Time to recurrence and survival in serous ovarian tumors predicted from integrated genomic profiles

Abstract

Background: Serous ovarian cancer (SeOvCa) is an aggressive disease with differential and often inadequate therapeutic outcome after standard treatment. The Cancer Genome Atlas (TCGA) has provided rich molecular and genetic profiles from hundreds of primary surgical samples. These profiles confirm mutations of TP53 in ~100% of patients and an extraordinarily complex profile of DNA copy number changes with considerable patient-to-patient diversity. This raises the joint challenge of exploiting all new available datasets and reducing their confounding complexity for the purpose of predicting clinical outcomes and identifying disease relevant pathway alterations. We therefore set out to use multi-data type genomic profiles (mRNA, DNA methylation, DNA copy-number alteration and microRNA) available from TCGA to identify prognostic signatures for the prediction of progression-free survival (PFS) and overall survival (OS).

Methodology/principal findings: We implemented a multivariate Cox Lasso model and median time-to-event prediction algorithm and applied it to two datasets integrated from the four genomic data types. We (1) selected features through cross-validation; (2) generated a prognostic index for patient risk stratification; and (3) directly predicted continuous clinical outcome measures, that is, the time to recurrence and survival time. We used Kaplan-Meier p-values, hazard ratios (HR), and concordance probability estimates (CPE) to assess prediction performance, comparing separate and integrated datasets. Data integration resulted in the best PFS signature (withheld data: p-value = 0.008; HR = 2.83; CPE = 0.72).

Conclusions/significance: We provide a prediction tool that inputs genomic profiles of primary surgical samples and generates patient-specific predictions for the time to recurrence and survival, along with outcome risk predictions. Using integrated genomic profiles resulted in information gain for prediction of outcomes. Pathway analysis provided potential insights into functional changes affecting disease progression. The prognostic signatures, if prospectively validated, may be useful for interpreting therapeutic outcomes for clinical trials that aim to improve the therapy for SeOvCa patients.

Figure 1. Correlation of TCGA clinical outcome measures.

(A) PFS and PFI are strongly correlated and do not need to be predicted separately. (B) PFS and OS are not well correlated, so we derived separate predictive signatures for each (data only for un-censored patients).

Figure 2. Integration Procedure and CoxPath Methodology.

Integration combines multiple data types for the multivariate Cox Proportional hazards model.

Figure 3. Quality of outcome prediction for survival time (A, B) and discrete risk categories (C, D).

(A) Prediction of time-to-event (PFS; un-censored data); (B) prediction of time-to-event (OS; un-censored data); (C) statistically significant stratification into low-, intermediate- and high-risk patients using the prediction method for TCGA test data based on c-score (Integrated PFS signature); and (D) stratification for the TCGA test data based on t-score (Integrated PFS signature).

Figure 4. Canonical pathway analysis of 156 genes from the integrated PFS gene signature.

IPA identified 23 statistically significant canonical pathways (p<0.1 and ≥3 genes).

Figure 5. Overrepresented GO categories for genes in the integrated PFS signature.

Six biological processes categories and two molecular function categories were indentified by Bingo containing (3

Figure 6. Network derived from the integrated PFS signature using IPA.

The top four networks identified were merged using IPA analysis. The features most discriminative between short and long-recurrence times are shown on larger scale. The nearest neighbor interactions of these nodes are highlighted in different colors. Nodes are colored based on the mRNA expression profile of different genes (green: down-regulated in short recurrence patients (PFS<6mo) compared to long recurrence (PFS>40mo), and red: up-regulated).

Figure 7. Netbox modules identified using the integrated PFS signature.

Different modules are spatially separated for visualization. The genes present in our signature are shaped as octagons (mRNA features), diamonds (methylation features) and rectangles (copy number feature). The linker nodes are represented as small circles. Nodes are colored based on the mRNA expression profile of different genes (green: down-regulated in short recurrence patients (PFS<6mo) compared to long recurrence (PFS>40mo), and red: up-regulated).

Figure 8. Features from the integrated PFS signature ranked based on their stratification performance.

Top ranked features (categorized based on their values from the respective data type as low [bottom 15%], intermediate and high [top 15%]) could potentially act as biomarkers and therapeutic targets.