animal procedures were approved by the Institutional Animal Care and Use Committee at the Duke University and Division of Laboratory Animal Resources (IACUC protocol# A011–23-01). Female, 4–6 week old NOD.Cg-Prkdcscid Il2rgtmlWjl/SzJ (NSG) mice and C57BL/6J (B6) mice, originally procured from The Jackson Laboratory, were maintained by the Duke School of Medicine's animal facility. Female, 4–6 week old, All animals were housed in cages of up to 5 mice per cage, in a temperature controlled (18–23°C), 40–60% humidity, 12 h light/dark cycle facility. Each experiment had at least 3 replicates per condition. Presented data are representative pooled data from at least 2 experiments representing 6 mice.

Cell lines.

Human cell lines Calu-6, A549, H1703, A427 (NSCLC), OvCar3 (ovarian), Hec1a (uterine carcinoma) were sourced from the American Type Culture Collection (ATCC). The murine KPM (lung cancer) and UPK10 (ovarian carcinoma) cell lines were provided by the Conejo-Garcia laboratory ^69,70. Phoenix-AMPHO and 293T cells were sourced from ATCC. Expi293F cells were sourced from ThermoFisher. Calu-6, A549, H1703, and OvCar3 cell lines were routinely cultured in R10 (RPMI-1640, 10% FBS, penicillin (100 IU ml⁻¹), streptomycin (100 μg ml⁻¹), L-glutamine (2 mM), sodium pyruvate (0.5 mM)) media (Thermo). A427, Hec1a, HEK293T, and Phoenix-AMPHO cell lines were cultured in D10 medium (DMEM, 10% FBS, penicillin/Streptomycin (100 μg/mL Lonza), l-glutamine (2 mM, Gen Clone), sodium pyruvate (0.5 mM)). Cells were incubated at 37°C in a humidified 5% CO₂ atmosphere.

METHOD DETAILS.

Tumor sample collection and processing.

Following the Miltenyi Biotec Human Tumor Dissociation Kit protocol, tumors were dissociated into single-cell suspensions using human tumor dissociation enzymes (H, R, and A, Miltenyi Biotec) and GentleMACs equipment (Miltenyi Biotec). After two 30-minute incubations at 37°C, the samples were filtered through a 70uM filter (Corning) and washed in RPMI 1640 (ThermoFisher). After centrifugation, the cells were resuspended in a 90% heat-inactivated fetal bovine serum (FBS, Gem-Cell) and 10% dimethyl sulfoxide (DMSO, Sigma-Aldrich) solution and aliquoted into cryogenic vials. The cells were then cooled in a CoolCell cell freezing container (BioCision) at −80°C. The following day the cells were placed in vapor phase liquid nitrogen until later testing.

CODEX multiplexed imaging.

The tissues were sectioned onto superfrost plus slides and stained with the validated panels of CODEX antibodies and imaged following a previously established protocol ⁷¹. Briefly, this involved cyclic stripping, annealing, and imaging of fluorescently labeled oligonucleotides complementary to the oligonucleotide conjugated to the antibody. Each array underwent CODEX multiplexed imaging. Raw imaging data were processed using Akoya PhenoCycler Fusion 2.2.0 for image stitching, drift compensation, deconvolution, and cycle concatenation. Following processing, the imaging data were evaluated for signal specificity, with markers that displayed low signal-to-noise ratios or untenable patterns being excluded from further analysis. Uploaded images were visualized in ImageJ (https://imagej.net/software/fiji/).

Transcriptome data generation.

Cryopreserved PBMC and dissociated tumor single-cell suspensions were thawed and washed in RPMI 1640 (ThermoFisher) with 10% heat-inactivated fetal bovine serum (FBS, Gem-Cell), and 1% Penicillin, Streptomycin and L-Glutamine (ThermoFisher). Benzonase (2μL/mL, Novagen) was added to the initial thaw media. The cells were resuspended, counted, and centrifuged prior to magnetic labeling. The cells were labeled with either CD138 MicroBeads (Miltenyi Biotec) or CD19 MicroBeads (Miltenyi Biotec) and buffer that consisted of a 1:20 dilution of MACS BSA Stock and MACS Rinsing Solution (Miltenyi Biotec). Cells were then incubated at 2–8°C for 15 minutes. After incubation, the cells were washed and resuspended in MACS buffer. The cells were magnetically separated using MS columns and an OctoMACs Separator (Miltenyi Biotec). After adding the cell suspension to the MS columns, the columns were washed three times with MACS buffer. The magnetically labeled cells were flushed from the column with 1mL of MACS buffer and transferred on ice to a separate lab library construction. Library generation from the purified cells was performed using the 10x Chromium Single-cell 5’ and V(D)J library construction kits (10x Genomics) according to manufacturer’s instructions. Purified libraries were sequenced by the Duke Sequencing and Genomics Technologies (SGT) Core using Illumina NextSeq 1000 or NovaSeq X Plus instruments via short read paired end sequencing, with a minimum of 5k read pairs for VDJ libraries, and 50k read pairs for GEX libraries.

Recombinant antibody production.

Protein-A purified recombinant monoclonal antibodies (mAbs) were commercially obtained from GenScript (Piscataway, NJ) using a high throughput TurboCHO^™ transfection pipeline. MAbs encoded clonotype consensus sequence for each BCR as described above. Sodium dodecyl sulfate–polyacrylamide gel electrophoresis (SDS-PAGE) and SDS-PAGE western blot for heavy and light chain protein expression were used for quality control and each recombinant mAb was verified endotoxin free.

Enzyme-linked immunosorbent assays (ELISAs).

Binding of recombinant mAbs to purified proteins was tested in ELISA as previously described ⁷². In brief, proteins at a concentration of 2 μg/mL were directly coated to Nunc-absorb (ThermoFisher) plates overnight at 4°C. Proteins not bound to the plate were washed away and the plates were blocked with 3% BSA in PBS ⁷². Serial dilution of antibodies starting at 100 μg/mL was added to the plate for a 90-minute incubation at room temperature. Binding antibodies were then detected with goat anti-human IgG (Thermo Fisher Cat#A18815) for a 60-minute incubation at room temperature, and these detected with streptavidin-HRP (Pierce, Cat#21132) with a 60-minute incubation at room temperature. HRP detection was quantified with 3,3′,5,5′-tetramethylbenzidine (TMB). Binding levels were measured at an optical density of 450 nm (OD_405nm).

Immunofluorescence staining of immortalized cell lines.

