Peritumoral B cells drive proangiogenic responses in HMGB1-enriched esophageal squamous cell carcinoma

Within tumor tissue, the spatial characteristics of immune cells are one of the important determinants of immune cell function. We used a tissue microarray (TMA) containing 89 paired tumor/adjacent normal formalin-fixed paraffin-embedded (FFPE) ESCC tissues to evaluate the density and spatial distribution of CD20⁺ B cells by multiplex immunohistochemistry (mIHC). Using the tissue segmentation function, the intra-(tumor or epithelial) region was defined by cytokeratin-5 (CK5⁺ epithelial cells) and the peri-(tumor or epithelial) region in stroma comprised CK5⁻ non-epithelial cells (Fig. 1a). We observed a higher density of B cells in stroma than in intra-epithelial locations in normal tissues (90%; 79/87, two cases were excluded due to a deficit of data on cytokeratin-5; Fig. 1b). Similar findings were observed in tumors where B cells were frequently found in peritumoral location instead of intratumoral region (88%; 79/89; Fig. 1c). For 15 of the 89 study patients, CD20 immunostaining was not evident in the tumor cores (Fig. 1d). For the 74 cases with peritumoral B-cell infiltration, we examined the proliferation marker Ki67 in B cells (hereafter referred to as proliferating B cells). The proportion of proliferating B cells in peritumoral compartment was significantly higher than that in corresponding normal tissues (median: 24.16 ± 4.50 vs. 13.59 ± 3.18, respectively, Fig. 1e–f and Supplementary Fig. 1a).

To further assess the clinical significance of non-T cell (CD3⁺) abundance, we evaluated the density of tumor-infiltrating lymphocytes (TILs) derived from a cell suspension (via flow cytometry) in four paired resected tumors and adjacent cancer-free normal tissues. The analysis showed that the percentages of total CD3^– cells were significantly higher than CD3⁺ cells, (2.71-fold for normal; p = 0.0267 vs. 3.68-fold for tumor; p = 0.0185), but no significant differences were found in the frequencies of total CD3⁺ and CD3^– cells between tumor and normal tissue (median: 18.48 ± 12.28 vs. 25.13 ± 7.51 and 67.93 ± 12.89 vs. 71.00 ± 6.99, respectively; Supplementary Fig. 1b). This finding was confirmed by RNA sequencing (RNA-seq) of three paired ESCC tumor and normal tissues (Supplementary Fig. 1c), indicating that CD3^– immune cells (non-T cells) are enriched in ESCC irrespective of whether tissues are malignant or benign. We previously reported that 22 types of TILs were identified in The Cancer Genome Atlas (TCGA) database of ESCC patients [8]. Here, we estimated the proportions of distinct immune cell subpopulations in our RNA-seq data using CIBERSORT and the frequencies of different B-cell subsets (naïve and plasma cells) were found to be higher in normal tissue compared to tumor (Supplementary Fig. 1d). We also determined that unswitched memory B cells could be a predominant subset present within the population of proliferating tumor-infiltrating B cells (TIL-Bs; Supplementary Fig. 1e–f). In contrast with the previously described higher proliferating B-cell densities in peritumoral regions than in adjacent normal tissue, the total density of TIL-Bs and proliferating TIL-Bs in tissue mass was lower in tumor tissue than normal tissue (Supplementary Fig. 1g–h).

Tumor-derived HMGB1 has been previously reported to resuscitate macrophage function in patients with ESCC [36]; we here aim to assess the clinical relevance of HMGB1 in B cells. We performed chromogenic staining of two TMAs incorporating total 125 cancer cases. Archival tissues from a total of 118 paired tumor/adjacent normal cases were selected to evaluate HMGB1 expression (Supplementary Fig. 2a). As expected, HMGB1 was detected in the intratumoral locations of tumor samples (Fig. 2a). We performed quantitative scoring based on HMGB1 staining intensity by Fuji software and found that HMGB1 expression was significantly higher (p < 0.001) in tumor relative to normal tissues, representing about 86% (102/118) of all ESCC-paired cases (Fig. 2b and Supplementary Fig. 2b). From patients with tumor RNA-seq data and an additional validation set of 38 unpaired ESCC samples, we further confirmed that HMGB1 mRNA expression was frequently upregulated in tumor tissues compared to the non-tumor counterpart (Supplementary Fig. 2c, d, respectively).

