Microvesicle Proteomic profiling of Uterine Liquid Biopsy for Ovarian Cancer Early Detection

: High-grade ovarian cancer (HGOC) is the leading cause of mortality from gynecological malignancies, due to diagnosis at a metastatic stage. Current screening options fail to improve mortality due to the absence of early-stage-specific biomarkers. We postulated that a liquid biopsy, such as utero-tubal lavage (UtL), may identify localized lesions better than systemic approaches of serum/plasma analysis. Furthermore, while mutation-based assays are challenged by the rarity of tumor DNA within non-mutated DNA, analyzing the proteomic profile, is expected to enable earlier detection, as it reveals perturbations in both the tumor as well as in its microenvironment. To attain deep proteomic coverage and overcome the high dynamic range of this body fluid, we applied our method for microvesicle proteomics to the UtL samples. Liquid biopsies from HGOC patients (n=49) and controls (n=127) were divided into a discovery and validation sets. Data-dependent analysis of the samples on the Q-Exactive mass spectrometer provided depth of 8,578 UtL proteins in total, and on average ~3,000 proteins per sample. We used support vector machine algorithms for sample classification, and crossed three feature-selection algorithms, to construct and validate a 9-protein classifier with 70% sensitivity and 76.2% specificity. The signature correctly identified all Stage I lesions. These results demonstrate the potential power of microvesicle-based proteomic biomarkers for early cancer diagnosis. fresh-frozen advanced HGOC


9
output tables from the separate MaxQuant analyses for the discovery and validation cohort are available as Supplementary Tables 2-5. Bioinformatic analysis of the discovery cohort was performed on the log2-LFQ-intensities. Data were filtered to include proteins with valid values in at least 75% of the samples.
Missing values were then imputed by replacing them with random, low intensity values that form a normal distribution with a width of 30%, and downshift of 1.8 standard deviations of the general data distribution. The imputed LFQ intensities of the discovery cohort are provided in Supplementary Table   6. Machine learning was performed on the imputed LFQ intensities of the discovery cohort. Support vector machines (SVM) algorithm using linear kernel function was employed to extract a predictive signature that can discriminate between the control and ovarian cancer patients. We combined three feature selection algorithms: recursive feature elimination (RFE)-SVM, SVM and ANOVA (30). In each of these processes, cross validation was performed on the discovery set with 250 iterations of random sampling of 85% of the samples as test and 15% as validation. The optimal number of overlapping features of these three analytic methods was calculated to provide highest predictive accuracy with the lowest possible error rates in the discovery set. Filtration and imputation of the validation set was performed in the same manner as for the discovery set. The performance of the extracted classifier was then blindly examined on the (log2) LFQ intensities of the 9-signature proteins in the validation cohort (Supplementary Table 7). ROC curve and the AUC calculations were performed in MATLAB. Hierarchical clustering was performed on the z-scored log2 intensities using Euclidean distances between averages.
The FDR of the signature proteins was estimated by permuting the sample labels of the discovery cohort 100 times, followed by the same SVM classification and feature selection procedures. FDR calculation was based on the number of times the protein reached the top 15 ranks out of the 100 permutations.

RNA Extraction and RT-PCR
Fresh-frozen HGOC tumors and fresh grossly benign FT fimbriae were obtained from the Chaim Sheba Institutional Tumor Bank. H&E staining was performed to ensure >80% tumor cellularity. The fimbriae were processed as previously described (31,32). Total RNA was extracted from primary fresh frozen HGOC tumors and dissociated normal FTE cells using QIAzol reagent (Qiagen, Valencia, CA, USA) followed by RNeasy clean-up kit (Qiagen) according to manufacturer's protocol. Gene expression was assessed using FastStart Universal SYBR Green Master (ROX) (Roche). Primers for the signature-genes are listed in Supplementary Table 8 (Sigma-Aldrich).