Calu-6 (ATCC HTB-56) or A549 (ATCC CCL-185) cell lines were grown to confluency on eight well chambered slides at 37°C (LAB-TEK, Thermo Fisher, Cat#154534). Per basic immunodetection protocol, cells were washed thrice with PBS containing CaCl₂ MgCl₂, fixed with 4% paraforlmaldehyde for 15 minutes at 4°C and then blocked for 30 minutes at 37°C with 3% fetal bovine serum (FBS)-PBS containing CaCl₂ and MgCl₂. Antibodies were diluted to 50 μg/mL in the blocking buffer and incubated overnight at 4°C. Binding antibodies were detected with AlexaFluor-488 tagged anti-human IgG (Invitrogen, Cat#A11013) with a 30-minute incubation at 37°C. A 2-minute 37°C incubation with 1:1500 dilution of CellMask Deep Red Plasma Membrane Stain (Invitrogen C10046) was used to visualize the cell membrane. Cells were then fixed with 4% formaldehyde for 15 minutes at 4°C, allowed to dry, and then counterstained and mounted with 4′,6-diamidino-2-phenylindole (DAPI) Fluouromount-G (SouthernBiotech Cat#0100–20).

Confocal microscopy and image analysis.

Images were acquired on Leica SP5 or SP8 confocal microscopes using 40x oil immersion objective. Single confocal sections were taken every 0.5 μm using sequential mode.

Expression of proteins used for structural studies.

The plasmids in this study were obtained from GeneImmune Biotechnology (Rockville, MD) with quality controls. The gene encoding the full-length p53 protein (UniProt: P04637) was synthesized with an N-terminal 6x histidine-SUMO tag with a HRV3C cleavage site. The construct was placed into a pET32 based plasmid for expression in E. coli. The p53 core domain (residues 94 – 311) was obtained by site directed mutagenesis of the full-length plasmid. The gene encoding the full-length CALR protein (UniProt: P27797) was synthesized with a C-terminal 10x Histidine tag and placed into a pVRC8400 vector for expression in mammalian cell lines. CALR-NP (1–302) and CALR-P domain (188–302) constructs were prepared by site directed mutagenesis of the full-length plasmid. The CALR-NC (1–187, 303–417) domain construct was prepared by removal of the P domain and insertion of a GSG linker between the N and C domains. His-tagged PC-1 Fab heavy chain (1–230) was produced via site-directed mutagenesis of the HC chain plasmid of PC-1 with the insertion of a His-tag position 230.

Protein purification.

Full-length p53 and the p53 core domain were produced in BL21(DE3) competent cells transformed with the appropriate plasmid. Cells were grown to an O.D. 600 of 0.8 in Luria broth supplemented with ampicillin to a final concentration of 100 μg/mL. Protein production was induced with the addition of Isopropyl β-D-1-thiogalactopyranoside to a final concentration of 1 mM. Cells were induced at 25 °C for 24 hours. Cells were harvested via centrifugation and resuspended in buffer A containing 50 mM Tris pH 7.5, 500 mM NaCl, 10% glycerol, and 20 mM Imidazole. β-mercaptoethanol (5 mM), Lysozyme (10 μM), and Benzonase (25U/mL) were added to the cell resuspension, and cells were lysed via sonication. The resulting lysate was centrifuged at 30000g for 30 min. The supernatant was filtered through a 0.22 μm filter and applied to a 5 mL HisTrap Excel column (Cytiva, MA) previously equilibrated with buffer A. Proteins were eluted from the column by applying a linear gradient of imidazole from 20–500 mM over 20 column volumes. Fractions containing p53 were pooled and applied to a Superdex 200 Increase 10/300 GL column (Cytiva, MA) for further purification. p53 was eluted from the column isocratically with a buffer containing 50 mM Tris pH 7.5, 500 mM NaCl, and 10% glycerol. Fractions containing p53 were subsequently pooled. The purity of the protein was assessed by SDS-PAGE using NuPage 4–12% (Invitrogen, CA). Purified proteins were flash frozen and stored at −80°C in single-use aliquots. Aliquots were thawed at 4°C before use.

CALR and CALR domains were produced in Expi293 cells and harvested from the supernatant on the 6th day post transfection. The CALR and domains containing a C-terminal 10x Histidine tag were purified using Talon cobalt affinity resin (Takara Bio, CA). Supernatant was loaded on to the column and washed with buffer A (50 mM Tris pH 7.5, 500 mM NaCl, and 10% glycerol). Proteins were eluted from the column with buffer B (50 mM Tris pH 7.5, 500 mM NaCl, and 10% glycerol, 500 mM Imidazole). Fractions containing CALR or CALR domains were pooled for further purification. Size exclusion chromatography and purity assessment were performed identically to the p53 proteins.

His-tagged PC-1 Fab was produced 293F cells and harvested from the supernatant on the 6th day post transfection. The Fab containing a 6x Histidine tag was purified using a 5mL HiTrap Nickel Excel column (Cytiva, MA) for affinity purification. The supernatant was applied to the column previously equilibrated with buffer A. (1x PBS, 20 mM Imidizole). Proteins were eluted from the column by applying a linear gradient of imidazole from 20–500 mM over 20 column volumes. Fractions containing Fab were pooled and applied to a Superdex 200 Increase 10/300 GL column (Cytiva, MA) for further purification. Fab was eluted from the column isocratically with a 1x PBS buffer. Fractions containing Fab were subsequently pooled. The purity of the protein was assessed by SDS-PAGE using NuPage 4–12% (Invitrogen, CA). Purified proteins were flash frozen and stored at −80 °C in single-use aliquots. Aliquots were thawed at 4°C before use. PC-1 antibody was produced in Expi293F cells and purified by Protein A affinity chromatography. Antibody was further purified using SEC using a Superose 6 Increase 10/300 GL column (Cytiva, MA).

Surface Plasmon Resonance.

Binding experiments were performed using SPR on a Biacore T-200 (Cytiva, MA) with HBS buffer supplemented with 3 mM EDTA and 0.05% surfactant P-20 (HBS-EP⁺, Cytiva, MA). All binding assays were performed at 25 °C.

Antigen binding to PC-1 IgG was assessed using a Series S CM5 chip (Cytiva, MA) which was labeled with anti-human IgG (fc) antibody using a Human Antibody Capture Kit (Cytiva, MA). IgG’s were coated at 200 nM (120s at 5 μL/min) on the chip surface. For single injection experiments, antigen domains were injected at 1 – 10 μM over the antibody using the high-performance injection mode (120 sec on, 240 sec off, 50 μL/min). The surface was regenerated three pulses of a 3 M MgCl2 solution for 10 seconds at 100μL/min. Kinetics experiments were performed using single cycle kinetics mode with five different concentrations of analyte serially diluted two-fold. CALR and Vimentin concentrations ranged from 4 – 64 nM, while p53 ranged from 1 – 16 nM. Injections lasted 120 seconds with a final off time of 240 seconds at 50 μL/min. Kinetics for the CALR-NP domain and IgG were determined similarly but using four separate injections at differing concentrations. Concentration range for the CALR-NP domain was 108 μM to 863 nM. Kinetics for p53 core domain and PC-1C IgG were determined using a Series S NTA chip (Cytiva, MA). The chip was activated with a 60 s pulses of 0.5 mM nickel chloride over the chip at 10 μL/min. p53 core domain at 100 nM was injected over the surface at 10 μL/min for 60 seconds. PC-1 IgG was then injected five different concentrations (10 nM – 0.625 nM, serially diluted two-fold) at 50 μL/min with an association time of 60 s and a dissociation time of 120 s. Sensogram data were analyzed using the BiaEvaluation software (Cytiva, MA).