To investigate whether the presence of HMGB1 within the tumor nest could have an impact on predicting patients’ survival, we stratified patient samples into low or high variance groups. Using the median fold change in HMGB1 expression as the cut-off point (Tumor to Normal ratio, T/N = 1.374), patients were divided into a high HMGB1 expression group (T/N fold change > 1.374) and a low HMGB1 expression group (T/N fold change < 1.374). The Kaplan–Meier curves showed that HMGB1 expression was not an independent prognostic factor for overall survival (OS) in ESCC (p = 0.865, Fig. 2c). Also, age, sex, pathologic T and N stages, and tumor differentiation showed no significant correlation with HMGB1 expression (Supplementary Table 1).

We next evaluated the relationship between peritumoral proliferating B cells and intratumoral HMGB1 expression and their impact on patients’ survival. We divided the density of peritumoral proliferating B cells and intratumoral HMGB1 expression into groups of high/low based on the medians as cut-offs and observed a significant correlation (p = 0.0244) between these two variables (Fig. 2d). As shown in Fig. 2d, a high proportion of proliferating B cells and high expression of HMGB1 were found in 49% (62/125) and 53.6% (67/125) of cases, respectively. Of the cases that were highly positive for proliferating B cells, 24.8% (31/125) were also highly positive for HMGB1. A positive association was observed between intratumoral HMGB1 and peritumoral proliferating B cells, but not with total B cells density (Supplementary Fig. 2e). Using these two markers (HMGB1 high/low and proliferating B cells high/low) in combination, patients were divided into four groups (Supplementary Table 2). There was no significant difference between the groups with respect to their clinicopathologic characteristics. Strikingly, by Kaplan–Meier analysis, the OS of cases that were highly positive for both proliferating B cells and HMGB1-expressing tumor cells was shorter than for those with low expression of both factors (p = 0.045; Fig. 2e). Overall, the findings suggest that intratumoral HMGB1 may interact with peritumoral proliferating B cells to modulate ESCC biology.

To understand how B cells are regulated within the HMGB1-enriched tumor nest, we initially investigated if HMGB1 binds to the surface of B cells. We found that HMGB1 is bound to the surface of B cells, with an average Mander’s overlap coefficient of 0.914 and Pearson correlation coefficient of 0.79 (Supplementary Fig. 3a). We next provided evidence that HMGB1 acts directly as a proliferative signal for B cells (Supplementary Fig. 3b). These findings prompted us to examine the functions of cancer-derived HMGB1 on B cells. Stable cell lines overexpressing HMGB1 were established using the EC18 and K510 cell lines (thereafter referred as EC18H and K510H; Supplementary Fig. 3c–e) and co-cultured with peripheral blood mononuclear cells (PBMC) or freshly isolated CD20⁺ B cells (Supplementary Fig. 3f). The carboxyfluorescein succinimidyl ester (CFSE)-labeled PBMC (Supplementary Fig. 4a) and B cells (Fig. 3a–b and Supplementary Fig. 4b) were co-cultured with HMGB1-overexpressing (EC18H and K510H) or vector control (EC18pc and K510pc) cells for 6 days. We observed that the proliferation rate of B cells co-cultured with HMGB1-overexpressing tumor cells was higher than that of their vector control cells. Similar findings, although to a lesser extent, were observed on B cells in co-culture with tumor cells for 72 h, as revealed by Ki67 expression (Supplementary Fig. 4c). Our subsequent in vitro analyses therefore utilized 6 days co-culture systems to further characterize the mechanistic link between B cells and cancer-derived HMGB1, as well as their potential roles in orchestrating tumorigenesis and progression in ESCC.

In the majority of clinical samples, B-cell aggregates were enriched peritumorally (i.e., in the stroma at the invasive border of tumor nests; Supplementary Fig. 4d). This may suggest that B cells are attracted toward HMGB1-expressing tumor front and we consequently focused on the chemotactic capacity of the B cells in response to HMGB1. Transwell experiments were used to study migration of B cells in response to recombinant HMGB1 (rHMGB1; Supplementary Fig. 4e–f) or in the presence of HMGB1-overexpressing tumor cells (Fig. 3c). Increased chemotactic response toward the presence of HMGB1 was observed in B cells, suggesting that cancer-derived HMGB1 could be an important factor favoring B-cell retention within the TME.