Statistical Analysis
Statistical significance (p < 0.05) was assessed by Student t-test for RT-PCR data or by Fisher exact test for IHC intensity scores. Binomial model analysis was used to evaluate the correlation of actual diagnosis, age and menopausal status confounders with the prediction. Pearson correlation test was used to examine the correlation of individual protein expression with age, and Chi square test was used to test the correlation of individual protein expression with menopause.

Results
Our approach to identify early-stage biomarkers with high sensitivity and specificity combines liquid biopsies from the lumen of the gynecologic tract, with deep microvesicles proteomics of the samples. To profile the proteome of the complex utero-tubal body fluid and extract diagnostic biomarkers, we adapted the previously described method for microparticle proteomics, which overcomes masking by highly abundant proteins, and followed with high-resolution MS analysis (22). Briefly, we isolated microvesicles from 1ml of UtLF by high-speed centrifugation, and followed by urea-based denaturation and in-solution digestion (Fig. 1B). Peptides were analyzed on the Q-Exactive Plus or Q-Exactive HF MS, and proteins were quantified using the label-free algorithm in MaxQuant. Patient cohorts included samples from HGOC patients and controls (with non-malignant gynecological conditions) from three medical centers.
Initial data analysis included all samples, in a combined MaxQuant analysis, to evaluate the data quality, and examine whether there are any technical artifacts associated with the sample origin and batch.
Combined analysis identified a total of 8760 proteins. Among them, we found known lineage markers of FTE/HGOC, such as MUC16 (CA125), WFDC2 (HE4), and OVGP1 (MUC9), as well as very low-abundance proteins, including cytokines and growth factors, such as IGF1, CXCL12, IL18 and HGF (Fig.1C). The dynamic range of relative abundance of the microvesicle proteome spanned eight orders of magnitude.
Moreover, the concentration of CA125 in unfractionated UtL was measured with a commercial assay (Access Immunoassay OV Monitor, Beckman Coulter), and demonstrated no significant difference between patients and controls (data not shown). Next, we examined potential 'batch effect' or differences in composition of samples (surrogate for UtL sampling technique variations and analysis batches). Principal component analysis (PCA) showed no clear separation between the groups of samples, implying low technical variation between the batches and between the three medical centers (Supplementary Fig. 1A and B). Additionally, correlation analysis between samples showed an average correlation of 0.67 within each center and correlation of 0.66 between centers. Reassuringly, we found higher correlations between controls from different centers, than between patients and controls from the same center ( Supplementary Fig. 1C). We therefore concluded that the batch effects and interinstitutional differences are negligible and did not require any correction. We then investigated whether we can identify significantly different proteins between patients and controls.

