Patients with esophageal cancer at clinical stages T1b-T4a and N0 or N+ were evaluated for surgery by a multidisciplinary team that included a medical oncologist, a radiation oncologist, a radiologist, and a thoracic surgeon. The surgical candidates (n = 40) underwent neoadjuvant chemoradiotherapy. Radiation therapy was delivered with a dose of 4300–4400 cGy in daily 210–215 cGy per fraction over 4 to 5 weeks. Concurrent chemotherapy of 5-FU (4000 mg/m² over 4 days) and cisplatin (60 mg/m² on day 1) were administered every 3 weeks for up to two cycles during RT. Endoscopic biopsies were performed at the time of initial diagnosis and 2 to 3 weeks after the start of preoperative CCRT (Fig. 1a). This genomic analysis study was approved by the institutional review board of the Samsung Medical Center (IRB no. SMC-2013-10-112) and written informed consent was obtained from all enrolled patients.Fig1Fig. 1

Study flow and CONSORT diagram of the study population. a The first endoscopic sample was obtained at the time of diagnosis. The second sample was acquired at 2 to 3 weeks after initiation of concurrent chemoradiation therapy. b, Samples were acquired from 40 patients with signed informed consent and prepared for genomic analyses. Twenty-nine patients were available for the paired whole-exome sequencing (WES), and 23 patients were available for the paired whole-transcriptome sequencing (WTS). A total of 25 patients underwent curative surgery. Abbreviations: SCC, squamous cell carcinoma; QC, quality control

Genomic DNA (gDNA) and RNA were purified from cancer tissues using the AllPrep DNA/RNA Mini Kit (Qiagen, USA). gDNA from peripheral blood was extracted using the QIAamp DNA Blood Mini Kit (Qiagen). gDNA concentration and purity were measured using a NanoDrop 8000 UV-Vis Spectrometer (Thermo Scientific Inc., USA) and a Qubit 2.0 Fluorometer (Life Technologies Inc., USA). To measure DNA degradation, median DNA size and ΔCt values were measured using a 2200 TapeStation Instrument (Agilent Technologies, USA) and real-time polymerase chain reaction (PCR; Agilent Technologies), respectively. RNA concentration and purity were measured using the NanoDrop and Bioanalyzer (Agilent Technologies).

High quality gDNA in each sample was sheared with an S220 ultra-sonicator (Covaris, USA) and used to construct a library with the SureSelect XT Human All Exon v5 and SureSelect XT reagent kit, HSQ (Agilent Technologies), according to the manufacturer’s protocol. This kit is designed to enrich 335,756 exons from 21,058 genes, covering ~ 71 Mb of the human genome. Enriched exome libraries were multiplexed and sequenced on the HiSeq 2500 platform (Illumina, USA). A paired-end DNA sequencing library was prepared via gDNA shearing, end-repair, A-tailing, paired-end adaptor ligation, and amplification. After hybridizing the library with bait sequences for 16 h, the captured library was purified and amplified with an indexing barcode tag, and library quality and quantity were assessed using a 2200 TapeStation Instrument and Qubit 2.0 Fluorometer, respectively. The exome library was sequenced using the 10-bp paired-end mode of the TruSeq Rapid PE Cluster kit and the TruSeq Rapid SBS kit (Illumina).

The University of California Santa Cruz hg19 reference genome (downloaded from http://genome.ucsc.edu) was used to align sequencing reads via the Burrows-Wheeler Aligner v. 0.6.2(15) with default settings. PCR duplications were marked using Picard-tools-1.8 (http://broadinstitute.github.io/picard/) and data cleanup was achieved using GATK-2.2.9 [25]. Point mutations were identified with the MuTect tool (https://github.com/broadinstitute/mutect) in paired samples. Indels were detected by VarScan2 (http://varscan.sourceforge.net/). Annovar was used to annotate variants. Signature analysis of mutational processes was carried out using the deconstructSigs tool [26]. Copy number variations were identified by EXCAVATOR [27]. Significant focal somatic copy number alterations were summarized by GISTIC analysis [28]. Tumor purity was estimated by PyLOH [29].

