Integration of proteomics with CT-based qualitative and radiomic features in high-grade serous ovarian cancer patients: an exploratory analysis

This was a multi-institutional, institutional review board–approved, and Health Insurance Portability and Accountability Act (HIPAA)–compliant retrospective study, with waiver of informed consent from all institutions that participated in The Cancer Genome Atlas-Ovarian Cancer (TCGA-OV) Imaging Research Group [21].

The eligibility criteria included the following: (1) confirmed diagnosis of HGSOC, (2) HGSOC tissue submitted and analyzed by TCGA, (3) no neoadjuvant chemotherapy administered, (4) intravenous contrast-enhanced CT of the abdomen and pelvis performed prior to primary cytoreductive surgery, and (5) protein relative abundance measurements available. The final patient cohort consisted of 20 patients. Figures 1 and 2 summarize the experimental design. All patients were included in two prior studies that evaluated the association between the Classification of Ovarian Cancer (CLOVAR) subtype signatures and CT features in a single- [17] and multi-institutional setting [21], respectively. All patients were also included in two other studies that evaluated the feasibility of CT-based texture measures in the quantification of inter-site tumor heterogeneity and combined intra- and inter-site tumor heterogeneity [17, 22] in patients with HGSOC. None of these prior studies involved any analysis of protein abundance data.

Thirty-six image features, including the 33 radiologist-scored CT imaging traits used in the study by Vargas et al, such as lesion size and laterality, locations of peritoneal disease, nodal stations involved, and locations of metastases, and three computer-extracted texture metrics were obtained [21]. All CT-derived variables are shown in Table 1 and Table 2.

The CT images were a part of the TCIA TCGA-OV [23] collection in The Cancer Imaging Archive (https://​cancerimagingarc​hive.​net) [24], an NCI-supported archive of anonymized medical images. The cases were reviewed on a cloud-based virtual machine using the ClearCanvas with Annotation and Imaging Markup viewing software (Northwestern University) [25]. Image data in DICOM format, and features and segmentations generated for this paper, are available at https://​doi.​org/​10.​7937/​TCIA.​2019.​9stoinf1 [26].

The CT imaging traits were assessed by two radiologists (E.S. and H.A.V.; 17 and 12 years’ experience in ovarian cancer imaging) independently, who were blinded to the patients’ clinical, pathological, and proteomic data. The details of the CT imaging trait evaluation are provided in the study by Vargas et al [17]. Briefly, each reader individually recorded the imaging interpretation on an electronic case report form, which was uploaded to a central server, compiled, and submitted for statistical analysis. Both ovarian masses and peritoneal disease were evaluated. If a definable ovarian mass was present, its features were recorded, including laterality, maximum size, internal architecture, and presence of calcifications and/or septations. The presence or absence of definable peritoneal implants and their specific location (small bowel mesentery, omentum, paracolic gutters, Pouch of Douglas, perisplenic/left upper quadrant region, lesser sac, and perihepatic/right upper quadrant region) was also recorded. Lymphadenopathy was defined as a short-axis dimension above a predefined size cutoff specific to each location or specific imaging appearance, as detailed in Vargas et al [21].

In the event, the two radiologists disagreed on a categorical feature (e.g., presence of peritoneal disease in the mesentery), and a third radiologist (L.B. 5th year of training) served as an arbitrator. For quantitative features, such as lesion size, the measurements of the two radiologist were averaged. The number of sites in which peritoneal disease was present was also recorded and was considered an additional human-read feature. This was also done for the presence of lymphadenopathy and intra-parenchymal metastases.

A detailed description of image segmentation and texture feature extraction are provided in Veeraraghavan et al [22]. A slice thickness of 5 mm was used for analysis. kVp were variable, as CT machines from different vendors in different institutions were used. In brief, two oncologic imaging research fellows with 4 and 6 years of experience (M.M., S.N.), respectively, in consensus, manually segmented primary ovarian tumor(s) and all metastatic tumor implants in the abdomen and pelvis. Segmentation was performed using 3DSlicer [27] by tracing the contour of each lesion on each slice to produce the volumes of interest (VOI). In total, tumor burden in 14 distinct anatomical regions was manually segmented. Voxel-wise Haralick textures (energy, entropy, contrast, and homogeneity) were computed from within the manually delineated VOIs using in-house software implemented in C++ using the Insight ToolKit [28]. Site-specific sub-regions were computed by voxel-wise clustering of the Haralick textures using the kernel K-means method [29]. Following clustering, tumor sites were divided into distinct sub-regions of similar texture features. The intra- and inter-site heterogeneity (IISTH) measures (i.e., cluster site entropy, cluster standard deviation, and cluster dissimilarity) were computed. IISTH measures summarized the heterogeneity across all tumor sites in each patient. The more different each site of disease, the higher are the IISTH measures.

Given that the molecular profiles of TCGA patients were based on primary ovarian tumor specimens and our radiologic measurements were made on metastatic patterns of disease, we sought to identify proteins conserved between primary and metastatic tumors in an ovarian cancer patient and then cross-reference these with proteins associated with radiologic patterns of metastasis and proteomic profiles in TCGA. Protein relative abundance measurements from the CPTAC analysis of 107 TCGA HGSOC patients were available for 3377 proteins (normalized log-ratios, standardized across two institutions), which were obtained from Zhang et al [11]. When multiple measurements were available, their abundances were averaged.