Diagnostic UtL-Based Proteomic Classifier
Aiming to identify diagnostic markers, we divided the data into a discovery set and a validation set. To eliminate any dependence between the discovery and the validation cohorts, we analyzed each of these sets separately in MaxQuant. A discovery cohort, including a total of 24 patients and controls, was used to construct a protein classifier for HGOC diagnosis. It was designed to include patients from all three medical centers, exclude any previously treated patients and BRCA carriers, and have equal numbers of cases and controls. MaxQuant analysis of these samples identified a total of 5565 UtL microvesicle proteins, and an average number of ~2500 proteins per sample (range: 1100-3600; Supplementary Fig Table 2). To obtain a signature of minimal number of proteins with highest accuracy and robustness, we tested three feature selection algorithms: Support vector machine (SVM), recursive feature elimination (RFE)-SVM and ANOVA. The entire analytical workflow was embedded in a cross validation procedure to reduce over-fitting. To minimize the dependence on the feature selection algorithm, we tested the performance of several sets of top-ranked overlapping signatures, ranging in size from 6 to 19 features ( Fig. 2A and B). Optimal sensitivity, specificity, and area under the curve (AUC) of Receiver Operating Characteristic (ROC) curve of sensitivity vs. 1-specificity were obtained with a 9protein classifier, 6 of which were higher in the HGOC patients, and 3 that were higher in controls (  Table 2). T-test showed that five of the signature proteins (S100A2, S100A14, SERPINB5, IVL and CLCA4) were also statistically significant (FDR 0.05; s0=0.5) between the control and patient samples in the discovery cohort ( Fig. 2D). This signature demonstrated 83% sensitivity (95% confidence interval: 51.6 -98%), at a specificity of 100% (95% confidence interval: 73.5 -100%), and an AUC of 0.99 in the discovery set (Fig. 2E). The permutation-based FDR for the signature proteins ranged between 0-0.11 (Supplementary Table 9). The coefficients of variation of the nine signature proteins were below 25% ( Supplementary Fig.2D). Importantly, this signature correctly predicted all three stage IA HGOC cases included in the discovery set. Intensities of seven of the nine proteins discriminated them from control samples better than they discriminated advanced stage HGOC samples from controls ( Supplementary Fig. 3A), suggesting the potential strength of this signature in Given the long-standing clinical use of CA125 and HE4 as diagnostic markers, we examined whether their combination with our signature has any predictive advantage. The performance of a combination of the 9-classifier and the best-validated protein biomarkers (CA125 and HE4) was calculated in the discovery and validation cohorts. Both sensitivity and specificity were reduced compared to the 9-protein signature alone: sensitivity of 83% and 68% and specificity of 75% and 67.5% in the discovery and validation cohort, respectively (Fig. 3B). The performance of CA125+HE4 alone was even worse, with sensitivity of 33% and 5.2% and specificity of 50% and 94.7% in the discovery and validation cohorts, respectively (Fig. 3C).
Since the patient age correlated with the prediction, and most HGOC patients were postmenopausal, we tested whether age and menopausal status affect the signature protein expression.
Since hormonal status information was not available for all patients, we divided the cohort into age<=50 (pre-menopausal) vs. age>50 (post-menopausal). Binomial model multivariate analysis demonstrated no correlation of the signature with age (p-value=0.414). P-value for regression correlation of 1.45 with menopausal status was 0.01, since diagnosis of HGOC directly and strongly correlates with menopause (p-value=2.5E-06). Reassuringly, the actual diagnosis strongly correlated with the signature prediction (p-value=3.9E-07). Moreover, the LFQ intensities of the individual signature proteins did not significantly correlate with menopausal status. Only one protein, Ectonucleotide Pyrophosphatase/ Phosphodiesterase 3 (ENPP3), inversely correlated with age (p-value=0.00078; Supplementary Table 9).

Real-time PCR Validation of Differential Expression of Signature Proteins
The UtL liquid biopsy samples proteins that are not necessarily exclusively expressed by the cancer cells, but can also capture stromal response to tumor development, or can result from an increase in specific tissue mass. Some known tumor markers (e.g. CA125) directly reflect an increase in mass of a specific tissue type, and are not uniquely expressed by malignant cells, nor do they possess cancer-promoting biological functions. Such markers are expected to detect tumors at an advanced stage, and may not be appropriate for early cancer diagnosis, whereas cancer-specific expression may increase the sensitivity of signature biomarker and increase the chances of diagnosing the disease at an early stage. We therefore examined the expression patterns of the signature proteins at the RNA level, by RT-PCR, and the protein level, by IHC. We measured the mRNA expression of all signature genes in HGOC tumors vs.
normal FTE on an independent set of unmatched samples: fresh-frozen advanced HGOC tumors (n=10) and unmatched benign FTE cells harvested from normal fimbriae (n=10). Our results indicate statistically significant transcriptomic differential expression (DE) in accordance with the proteomic analysis of five of the nine genes (Fig. 4). The fact that not all transcripts are DE suggests that some proteins remain relatively consistent through malignant transformation and may also stem from the profound differences in the type of biological materials examined (extracellular microvesicle proteins vs. cellular mRNA), and the methodologies used (MS vs. RT-PCR).

