Risk prediction for late-stage ovarian cancer by meta-analysis of 1525 patient samples

Abstract

Background: Ovarian cancer causes more than 15000 deaths per year in the United States. The survival of patients is quite heterogeneous, and accurate prognostic tools would help with the clinical management of these patients.

Methods: We developed and validated two gene expression signatures, the first for predicting survival in advanced-stage, serous ovarian cancer and the second for predicting debulking status. We integrated 13 publicly available datasets totaling 1525 subjects. We trained prediction models using a meta-analysis variation on the compound covariable method, tested models by a "leave-one-dataset-out" procedure, and validated models in additional independent datasets. Selected genes from the debulking signature were validated by immunohistochemistry and quantitative reverse-transcription polymerase chain reaction (qRT-PCR) in two further independent cohorts of 179 and 78 patients, respectively. All statistical tests were two-sided.

Results: The survival signature stratified patients into high- and low-risk groups (hazard ratio = 2.19; 95% confidence interval [CI] = 1.84 to 2.61) statistically significantly better than the TCGA signature (P = .04). POSTN, CXCL14, FAP, NUAK1, PTCH1, and TGFBR2 were validated by qRT-PCR (P < .05) and POSTN, CXCL14, and phosphorylated Smad2/3 were validated by immunohistochemistry (P < .001) as independent predictors of debulking status. The sum of immunohistochemistry intensities for these three proteins provided a tool that classified 92.8% of samples correctly in high- and low-risk groups for suboptimal debulking (area under the curve = 0.89; 95% CI = 0.84 to 0.93).

Conclusions: Our survival signature provides the most accurate and validated prognostic model for early- and advanced-stage high-grade, serous ovarian cancer. The debulking signature accurately predicts the outcome of cytoreductive surgery, potentially allowing for stratification of patients for primary vs secondary cytoreduction.

Figure 1.

Flowchart of the study. This outlines the steps for training and validating the prognostic models presented in this study. * Including 27 TCGA early stage samples. †Samples that were included in multiple studies were identified and removed (see Supplementary Table 1, available online). qRT-PCR = quantitative reverse-transcription polymerase chain reaction.

Figure 2.

Leave-one-dataset-out validation of performance of the new gene signature in predicting overall survival in late-stage ovarian cancer. We used our database of the ovarian transcriptome (5) for developing a novel overall survival gene signature by a meta-analytic signature development method. In total, six large datasets passed our training criteria (see Methods). To first test our methodology on the six large datasets we used for training (6,8,9,10,15,63), we applied a leave-one-dataset-out approach (Supplementary Figure 1, available online). Specifically, for each of the six datasets, we trained a prediction model using the remaining five datasets only and then stratified the patients of the dataset not used for training into high- and low-risk groups. P values were calculated with the log-rank test, and cutoffs for patient stratification were the medians of predicted risk scores in the combined training cohorts. All statistical tests were two-sided. TCGA = The Cancer Genome Atlas.

Figure 3.

Validation of the survival signature in independent data (9,13,16,19,28–30). A–E) Risk stratification in five validation microarray datasets of late-stage, high-grade, serous ovarian cancer by a model trained on all six training datasets (Figure 2). The Gillet et al. reverse-transcription polymerase chain reaction validation dataset (28) assayed only 12 of the 200 genes in the signature, which included, however, well-characterized cancer genes such as APC, RB1, and PDGFRB. The remaining four datasets had less than the 75 samples we required for training (13,29,30) or were published after our model was finalized (16). Cutoffs for high and low risk were again determined by calculating the median of the combined patient risk scores in training data. F) Performance in all early-stage, high-grade, serous samples from The Cancer Genome Atlas (TCGA) data. P values were calculated with the log-rank test. All statistical tests were two-sided. G) We further tested the model in a dataset in which survival information was only available as binary outcome (long-term vs short-term survivors) (19). The prediction model here estimated the probability of short term survival and its accuracy is shown with a receiver operating characteristic curve. This curve shows the true- and the false-positive rates for all possible cutoffs of the continuous prediction score. True positives are correctly classified short-term survivors. *Area under the curve (AUC) statistically significantly greater than 0.5 [DeLong test (23)].

Figure 4.

