A total of 62 formalin-fixed paraffin-embedded (FFPE) samples from 31 Japanese patients histologically diagnosed with CAS based on biopsy or surgical resection samples were obtained at Nagoya City University Hospital. The cohort is summarized in Supplementary Table 1 . This is the cohort of our previous report with one additional lymph node metastasis sample (4). Immunostaining was performed in this cohort. Comprehensive RNA sequencing using next generation sequencer (NGS) was performed on 15 samples from 7 cases selected from this cohort ( Table 1 ). Patients with primary and recurrent lesions were selected, including one case with multiple recurrences. The cohort included 6 men and 1 woman with a median age of 73.29 (range 62-81) years. 6 cases (85.7%) occurred in the head and neck, and 1 case (14.3%) occurred in the trunk. According to the pathological differentiation classification, 12 samples were highly differentiated, and 3 samples were poorly differentiated. These patients were treated with one or a combination of the following: surgery, radiation therapy, taxanes, eribulin, or pazopanib.

RNA extraction and sequencing were performed as previously described (8). Tumor tissue was carefully dissected from 3 to 10 undyed FFPE tissue sections (7 µm thickness) using a scalpel blade and deparaffinized in 640 µl deparaffinization solution (Qiagen, Hilden, Germany). According to the supplier’s instructions, total RNA was refined using an AllPrep DNA/RNA FFPE Kit (Qiagen). The concentration and quality of the extracted RNA were evaluated by Qubit RNA HS Assay Kit using Qubit 3.0 Fluorometer and by a functional RNA quantitation (FRQ) assay using StepOne Plus (Thermo Fisher Scientific, Waltham, MA). RNA samples confirmed to be of sufficient quality were reverse-transcribed to cDNA samples, which were amplified using Ion AmpliSeq Library Kit 2.0 (Thermo Fisher Scientific). Ion Library TaqMan Quantitation Kit measured the concentration of library DNA samples using ABI7500FastDx (Thermo Fisher Scientific). Amplified library DNA samples were applied to the NGS using Oncomine Immune Response Research Assay (Thermo Fisher Scientific). NGS analysis was performed using the Ion S5 XL System (Thermo Fisher Scientific). Data were analyzed on the software application Transcriptome Analysis Console and Gene spring analysis (Thermo Fisher Scientific). All data were uploaded to the national center for biotechnology information (NCBI) gene expression omnibus (GEO) database (GSE203215).

Gene set enrichment analysis (GSEA) was performed using the c5 Gene Ontology gene set collections as provided by the Molecular Signatures Database (MSigDB) (9) and GSEA software (https://www.gsea-msigdb.org/gsea/) (10).

FFPE tissue sections of 62 CAS samples were processed for indirect immunofluorescence to detect the expression of signal transduction proteins using primary antibodies: anti-NY-ESO-1 antibody (1:100, ab223498, Abcam, Cambridge, UK), anti-XAGE-1 antibody (1:130, ab134805, Abcam), anti-PD-L1 antibody (1:100, ab205921, Abcam), anti-CD3 antibody (1:25, ab17143, Abcam), anti-CD8 antibody (1:50, ab17147, Abcam), and anti-CD20 antibody (1:50, ab78237, Abcam). Bound antibodies were visualized with the appropriate secondary antibodies (1:500, Alexa Fluor 488, Alexa Fluor 594, Invitrogen, Waltham, Massachusetts, USA) at 37°C for 30 min at 1:500 dilution with 5% goat serum. 4’,6-diamidino-2-phenylindole (Vector Laboratories, Burlingame, California, USA) was used as a counterstain. The green fluorescence produced by Alexa 488, red fluorescence produced by Alexa 594, and blue fluorescence produced by 4’,6-diamidino-2-phenylindole were observed and captured using a fluorescence microscope BZ-X800 (Keyence, Osaka, Japan). TLSs were identified as clusters of 10 or more CD20-positive cells surrounded by CD3-positive cells approximately half the circumference, and the number of TLSs in a section was counted. The fluorescence intensities of NY-ESO-1 (CTAG1B), XAGE-1, and PD-L1 were calculated from binarized images after channeling into red, blue, and green using ImageJ software (NIH, Bethesda, Maryland, USA) from 10 randomly selected fields as previously described (11). The values higher than the mean of 62 samples of 31 cases were defined as “high”, and those lower were defined as “low”. CD8-positive cells were counted in several locations with a high infiltrating cell density. Samples with at least one CD8-positive cell in a 200x high-power field (0.40mm² field area) were defined as ‘CD8-positive’, as previously described (4).

NGS data were analyzed on the software application Transcriptome Analysis Console and Gene spring analysis (Thermo Fisher Scientific). A clustered heatmap of all samples was generated using the online tool iDEP.94 (http://bioinfomatics.sdstate.edu/idep/). A paired t-test was used to compare primary and recurrent lesions in immunohistochemistry results. Disease-specific survival was calculated as the time elapsed from definite diagnosis to death from CAS and analyzed using the Kaplan-Meier method and log-rank test. Statistical analyses were performed using Graph Pad Prism 8 (Graph Pad Software, San Diego, CA, USA). Probability values (p-values) of less than 0.05 were considered statistically significant.

