Genomic DNA and total RNA were extracted from frozen tumor samples after cryostat sectioning, using DNA and AllPrep DNA/RNA Mini Kits (Qiagen, Hilden, Germany). Genomic DNA was isolated from matched peripheral blood samples using QIAamp DNA Mini Kits (Qiagen). Eighty HGSC samples were analyzed in this study.

Paired tumor and blood genomic DNA libraries were constructed according to the protocol provided with the KAPA Hyper Prep Kit (Kapa Biosystems). Whole-exome capture was performed with the SureSelect Human All Exon kit V.4 and V.5 (Agilent Technologies) following the manufacturer's protocols. We sequenced exome capture libraries on the HiSeq 2000 platform according to the manufacturer's instructions, and 2×100 bp paired-end reads were generated. Image analysis and base calling were performed using the Illumina pipeline with default settings.15 Exome reads were independently mapped to the human genome (GRCh37/hg19) using Burrows-Wheeler Aligner and Novoalign software. Reads with a minimal editing distance to the reference genome were taken to represent optimal alignments. Bam files were then locally realigned with short-read micro re-aligner (SRMA). Normal-tumor pair bam files were processed using the in-house genotyper karkinos (https://sourceforge.net/projects/karkinos/). OxoG artifacts were removed by the D-ToxoG program.16 RefSeq gene annotation was performed by ANNOVAR.17

RNA sequencing was performed as previously described18 for 80 HGSC samples that had RNA of sufficient quality and quantity. An RNA-sequencing library was prepared using the TruSeq Stranded mRNA LT Sample Prep Kit (Illumina) according to the manufacturer’s protocol. Briefly, 1 µg of total RNA was purified using oligo dT magnetic beads and poly A+RNA were fragmented at 94°C for 2 min. complementary DNA (cDNA) was synthesized using SuperScript II (Invitrogen) and adapter-ligated cDNA was amplified with 12 cycles of PCR. Each library was sequenced using HiSeq 2000, loading four libraries per lane of the flowcell, which produced an average of 59.2 million reads of 101-cycle reads for each sample. RNA-sequencing reads were aligned to a human transcriptome database (University of California Santa Cruz (UCSC) genes) and the reference genome (GRCh37/hg19) using the STAR algorithm (V.2.5.2a).19 Finally, fragments per kilobase of exon per million fragments mapped (FPKM) values were calculated for each UCSC gene while considering strand-specific information.

DNA bisulfite conversion was performed with EZ DNA Methylation Kits (Zymo Research）for tumor genomic DNA (500 ng). All samples were run with Infinium HumanMethylation450 BeadChip Kits (Illumina), which target >4,50,000 methylation sites, according to the manufacturer's protocol. To determine BRCA1 and RAD51C promoter methylation, eight and 15 probes for the promoter regions in the corresponding genes respectively were used as reported previously.20

Gene set enrichment analysis (GSEA) was performed with GSEA V.3.0 software with default parameters. The immune cell-related gene sets provided by Angelova et al were used and the association with each immune cell gene set was represented by a normalized enrichment score.21 An immune cell type was considered significantly enriched when the familywise error rate (FWER, p value) was <0.05. Gene sets for ‘immune_responses’ in the gene ontology (GO) biological process (BP) retrieved from c5 in the Molecular Signature Database (MSigDB) V.7.0 and hallmark gene sets (MSigDB V.7.0) were used for the analysis.22

To quantify tumor-infiltrating immune cells (TICs), we analyzed RNA-seq data with CIBERSORT algorithm with default parameters.23 Twenty-two immune cell subtypes were analyzed.

Human leukocyte antigen (HLA)-typing was performed on the 80 patients from whole exome sequencing data of normal tissues or peripheral blood mononuclear cells (PBMCs) using HLA typing software (Omixon target HLA) (online supplementary table S1). Mutated peptides derived from missense mutations were used for major histocompatibility complex (MHC) class I binding prediction as previously described.24 25 The missense, insertion/deletion and non-stop mutations in genes with low expression that had FPKM values<1 were eliminated. Long peptides containing the predicted mutation or of wild-type were assessed using the Immune Epitope Database and Analysis Resource (http://www.iedb.org/) offline; 9-mer and 10-mer peptides were selected, each predicted to bind to a specific HLA allele for each patient. Mutated peptides (9-mer and 10-mer) which had an IC₅₀ value below 500 nM were regarded as candidate neoepitopes from missense mutations (online supplementary table S2) and insertion/deletion (indel) and non-stop mutations (online supplementary table S3). We defined one neoantigen as a mutation producing any number of HLA-A, HLA-B, and HLA-C-restricted potential neoepitopes (online supplementary table S4).

Overall survival (OS) times were calculated as the number of days from surgery to death, or the last time the patient was known to be alive. Progression-free survival (PFS) times were calculated as the time elapsed between surgery and tumor progression or death. Survival was plotted according to the Kaplan-Meier method. The log rank test was used to examine the significance of differences in survival between groups. Comparison of results was by an unpaired, two-tailed Student t-test using GraphPad Prism 5 (GraphPad Software). A value of p<0.05 was considered significant.

