Discovery of candidate tumor biomarkers for treatment with intraperitoneal chemotherapy for ovarian cancer

Tumor mRNA expression was used to discover genes associated with worse survival or no survival benefit after intraperitoneal (IP) chemotherapy. Data for high grade serous ovarian cancer patients treated with IP (n = 90) or IV-only (n = 398) chemotherapy was obtained from The Cancer Genome Atlas. Progression free survival (PFS) and overall survival (OS) were compared between IP and IV groups using Kaplan-Meier analysis and Cox regression. Validations were performed by analyses of microarray and RNA-Seq mRNA expression data. PFS and OS were compared between IP and IV groups by permutation testing stratified by gene expression. P-values are two-tailed. IP chemotherapy increased PFS (26.7 vs 16.0 months, HR 0.43 (0.28–0.66), p = 0.0001) and OS (49.6 vs 38.2 months, HR 0.46 (0.25–0.83), p = 0.01). Increased expression of NCAM2 and TSHR and decreased expression of GCNT1 was associated with decreased PFS and OS after IV chemotherapy (p < 0.05). High tumor expression of LMAN2, FZD4, FZD5, or STT3A was associated with no significant PFS increase after IP compared to IV chemotherapy. Low expression of APC2 and high expression of FUT9 was associated with 5.5 and 7.2 months, respectively, decreased OS after IP compared to IV chemotherapy (p ≤ 0.007).

Survival analysis by chemotherapy exposure. We used the R platform for statistical computing 18,19 .
P-values were two-tailed. Demographic and outcome information were compared using statistical tests indicated with Results. Incomplete clinical data reporting to TCGA was common. Cases with missing data were excluded from statistical comparisons. Cytoreduction was considered optimal if residual disease was ≤ 10 mm after surgery. PFS and OS data was truncated at 60 months to prevent biasing results by long surviving individuals and to match a restriction time of 60 months set for restricted mean survival (RMS) calculations. Empiric Kaplan-Meier (KM) survival analysis and Cox proportional-hazards regression adjusted for covariates age, surgical stage, histologic grade, cytoreduction status and race were used to compare PFS and OS by chemotherapy route (IP versus IV). For KM analysis of all cases (n = 488), 96 and 42 cases were omitted due to missing OS and PFS data, respectively. For KM analysis of optimally cytoreduced cases (n = 282), 47 and 18 cases were omitted due to missing OS and PFS data, respectively. For multivariate Cox regression of OS and PFS for all cases (n = 488), 178 and 144 cases, respectively, were omitted due to missing data. For multivariate Cox regression of OS and PFS for optimally cytoreduced cases (n = 282), 52 and 22 cases, respectively, were omitted due to missing data.
Discovery and validation of differentially expressed genes. Within each adjuvant chemotherapy group, cases were stratified by PFS time < or ≥ 12 months. Twelve months was used because cancers that recurred or progressed before 12 months were likely chemoresistant. Patients who undergo surgical staging and six cycles of adjuvant chemotherapy often reach 6 months after completion of chemotherapy, the time at which platinum sensitivity is designated, about 12 months after surgery. A diagram of the exploratory gene analysis and validation steps is shown in Supplementary to discover differentially expressed genes. Standard t-testing was used because the results are fully reproducible unlike permutation methods of microarray data analysis, which yield a larger number of differentially expressed genes with a higher number of false positive discoveries. Microarray data was used for exploratory analysis because too few patients had RNA-Seq data for gene discovery. Genes evaluated were listed under KEGG pathways identified by the microarray annotation 14,20    OS and PFS at all expression levels of each gene. This null hypothesis is consistent with survival analysis of the clinical data (see Results). Findings from permutation testing with microarray expression data were validated by identical permutation testing using RNA-Seq data. Any gene for which the null hypothesis is rejected by both methods is considered a candidate biomarker that warrants experimental validation.

Results
Patient and chemotherapy information.  Figure 1).

Figure 1. Kaplan-Meier survival curves for patients treated with intraperitoneal (IP) and intravenous (IV) chemotherapy.
Log rank is log rank p-value from Kaplan-Meier analysis. Cox PH is multivariate Cox proportional hazards regression p-value for chemotherapy route with adjustment for age, surgical stage, histologic grade, cytoreduction status, and race.  (Fig. 2). Comparing univariate RMS curves and multivariate RMS times between IP and IV groups evaluates differentially expressed genes as biomarkers for treatment benefit from IP chemotherapy. For instance, according to microarray data, patients whose tumors had low expression of APC2 or high expression of BCAT1 experienced less survival benefit from the addition of IP chemotherapy to their treatment (Fig. 2). Patients whose tumors had high expression of PER1 or TSHR experienced decreased survival compared to patients whose tumors did not have high expression of these genes, with the associations of gene expression and survival outcomes reaching statistical significance in the larger cohort of patients who received IV-only adjuvant chemotherapy (Fig. 2). Table 4 lists differentially expressed genes for which there is evidence to reject the null hypothesis based on microarray data and designate the gene as a candidate biomarker for patient selection for IP chemotherapy (see Methods above). For example, patients with high expression (upper 10 th percentile) of FZD5 showed an expected overall survival benefit among the IP chemotherapy group but did not show a significant PFS benefit (no significant difference between IP and IV-only groups) ( Table 4).  1.34 (1.10-1.64), 0.004 1.64 (1.04-2.60) 1.19 (1.03-1.38), 0.020  1.17 (1.03-1.34), 0.018  1.21 (1.02-1.42), 0.025 1.23 (1.05-1.45), 0.011 1.21 (1.05-1.40), 0.008  1.17 (1.04-1.33), 0.009 1.71 (1.31-2.22), <0.001 1.64 (1.22-2.21), 0.001 1.16 (1.01-1.35), 0.041  1.18 (1.03-1.35)