X-ray crystallography.

Crystals were grown via the vapor diffusion method at room temperature in siting drop well format. His-tagged PC-1 Fab crystals were formed by mixing 0.25 μL of Fab at 40 mg/mL with 0.15 μL of well solution containing 0.2 M calcium acetate hydrate, 0.1 M sodium cacodylate trihydrate pH 6.5, and 18% w/v PEG8000. Crystals appeared within four weeks. Cryoprotection of the crystals was done by mixing glycerol to a final concentration of 50% with well solution. 0.5 μL of this solution was added to the well containing the crystals. Crystals were flash-cooled in liquid nitrogen and data were collected at the Advanced Photon source, Beamline 22-ID at 100 K and a wavelength of 1 Å. All data were indexed and integrated and scaled using HKL2000 ⁷³.

Phaser in the PHENIX suite was used to perform molecular replacement using a model generated from the Alphafold3 prediction server using the Fab sequence ^44,74. Phenix.refine was used for data refinement, and manual refinement was done in Coot ^75–78.

Flow Cytometry of Live Cells.

Adherent cells were detached from the tissue culture plate using a cell scraper or enzyme free cell dissociation buffer (ThermoFisher #13151014) and washed with PBS without CaCl2 and MgCl2. Cells were incubated with primary antibodies at a final concentration of 5ug/mL at 4°C for 20 minutes, washed, and incubated with species specific fluorescent secondary antibodies. Flow cytometry was conducted using the BD Biosciences FACSCanto and analyzed using Flowjo.

Western Blot.

Cells were harvested and lysed in ice-cold RIPA buffer containing Roche Complete Protease Inhibitor Cocktail (Sigma #11697498001) for 30 minutes. Lysates were sonicated, pelleted, and debris removed. Protein content was quantified using the Pierce BCA Protein Assay Kit (ThermoFisher #23227). Lysates were aliquoted, incubated with 4x Laemmli Sample Buffer (Biorad #1610747), and loaded onto a 4–20% Mini-PROTEAN TGX Stain-Free Protein Gel (Biorad #17000436). Gels were ran at 100V for 2 hours and transferred unto a Biorad Immun-Blot PVDF Membrane using a Biorad semi-dry transfer apparatus. Membranes were blocked in Intercept Blocking Buffer (Li-Cor #927–70001) and incubated with PC-1, diluted 1:1000, overnight at 4°C with gentle shaking. GAPDH (Cell Signaling) was used as a loading control. Membranes were washed with PBS containing Tween-20 and incubated with donkey anti-rabbit IRDye680RD (Li-Cor #926–32232) or goat anti-human IRDye88CW (Li-Cor #926–32232) at room temperature with gentle shaking for 1 hour. Membranes were imaged using the Li-Cor Odyssey Imager.

Immunoprecipitation.

PC-1 antibody was covalently conjugated to Dynabeads using the Dynabeads Antibody Coupling Kit (ThermoFisher #14311D). Tthe antibody was incubated with the beads at 37°C overnight with gentle shaking and then washed according to the kit instructions. To perform the immunoprecipitation, PC-1-antibody-bound beads were incubated with HEK293T cell lysate overnight at 4°C with gentle shaking. A parallel experiment using unbound beads was used as a control. Beads were washed with ice-cold RIPA buffer three times on a magnetic column and the eluted with 4x Laemmli Sample Buffer for 15 minutes. Eluates were loaded and ran on a gel, mass spectrometry-compatible stain was performed, and the visible band was excised for mass spectrometry analysis.

Mass Spectrometry.

Immunoprecipitation samples were loaded on the Invitrogen’s NuPAGE Bis-Tris mini-gel system. An approximately 57kDa band was excised from both the PC-1-conjugated-Dynabead sample as well as the Dynabead control sample. Samples were digested and subjected to chromatographic separation on a Waters MClass UPLC coupled with mass spectrometry via ThermoFisher Fusion Lumos mass spectrometer. Data was processed using Proteome Discoverer 2.5 followed by Mascot database searches against a Homo Sapiens protein database. Proteins of ~50–65kDa in size as well as proteins overlapping between the PC-1 and control samples were excluded to generate PC-1 antigen candidates.

Citrullinating Recombinant Proteins.

Recombinant vimentin (Cayman Chemical #11234) and citrullinated vimentin (Cayman Chemical #21942) were used for the ELISA. Recombinant p53 (Abcam #ab237007), recombinant calreticulin (Abcam #ab91577), and HSP90α (R&D Systems #AP-160–050) were incubated with 200g recombinant human PAD2 (Cayman Chemical #10785) in 50 mM Tris-HCl at pH 7.5 containing 10 mM CaCl2 and 5 mM dithiothreitol (DTT) for 5 hours at 37°C to generate their citrullinated protein counterparts for the ELISA.

Cyclic Citrullinated Peptide (CCP) ELISA.

The Immunoscan CCPlus test kit (SVAR #RA-96PLUS) was used to determine whether PC-1 would bind cyclic citrullinated peptides (CCPs). The ELISA was performed according to the kit instructions at room temperature. Briefly, the positive human plasma pool standards, plasma negative control, plasma positive control, serum from healthy donors, and 10μg/mL of PC-1 was added to CCP peptide-coated ELISA plate and incubated for 1 hour. Wells were washed and the conjugate was added and incubated for 30 minutes. Wells were washed and the substrate solution was added and incubated for 30 minutes. Stop solution was added to the wells and the absorbance values of the samples were read by a spectrometer at 450 nm.

Nuclear antigen antibody reactivity.

Antibody autoreactivity was measured using the AtheNA Multi-Lyte ANA kit (Zeus scientific, Cat#A21101) as previously described ^41,79. Briefly, we tested recombinant mAbs at 2-fold serial dilutions starting at 50 mcg/mL, and the secondary antibody Phycoerythrin-conjugated goat anti-human IgG was provided with the kit. More specifically, we tested experimental mAbs at 50, 25, 12.5 and 6.25 μg/ml and included an external negative control, CH65, and an external positive control, 4E10 (Polymun). The manufacturer’s protocol was followed where positivity is defined as a value >120 MFI.