To analyze the impact of HMGB1 on migratory B cells, RNA sequencing on B cells migrated toward rHMGB1 or a medium control revealed that 76 genes were differentially expressed, of which 66 genes were upregulated. Among those 66 upregulated genes, the gene ontology analysis outcomes confirmed cell migration as one of the central processes for the HMGB1-educated migratory B cells. Most notably, angiogenesis was an unexpected category from the top ten enriched biological processes identified in enhanced migrated B cells, based on biological processes level 5 (Fig. 4a).

Next, we investigated the functional roles of HMGB1-transmigrated B cells on endothelial cells (ECs). Human umbilical vein endothelial cells (HUVECs) were overlaid on Matrigel and cultured with migrated or non-migrated B cells toward rHMGB1/medium alone, and the angiogenic ability of HUVECs was examined by tube formation and migration assays (Fig. 4b). As shown in Fig. 4c, enhanced EC capillary tube and network formation were observed on the Matrigel containing migrated B cells toward medium containing rHMGB1 (subgroup 4), as compared to medium alone (subgroup 2; increased by 1.62-, 1.33-, and 1.49-fold, with respect to number of branches, junctions, and total length of segments). To exclude the effect of direct contact between HUVECs and B cells on angiogenesis, HUVECs were incubated with conditioned medium (CM) from B cells pre-treated with/without rHMGB1. Figure 4d reveals that CM from rHMGB1-pre-treated B cells enhanced HUVEC angiogenesis in vitro. Also, no significant difference was observed in activated B cells without the stimuli of HMGB1, suggesting that only activated B cells in response to HMGB1 provoke tumor angiogenesis.

We demonstrate here that HMGB1-conditioned B cells promote angiogenesis in HUVECs in vitro. However, the TME is complicated. We therefore further investigated whether ESCC-derived HMGB1 could drive B cells into a proangiogenic state. To exclude the effect of cell contact, B cells were co-cultivated with tumor cells via transwell assay. qRT-PCR analysis of genes implicated in angiogenesis process demonstrated that angiogenic factors, VEGF, IL8, TGFβ, and MMP9 were increased in the B cells co-cultured with HMGB1-overexpressing tumor cells as compared to control cells (Fig. 5a). VEGF is an important angiogenic factor and B cells are the possible source of this growth factor. In fact, immunocytochemical analysis (Fig. 5b) revealed upregulated VEGF production in B cells when co-cultured with HMGB1-overexpressing tumor cells.

Next, we performed flow sorting according to CFSE signals and confirmed that low CFSE signal (high proliferative signal) attributed to increased VEGF expression in B cells (Supplementary Fig. 5a). Quantification of VEGF and Ki67 expression by FACS in B cells confirmed the increased frequencies of proliferating VEGF⁺ B cells upon co-culture with HMGB1-overexpressing tumor cells (Fig. 5c and Supplementary 5b). Accumulation of Ki67⁺VEGF⁺ B cells was observed in the peritumoral area of ESCC tissue (Fig. 5d). Further analysis revealed that 51.2% (22/43) of the cases showed peritumoral proliferating VEGF⁺ B cells (Fig. 5e and Supplementary 5c), where the proportion of VEGF⁺ B cells subsets in total B cell counts was 56.3%, in which 14.3% was proliferative cells. Importantly, when CD31⁺ blood vessels are present, CD20⁺ B cells tended to accumulate around blood vessels rather than distributed evenly throughout tumor tissues, with the majority of B cells being < 10 µm from a blood vessel (Fig. 5e–f and Supplementary 5e). We considered two possible explanations for the close proximity of B cells to vasculature: 1) blood vessels are an important route for B cells trafficking to tumor sites or 2) B cells within the TME are critical for, or associated with, angiogenesis.