Discussion
Adjuvant IP chemotherapy is associated with increased PFS and OS in TCGA data. A large and increasing number of publications reported survival outcomes of TCGA ovarian cancer patients without adjusting for chemotherapy route as a potential confounder. Reanalysis of previously reported survival outcomes to adjust for chemotherapy route as a potential confounder may be needed given that IP chemotherapy use was common among TCGA HGS OvCa patients and that receiving IP chemotherapy was associated with large and highly significant increases in PFS and OS. A detailed review of the tumor biology of all candidate biomarkers is beyond the scope of this report. DAVID gene functional classification failed to reveal any clusters among differentially expressed genes 23 . Three differentially expressed genes, APC2, FZD4, and FZD5, which had mRNA levels that were significantly associated with OS and PFS, are members of the Wnt signaling pathway. The observation that increased expression of APC2 and decreased expression of FZD4 and FZD5 was associated with decreased OS and PFS among patients treated with IV chemotherapy suggests that downregulation of Wnt signaling may play a role in chemoresistance 24,25 . The Wnt pathway may be targeted in cancer trials 25,26 . Recently, advanced ovarian cancers with increased membrane β -catenin expression by immunohistochemistry were shown to have decreased PFS and increased platinum-resistance, consistent with our findings 27 .
Limitations of our study include incomplete clinical data reporting to TCGA, a relatively small number of patients who received IP chemotherapy, and also small numbers of tumors that were analyzed by RNA-Seq, especially among the IP group. During the RNA-Seq validation, many survival associations trended (p = 0.05-0.10) but failed to reach significance due to few cases having RNA-Seq data. Clinical and chemotherapy information reported to TCGA was also not verified for correctness. We could not perform validation in another patient cohort because no other suitable data is available. To our knowledge, TCGA provides the only cohort of ovarian cancer patients treated with IP chemotherapy for which there is currently transcriptome or proteome data. Also, tumor protein expression was not confirmed experimentally by immunohistochemistry of primary tumor specimens. The TCGA reverse phase protein array does not include any of the differentially expressed genes and therefore the TCGA proteomics data could not be used as a potential protein-level validation of our transcription-level findings. In addition, the process of gene discovery by t-testing and gene validation with limited RNA-Seq data is very stringent. Thus, there are likely additional genes that are significantly associated with survival outcomes among this patient cohort. Targeted rather than exploratory analyses of well-known molecular pathways driving cancer development or chemotherapy resistance may lead to discovery of additional genes with similar survival associations after adjusting for chemotherapy exposure.
Strengths of our study include use of a standard and fully reproducible statistical test (the t-test) to compare gene expression levels between groups. We also analyzed gene mRNA expression levels as a continuous variable in all regression models of PFS or OS. Dichotomizing or otherwise arbitrarily subdividing cases by gene expression thresholds is very commonly reported in the literature but is less informative than RMS analysis. RMS curves illustrate the relationships of gene expression and survival. We provide TCGA case identifiers with chemotherapy route to ease reproduction of our findings and reanalysis of previous studies. Adjuvant IP chemotherapy of TCGA patients was associated with increased PFS and OS similar to randomized trials [3][4][5] . Our study is the second study, after one other recent report, to demonstrate off-trial survival advantage associated with adjuvant IP chemotherapy use that included some common "modified regimens" that are often individual provider-designed to decrease toxicity 28 . Our study reports exploration of transcriptome data motivated by a practical, clinically-oriented question: Given the increased morbidity of IP chemotherapy, are there biomarkers for response to IP chemotherapy that may aid selection of patients for IP chemotherapy? We discovered candidate biomarkers that are significantly associated with chemoresistance and decreased survival, or lack of benefit from IP chemotherapy. Our findings generate hypotheses regarding route-specific chemoresistance that may be tested using in vivo models. A targeted clinical assay for efficient measurement of primary tumor gene expression of these candidate biomarkers may be developed. Some gene products may be amenable to measurement in serum or ascites, or by primary tumor immunohistochemistry. Our findings may be corroborated at the level of protein expression by performing immunohistochemical analysis of primary tumor specimens, if made available, from a previous randomized trial of IP versus IV chemotherapy [3][4][5] . If similar results are obtained, a study of specimens from a previous randomized trial may efficiently provide validation of these candidate biomarkers as clinical biomarkers for patient selection for IP chemotherapy.