HuProt Microarray for Antibody Polyreactivity and Candidate Antigen Binding:

PC-1 mAb was tested at a concentration of 1μg/mL for binding to a 23,004 full-length human proteins recombinantly produced in the yeast S. cervesiae that were purified and printed in duplicate on glass slides (HuProt microarray from CDI laboratories, microarray barcode CDI91333288). Briefly, the array was incubated with blocking buffer (1xTBS / 0.1% Tween 20 / 5% BSA) at room temperature for 1 hr. with gentle shaking. Then, the antibody was diluted to 1 μg/mL in blocking buffer and probed on the array at room temperature for 1 hour with gentle shaking. After probing, the array was washed with TBST (1xTBS / 0.1% Tween 20) for 10 min X 3, and then probed with Alexa647-anti-human IgG Fc secondary antibody (0.25 ug/mL) at room temperature for 1 hour in a light-proof box with gentle shaking, followed by 3 washes with TBST for 10 min each and 3 rinses with ddH2O. The array was then dried with an air duster and scanned using a GenePix 4000B scanner for data collection. CDI software was used to quantify the specificity of each individual sample to specific proteins on the array based on Z-scores after elimination of non-specific hits that directly bound to the secondary antibody. Reported Z score is the average Z score of duplicate spots of a given protein. A Z-score cut-off of three and above was used to determine positivity.

Immunohistochemistry:

Immunohistochemistry of a normal donor tissue microarray was performed using the Roche Discovery Ultra. Antigen retrieval was performed with pH 8.0 EDTA at 99°C for 56 minutes. Sections were incubated with PC-1 at 1:100 dilution for 60 minutes. Hydroquinone-labeled anti-mouse multimers and anti-hydroquinone-HRP were incubated at 12 minutes each for detection. 3,3'-diaminobenzidine (DAB) was used for visualization.

CAR T Constructs and Transduction.

Constructs encoding citrullinated-vimentin human or murine chimeric antigen receptors with either CD28 or 41BB co-stimulatory domains or mock PBMN were generated in our laboratory and confirmed by sequencing services from GeneWiz. Retroviral particles were generated by transfecting Phoneix-AMPHO cells with these constructs using Lipofectamine 3000 (Invitrogen). 5 × 10⁷ human PBMCs (Stem Cell Technologies) from one healthy human donor per experiment were cultured in AIM-V media supplemented with 300 U/mL human IL-2 and 1 ng/mL of OKT3 to activate ab T cells. Viral supernatants were collected 48 hours after transfection into a 24 well plate coated with retronectin overnight. The plates were spun for 2 hours at 32°C and 2000g. 2 × 10⁶ human ab T cells were seeded into the viral supernatant. The plates were spun for 10 minutes at 32°C and 1000g. T cells were collected after 24 hours and cultured in AIM-V media supplemented with 300U/mL of human IL-2. Target gene expression was analyzed by FACS for GFP expression. Pooled CAR T cells were utilized if GFP expression was ≥40%.

Murine CAR T cells were generated by harvesting the spleen of a B6 mouse, generating a single cell suspension, lying red blood cells, and culturing remaining cells in RPMI media supplemented with 1 ng/mL IL-7 and 50 U/mL IL-2 to activate T cells. Retroviral particles were generated by transfected Phoenix-ECO cells with murinized CAR or mock PBMN constructs using Lipofectamine 3000 (Invitrogen). Viral supernatant generation was conducted as per human CAR T cell protocol above.

Cell Culture for Cytotoxicity Assays.

Calu-6, A427, and Phoenix-AMPHO cell lines were purchased from the ATCC. KPM and UPK10 cells lines were generated previously in our lab. Cell lines were routinely cultured in media (Calu-6: RPMI, A427: EMEM, Phoenix-AMPHO: DMEM) supplemented with 10% FBS, penicillin (100 IU ml−1), streptomycin (100 μg ml−1), L-glutamine (2 mM), sodium pyruvate (0.5 mM)) media (Thermo). Human frozen PBMCs were purchased from STEMCELL Technologies. Human ab T-cells were isolated from frozen PBMCs by culture of in AIM-V media supplemented with 5% human serum, penicillin (100 IU ml−1), streptomycin (100 μg ml−1), L-glutamine (2 mM), 300U/mL human IL-2, and 1 ng/mL human OKT3 for 72 hours prior to transduction.

Cytotoxicity.

The xCELLigence system (Roche/ACEA Biosciences, San Diego, CA), allowing for continuous real-time monitoring of cellular adhesion properties in vitro was used according to manufacturer’s instructions. Firstly, we added complete media (50 μL) to wells of E-96 xCELLigence plates. After equilibration to 37 °C, plates were inserted into the xCELLigence station to measure the baseline impedance. This ensured that all wells and connections were working within acceptable limits. Cell lines were firstly seeded into the wells at 5×10⁵ cells/well. After 24h of culture, citrullinated-vimentin chimeric antigen receptors with either CD28 or 41BB co-stimulatory domains or mock transduced CAR T cells were added to each well at a concentration of 1:5 target to effector cellsCAR T (2.5 × 10⁶ pooled CAR T cells per well). We used xCELLigence software to provide an electronic record of the cell index which represents the measure of cellular adhesion across each individual well. In the absence of living cells (media only) or with a suspension of dead cells, the cell index values will be close to zero. After cellular attachment onto the electrode, the measured signal correlates linearly with cell number throughout the experiment.

Murine Tumor Models.

Flank tumors were initiated by injecting 2×10⁶ Calu-6 or A427 cells in PBS into flanks for NSG mice. Flanks were palpated daily to ensure a palpable tumor was present prior to injection of CAR T cells. Once a palpable tumor was present, 1×10⁶ transduced CAR T cells were injected once intraperitoneally 4–10 days after inoculation of tumor cells. Tumors were measured twice weekly. No lymphodepletion was pursued. Tumor volume was calculated as 0.5 × L × W × H where L is length, W is width, and H is height.

Intraperitoneal tumors were initiated by injecting 4×10⁶ UPK10 cells in PBS into the peritoneum of B6 mice. After 4 days, 1×10⁶ transduced CAR T cells were injected once intraperitoneally into these mice. Mice were monitored daily for signs of distressed and sacrificed at the time that any distress including cachexia, hunching, diminished activity level, significant ascites, or fur changes were noted. Serum and vital organs were harvested at the time of sacrifice. Toxicity testing was performed through the Duke University Research Histology Core Facility and Research Animal Pathology Core Facility.