As we included only 20 patients in this study, a simple correlation analysis between the protein abundance of 3377 proteins and the CT imaging traits and texture features would result in a high number of false-positive results. Therefore, the purpose of the workflow described below was to reduce the number of proteins used for the final analysis. First, the 3377 proteins were prioritized for inclusion in the analysis, based on evidence of conserved transcript expression between primary and metastatic sites and on agreement between transcript and protein abundance levels. Transcript expression (Affymetrix) data (Supplemental Table 2 in reference [30]) from the primary tumor and multiple metastatic sites (spleen, right-upper-quadrant, liver, and vaginal cuff) from an HGSOC patient [30], along with transcript expression (mRNA-seq) data from primary HGSOC tumors (n = 107) [31] with matching protein expression data from the same case set (n = 107) [11], were merged by gene name, yielding a final matrix of 5504 co-measured proteins and transcripts (Supplemental Table 1, in reference [11]). Second, coefficients of variation (CV) values in transcript expression between the primary and metastatic sites, as reported by Jimenez-Sanchez et al [30], and Spearman correlations calculated for transcript and protein expression from the n = 107 patients, were normalized using the boxCox function in R (version 3.3.2). Transcripts that exhibited a low CV between the primary and metastatic sites (i.e., stable expression independent of metastatic loci), and with expression levels highly correlated to their proteins, were selected by comparison with normal distributions using T test analyses and a significance threshold of a p value < 0.1. These analyses identified 529 candidates with conserved expression between primary and metastatic sites and 569 candidates highly correlated at the protein and transcript level; 47 of these intersected between these two analyses. To further reduce the number of proteins, we prioritized proteins based on their molecular function and their interaction with each other using MetScape [32]. Sixteen proteins were associated with the regulation of the amino acid metabolism (p < 0.001). The selection of these 16 proteins was done solely on the basis their summary statistics (coefficients of variation in transcript expression between primary and metastatic sites, correlations between transcript and protein expression) and molecular function while correlations of imaging features were not considered in any way in selecting these proteins. Finally, these 16 candidates were included in subsequent analyses to identify those that further correlated with clinical imaging features (Fig. 1). The 16 proteins are given in Supplementary Table 1.

The percentage agreement and the Cohen’s kappa coefficient used to assess the agreement between the two radiologists were calculated for all CT-based imaging features. Univariate associations were assessed between the expression levels of the proteins and CT-based image features in which previous results provided evidence of prognostic power (presence of peritoneal disease in the mesentery and diffuse—non-mass forming—peritoneal involvement, as found in the study by Vargas et al [21]: number of sites with peritoneal disease; presence of peritoneal disease in the liver; presence of Pouch of Douglas implants; presence of supradiaphragmatic lymphadenopathy; and non-visualization of a discrete ovarian mass). For binary image features, these associations were assessed using a Mann-Whitney U test to compare the expression levels of a specific protein among patients with the feature to those among patients without the feature. The area under the receiver operating characteristic curve (AUC) was reported as a metric of the strength and direction of this association. For quantitative features, these associations were assessed through inferences on the Kendall tau rank correlation coefficient.

The patients’ characteristics are shown in Table 3. The median number of tumor sites was six (range, 1–11). The median tumor volume was 192 cm³ (range, 12–2574 cm³), the highest being the tumor volume of the ovarian masses, with a median of 80 cm³ (range, 0–600 cm³). Figure 3 illustrates the tumor volumes for each anatomical sub-region.

There was a good to excellent inter-reader agreement for CT-based imaging traits, with an absolute agreement ranging between 75% (κ = 0.32) for the variable mass internal architecture to 100% (κ = 1.0) agreement for the majority of variables (Table 1).

Four CT imaging traits were associated with an abundance of several proteins (Table 4). For example, disease in the mesentery was associated with reduced levels of CRIP2 (p = 0.0002, AUC = 0.05) and GPI (p = 0.004, AUC = 0.22). Supradiaphragmatic lymphadenopathy was associated with decreased abundance of ALDH3A2 (p = 0.02, AUC 0.192) and increased abundance of MAGEA4 (p = 0.046, AUC = 0.768). The number of sites with peritoneal disease was correlated with increased levels of ALDH3A2 (p = 0.03, τ = 0.36) and CRIP2 (p = 0.01, τ = 0.378). The shape of peritoneal disease was associated with the abundance of MAGEA4 (= 0.04, τ = − 0.343). After correction for multiple testing through the Benjamini-Hochberg procedure, the correlation between CRIP2 and the presence of supradiaphragmatic lymphadenopathy remained significant (corrected p = 0.03). Patients with without or diffuse peritoneal disease had higher levels of MAGEA4 (median 2.31 and 2.31 respectively) than patients with predominantly diffuse or predominantly nodular disease (median − 0.233 and − 1.51 respectively).

The protein abundance of three proteins was associated with intra- and inter-site tumor heterogeneity texture metrics (Table 5). For example, cluster site entropy was positively correlated with the abundance of STXBP2 (p = 0.007, τ = 0.432) and negatively with ASS1 (p = 0.011, τ = − 0.364). Cluster standard deviation was positively correlated with an STXBP2 (p = 0.05, τ = 0.453). Cluster dissimilarity was positively correlated with STXBP2 (p = 0.03, τ = 0.368) and negatively with ASS1 (p = 0.009, τ = − 0.427) and CBK (p = 0.047, τ = − 0.326). However, none of these associations were significant after correction for multiple testing through the Benjamini-Hochberg procedure.