The striking observation that soluble HMGB1 derived from tumor cells mediates the production of proliferating VEGF⁺ B cells, prompted us to ask if suppressing HMGB1 could interfere with this proangiogenic phenotype. We challenged this hypothesis by studying the crosstalk among tumor, B cells, and HUVECs through the use of CM. In brief, we first neutralized the secretion of HMGB1 from tumor cells by a direct HMGB1 antagonist, glycyrrhizin (GL; Fig. 6a). We found no angiogenic ability of HUVECs in the absence of B cells (Fig. 6b), suggesting that HMGB1 derived from ESCC cells alone was not sufficient to elicit tumor angiogenesis in vitro. We next analyzed the effect after co-culture of tumor with B cells in the presence or absence of GL. The administration of GL significantly inhibited capillary-like tube formation, including total mesh area (99.26% and 90.90% reduction in EC18H/B and K510H/B, respectively) and total length of segments (94.14% and 52.36% reduction in EC18H/B and K510H/B, respectively) (Fig. 6c). To better mimic in vivo blood vessel formation, we conducted HUVEC spheroid sprouting assays in 3D culture (Supplementary Fig. 6a). The results showed that the average sprout length in HUVECs treated with CM from co-culturing B cells with HMGB1-overexpressing tumor cells at 72 h was significantly longer (3.716 ± 5.062 μm/spheroid) than in those treated with control medium (1.692 ± 2.535 μm/spheroid). The spheroid sprouting abilities of HUVECs were suppressed by the presence of GL (Supplementary Fig. 6b,c). Likewise, CM from co-culturing B cells with GL-treated HMGB1-overexpressing tumor cells significantly reduced HUVEC transmigration via transwell assay (77.42% and 52.03% reduction in EC18H/B and K510H/B, respectively; Fig. 6c) and live-cell imaging (Supplementary Fig. 6d). In addition, qRT-PCR analysis revealed that the expression of genes implicated in vascular growth (CD31, VEGFR, and ENDOTHELIN-1) was induced in HUVECs after incubation with CM from EC18H/B and K510H/B co-cultures. This effect was dramatically reversed by the addition of GL (Fig. 6d). The observed angiogenic behavior was specific for HMGB1 since comparable in vitro neutralizing efficiency was obtained with the administration of an anti-HMGB1 monoclonal antibody (anti-HMGB1-mAb; Supplementary Fig. 7a,b). Together, our data highlight the relevance of tumor cell-derived HMGB1 in inducing proangiogenic B-cells activities.

Next, we demonstrated that ERK and p38 MAP kinase were activated (by phosphorylation) and involved in the promotion of VEGF expression in B cells via tumor-derived HMGB1 and the induction of VEGF was attenuated by the presence of GL (Supplementary Fig. 7c). To validate these results, we investigated the angiogenic components in the CM after co-culture with HMGB1-overexpressing tumor cells, using human angiogenesis arrays. As shown in Fig. 6e, several angiogenic factors were induced when B cells were co-cultured with tumor cells, and their levels were further upregulated in co-culture with HMGB1-overexpressing tumor cells. Treatment with GL significantly suppressed the secretion of several proangiogenic factors, such as VEGF, thrombospondin-1, IL-8, and monocyte chemoattractant protein-1. By contrast, the anti-angiogenic factor serpin E1 was restored by GL. Finally, we assessed whether VEGF was the main angiogenic factor produced by B cells upon HMGB1 stimulation. To this end, we investigated the chemotactic migration and tube formation in HUVECs via VEGF blockade (Supplementary Fig. 7d). A significant reduction of HUVEC functions was observed upon anti-VEGF antibody administration (Supplementary Fig. 7e). These data unequivocally demonstrate that cancer-derived HMGB1-induced angiogenesis in B cells is dependent on VEGF.

To better mimic the in vivo situation of microenvironmental interplay between HMGB1-expressing tumor cells and B cells, a mixture of the indicated cells was subcutaneously injected into NOD/SCID mice (Fig. 7a). As shown in Fig. 7b, tumor growth in HMGB1-overexpressing cells was similar to the vector control group (labeled as EC18H in blue and EC18pc in green, respectively). Compared with tumors in mice injected with EC18H tumor cells alone, higher tumor growth rate was observed in mice injected with B cells mixed with HMGB1-overexpressing tumor cells (labeled as EC18H/B in red). Moreover, the tumors of EC18H/B mice had more blood vessels, were larger in size (an average of 212% larger; Fig. 7c), and showed increased CD31 staining in tumor tissue compared with mice injected with vector cells mixed with B cells (labeled as EC18pc/B in black; Fig. 7d). To determine whether the different number of blood vessels attributed to tumor angiogenesis was due to the presence of proangiogenic B cells, the B cells from tumors were FACS sorted based on Epcam (epithelial cell marker), CD45 (immune cell marker), and CD20 (B-cell marker) (Supplementary 7f). We observed higher mRNA expression of VEGF and IL-8 in the B cells isolated from the EC18H/B group than the control vector EC18pc/B group (Fig. 7e). An enriched VEGF⁺ B-cell phenotype was further demonstrated by immunofluorescence (Fig. 7f).