In total, 80 cases of HGSC were analyzed in this study. The clinicopathological characteristics of these patients including age, stage, residual tumor after surgery, and platinum-sensitivity are summarized in table 1. All patients received standard chemotherapy such as carboplatin plus paclitaxel; no patients received PARP inhibitors or immune checkpoint blockade such as programmed cell death-1/programmed cell death-ligand 1 (PD-1/PD-L1) blockade therapies.

HGSC patients with HR-deficient tumors may derive more clinical benefit from treatment with PARP inhibitors, relative to patients with HR-proficient tumors. Therefore, we here focused on HR-proficient patients for whom this treatment modality is less effective. We first determined the HR-deficient or HR-proficient status of the 80 HGSC patients using exome sequencing and methylation arrays. We defined tumors with BRCA1/2, RAD51C/D, ATM, ATR, BARD1, CHEK1, CHEK2, MRE11A, NBN, PALB2, RAD50 and BLM mutations, or BRCA1 and RAD51C promoter methylation, as HR-deficient in this study. In this way, we classified 34 (43%) of patients as having HR-deficient tumors (figure 1A and table 1). Older patients (≥55 years) tended to have HR-proficient tumors but younger patients’ tumors were more likely to be HR-deficient. Although platinum-sensitivity is significantly higher in HR-deficient than HR-proficient tumors (table 1), there was no difference in PFS or OS between the patients in this cohort (figure 1B).

Neoantigens derived from somatic mutations are major targets for antitumor immune responses. Recent reports documented that HR-deficient tumors have higher neoantigen loads than HR-proficient HGSCs.26 Therefore, we surveyed our cohort of 80 HGSC and compared the number of predicted neoantigens derived from non-synonymous mutations including missense, indel and non-stop mutations in HR-deficient versus HR-proficient tumors. The number of non-synonymous mutations from whole-exome sequencing data ranged from nine to 286 (median value: 82) (online supplementary table S1), suggesting that most patients in this cohort were categorized into low/medium TMB group, as previously reported.5 We next identified candidate neoepitopes derived from non-synonymous mutations in genes with FPKM values≥1 in each tumor based on MHC class I binding prediction scores (IC₅₀ <500 nM) according to NetMHCpan V.2.8, as previously reported24 25 (online supplementary tables S2 and S3). We defined antigenic mutations potentially generating one or more neoepitopes as predicted neoantigens (online supplementary table S4). The number of predicted neoantigens ranged from one to 75 (median number, 22) in these 80 cases (online supplementary table S1). As expected, this was lower in HR-proficient than in HR-deficient tumors (Student t-test, p=0.0018) (figure 2A). However, 39% of patients with HR-proficient tumors still had numbers of neoantigens higher than the median values of all 80 cases, falling into the neoAg^hi group within this HGSC cohort (figure 2B).

HLA expression is also important for antitumor immune responses because neoepitopes are recognized by antigen-specific T cells as peptide-HLA complexes on the cell surface. It is widely accepted that tumors can evade antitumor immune responses by downregulating MHC class I expression, as well as by antigen loss. A subset of HGSC cases with decreased expression of MHC class I has been reported to have worse clinical outcomes.27 Therefore, we evaluated the expression of HLA-A, HLA-B and HLA-C using RNA-seq data in HR-deficient and HR-proficient tumors. The mean FPKM values for HLA-A, HLA-B, and HLA-C ranged from 54.1 to 2300.4 (median, 613.4) in all 80 cases (figure 2C), but there was no significant difference between HR-deficient and HR-proficient patients (p=0.1122). Nonetheless, 41% of patients with HR-proficient tumors did have higher than median MHC class-I expression (HLA^hi, above the median value of 613.4), whereas >60% of HR-deficient tumors were HLA^hi (figure 2D).

We next investigated neoantigen load and HLA-class I expression, and their relevance for prognosis in HR-proficient HGSC (n=46). There was no difference between the neoAg^hi group (n=18) and neoAg^lo group (n=28) in clinical outcomes (both PFS and OS) (log rank test, p=0.1913 and p=0.1397, respectively) (figure 3A). We also investigated HLA-class I expression and its relevance for clinical outcome. The HLA^hi group (n=19) had a significantly more favorable PFS but not OS, relative to the HLA^lo group (n=27) (log rank test, p=0.0277 and p=0.5707, respectively) (figure 3B).

We further stratified these patients according to a combination of the number of neoantigens and the levels of HLA-class I expression. For this, patients were divided into four groups: (1) neoAg^hi/HLA^hi tumors, (2) neoAg^hi/HLA^lo, (3) neoAg^lo/HLA^hi and (4) neoAg^lo/HLA^lo. As shown in figure 3C, patients with neoAg^hi/HLA^hi tumors had notably longer PFS but not OS than others (log rank test, p=0.0087 and 0.0567, respectively). The same analysis was applied to HR-deficient HGSC (n=34). However, patients in the neoAg^hi (n=24) (figure 3D), HLA^hi (n=21) (figure 3E), or neoAg^hi/HLA^hi groups (n=13) (figure 3F) did not have longer PFS or OS. Therefore, these results show that patients with neoAg^hi/HLA^hi tumors have a better clinical course relative to the other neoAg/HLA groups, but only if their tumors are HR-proficient.