Fluorescence-Activated Cell Sorting (FACS).

Sorting of antibody transduced human T cells (GFP expressing) was performed by staining with DAPI (Thermo Scientific) viability dye and gated for DAPI–GFP⁺ cells. Samples were fluorescence-activated cell sorting (FACS) analyzed using BD FACS ARIA. Samples with at least 50% GFP⁺ cells were utilized for cytotoxicity and tumor models.

CRISPR and Generation of PAD2^−/− Cell Line.

Guide RNA complexes were generated by mixing tracrRNA (IDT Technologies) and crRNA specific to PAD2 generated in our lab and ordered through IDT technologies. CtRNA complexes were generated by mixing the tracrRNA:crRNA complex with GFP labeled Cas9 protein (IDT technologies). 1×10⁶ A427 cells were electoporated with 1300V in a 30 ms pulse using the Neon NxT Electoporation system (ThermoFischer). Cells were cultured in EMEM medium for 48 hours and sorted by FACS for GFP positive, citrullinated vimentin negative cells. After reaching confluence, cells were negatively selected to eliminate residual citrullinated vimentin expressing cells using EasySep Biotin bead selection kit (StemCell Technologies) and Ab PC-1 labeled with streptavidin. Negative selection was repeated every 3–4 passages of cells to prevent outgrowth of citrullinated vimentin expressing clones.

Quantification and Statistical Analysis

TLS Neighborhood Clustering and Patch Delineation

Cell-type neighborhood analysis was performed across 4 regions to infer 30 spatial neighborhood clusters. TLS-associated neighborhoods were defined by enrichment of B cells and CD4⁺ T cells; only cells assigned to the TLS neighborhood label were used for patch construction. TLS cells were grouped into spatially coherent patches using density-based hierarchical clustering, with local density estimated from 6-nearest-neighbor distances and cluster membership constrained to exclude very small structures (minimum 80 cells per cluster, with a stability requirement of 50 supporting neighbors). Spatial scale was adaptively determined using the 85th percentile of local distances with a conservative scaling fraction of 0.3 and a small numerical floor (10⁻⁶) to prevent degenerate radii.

Patch boundaries were delineated using a concave silhouette approach by taking unions of discs centered on TLS cells, with disc radius set to 0.95x the local scale estimate and a minimum absolute radius of 1.5. To prevent artificial “bridges” between nearby structures, boundaries were refined using a neck-breaking operation with strength 4.0, followed by contour smoothing with strength 1.0. Clusters with fewer than 2 points were discarded before hull construction. All disconnected polygon parts were retained, but small fragments were filtered (minimum part area 3.0; minimum total patch area 12.0), and highly elongated shapes were excluded using a compactness threshold of 0.01.

Resulting patches were then filtered to retain well-supported TLS structures using subpart-aware thresholds: candidate TLS subparts required at least 100 cells, retained TLS patches required at least 150 total cells, and TLS identity within each patch required ≥20% TLS-labeled cells.

TLS Maturity Classification

TLS maturity was defined based on B-cell enrichment and proliferative activity, reflecting established biological features of tertiary lymphoid structures including follicular B-cell organization and germinal center-like proliferation. For each TLS patch, the fraction of B cells and the proportion of Ki67⁺ B cells were quantified. TLS were classified as immature when B cells comprised <40% of cells within the patch, indicating a lymphoid aggregate lacking clear B-cell follicular organization. Patches containing ≥40% B cells were considered mature TLS, consistent with a B-cell–enriched follicular structure. Among mature TLS, proliferative status was determined using Ki67 expression: patches with <10% Ki67⁺ B cells were classified as quiescent mature TLS, whereas those with ≥10% Ki67⁺ B cells were classified as proliferating mature TLS, reflecting germinal center–like activity. Ki67 positivity was defined using a fluorescence intensity threshold (raw intensity ≥1.0) determined from the distribution of Ki67 signal to distinguish true positive staining from background.

Single-cell RNA-seq data collection, quality control, integration, and annotation.

Raw scRNA-seq datasets for three tumor samples were pre-processed using either CellRanger V5 (patient 004) or V7 (patient 201 and 301) against the GRCh38 human reference genome prior to analysis using Seurat (version 5.1.0) ⁸⁰. Cells were merged into a single object retaining metadata of patient identifier. Cells with fewer than 200 genes or more than 25% of mitochondrial genes were removed. We applied SCTransform to integrate heterogeneous data in our dataset and split the data based on patient identifier. Genes classified as highly variable genes within the dataset were ranked based on frequency and the top 3,500 consensus genes were selected. We removed artifactual or non-informative genes by filtering out long non-coding RNAs and processed pseudogenes based on Ensemble Biomart annotations, and excluded any gene present in a previously published blacklist ²¹ including immunoglobulin, T cell receptor, ribosomal, heat-shock, and mitochondrial genes. Prior to proceeding with principal component analysis (PCA), we further subsetted B cells by retaining only cells with CD79a expression and lacking CD3E, CD3G, CD4, and CD8a expression. Harmony (V1.2.0) was applied after PCA with default parameters ⁸¹ and the harmony matrix used for unsupervised clustering by building a nearest neighbor graph with a resolution of 2. This filtered and integrated object was then annotated using Azimuth (V0.5.0) with the human Tonsil reference ²⁵. To focus on B-lineage populations, we excluded any cells lacking immunoglobulin constant region expression. expression cells enriched for ITGAX expression were further investigated as potential members of an atypical memory B cell pool ^26,27. Two plasma cell clusters were identified by Azimuth and relabeled as “ASC” for antibody-secreting cell. Using this methodology, seven major B cell clusters were identified (naïve B, activated naïve B, cycling B, germinal center (GC) B, memory B (MBC), atypical memory B, and ASCs). These populations were validated using enrichment of canonical markers including TCL1A, YBX3, FCER, and IGHD (naïve-like B cells), CD83, MIR155HG, NFKB1, CD69, and CD83 (activated naïve B cells), MKI67, STMN1, and HMGB2 (cycling B cells), BCL6 and AICDA (GC B cells), CD27 and TNFRSF13B (MBCs), DUSP4, FCLR5, ZEB2 and ITGAX (aka CD11c) (AtM B cells), and MYDGF, JCHAIN, XBP1, and MZB1 (ASCs) (Figure S2D and SE) and using gene set-based analyses against gene sets found in Table S4.

BCR Repertoire Analysis.