Library construction for RNA sequencing (RNA-seq) was performed using a Truseq RNA Sample Preparation v2 Kit (Illumina). Isolated total RNA was used in a reverse transcription reaction with poly (dT) primers, using SuperScriptTM II Reverse Transcriptase (Invitrogen/Life Technologies, USA), according to the manufacturer’s protocol. An RNA-seq library was prepared via cDNA amplification, end-repair, 3′ end adenylation, adapter ligation, and amplification. Library quality and quantity were measured using the Bioanalyzer and Qubit. Sequencing of the RNA library was carried out using the 100-bp paired-end mode of the TruSeq Rapid PE Cluster Kit and the TruSeq Rapid SBS Kit (Illumina).

The reads from the FASTQ files were mapped against the hg19 human reference genome using TopHat version 2.0.6 (http://ccb.jhu.edu/software/tophat/index.shtml). Raw read counts mapped to genes were measured using the BAM format file by HTSeq version 0.6. 1 [30]. A total of 18,161 coding genes were analyzed for transcript abundance and poorly expressed genes were eliminated based on the criteria of a maximum read count > 20 for all samples. Read counts were normalized using the Trimmed Mean of M-values normalization method. Differentially expressed genes were identified using the DESeq R package (www.huber.embl.de/users/anders/DESeq/). A gene set enrichment analysis (GSEA) [31] was conducted to analyze up- or down-regulated gene sets in specific groups of ESCC samples. Stromal and immune scores based on WTS were calculated using ESTIMATE [24]. Fractions of immune-associated cell types were calculated by CIBERSORT using RNA-seq expression profiles [23]. The immune cytolytic activity (CYT) was measured by the geometric mean of GZMA and PRF1 expression values in TPM [13].

Cancer cell fractions (CCF) were measured by PyClone, which de-convolves tumor sequences into sub-clones based on a hierarchical Bayesian clustering model [32]. Input data were generated from somatic single-nucleotide variants (SNVs) detected by MuTect and corresponding copy number variations. SNVs were grouped into clusters with similar CCF values.

TMB was measured by the number of somatic single nucleotide variants and indel mutations per megabase in the coding region [33]. Somatic single nucleotide variants include nonsynonymous as well as synonymous mutations. Non-coding alterations were not counted. In addition, known somatic alterations in COSMIC and truncations in tumor suppressor genes were excluded from the count.

Neoantigens were predicted using MuPeXI v.1.1.3 [34]. The three types of human leukocyte alleles (HLA-A, −B, and -C) were identified from the WTS data of each patient using seq2HLA [35]. Somatic mutations, gene expression counts, HLA types for each patient, and peptide lengths (8–11 mer) were provided as input for MuPeXI. Peptides with a half maximal inhibitory concentration (IC50) value ≤500 nM were considered to have a high binding affinity for the major histocompatibility complex (MHC). Expressed mutant peptide sequences with an IC50 value of ≤500 nM that showed binding affinity above normal were picked as candidate neoantigens.

4-μm sections of specimens were cut from formalin-fixed paraffin-embedded (FFPE) blocks. Slides were heated for at least one hour in a dry oven at 60 °C and dewaxed using xylene, then followed by multiplex immunofluorescence staining with a Leica Bond Rx™ Automated Stainer (Leica Biosystems, Newcastle, UK). Briefly, the slides were baked for 30 min and dewaxed with Leica Bond Dewax solution (#AR9222, Leica Biosystems), followed by antigen retrieval with Bond Epitope Retrieval 2 (#AR9640, Leica Biosystems) in a pH 9.0 solution for 30 min. The first primary antibodies for CD4 (ab133616, Abcam, dilution 1:100) were incubated for 30 min, followed by detection using the Polymer HRP Ms. + Rb (ARH1001EA, Perkin-Elmer) for 10 min. Visualization of CD4 was accomplished using Opal 570 TSA Plus (dilution 1:150) for 10 min, after which the slide was treated Bond Epitope Retrieval 1 (#AR9961, Leica Biosystems) for 20 min to remove bound antibodies before the next step in the sequence. In a serial fashion, CD20 (ab9475, Abcam, ARH1001EA, Perkin-Elmer, Opal 520 TSA Plus), FOXP3 (ab20034, Abcam, ARH1001EA, Perkin-Elmer, Opal 690 TSA Plus), PD-L1 (13684S, Cell Signaling, ARH1001EA, Perkin-Elmer, Opal 620 TSA Plus), CD8 (MCA1817T, BIO-RAD, ARH1001EA, Perkin-Elmer, Opal 480 TSA Plus) and CK (M3515, Dako, ARH1001EA, Perkin-Elmer, Opal 780 TSA Plus) was stained. Nuclei were subsequently visualized with DAPI, and the section was coverslipped using HIGHDEF® IHC fluoromount (ADI-950-260-0025, Enzo).