To ascertain whether soluble factors, rather than direct cell–cell contact, contribute to angiogenesis, we focused on the CM collected from co-culture of EC18H/B in the absence or presence of GL/anti-HMGB1. As shown in Fig. 8a and b, mice in the anti-HMGB1 mAb group (labeled EC18H + B + αHMGB1 in green) and GL group (EC18H + B + GL in purple) had smaller tumor volumes and weights in comparison to their control tumors (labeled EC18H + B in black). Although treatment with GL alone (labeled EC18H + GL in pink) reduced basal tumor growth (compared with EC18H in blue), we showed that the addition of anti-HMGB1-mAb, and to a greater extent with GL, reversed the HMGB1-induced B-cell activation in tumor growth to a basal level (150% and 215% reduction in the effect of B cells mixed with EC18H, Fig. 8c). Consistently, loss of HMGB1 function in the GL-treatment group reduced the expression of genes associated with angiogenesis (Fig. 8d).

Collectively, our results suggest that B cells within the TME, being primed by HMGB1 released from cancer cells, migrate to the tumor. Here, they proliferate and activate a proangiogenic phenotype, leading to the activation of endothelial cells and ultimately tumor progression (our tentative synthesis is depicted in Fig. 8e).

ESCC cell lines KYSE140 (K140), K180, K410, K510, and K520 were obtained from DSMZ, the German Resource Center for Biological Material. ESCC cell lines EC18, HKESC1, and immortalized esophageal epithelial cell line NE1 were provided by Professors G. Srivastava and G.S. Tsao (The University of Hong Kong). HUVECs were provided by Dr Stephanie Ma (The University of Hong Kong). All experiments were done using endothelial cells between passages 3 and 8.

Four TMAs (two slides each for TMA-1 and TMA-2) composed of samples from a total of 125 ESCC patients from the surgical pathology archives of Linzhou Cancer Hospital (Henan, China) were used for histological staining. Of these, 89 paired ESCC and corresponding normal tissue samples from TMA-1 were used to evaluate B-cell spatial analysis by mIHC. One hundred and eighteen matched normal/tumor pairs from a consecutive slide of TMA-1 (n = 89) and TMA-2 (n = 29) were chosen to perform HMGB1 chromogenic immunohistochemistry. All tumor cases positive for B cells (n = 74 from TMA-1 and n = 51 from TMA-2) were chosen for cooperative biomarkers prognostic analysis. A summary of cases used for TMA is shown in Supplementary Fig. 2a.

Three paired ESCC tissues were used for RNA sequencing. An additional 38 paired ESCC tissues were used for qRT-PCR analysis. No patient in the study had received preoperative radiation or chemotherapy. Studies using human tissues were approved by the committees for ethical review of research involving human subjects of Zhengzhou University (Zhengzhou, China) and the Institutional Review Board of The University of Hong Kong/Hospital Authority Hong Kong West Cluster (HKU/HA HKW IRB).

Seven ESCC patients who had undergone upfront esophagectomy at the Department of Surgery, Queen Mary Hospital, Hong Kong were also recruited for tissue dissociation and subsequent flow cytometric analysis.

PBMC were isolated from healthy donors (buffy coats, Hong Kong Red Cross Blood Transfusion Service) for subsequent culture.

The tumor and normal tissue samples were processed into single-cell suspensions by mechanical disaggregation followed by enzymatic digestion using 1.5 μg/mL collagenase IV (Roche), 0.8 mg/mL dispase (Invitrogen), and 0.1 mg/mL DNase I (Sigma). Tissues were incubated in digestion medium at 37 °C for 30 min; released cells were collected and filtered, and the remaining tissue was further processed.