Raw scBCR-seq data were assembled and quantified using the Cell Ranger VDJ pipeline against the GRCh38 human reference genome and low-quality B cells removed. Analysis of clonally-expanded plasma cells (as shown in Figure 3A–F and Figure S3) was performed as follows. Clonotypes were defined as variable heavy (VH) and variable light (VL) chain pairings containing exact match productive CDR3 nucleotide sequences. For each clonotype and each CDR3, contigs were assembled to produce a clonotype consensus sequence that was then used for recombinant antibody generation. Analysis of BCR sequences derived from clonally-expanded plasma cells was performed using the scRepertoire R package ⁸² and Cloanalyst software ⁸³.

Analysis of BCR sequences from annotated B cell populations (as shown in Figure 2 and Figure 3G–L) was performed as follows. Using the filtered contig data from CellRanger, the C, V, D, J gene, and CDR3 sequences were annotated using the Change-O toolkit ⁸⁴ with alignments performed using IgBlast ⁸⁵. Non-productive sequences as well as those with multiple heavy chains or no heavy chains were removed from analysis. The alphaDiversity and estimateAbundance functions in Alakazam (V1.3.0) were used to define BCR clonal diversity and abundances, respectively ⁸⁴. A consensus threshold averaging three calculated patient-specific thresholds was calculated at 0.187 and used to cluster BCR sequences with the same junction length and same V and J heavy chain genes ⁸⁶. Hierarchical clustering of heavy chain sequences defined clones using Shazam (V1.2.0). Germline V(D)J sequences were reconstructed using references obtained from IMGT (downloaded 2024–06-20) and somatic hypermutation frequency calculated using the observedMutations function in Shazam ⁸⁶. ScRNA-seq was integrated with paired scBCR-seq data based on matching cell barcodes.

Gene set based analysis.

Curated published gene signatures (detailed in Supplemental Table S4) were imported from a custom gene set file, with each column representing a distinct gene program. Genes present in the Seurat object were retained, and module scores were computed for individual cells using the Seurat AddModuleScore function. Average module scores were calculated for each annotated B-cell population (Naive B, Activated Naive B, Memory B, Atypical Memory B, Germinal Center B, Cycling B, and ASC) and visualized as heatmaps using the pheatmap R package. Gene signatures were displayed in their original order and hierarchical clustering was disabled.

Epitope Model Prediction.

AlphaFold3 ⁴⁴ was used to produce (n = 40) structures of the predicted binding site for both the PC-1 calreticulin complex and the PC-1 p53 complex using eight different seeds with each producing five models. AbEpiTope-1.0, an inverse folding model for antibody-specific B-cell epitopes, was subsequently used to assess the accuracy and identify the structure with the most realistic binding site ⁸⁷. AbEpiScore-1.0 is a learned score that predicts the interface accuracy and AbEpiTarget-1.0 is a score that, across multiple complexes, identifies which antibody most likely binds a given antigen ⁸⁷. For the calreticulin P-domain, the best predicted model was seed 4, model 1 with an AbEpiScore-1.0 of 0.457 and an AbEpiTarget-1.0 score of 0.563, and for p53 core domain, the best predicted model was seed 5, model 0 with an AbEpiScore-1.0 of 0.295 and an AbEpiTarget-1.0 score of 0.481.A.

The predicted epitope was determined at a 5.0Å cutoff between the PC-1 heavy and light chains and the antigen. Within the other 39 calreticulin models, 9 had over 80% epitope homology with seed 4, model 1. Furthermore, epitope homology with seed 4, model 1 changed proportionally with AbEpiScore-1.0 and AbEpiTarget-1.0 scores. This is shown through the second best AbEpiTope model, seed 4, model 3, which had an AbEpiScore-1.0 of 0.415, an AbEpiTarget-1.0 score of 0.557, and had 93.75% epitope homology with seed 4, model 1. A similar pattern was noticed for the other 39 p53 core domain models, where 6 models had over 80% epitope homology with seed 5, model 0. Here, the second best AbEpiTope model, seed 2, model 1, had an AbEpiScore-1.0 of 0.232, an AbEpiTarget-1.0 score of 0.402, and had 85.2% epitope homology with seed 5, model 0. As a whole, the p53 core domain models had worse AbEpiTope scores, which is reflected in their epitope homology with the top model, compared to the calreticulin models. Overall, these findings further reinforce the validity of the predicted epitopes.

Statistical analysis.

Statistical analyses were conducted using GraphPad Prism version 10 software. Survival analyses were performed using Kaplan-Meier estimation and log rank tests for comparison. For comparisons of cell survivability in cytotoxicity assays significance was determined by two-way repeated measures ANOVA. For comparisons of tumor growth in mouse models significance was determined by two-way repeated measures ANOVA, followed by a post-hoc Šídák’s multiple comparisons test.

Supplementary Material

1

Document S1. Figures S1–S7, Table S2. Table S7.

2

Table S1. Clinical characteristics of studied patients. Related to Figures 1 and 4.

3

Table S3. Quantitative data from analysis of TLSs. Related to Figure 1.

4

Table S4. Gene lists for gene set based analyses. Related to Figure 2.

5

Table S5. Protein list from PC-1 immunoprecipitation and mass spectrometry analyses. Related to Figure 4.

6

Table S6. Protein list from PC-1 proteome microarray screening. Related to Figure 4.

Key resources table.