Slides were scanned using the PerkinElmer Vectra 3.0 Automated Quantitative Pathology Imaging System (Perkin-Elmer, Waltham, MA), and images were analyzed using the inform 2.2 software and TIBCO Spotfire™ (Perkin-Elmer, Waltham, MA). Each cell was identified by detecting nuclear spectral elements (DAPI). All the immune cell populations from each panel were characterized and quantified using the cell segmentation tool by the InForm image analysis software. All cells in each slide were designated as positive or negative for each antibody, and the data were categorized and exported to an xls file for analysis. We used the Spotfire™ program after tissue and cell segmentation and expression intensity was compared and then judged based on the cut-off value. The numbers of CD4, CD20, FOXP3, PD-L1, CD8, and CK positive cells were counted in each slide.

The two-sided t-test was used for comparisons of tumor purity, mutation burden, immune and stromal score, cytolytic score, fraction of immune cell from the pre- and post-CCRT samples. The Kaplan-Meier curves and log-rank test was used for the survival analysis. P-value less than 0.05 was considered as a statistically significant.

From a total of 40 study participants, 29 and 23 patients participated in paired DNA analysis and whole-transcriptome analysis, respectively (Fig. 1b). The pre-treatment clinical stages were IVA (n = 4), III (n = 21), and II (n = 4). Of the 29 patients that enrolled for paired genomic analysis, four patients did not undergo surgery due to disease progression (n = 2), refusal of surgery, and failure to follow-up. Surgical samples after CCRT (n = 25) showed post-neoadjuvant stages of IVA (n = 2), IIIB (n = 3), IIIA (n = 7), II (n = 1), I (n = 7) and 5 cases of pathologic complete response (pCR). The first biopsy was conducted at a median time of 4 days (range 1–14) before CCRT and the second biopsy was conducted at a median of 18 days (range 4–24) after CCRT. The resection margins were negative for all samples. Seventeen patients are currently enrolled in other clinical trial, which aims to study the effects of randomized adjuvant durvalumab treatment versus placebo (NCT02520453; Additional file 1: Table S1) [36].

We compared the somatic mutation landscape between pre- and post-CCRT samples, and analyzed representative genes related to cell cycle, histone modification, Hippo pathway, Notch pathway, and the PIK3CA pathway (Fig. 2a). Twenty-four pre-CCRT samples had missense, nonsense, or splicing mutations in the tumor suppressor gene, TP53, and these alterations were maintained in 11 samples after CCRT. Although nine pre-CCRT samples had a missense mutation in the nuclear factor erythroid 2 like 2 gene, NFE2L2, only two post-CCRT samples retained that mutation. SNV analysis showed that transition mutations, specifically C to T or G to A, were prominent in all samples. Comparison of copy number amplification and deletion regions between genomes of pre-CCRT and post-CCRT samples (Fig. 2b), showed that copy number at chromosome 7p14.1 was amplified only in post-CCRT samples. Additionally, the region harboring the cyclin D1 (CCND1) oncogene was amplified and the region harboring the tumor suppressor genes, Cyclin Dependent Kinase Inhibitor 2A (CDKN2A) and CDKN2B, was deleted in both pre- and post-CCRT samples.Fig2Fig. 2

Tumor mutation profile and changes in copy number, mutation burden in pre- CCRT and post-CCRT biopsy samples. a, Heatmap of somatic mutations (including missense, nonsense, and splicing) detected in 29 study patients. Each patient had a paired column (1st: Pre-CCRT and 2nd: Post-CCRT). Samples were annotated for mutation quantity, immune or stromal score changes, and surgical pathology. Each bar portrayed in the bottom represents the proportion of each type of substitution. b, Copy number alteration was analyzed from pre- and post-samples. c, Tumor purity was calculated by whole-exome sequencing. d, Comparison of mutation burden and neoantigen frequency between pre-CCRT and post-CCRT samples (n = 29). Abbreviations: CCRT, concurrent chemoradiation therapy; pCR, pathologic complete response

