Association of immune-related expression profile with sensitivity to chemotherapy in esophageal squamous cell carcinoma

2.1 Patients, neoadjuvant chemotherapy and response evaluation

Details of patients enrolled in this study were described in our previous study.⁵ Briefly, this cohort study included ESCC patients who received DCF-NAC followed by surgery at Oita University Hospital between June 2013 and October 2016. In total, 39 patients met the following inclusion criteria: (i) histological diagnosis of primary ESCC; (ii) stage IB/II/III according to UICC 7th edition⁸; (iii) age ≤80 years; (iv) performance status 0–1; and (v) no previous chemotherapy, thoracic radiotherapy (RT), or thoracic surgery. ESCC samples were collected by biopsy during endoscopic examination before administration of the first course of chemotherapy. The biopsy specimens were each collected from an elevated part at the proximal side of the tumor in a uniform manner. Matched non-tumor tissues (at least 5 cm from the tumor site) from five of the 39 cases were also collected as control samples. The specimens were then frozen and preserved in a freezer maintained at −80°C. This study was approved by the Ethics Committee of Oita University Faculty of Medicine (Approval No. 1081), and all patients included in the study provided written informed consent.

The DCF-NAC regimen consisted of a 1-h i.v. infusion of docetaxel (70 mg/m²) on day 1, a 2-h infusion of cisplatin (70 mg/m²) on day 1, and continuous i.v. infusion of 5-FU (750 mg/m²) on days 1–5. This regimen was administered every 3 weeks, and three scheduled courses were administered before esophagectomy. Surgery was scheduled to be carried out within 4–6 weeks after the last day of preoperative chemotherapy, when curative resection was considered possible. Clinicopathological characteristics are shown in Table 1.

The pathological response was evaluated according to the Japanese Classification of Esophageal Cancer 11th edition as follows: grade 0, no recognizable cytological or histological therapeutic effect; grade 1a, viable cancer cells accounting for two-thirds or more of the tumor tissue; grade 1b, viable cancer cells accounting for between one-third and two-thirds of the tumor tissue; grade 2, viable cancer cells accounting for less than one-third of the tumor tissue; grade 3, no viable cancer cells apparent (pCR).^{6, 7} Patients were divided into highly resistant (n = 12; grade 0 and 1a) and sensitive (n = 27; grade 1b, 2, and 3) groups according to the pathological response (Table 2).

2.2 Gene expression microarray and data analysis

Details are described in our previous study.⁵ Frozen specimens were homogenized, and total RNA was extracted using a QIAGEN RNeasy mini kit (QIAGEN) in accordance with the manufacturer's protocol. Total RNA (200 ng) was reverse transcribed to cDNA using murine leukemia virus reverse transcriptase (Invitrogen Crop.).

A human 8 × 60 K whole-genome oligo DNA microarray chip (SurePrint G3 Human Gene Expression v3 Microarray Kit, G4851C, Agilent Technologies, Santa Clara, CA) was used for global gene expression analysis, in accordance with the manufacturer's protocol. Cyanine (Cy)-labeled cRNA was prepared using T7 linear amplification, in accordance with the Agilent Low RNA Input Fluorescent Linear Amplification Manual (Agilent Technologies). Labeled cRNA was fragmented and hybridized to the oligonucleotide microarray (Agilent Technologies). The fluorescence intensities were determined with an Agilent DNA Microarray Scanner and analyzed using Feature Extraction v.10.7.3.1 (Agilent Technologies). Expression levels were converted to log₂ values and normalized to the median of the entire spot array using GeneSpring GX11 (Agilent Technologies). All data are available at the Gene Expression Omnibus (GEO) via NCBI under Accession No. GSE225178 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE225178). The probe set data were filtered based on signal intensity and flagged values, resulting in 33,632 remaining probes. Genes showing differential expression between the three groups were identified by one-way ANOVA (p < 0.05) and post hoc testing correction (Tukey). The Benjamini and Hochberg correction was used for false-positive reduction (multiple testing correction). The genes differentially expressed between the highly resistant and sensitive groups were identified as DEGs, and these comprised 204 genes. Hierarchical clustering with Euclidean distance was performed to stratify patients according to their expression profiles of the 204 DEGs. GeneSpring GX11 was used for these statistical and clustering analyses.

DAVID, a web bioinformatic resource system (https://david.ncifcrf.gov/home.jsp)^{9, 10} was used to reveal enriched functional annotations among the DEGs. The data from the Functional Annotation Chart with the default setting was expressed in a table format and modified.

To identify the molecular signatures in each sample group, we performed GSEA (https://www.gsea-msigdb.org/gsea/index.jsp).^{11, 12} For this, microarray data for the 33,632 filtered probes in 39 cancer samples were ranked based on differential expression and compared with the Hallmark gene sets, composed of predefined 50 gene sets. Enriched gene sets with an FDR of less than 0.05 were considered significant.

To estimate the proportion of immune cell infiltration in each sample, microarray data were analyzed using CIBERSORTx (https://cibersortx.stanford.edu/index.php)^13-15 with a validated LM22, representing the expression profiles of 547 immune-related genes in 22 leukocyte lineages.¹³ We set the job parameters for Batch correction, Disable quantile normalization, Run mode and Permutations as “disabled,” “true,” “absolute,” and “100,” respectively.