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Goat anti-human IgG for ELISA	Thermo Fisher	#A18815
AlexaFluor-488 anti-human IgG	Invitrogen	#A11013
CH65 (anti-flu hemagluttinin mAb)	Saunders Laboratory	Lot 277MRM
4E10 (anti-HIV gp41 mAb)	Polymun	Lot: T531002-A
17b (anti-HIV gp120 mAb)	Saunders Laboratory	Lot: 195EAM
anti-rabbit IRDye680RD	LI-COR	#926-68071
goat anti-human IRDye88CW	LI-COR	#926-32232
Functional anti-human CD3 Antibody (OKT3)	BioXcell	# BE0001-2
Anti-human CD3 antibody for flow cytometry	Biolegend	Clone UCHT1
Anti-human CD4 antibody for flow cytometry	Biolegend	Clone OKT4
Anti-human CD8 antibody for flow cytometry	Biolegend	Clone SK1
Anti-human CD14 antibody for flow cytometry	Biolegend	Clone HCD14
Anti-human HLA-DR antibody for flow cytometry	Biolegend	Clone L243
Anti-human CD45 antibody for flow cytometry	Biolegend	Clone HI30
Anti-human FAP antibody for flow cytometry	Thermo Fisher	Clone F11-24
Mouse IgG2b isotype control	Biolegend	Clone 27-35
Mouse IgG1 control	Biolegend	Clone MG1-45
Anti-acetyl-alpha tubulin (Lys40) (Clone: 6-11B-1)	Thermo Fisher Scientific	#32-2700; RRID: AB_2533073
Anti-Annexin A1 (Clone: MRQ-3)	Cell Marque	#221M-14; RRID: AB_1516714
Anti-Arginase 1 (Polyclonal)	Novus	#NBP1-32731; RRID: AB_10003985
Anti-alpha smooth muscle actin / α-SMA (Polyclonal)	Abcam	#ab5694; RRID: AB_2223021
Anti-Aquaporin 5 / AQP5 (Polyclonal)	Thermo Fisher Scientific	#PA5-36529; RRID: AB_2553573
Anti-beta-catenin / β-catenin (Clone: 14)	Cell Marque	#224M-14; RRID: AB_1158971
Anti-BCL2 / B-cell lymphoma 2 (Clone: 124)	Cell Marque	#226S; RRID: AB_1158101
Anti-CD3 (Clone: MRQ-39)	Cell Marque	#103R; RRID: AB_2864399
Anti-CD4 (Clone: EPR6855)	Abcam	#ab181724; RRID: AB_2864377
Anti-CD8 (Clone: C8/144B)	Thermo Fisher Scientific	#MA5-13473; RRID: AB_11000353
Anti-CD11b / Integrin alpha M / ITGAM (Clone: EPR1344)	Abcam	#ab209970; RRID: AB_2915959
Anti-CD15 / Lewis X / SSEA-1 (Clone: MMA)	BD Biosciences	#559045; RRID: AB_397181
Anti-CD20 (Clone: rIGEL/773)	Novus	#NBP2-54591-100ug; RRID: AB_2864380
Anti-CD25 / IL-2 receptor alpha (Clone: 4C9)	Cell Marque	#125M; RRID: AB_2864401
Anti-CD31 / PECAM-1 (Clone: C31.3+C31.7+C31.10)	Novus	#NBP2-47785-0.1mg; RRID: AB_2864381
Anti-CD34 (Clone: QBEnd/10 + HPCA1/763)	Novus	#NBP2-47909-0.1mg; RRID: AB_2864382
Anti-CD36 (Clone: D8L9T)	Cell Signaling Technology	#14347; RRID: AB_2798458
Anti-CD38 (Clone: E7Z8C)	Cell Signaling Technology	#43382; RRID: AB_3675778
Anti-CD45 / Leukocyte common antigen / PTPRC (Clone: 2B11 + PD7/26)	Novus	#NBP2-34287; RRID: AB_3285704
Anti-CD56 / NCAM1 (Clone: MRQ-42)	Cell Marque	#156R; RRID: AB_2864402
Anti-CD57 (Clone: HCD57)	BioLegend	#322325; RRID: AB_2563757
Anti-CD68 (Clone: KP1)	BioLegend	#916104; RRID: AB_2616797
Anti-CD73 / NT5E (Polyclonal)	R&D Systems	#AF5795; RRID: AB_1964487
Anti-CD90 / Thy-1 (Clone: EPR3132)	Abcam	#ab92574; RRID: AB_10563647
Anti-CD138 / Syndecan-1 (Clone: DL-101)	BioLegend	#352302; RRID: AB_10915555
Anti-CD163 (Clone: EDHu-1)	Novus	#NB110-40686; RRID: AB_714951
Anti-CD206 / Mannose receptor / MRC1 (Polyclonal)	R&D Systems	#AF2534; RRID: AB_2063019
Anti-cytokeratin 5 / CK5 (Clone: EP1601Y)	Abcam	#ab214586; RRID: AB_3717733
Anti-cytokeratin 10 / CK10 (Clone: SPM623)	Novus	#NBP2-53134; RRID: AB_3332345
Anti-cytokeratin 14 / CK14 (Clone: SPM263)	Novus	#NBP2-34355; RRID: AB_3285753
Anti-Cytokeratin, pan reactive (Clone: C-11)	BioLegend	#628602; RRID: AB_439775
Anti-Decorin / DCN (Polyclonal)	Thermo Fisher Scientific	#PA5-27370; RRID: AB_2544846
Anti-EGF Receptor / EGFR (Clone: D38B1)	Cell Signaling Technology	#5588; RRID: AB_10694773
Anti-FAP alpha / Fibroblast activation protein alpha (Clone: EPR20021)	Abcam	#ab271976; RRID: AB_3676372
Anti-FOXP3 / Forkhead box P3 (Clone: 236A/E7)	Thermo Fisher Scientific	#14-4777-82; RRID: AB_467556
Anti-Ki67 (Clone: B56)	BD Biosciences	#556003; RRID: AB_396287
Anti-LYVE1 / Lymphatic vessel endothelial hyaluronan receptor 1 (Polyclonal)	R&D Systems	#AF2089; RRID: AB_355144
Anti-Lysozyme (Clone: BGN/06/961)	Abcam	#ab36362; RRID: AB_776115
Anti-Mast Cell Tryptase / TPSAB1/TPSB2 (Clone: EPR9522)	Abcam	#ab151757; RRID: AB_2909419
Anti-MMP9 / Matrix metalloproteinase-9 (Clone: L51/82)	BioLegend	#819701; RRID: AB_2564833
Anti-MUC5AC / Mucin 5AC (Clone: CLH2)	Novus	#NBP2-44455-0.1mg; RRID: AB_2864386
Anti-p53 / Tumor protein p53 (Clone: D07)	Cell Marque	#453M; RRID: AB_2864403
Anti-p63 / Tumor protein p63 (Clone: D2K8X)	Cell Signaling Technology	#13109; RRID: AB_2864410
Anti-PD-1 / Programmed cell death protein 1 (Clone: D4W2J)	Cell Signaling Technology	#63815; RRID: AB_3675993
Anti-phospho-Histone H2AX (Ser139) / γH2AX (Clone: EP854(2)Y)	Abcam	#ab215967; RRID: AB_2864385
Anti-Podoplanin / PDPN (Clone: D2-40)	BioLegend	#916605; RRID: AB_2565819
Anti-SFTPC / Surfactant protein C (Polyclonal)	Thermo Fisher Scientific	#PA5-119603; RRID: AB_2913176
Anti-TRPC6 / Transient receptor potential cation channel subfamily C member 6 (Polyclonal)	Proteintech	#18236-1-AP; RRID: AB_10859822
Anti-Vimentin (Clone: RV202)	Novus	#NBP1-97672; RRID: AB_2915945
Bacterial and virus strains
E. coli BL21(DE3) expression strain	New England Biolabs	C2527
Biological samples
TOP1501 clinical specimens – tumors and serum/plasma (NCT02818920).	Duke University BioRepository and Precision Pathology Center	https://brpc.duke.edu/