The characteristics and the immunochemical statuses of the 15 samples are summarized in Table 1 . Hierarchical cluster analysis of the top 25 differentially expressed genes (DEGs) between TLS positive and TLS negative groups (total 395 genes) was performed ( Figure 1A ). NY-ESO-1 (CTAG1B) is the only gene that significantly upregulated in the TLS positive group (log₂ fold-change [FC]=4.426 p-value=0.0296) ( Figure 1B ). There were no other CTAs upregulated in log₂FC>4.4, p-value<0.1.

PD-L1 was the upregulated gene in the TLS-negative group with the smallest p-value among 395 genes (log₂FC=-6.707, p-value=0.0006, Figure 1B ). Hierarchical cluster analysis of top 25 DEGs between TLS≥5 and TLS=1-4 groups (total n=395 DEGs) was performed ( Figure 1C ). PD-L1 and PD-1 were upregulated in TLS=1-4 group (n=5) compared to TLS≥5 group (n=6). (PD-L1, log₂FC=9.427, p-value=0.0033, PD-1, log₂FC=6.335, p-value=0.0373) ( Figure 1D ). Immune checkpoint molecules seem to be deeply involved in the suppression of the TLS formation. Gene set enrichment analysis (GSEA) was performed using the Gene Ontology (GO) resource including 10, 192 gene sets. Compared between TLS-positive and TLS-negative group, 5 gene sets were upregulated in TLS-positive group, and 13 gene sets were upregulated in TLS-negative group, with a criterion of p-value < 0.05. The top four ranked gene sets with the highest and smallest normalized enrichment scores (NES) with a p-value<0.05 are shown in Figure 1E . ‘T cell-mediated cytotoxicity’ set, ‘Antigen binding’ set, ‘Negative regulation of defense response’ set, and ‘Positive regulation of cell killing’ set were upregulated in TLS positive group, suggesting immune activation.

For comparison of the primary and recurrent lesions, hierarchical cluster analysis between primary and recurrent lesions was performed. The top 40 DEGs in the primary or recurrent lesions are shown in Figure 2A . CTAs, including XAGE-1B, MAGEA4, GAGE1, GAGE2C, GAGE12J, GAGE13, and SSX2 gene were upregulated in the recurrent lesions. Only XAGE-1B gene was significantly upregulated in the recurrent lesions compared to primary lesions. (log₂ FC=8.042, p-value=0.0346) ( Figure 2B ). Total of 12 genes, such as CD44, TLR3, EGFR, FAS, and CCL20, were significantly upregulated in the primary lesions. The top four ranked gene sets with the highest and smallest NES with a p-value<0.05, which were revealed by GSEA, were shown in Figure 2C . CTAs accounted for 7 of the top 20 genes that were upregulated in recurrent lesions. However, their role remains unclear.

Eight patients who had both primary and recurrent lesions sampled were selected to analyze intra-patient heterogeneity of potential biomarkers with immunohistochemical staining. The first recurrent lesion was picked up in a patient with multiple recurrent lesions. Biomarkers, including PD-L1 expression, the number of infiltrating CD8-positive cells, and the number of TLSs have changed dramatically in the recurrent lesions from the primary lesions. PD-L1 expression in recurrent lesions was significantly suppressed compared to primary lesions (paired t-test, p-value=0.0310) ( Figure 3A ). There was no significant tendency in the number of CD8-positive cells (paired t-test, p-value=0.5830) ( Figure 3B ) or TLSs (paired t-test, p-value=0.6095) ( Figure 3C ). The number of TLS varied widely in two cases, but only one case changed from nonexistent to present, and none of the cases had TLS disappeared later. Representative immunostaining images for PD-L1 ( Figure 3D ), CD8 ( Figure 3E ), and TLSs ( Figure 3F )were shown.

Additional analysis was performed on 28 chemokines included in the Oncomine Immune Response Research Assay. The number of TLS and the expression levels of chemokines analyzed by clustering analysis for each sample are shown as a heat map ( Figure 4A ). The only chemokine correlated with the number of TLS was CCL21 (liner regression, p-value=0.0273, Figure 4B ). Two chemokines, CCL21 (p-value=0.0037) and CXCL8 (p-value=0.0242), positively correlate with NY-ESO-1 expression in immunohistochemical staining ( Figures 4C, D ). And no chemokines correlated with XAGE-1 expression.

NY-ESO-1 and XAGE-1 are detectable by immunohistochemistry. NY-ESO-1 pixel values greater than 74.8 (average of 62 samples) were defined as “high” (14 cases) and those less than 74.8 were defined as “low” (17 cases). Representative images of each are shown in Figures 5A, B . Kaplan-Meier curve showed that expression of NY-ESO-1 in the primary lesions does not significantly correlate with the prognosis (n=31, log-rank, p-value=0.5527) ( Figure 5C ). XAGE-1 pixel values greater than 98.2 were defined as “high (18 cases) and those less than 98.2 were defined as “low” (13 cases). Representative images of each are shown in Figures 5D, E . Kaplan-Meier curve showed that expression of XAGE-1 in the primary lesions does not significantly correlate with the prognosis (n=31, log-rank, p-value=0.2559, Figure 5F ). These CTAs were significantly different in each group only when stratified by presence or absence of TLS. The combination of TLS and NY-ESO-1 showed no difference in high or low NY-ESO-1 expression in TLS-positive cases, but the NY-ESO-1 positive group tended to have a better prognosis in TLS-negative cases (n=31, log-rank, p-value=0.021, Figure 6A ). In the combination of TLS and XAGE1, the prognosis was better in the group of TLS-positive cases with low XAGE1 levels (n=31, log-rank, p-value=0.027, Figure 6B ).