The impact of the expression of the top 50 DEGs on patient survival was investigated using the UALCAN web tool (http://ualcan.path.uab.edu/)^{16, 17} with an HNSCC dataset from TCGA. We used the HNSCC dataset because all of our cases were squamous cell carcinoma while the EC dataset in TCGA included many adenocarcinomas. Furthermore, fluoropyrimidine-based and platinum-based chemotherapies are also standard in HNSCC.¹⁸ High-expression patients showed more than one-third quartile expression levels.

2.3 RNAScope in situ hybridization

The distribution of CXCL9 mRNA on formalin-fixed paraffin-embedded sections of biopsied tissues was investigated using RNAScope in situ hybridization in accordance with the manufacturer's instructions (Advanced Cell Diagnostics).

2.4 Immunohistochemistry

After using RNAScope for analysis of CXCL9, the distribution of epithelial cells and macrophages was detected by immunohistochemistry using antibodies against pan-cytokeratin (clones AE1/AE3, DAKO) or CD68 (clone KP1, Arigo Biolaboratories), respectively. The sections were visualized using Histofine Simple Stain AP (Nichirei) and Perma Red substrate (Diagnostic Biosystems) in accordance with the manufacturer's instructions.

2.5 Statistics

All statistical information is indicated in the figure legends for each set of data.

3.1 Involvement of immune-related genes in sensitivity to DCF chemotherapy

In this study, we re-analyzed our previous dataset (Case ID: 1–35) after the addition of the expression profiles for five non-tumor tissues (Case ID: 25 N–35 N) and seven tumor tissues (Case ID: 36–43; Table 2). The expression profiles of, in total, 44 samples were then divided into three groups; non-tumor group (5 cases), sensitive group (27 cases, therapeutic effect grades 1b, 2, and 3) and highly resistant group (12 cases, therapeutic effect grades 0 and 1a). Genes differentially expressed among the three groups were identified by one-way ANOVA (p < 0.05) and post hoc testing correction (Tukey) (Figure 1A). The Benjamini and Hochberg correction was used for false-positive reduction (multiple testing correction). We found that 4347 and 6282 probes, including redundant and non-coding genes, exhibited differential expression in the highly resistant and sensitive groups compared with the non-tumor group, respectively (Figure 1A). We identified 204 probes, corresponding to 172 genes, as DEGs between the sensitive and highly resistant groups (Figure 1A and Table S1). Among the 204 probes, the top 50 DEGs showing the most differential expression are listed in Table 3. Interestingly, most of these DEGs were upregulated in the sensitive group (184 upregulated and 20 downregulated). To investigate the impact of expression of the top 50 DEGs on patient survival, we used the UALCAN online tool for prognostic analysis in TCGA HNSCC datasets. We chose the HNSCC datasets because all of our cases were squamous cell carcinomas while the EC dataset in TCGA included many adenocarcinomas. Furthermore, fluoropyrimidine-based and platinum-based chemotherapies are also standard for HNSCC.¹⁸ Among the top 50 DEGs, 28 showed a tendency for higher expression to be associated with better prognosis (p < 0.25; Table 3), and this association was significant (p < 0.05) for 18 of them (Figure S1), suggesting that these genes could be used as biomarkers of patient survival as well as sensitivity to chemotherapy. Although the sample size was small, we also investigated the impact of the top 50 DEGs on patient survival in our cohort. We found that most of the genes showed a tendency for higher expression to be associated with better outcomes (data not shown), although the p-value was less than 0.3 for only 10 of those genes (CXCL13, LGALS2, linc-LRRC10-1, IKZF3, SEMA5B, MIR4697HG, ZNF831, PTPRC, TIGIT, and LAMP3; Table 3). Furthermore, in the DEG list, there are some genes functionally related to immune response such as chemokines, complement components and Toll-like receptors (Table 3). For example, CXCL13 is a chemokine that preferentially attracts B lymphocytes in comparison with T cells and macrophages.¹⁹ CXCL9 and CXCL10 are induced by IFNγ and their increased expression is associated with infiltration of activated T cells through their cognate receptor, CXCR3,²⁰ which is also listed in Table 3. LGALS2 (galectin 2) was recently reported to be an immunotherapy target in triple-negative breast cancer.²¹ Granzyme K (GZMK) is one of the serine proteases from cytoplasmic granules of cytotoxic lymphocytes, such as CD8⁺ T cells and NK cells.²² As shown in Figure 1B and Table S1, most of the immune-related genes, including CXCL13, CXCL9, CXCL10, LGALS2, GZMK, CD8B, CYBB, IRF8, C1QB, TLR4, TIGIT, and CXCR3, were specifically upregulated in the sensitive group compared with the non-tumor group, while the expression levels of these genes were less changed in the resistant group. As a consequence, most of the resistant (11/12) and non-tumor (5/5) samples were clustered in the same branch, while seven of the 10 cases showing a complete response (therapeutic effect grade 3) were clustered in a different branch from resistant and non-tumor samples in hierarchical clustering using the expression profiles of 204 DEGs (Figure 2A). DAVID analysis indicated that the DEGs were strongly associated with the term “immunity” (Figure 2B). GSEA also showed that the Hallmark gene sets of “allograft rejection” and “interferon γ response,” both of which are related to immune response, were enriched in the sensitive group (Figure 2C). None of the gene sets in the Hallmark gene sets (ver. 7.5) were enriched in the highly resistant group (data not shown). Taken together, our data suggested that the immune response-related expression profile in preNAC biopsy samples is associated with sensitivity to DCF chemotherapy.