Human PBMCs	STEMCELL technologies	https://www.stemcell.com/products/human-peripheral-blood-mononuclear-cells-frozen.html
Human plasma	STEMCELL technologies	https://www.stemcell.com/products/human-peripheral-blood-serum-fresh.html
Cryopreserved viable single cell suspensions of freshly dissociated human ovarian carcinomas	Conejo-Garcia laboratory	De-identified
Cryopreserved viable single cell suspensions of human pleural effusions from patients with NSCLC	Conejo-Garcia laboratory	De-identified
Chemicals, peptides, and recombinant proteins
Streptavidin-HRP	Pierce	#21132
CellMask Deep Red Plasma Membrane Stain	Invitrogen	#C10046
4′,6-diamidino-2-phenylindole (DAPI) Fluouromount-G	SouthernBiotech	#0100-20
Enzyme Free Cell Dissociation Buffer	ThermoFisher	#13151014
Recombinant Vimentin Protein	Cayman Chemical	#11234
Recombinant Citrullinated Vimentin Protein	Cayman Chemical	#21942
Recombinant p53 Protein	Abcam	#ab237007
Recombinant Calreticulin Protein	Abcam	#ab91577
Recombinant HSP90α Protein	R&D Systems	#AP-160-050
Roche Complete Protease Inhibitor Cocktail	Sigma	#11697498001
IL-2	PeproTech	Cat# 200-02
Recombinant PAD2	Cayman Chemical	#10785
DAPI	Biolegend	# 422801
Zombie Yellow Fixable Viability Kit	Biolegend	# 423103
Ficoll-Paque gradient	GE Healthcare Systems	#17-1449-02
Critical commercial assays
Recombinant mAb production	GenScript, TurboCHO transfection	https://www.genscript.com/custom-antibody-production-services.html
Human Protein Microarray	CDI labs	https://www.cdilabs.com/services/huprot-microarray Microarray barcode CDI91333288
Human Tumor Dissociation Kit	Miltenyi Biotec	#130-095-929
Immunoscan CCPlus test kit	SVAR Life Science	#RA-96PLUS
AntheNA Multi-Lyte ANA kit	Zeus Scientific	#A21101
Dynabeads Antibody Coupling Kit	ThermoFisher	#14311D
Dynabeads™ Human T-Activator CD3/CD28 for T Cell Expansion and Activation	Thermo Fisher	# 11131D
Pierce BCA Protein Assay Kit	ThermoFisher	#23227
Deposited data
Raw and analyzed scRNAseq and scBCRseq data	This paper	BioProject: PRJNA1292770
PC-1 fab structure	This paper	Figure 5, PBD: 9ZQF
PC-1 IP-Mass spectrometry	This paper	Table S5
PC-1 protein microarray	This paper	Table S6
Experimental models: Cell lines
Calu-6	ATCC	HTB-56
A549	ATCC	CCL-185
H1703	ATCC	CCL-5889
KPM	Conejo-Garcia laboratory
OvCar3	ATCC	HTB-161
Hec 1a	ATCC	HTB-112
A427	ATCC	HTB-53
UPK10	Conejo-Garcia laboratory
Phoenix-AMPHO	ATCC	CRL-3213
293T	ATCC	CRL-3216
Expi293F cells	ThermoFisher	A14527
Experimental models: Organisms/strains
NOD.Cg-Prkdcscid Il2rgtm1Wjl/SzJ	Jackson Laboratories	#:005557
Oligonucleotides
Human PADI2-specific CRISPR guide RNA	IDT DNA	Custom RNA oligo
Recombinant DNA
pBMN-I-GFP	Addgene	# 1736
His-SUMO p53 (pET-32a)	GeneImmune Biotechnologies	Uniprot:P04367
His-SUMO p53 core domain (pET-32a)	GeneImmune Biotechnologies
Calreticulin-10xHis (pVRC8400)	GeneImmune Biotechnologies	Uniprot:P27797
CALR-NP domain	GeneImmune Biotechnologies
CALR-P domain	GeneImmune Biotechnologies
CALR-NC domain	GeneImmune Biotechnologies
PC-1 heavy chain (pVRC8400)	GeneImmun Biotechnologies
PC-1 light chain (pVRC8400)	GeneImmun Biotechnologies
PC-1 heavy chain His-Fab	GeneImmun Biotechnologies
Software and algorithms
Image J	Fiji	https://imagej.net/software/fiji/
CellRangerV5 and CellRanger V7	10x Genomics	https://www.10xgenomics.com/support/software/cell-ranger/downloads
Seurat v5.1.0	Hao et al. 80	https://satijalab.org/seurat/h
Harmony v1.2.0	Korsunsky et al. 81	https://github.com/immunogenomics/harmony
Azimuth v0.5.0 tonsil reference	Massoni-Badosa et al. 25	https://azimuth.hubmapconsortium.org/
scRepertoire	Borcherding et al. 82	https://github.com/BorchLab/scRepertoire
Cloanalyst	Ramesh et al. 83	https://www.bu.edu/computationalimmunology/research/software/
Change-O toolkit	Gupta et al 84	https://changeo.readthedocs.io/en/stable/
IgBlast	Ye et al. 85	https://www.ncbi.nlm.nih.gov/igblast/
Alakazam v1.3.0	Gupta et al. 84	https://alakazam.readthedocs.io/en/stable/
Shazam v1.2.0	Ye et al. 86	https://github.com/immcantation/shazam
FlowJo		https://www.flowjo.com/
GraphPad Prism		Version 11.0.1 (90)
BiaControl/BiaEvaluation v3.2.2	Cytiva	https://www.cytivalifesciences.com/en/us/products/items/biacore-t200-software-upgrade-31-p-05916
HKL2000 release 722	Otwinowski et al. 73	https://hkl-xray.com/hkl-2000
PHENIX v2.1	Adams et al. 75	Phenix-online.org
AbEpitope-1.0	Clifford et al. 87	https://services.healthtech.dtu.dk/services/AbEpiTope-1.0/
Alphafold3	Jumper et al. 44	Alphafoldserver.com
Other
Single cell sequencing data	SRA database	Accession #: PRJNA1292770
PC-1 Fab Crystal Structure	Protein Data Bank	PDB#: 9ZQF
PC-1 mass spectrometry raw data, metadata, database search results	ProteomeXchange Consortium, MassIVE Repository	ProteomeXchange Accession#: PXD079756 MassIVE Repository Accession#: MSV000102166
PC-1 human protein (HuProt) microarray data	Zenodo	https://doi.org/10.5281/zenodo.21015949

Highlights.

• PD-1 blockade is associated with tumor-specific antibody development.

• Single-cell BCR mining identifies tumor-selective antibody lineages.

• Tumor-binding antibodies target citrullinated antigens on cancer and tumor promoting myeloid cells.

• PC-1 CAR T cells target tumors and tumor-promoting myeloid cells without off-tumor toxicity.