PBMC were then re-suspended in chilled MACS buffer (PBS, 0.5% FCS and 2 mM EDTA), washed and incubated with CD20 microbeads (130-091-104; Miltenyi Biotec) for 15 min at 4 °C, and subsequently passed through magnetic separation columns (LS; Miltenyi Biotec). The bead-bound cells were collected as enriched, positively selected B cells. Purified CD20⁺ B cells were either re-suspended in FACS buffer for flow cytometry or in DMEM supplemented with 10% FCS, 50 IU/mL penicillin–streptomycin, and 10 mM HEPES buffer (Gibco/Invitrogen) for subsequent culture. Preparations were typically > 95% pure.

Directional migration of B cells was evaluated in Costar Transwell permeable polycarbonate supports (5 μm pores) in 24-well plates. HMGB1-overexpressing tumor cells/different concentration of recombinant HMGB1 (rHMGB1) were used to compare with parental cell lines/medium alone. B cells (0.25 × 10⁶ cells/ml) pre-treated with 100 ng/mL recombinant IL-4 and 5 µg/mL IgM for 24 h were washed, placed in the top chamber, and allowed to migrate for 4 h at 37 °C. After that, cells in the bottom chambers were collected, stained with FITC-labeled anti-CD20 antibody, and the number of CD20⁺ B cells was calculated by flow cytometry.

Supernatants from co-cultured cells were collected. Total IgG/M were detected using human IgG/IgM ELISA quantitation set (Bethyl Laboratories) according to the manufacturer’s instructions.

PBMC/purified B cells (1 × 10⁶/well) were labeled with 5 µM CFSE (Biolegend) in PBS/0.1% BSA for 8 min at 37 °C. Unbound dye was quenched by washing three times with ice-cold complete medium, followed by stimulation with 100 ng/mL recombinant IL-4 and 5 µg/mL IgM. 24 h later, pre-stimulated CFSE-labeled PBMC/B cells (1 × 10⁴) were washed and mixed with ESCC cell lines in 96-round bottom plate at effector-to-target ratios of 10:1. Proliferation of PBMC/B cells was monitored by CFSE partitioning 6 days post-co-culture. The total B-cell population from PBMC, stained with anti-CD20-APC antibodies, was analyzed with a FACSCanto II. In some experiments, pre-stimulated B cells were treated in the presence or absence of rHMGB1 (10 ng/mL) or anti-VEGF (10 µg/mL; Sino Biological), cells were collected for flow cytometric analysis, and conditioned medium (CM) was collected for subsequent functional assays.

CM was collected and centrifuged at 2000×g for 10 min. CM was either used undiluted for characterization by protein array or concentrated for an in vivo mouse model by centrifugation at 4000×g for 15 min at 13 °C, using ultrafiltration units (Amicon Ultra-PL 10, Millipore, Bedford, MA, USA). Filter units were used only once to avoid membrane saturation. Concentrated CM were then sterilized on 0.22 μm filters (Millipore), aliquoted, and stored at –80 °C until use.

HUVECs were maintained in endothelial cell growth medium M200 (Invitrogen) supplemented with 2% FBS and endothelial cell growth supplements (LSGS Medium). 2 × 10⁴ cells were re-suspended in 100 µL co-culture CM (1:5; diluted with serum-free M200) and seeded in 96-well plates pre-coated with 50 µL of growth factor-reduced Matrigel (BD Bioscience). The cells were incubated at 37 °C for 6–8 h to allow for the formation of tube-like structures. Enclosed networks of tube structures from three fields were photographed randomly in each well and quantified using ImageJ analysis.

Spheroids containing 6000 HUVECs were generated by incubating suspended cells in M200 in 96-well Corning® Spheroid Ultra-Low Attachment Microplates overnight, after which they were embedded into 3 mg/mL Matrigel for a fixed 3D cell culture. In brief, 70 µL hydrogel solution containing 3 mg/mL Matrigel and cells were dispensed into each well followed by 20 µL of CM. The plate was then warmed to 37 °C to induce gelation. Images of the spheroid within the polymerized gels were captured using InCell6500 (Perkin Elmer). Endothelial sprouts were characterized by measuring average branch lengths using Fiji distribution of ImageJ. Briefly, images were converted to 8-bit grayscale then converted to binary images with appropriate threshold values. Parameters for the binary mask, such as area and perimeter, were analyzed using the ‘Analyze Particles’ function. For end-point calculation, the areas covered by the spheroids were traced followed by ‘Skeletonize’ function of Fiji for analysis.