Combined comparison of our novel meta-analysis gene signature with existing prognostic factors and signatures proposed by The Cancer Genome Atlas (TCGA). This figure provides a summary view of both leave-one-study-out cross-validation and independent validation performance shown in Figures 2 and 3. A) Summary of the risk stratifications in Figures 2 and 3 in a single Kaplan–Meier plot [all studies (Table 1), excluding the TCGA cohort so comparison with the TCGA signatures could be made (6,8,10,13,15,16,28–30)]. This plot compares the survival of all patients classified as high risk in this meta-analysis with all low-risk patients. B) Risk stratification based on a fivefold cross-validated multivariable model using the gene signature risk group (high vs low risk) and the categorical covariables tumor stage (III vs IV) and debulking status (optimal vs suboptimal). The smaller sample sizes arise because of missing clinical annotations. C) Risk stratification based on a fivefold cross-validated multivariable model using tumor stage and debulking status only. D) Summary of risk stratifications by the TCGA gene signature (9) over all cohorts excluding TCGA, as in panel A. E) Summary of risk stratifications by the Verhaak et al. survival signature (18), as in panel A. F) Summary of risk stratifications by the multivariable model proposed by Verhaak et al. using their survival signature, continuous TCGA subtype scores, as well as categorical debulking status and tumor stage. P values were calculated with the log-rank test. G) Forest plot providing an alternative visualization of the hazard ratios (HRs) shown in the Kaplan–Meier plots in Figures 2 and 3 and including a comparison with the corresponding hazard ratios of the TCGA and Verhaak et al. models in all datasets separately. Squares show hazard ratio point estimates, with size corresponding to their weighting in the fixed-effects meta-analysis summaries (inverse of squared standard error); horizontal lines show 95% confidence intervals (CIs); and diamonds at the bottom show fixed-effects summaries of the hazard ratios over all shown datasets. Note that this fixed-effects summary corresponds to the hazard ratios shown in the summary Kaplan–Meier plots in panels A, D, and E. All statistical tests were two-sided.

Figure 6.

Validation of selected genes associated with debulking status by quantitative reverse-transcription polymerase chain reaction (qRT-PCR) in the Bonome et al. validation data (n = 78). A) Observed fold-changes in suboptimal vs optimal tumors and their standard deviations of the genes with statistically significantly (P < .05, two-sided Student t test) different expression between the two groups. B) The prediction accuracy of a multivariable model in which the qRT-PCR–validated genes were equally weighted. We stratified the samples into groups of high and low risk for suboptimal surgery based on the tertiles of the multivariable risk score: the 33% of patients with highest risk score were classified as high risk, the 33% with lowest risk score were classified as low risk, and all others were classified as medium risk. Between the high- and low-risk groups, 76.9% of samples were classified correctly. The accuracy of the multivariable risk prediction is further shown with a receiver operating characteristic curve. *Area under the curve (AUC) significantly greater than 0.5 [DeLong test (23)].

Figure 7.

Validation of POSTN, pSmad2/3, and CXCL14 in an independent cohort by immunohistochemistry (IHC). A–C) IHC staining of POSTN, pSmad2/3, and CXCL14. The four IHC classes are categorized by IHC intensity scores as described in the Supplementary Information (available online) and previously (27). Scale bar = 100 μm. D) Histogram visualizing the frequency of optimal and suboptimal tumors stratified by POSTN IHC class (A) in an independent cohort of 177 samples. The true- and false-positive rates of POSTN IHC intensity scores (27) used for classification are further shown with a receiver operating characteristic curve. E and F) Corresponding figures for pSmad2/3 and CXCL14. G) The prediction accuracy of the multivariable model in which the three IHC validated genes were equally weighted (as in Figure 6B). * Area under the curve (AUC) statistically significantly greater than 0.5 [DeLong test (23)].

Figure 5.

Pathway analysis of the debulking signature. Using the Pathway Studio 7.1 (Ariadne Genomics) software and a novel signature of 200 debulking-associated genes, we identified pathways statistically significantly associated with suboptimal debulking surgery. A gene is labeled in red when it is overexpressed in tumors that were subsequently suboptimally debulked. Conversely, genes overexpressed in tumors with optimal cytoreduction are labeled in blue. Genes with predictive power toward poor prognosis based on the meta-analysis are highlighted with pink borders. Red broken arrows indicate direct stimulatory modification. Green arrows indicate EGR-1–based transcriptional regulations. Orange arrows indicate TGF-β/Smad–based transcriptional regulations. Blue solid arrows indicate other direct regulations. Blue broken arrows indicate other indirect regulations. Purple sticks indicate binding.