We found no difference in tumor purity, between pre- and post-CCRT samples (Fig. 2c). This allowed us to analyze the TMB and neoantigen load using WES and WTS data. We found that the TMB and neoantigen load were significantly lower in post-CCRT samples (p < 0.001) compared with pre-CCRT samples (Fig. 2d). Using ESTIMATE (Estimation of STromal and Immune cells in MAlignant Tumor tissues using Expression data), we found that 14 post-CCRT samples (60.9%) showed an increased immune score compared with pre-CCRT samples. Of those 14 post-CCRT samples, 10 also showed a concurrent increase in the stromal score (Fig. 3a). Therefore, CCRT leads to increased immune and stromal scores in the tumor tissue (Figs. 3b and c). No difference in immune CYT was observed between pre- and post-CCRT samples (Fig. 3d). Using CIBERSORT, we analyzed the changes in the population of immune cell types before and after CCRT. We found that the numbers of resting natural killer (NK)/T-cells (p < 0.001), follicular helper T-cells (p = 0.049), and regulatory T-cells (p = 0.038) significantly decreased after CCRT. However, the population of neutrophils was significantly increased in the post-CCRT samples (p = 0.001; Fig. 3e). We found no difference in the abundance of major immune cells such as B-cells, CD4 T-cells, CD8 T-cells, and M0, M1, and M2 macrophages between pre- and post-CCRT cells.Fig3Fig. 3

Comparison of changes in immune cell fractions, and immune, stromal, and immune cytolytic activity scores between pre-CCRT and post-CCRT samples. a Details regarding the change in each sample based on pre-CCRT and post-CCRT samples. Comparison of the immune and stromal score (b) and (c), and immune cytolytic (CYT) activity (d), in samples available for paired whole-transcriptome sequencing (n = 23).e, The fraction of immune cells showing statistically significant changes (n = 23). Abbreviations: CCRT, concurrent chemoradiation therapy, CYT, cytolytic activity

We found a strong correlation between outcomes from gene expression profile and multiplex IHC. Multiplex IHC for CD8 was significantly correlated with CYT score (p = 0.012). Especially PRF1 (Perforin 1) expression among CYT genes was significant (p = 0.001) while GRZA (Granzyme A) was not significant (p = 0.105). IFN-r was also correlated with CD8 (p = 0.104), and PD-L1 expression was negatively correlated with CD8 (p = 0.022) (Fig. 4a). Multiplex IHC is conducted in four patients available for both pre- and post- CCRT samples (Fig. 4b, Additional file 2: Table S2). Interestingly, all the samples showed an incremental trend in CD4 cell proportion after the CCRT. The proportion of cell expressing CD8, CD20, FOXP3, PD-L1 showed no definitive trend. (Fig. 4c). TMB was also highly correlated with CD8, but it was not significant (r = 0.69, p = 0.059) (data not shown).Fig4Fig. 4

a Correlation between multiplex IHC result and expression of genes and gene sets. CYT score, PERF1, PD-L1 were correlated with CD8 positive cells and CYT was correlated with FOXP3 positive cell by multiplex IHC. Pearson’s r and P-value is calculated using both pre- and post- CCRT sample. b Represent images from multiplex IHC (PS-1740) showing result from pre-CCRT and post-CCRT samples. c The proportion of cells per total cell in shown based on cytokeratin positivity

The 14 post-CCRT samples that showed increased immune scores also showed an increased abundance of neutrophils and a decrease in the number of follicular helper T-cells, regulatory T-cells, and resting NK/T-cells compared with the baseline pre-CCRT samples. We found that pre-CCRT samples with a higher proportion of M0 macrophages and lower resting mast cells were likely to show an increase in immune score after CCRT (Additional file 3: Table S3). Survival analysis showed extended PFS and overall survival (OS) in patient subgroups with increased immune scores (12-month PFS rate: 83.1% vs. 57.1%, p = 0.248; 12-month OS rate: 92.3% vs. 85.7%, p = 0.702) compared with the remaining subgroups (Fig. 5a and b). To validate the strength of these potential prognostic factors, we analyzed patient survival using pCR, a well-known prognostic factor for ESCC [37]. Disease free survival (DFS) and OS were extended in patients that showed pCR after CCRT (12-month DFS rate: 100.0% vs. 62.2%, p = 0.210; 12-month OS rate: 100.0% vs. 82.2%, p = 0.465), which was similar to the difference in survival rates analyzed according to the change of immune score (Figs. 5c and d).Fig5Fig. 5