Starved HUVECs (2 × 10⁴) were seeded in the top chambers of Transwell plates (8 µm pore size) in 400 µL of M200 medium without serum. The bottom chambers were filled with 400 µL co-culture CM (1:5 dilutions). After 24 h, cells were fixed, stained with 1% crystal violet, and counted. For gene expression assays, HUVECs were plated in 24-well plates (5 × 10⁴cell/well) incubated with CM (1:5 dilutions) for 8 h and 24 h. For the μ-slide chemotaxis assay (Ibidi), HUVECs were cultured on 2 µg/cm² Matrigel-coated μ-slides and allowed to adhere for 3 h. One reservoir was filled with M200/2% FBS/LSGS and the second reservoir with the indicated 1:5 diluted CM collected from the co-cultures. Directional migration was assessed after 16 h and tracked with the aid of time-lapse microscopy.

The relative levels of human angiogenesis-related proteins in CM were measured using a Human Angiogenesis Array Kit (R&D Systems Inc.). Aliquots of CM (500 µL) were added to the array and the results were analyzed with the ImageJ software.

TMA sections were digitally scanned at an absolute magnification of × 20 using the Vectra 3.0 Vectra Polaris imaging system (Akoya Biosciences) and analyzed with inForm Tissue Finder software (Akoya Biosciences). Multispectral images were unmixed using the spectral libraries built from images of single-stained slides. Firstly, 15 TMA cores were selected to train machine learning algorithms for tissue segmentation, cell segmentation, and cell phenotyping, which were later applied on the whole TMA cohort. The software was first trained to segment tissue by manually to segment tumor tissues into carcinoma, the intra- (epithelial), and stromal, the peri-(non-epithelial), areas based on tumor marker, cytokeratin-5 (CK-5). To detect immune cells, an algorithm was designed based on pattern recognition that quantified CD20, VEGF, and Ki67 cells. After which the distribution of immune cells was analyzed and cell segmentation was based on the nuclear DAPI stain but assisted using membrane CD20 staining. Training sessions for tissue segmentation and phenotype recognition were carried out repeatedly until the algorithm reached the level of confidence recommended by the program supplier (at least 90% accuracy) before performing final evaluation. Each scanned image was examined by one observer under the supervision of an experienced pathologist. The area of each tissue category, carcinoma, and stroma was evaluated to assess the density of lymphocytes, represented by (number of lymphocytes)/(pixel area mm²) in each tumor cores. Spatial relationships between cellular phenotypes CD20⁺ and CD31⁺ cells in the peritumoral area were determined using the phenoptrReports package (Akoya Biosciences). Then, the distribution between CD20⁺ and CD31⁺ cells was identified in consecutive 10 µm steps (distance classes) within 200 µm.

HMGB1 immunohistochemistry was quantified in tissues using ImageJ (Ver. 1.52b). As tumorous samples consisted of mixed cell types including tumor cells and immune cells, image segmentation was first performed to isolate tumor and non-tumor regions. This allowed us to determine the regions and boundaries between the tumor cells and surrounding stroma. A supervised training of the Trainable Weka Segmentation (TWS) from Fiji (ver. 3.2.3) [56] was performed. A training set of five randomly selected 1604 × 1604 pixel images consisted of features from three classes: tumor, non-tumor, and background. The feature set used for training included a total of 80 attributes. A multithreaded implementation of random forest classifier with 200 trees and two features per node was used to build the model. The model was then applied to all images for classification. Representative results of segmentation are presented in Fig. 1A. Quantification of staining intensity of HMGB1 in tumorous or normal tissue were performed by first obtaining average pixel intensity of the identified tumor region in the tumorous samples or average intensity of the healthy tissue samples, respectively, followed by background subtraction.

K510 and EC18 ESCC cell lines were maintained in RPMI and DMEM supplemented with 10% FBS and antibiotics, respectively. To generate stable HMGB1-overexpressing clones, cells were transfected by lipofectamine 2000 with pcDNA3.1 vector. Stable cell lines were selected by adding 500 µg/mL G418 for EC18 tumor cells and 100 µg/mL for K510 tumor cells.