To investigate whether the better prognosis of the HR-proficient neoAg^hi/HLA^hi subgroup relative to all other HR-deficient subgroups is related to immune parameters in the tumor microenvironment, we next compared immunologic signatures between neoAg^hi/HLA^hi (n=9) and the others (n=37) using RNA-seq. We performed GSEA. We used the 72 gene sets for ‘immune_responses’ in the GO BP retrieved from c5 in the MSigDB V.7.0. We found 42 out of 72 gene sets such as ‘GO_ADAPTIVE_IMMUNE_RESPONSE’ and ‘GO_T_CELL_ACTIVATION_INVOLVED_IN_IMMUNE_RESPONSE’ were significantly enriched in the neoAg^hi/HLA^hi group (figure 4A) (online supplementary table S5). We also found that the gene sets for ‘T cells’ (FWER, p<0.001), ‘Th1’ (FWER, p<0.001), ‘TGD’ (FWER, p<0.001), ‘effector memory CD8’ (FWER, p<0.001), and certain others in 28 immune cell subpopulations (gene sets provided by Angelova et al) (figure 4B) (online supplementary table S6) and ‘HALLMARK_ALLOGRAFT_REJECTION’ (FWER, p<0.001), ‘HALLMARK INTERFERON GAMMA RESPONSE’ (FWER, p<0.001) in hallmark gene sets were significantly enriched in the neoAg^hi/HLA^hi group (figure 4C) (online supplementary table S7). CD274 (PD-L1) expression, a known candidate biomarker for ICIs, was also higher in the neoAg^hi/HLA^hi subgroup with statistical significance (Student t-test, p=0.0078) (figure 4D). We also tried to estimate TICs by CIBERSORT.23 Different types of immune cells with antitumor or protumor functions were observed as illustrated by absolute immune fraction scores (figure 4E left and online supplementary figure S1). Absolute scores of TICs were significantly higher in the neoAg^hiHLA^hi tumors (Student t-test, p=0.0009) (figure 4E right).

It has been shown that certain cancers known to be immunogenic exhibit preferential depletion of neoantigenic mutations within the totality of mutations in the tumor.25 28 29 To test whether the same may apply in an immunogenic subgroup of HGSC, we compared the ‘neoantigen frequency’ (the ratio of neoantigen per somatic mutation) between neoAg^hi/HLA^hi (n=9) and the others (n=37). However, there were no significant differences in the ratios of neoantigens per missense or indel/non-stop mutations (figure 4F).

The subgroup of patients with neoAg^hi/HLA^hi HR-proficient tumors had a better clinical outcome than patients with tumors of the same phenotype but with HR-deficiency, both for PFS and OS (log rank test, p=0.0015 and p=0.1356, respectively) (figures 3C,F and 5A). To investigate what causes this difference in clinical benefit between HR-proficient and HR-deficient patients despite their being in the same ‘high neoantigen load and high HLA expression’ subgroup, we next compared immune-related gene expression between them. We performed the same analysis as in figure 4, In GSEA using 72 gene sets for ‘immune_responses’ in GO BP from c5 in MSigDB V.7.0, we found 29 out of 72 gene sets such as ‘GO_ADAPTIVE_IMMUNE_RESPONSE’ and ‘GO_T_CELL_ACTIVATION_INVOLVED_IN_IMMUNE_RESPONSE’ were significantly enriched in the HR-proficient group (figure 5B and online supplementary table S8). We also found that the gene sets for ‘T cells’ (FWER, p<0.001), ‘Th1’ (FWER, p<0.0001), ‘TGD’ (FWER, p<0.001), ‘effector memory CD8’ (FWER, p<0.001), and others in 28 immune cell subpopulations were significantly enriched in patients with HR-proficient tumors (figure 5C) (online supplementary table S9). Furthermore, ‘HALLMARK_ALLOGRAFT_REJECTION’ (p<0.001), ‘HALLMARK_INTERFERON_ALPHA_RESPONSE’ (FWER, p<0.001) and ‘HALLMARK_INTERFERON_GAMMA_RESPONSE’ (FWER, p<0.001), in hallmark gene sets were significantly enriched in HR-proficient tumors (figure 5D) (online supplementary table S10). There was no statistically significant difference between HR-proficient and HR-deficient tumors for CD274 (PD-L1) expression (figure 5E). We estimated TICs by CIBERSORT again. Different types of immune cells with antitumor or protumor functions were observed (figure 5F left and online supplementary figure S2) and absolute scores of TICs were significantly higher in the HR-proficient tumors (Student t-test, p=0.0165) (figure 5F right).

We again compared the ‘neoantigen frequency’ (the ratio of neoantigen per somatic mutation) between HR-proficient (n=9) and HR-deficient (n=13) tumors. However, there were no significant differences in the ratios of neoantigens per missense or indel/non-stop mutations (figure 5G).