To exclude the possibility that biopsy samples in the highly resistant group accidentally contained far fewer tumor cells than the sensitive group, we compared the expression of genes upregulated in HNSCC, chosen based on the data set from GEPIA2,²³ between the non-tumor, highly resistant and sensitive groups in our microarray data (Table S2). We chose HNSCC datasets for the same reason as that described above. Among the eight most upregulated genes in the HNSCC data set, six were also significantly upregulated in tumor tissue relative to non-tumor tissue in our microarray data (Table S2). Furthermore, all of the six genes were upregulated in both the resistant and sensitive groups, and showed no significant difference in expression levels between the two groups (Figure S2), suggesting it was unlikely that biopsy samples from the highly resistant group contained far fewer tumor cells than the sensitive group.

3.2 Potential involvement of M1 macrophage infiltration in sensitivity to DCF chemotherapy

Because our data implicated immune cell infiltration in tumor tissues of DCF-NAC-sensitive cases, fractions of 22 immune cell types in each sample tissue were estimated using CIBERSORTx, which is a web bioinformatic tool that can assess the proportions of various types of infiltrating immune cells. The estimated-total absolute scores and absolute fractions of 22 immune cell types were compared between the non-tumor, highly resistant, and sensitive groups (Figure 3 and Table S3). The estimated-total absolute score was significantly higher in both of the tumor groups than in the non-tumor group (p = 0.02; and 0.0015 in the highly resistant and sensitive groups, respectively) but showed no significant difference between the highly resistant and sensitive groups (p = 0.14; Figure 3A,B), suggesting an immune-expression profile in both of the tumor groups. The estimated absolute cell fractions of CD4⁺ memory-activated T cells (p = 0.039) and M1 macrophages (p = 0.0052) in the sensitive group were significantly higher in those in the highly resistant group (Figure 3C), indicating possible infiltration of these immune cell types in the sensitive group. In particular, M1 macrophages exhibited the most significant difference between the highly resistant and sensitive groups (p = 0.0052), suggesting that M1 macrophage infiltration or expression of its related genes could be diagnostic biomarkers of DCF-NAC sensitivity.

3.3 Distribution of CXCL9-expressing cells in DCF-NAC-sensitive tumor tissues

Consistent with our microarray data, CXCL9 and CXCL10, two of the most upregulated genes in the sensitive group (Figure 1B and Table 3), were highly expressed in M1 macrophages among the 22 leukocyte types in the LM22 expression profile,¹³ which is often referred to as a representative profile of 22 individual leucocyte lineages in CIBERSORTx. However, in contrast with CIBERSORTx estimation, a few of the previous studies reveal overexpression of CXCL9 in cancer epithelial cells as well.^{24, 25} Therefore, we next sought to determine the source of CXCL9 by in situ hybridization using two sensitive cases with high CXCL9 expression (cases 11 and 33) and two highly resistant cases with low CXCL9 expression (cases 20 and 25) in the microarray data (Figure 4A). Consistent with the microarray data, CXCL9 signals were clearly detectable in both two sensitive samples but barely detectable in the resistant samples (Figure 4B). We also investigated the distribution of CXCL13 using two sensitive cases with high CXCL13 expression (cases 28 and 37) and two highly resistant cases with low CXCL13 expression (cases 5 and 38) in the microarray data, and demonstrated consistency of the data between microarray and in situ hybridization (Figure S4). CXCL13 was clearly expressed in the stromal cells.

Next, we compared the distribution of CXCL9 with that of an epithelial cell marker (anti-cytokeratin antibody) or a macrophage marker (anti-CD68 antibody) by double staining using RNAScope and immunohistochemistry. Most of the CXCL9 signals were not detected in cytokeratin-positive cells, although only a small proportion of cytokeratin-positive cells expressed CXCL9 (Figure 5A,B). In contrast, many, but not all of the CD68-positive cells expressed CXCL9 (Figure 5C). To determine whether or not each of the CXCL9-positive macrophages are really M1 macrophages, further studies will be required. Interestingly, we also found that some CD68-negative stromal cells (non-macrophage stromal cells) also expressed CXCL9 (Figure 5D–F). Considering that previous reports have suggested T cells as the source of CXCL9,^{25, 26} it is possible that T cells in the biopsy samples also expressed CXCL9. Further studies are required to address this possibility. Taken together, our data suggest that CXCL9 is expressed mainly by tumor-surrounding stromal cells, including macrophages, and also slightly expressed by a small proportion of cancer cells, while another stromal cell type(s) not identified here could also be involved in CXCL9 expression.