ESCC overexpressing the HMGB1 gene (K510H or EC18H) or vector (K510pc or EC18pc) were seeded into a 24-well plate (5 × 10⁴/mL/well) and used for co-culture experiments at 80% confluence. CD20⁺ B cells (1 × 10⁶ cells) were pre-treated with 5 µg/mL IgM and 100 ng/mL IL-4 for 24 h, washed, and co-cultured with the relevant ESCC cell line (5 × 10⁴ cells/well) through microporous cell inserts (1 µm pore size) in a 24-well plate for 6 days. Where indicated, the HMGB1 inhibitor GL (0.5 mM; Sigma), anti-HMGB1 monoclonal antibody (αHMGB1; 10 µg/mL; Arigo), or DMSO were added to ESCC cell lines before co-culture.

Total RNA was isolated using the RNeasy Mini kit (Qiagen). RNA was reversed transcribed (Takara) into cDNA, which served as a template for the amplification by qRT-PCR using the SYBR Green gene expression assay (Applied Biosystems 7900). Relative quantification was measured using the comparative Ct (threshold cycle) and the Ct values were normalized to β-actin or GAPDH, where appropriate. Primers used are shown in Supplementary Table 3.

CD20⁺ B cells isolated from two healthy donor PBMC samples were subjected to a migration study for 24 h. B cells that migrated toward recombinant HMGB1 or medium only as a control were collected and RNA was isolated for sequencing using the SMART-Seq™ v4 Ultra™ Low Input RNA Kit. PCR products were amplified and sequenced on an Illumina HiSeq™ 2500 platform by Novogene (Beijing China). High-quality clean reads from all two samples were merged together and mapped to the reference sequence. To determine the biological significance of the differentially expressed genes, which were defined as genes with log₂ expression fold change ≥ 0.5,or ≤ − 0.5, functional classification and gene enrichment analysis were performed using GO Term (Biological Process level 5) with DAVID Bioinformatics Resources. Top ten highly enriched functional categories were listed and arranged in descending order of p-value of enrichment.

B cells collected from the transwell of co-cultured systems were washed and re-suspended in FACS buffer (PBS, 0.5% BSA, and 2 mM EDTA). Cells were stained with antibodies to the surface markers CD19, CD27, and CD38 (BD Biosciences) for 30 min on ice, followed by intracellular staining for Ki67 and VEGFA (BD Biosciences) using a BD Cytofix/Cytoperm kit. Antibodies were diluted FACS buffer for surface staining and in BD Perm/Wash buffer for intracellular staining. Cells were analyzed on a NovoCyte Quanteon and the data were analyzed using FlowJo Software.

For co-implanting tumor cells with B cells, 6-week-old NOD/SCID mice were irradiated at 300 cGy, then 4 h later, purified CD20⁺ B cells pre-treated with IgM and IL-4 were mixed with tumor cells (1 × 10⁶ HMGB1-overexpressing/empty vectors in 100 µl of PBS, n = 6 per group) in growth factor-reduced Matrigel (BD Biosciences) at a 5:1 ratio, and then implanted subcutaneously into mice under anesthesia. B cells or tumor cells alone served as controls. Tumor growth was monitored every 2 days for the indicated time. Excised tissues were divided into portions for mIHC and FACS. To increase the yield of B cells, two tissues from the same group were pooled together for enzymatic dissociation and B cells were sorted based on Epcam (epithelial cell marker), CD45 (immune cell marker), and CD20 (B-cell marker), followed by qRT-PCR analysis. A loss-of-function analysis was performed by subcutaneous injection with 5 × 10⁶ tumor cells. Five days later, mice were treated with 50 μL CM collected from co-culture by intratumoral injection for consecutive 5 days. The control group received CM from co-culture in the absence of GL. Tumors were harvested to prepare paraffin tissue sections for immunofluorescent staining.

Differences in quantitative variables were analyzed by the Mann–Whitney U test when comparing two groups and by the Kruskal–Wallis with Dunn’s post hoc test when comparing more than two groups. All analyses were performed using GraphPad Prism software. Survival curves were generated according to the Kaplan–Meier method, and statistical analysis was performed using the log-rank test. The association between HMGB1 expression and clinicopathological characteristics was tested by Pearson’s Chi-Square test. p value < 0.05 was considered statistically significant.