Survival analysis based on immune score changes and pathologic complete response. Twenty-three patients with paired immune score showed (a), progression-free survival: hazard ratio (HR) 0.362, 95% confidential interval (CI) (0.060–2.184), p = 0.248; and (b), overall survival: HR 0.585, 95% CI (0.036–9.381), p = 0.702. c, Disease-free survival: HR 4.69e-16, 95% CI (0-NR), p = 0.210 (d), and overall survival: HR 4.84e-16, 95% CI (0-NR), p = 0.465 was shown in 25 patients who had curative surgery. HR and P calculated by log-rank test. Abbreviations: NR, not reached; pCR, pathologic complete response

Sixty genes that satisfied pre-defined criteria (> 2-fold change and adjusted p < 0.01) were identified by differentially expressed gene analysis in pre- and post-CCRT tissue samples (Fig. 6a and Additional file 4: Table S4). We found that the expression of cell cycle-related tumor suppressor genes, CCND2 and CDKN1A, was increased after CCRT. Gene set enrichment analysis showed that several immune-related gene sets, such as those involved in interferon gamma signaling, cytokine signaling, adaptive immune system, innate immune system, PD1 signaling, T-cell receptor signaling, and CD28 co-stimulation were enriched in post-CCRT samples (Figs. 6b and c).Fig6Fig. 6

Differentially expressed genes and gene set enrichment analysis in pre- and post-CCRT samples. a, Heatmap comparison of 60 significant genes (> 2-fold change & adjusted p < 0.01) between pre- and post-CCRT samples. b, Enriched differentially expressed gene sets between pre-CCRT and post-CCRT. c, Representative gene set enrichment analysis of differentially expressed genes in pre- and post-CCRT samples

To analyze the genomic profile of the ESCC tissue samples based on their surgical pathology, we compared pCR (n = 5) and non-pCR (n = 20) subgroups using pre- andpost-CCRT samples. We observed no significant difference in immune and stromal scores between the two subgroups in pre-CCRT samples. However, the pCR subgroup showed significantly lower immune CYT compared with the non-pCR subgroup (p = 0.011) in pre-CCRT samples. We found a prominent decrease in CYT in the non-pCR group (p = 0.047) after CCRT, but no change in CYT in the pCR group compared with pre-CCRT samples (Fig. 7a). We also did not find any predictive somatic mutation markers in the pCR subgroup. NFE2L2, a gene frequently found to be mutated in our pre-CCRT samples, mostly disappeared after CCRT (Fig. 3a). CCF estimation analysis showed that NFE2L2 p.D15E was the only unique variant in pCR samples. However, the proportion of cancer cells with specific mutation failed to specifically correlate with the pathologic response (Fig. 7b). Analysis of changes in the population of immune cell types showed a significant increase in activated dendritic cells after CCRT in the pCR subgroup compared with pre-CCRT samples. Neutrophils were significantly increased in number in both pCR and non-pCR subgroups after CCRT compared with pre-CCRT samples (Fig. 7c and Additional file 5: Table S5). However, analysis of individual gene expression patterns revealed no specific pattern in the pCR and non-pCR subgroups before and after CCRT (Additional file 6: Table S6 and Additional file 7: Fig. S1).Fig7Fig. 7

Comparison of changes in the immune, stromal, and immune cytolytic activity scores and immune cell fractions between pre- CCRT and post-CCRT samples based on their surgical pathologic profile. a, Immune and stromal scores, and immune cytolytic (CYT) activity in pathologic complete response (pCR), non-pCR, and no-operations (Op) samples. b, Somatic mutation changes in NFE2L2 using cancer cell fraction measurements based on the surgical pathologic profile.c, Changes in immune cell fractions based on the surgical pathologic profile