Immunohistochemistry (IHC) Validation of Tumor Expression of Signature Proteins
MS and RT-PCR methods lack spatial resolution, thus precluding disclosure of the specific cell-type that expresses each of the classifier's proteins. To explore the localization of selected signature proteins in HGOC tumors and normal FTE, and confirm the DE by tumor cells, we performed IHC for SERPINB5 and S100A14, two selected proteins that were over-represented in UtL of HGOC patients. IHC was performed on a tissue microarray (TMA) of HGOC tumors vs. four control-TMAs representing grossly normal FT fimbriae removed from women with: HGOC, tubal ectopic pregnancy (EP), leiomyomatous uterus (LM, benign condition of the uterus not affecting the fallopian tube), or BRCA-mutation carriers undergoing RRBSO.
SERPINB5 is an epithelial-cell-specific member of the SERPIN family that lacks serine protease inhibition activity. Not much is known about its cellular functions in cancer, yet it has been implicated as cancer susceptibility gene and a prognostic factor in several cancer types (33). It has been also attributed a role as an exosomal protein (34). In accordance with the proteomic analysis, IHC exhibits weak cytoplasmic staining in less than 50% of normal FTE specimens (intensity 0-1), and a stronger expression in a subset of HGOC tumors (p-value= 1.65E-09; Fig.5A, Supplementary Fig. 5). S100A14 is a member of the S100 family lacking calcium-binding function, known to be involved in the regulation of TP53 protein expression and of cellular motility (35). In FTE, it localized exclusively to the cytoplasm of ciliated cells, with very low staining in secretory cells (intensity 0-1) (Fig.5B,   Supplementary Fig. 6). In agreement with the proteomic analysis, its expression was significantly higher in HGOC tumor cells compared to the presumed cell-of-origin -secretory FTE (p-value=2.04E-06; We further obtained IHC evidence from the Human Protein Atlas database (www.proteinatlas.org(36)) for the expression of three additional proteins. According to publicly available histology images in the database, CLCA4, S100A2 and MYH11 had stronger cytoplasmic staining in HGOC tumor cells than in normal FTE. Overall, the IHC results confirm the DE of the five signature proteins in HGOC tumors compared to normal FTE, and localize their expression specifically to tumor cells.

Discussion
In this work, we present the discovery of potential early diagnostic markers for HGOC, using microvesicle proteomics of UtL liquid biopsies. Isolation of microvesicles enabled overcoming the large dynamic range of this body fluid, and untargeted identification of thousands of proteins per sample in single LC-MS runs.
As opposed to our original methodological study of plasma microparticle proteomics (22), in the current work we used LFQ rather than SILAC, since we did not find a suitable SILAC standard for this particular body fluid. The MaxQuant LFQ algorithm enabled general normalization that overcame all potential batch effects; however, even in the separate analysis of the discovery and validation sets, predictive ability of the signature was still high. We believe that this work is the first step towards translation of the signature proteins into a clinical test, and envision that such a test will use simpler MS-based targeted assays and shorter analytical times. MS-based clinical tests are expected to increase the accuracy and multiplexing capabilities compared to more commonly used antibody-based tests (e.g. ELISA).
Furthermore, these will reduce the cost and assay development times due to the high specificity of the MS results (37). Recent attempts to advance the applicability of the MS-based assays have already simplified the sample preparation and MS analyses, increased the throughput and implemented targeted MS methodologies, combined with absolute quantification (38-41). Future combination of our study with such technologies can potentially lead to implementation of the signature proteins to routine clinical labs.
Early diagnosis of HGOC is of highest importance to women with genetic predisposition, since they are currently counselled to undergo RRBSO around the age of 40, despite the incomplete penetrance and the highly variable age of presentation. This practice gains legitimacy from the exceedingly narrow window-of-opportunity for early-stage diagnosis and the unbearably high mortality rates, thus necessitating extremely cautious management. Recently, evolutionary mutation analyses revealed that the time gap between development of STIC and clinical appearance invasive HGOC is longer than 6 years (17), thus implying that early detection may, after all, be possible once new methodologies become available. UtL liquid biopsy, as opposed to blood, may potentially disclose localized HGOC lesions, which are curable.
Our 9-protein classifier has 70% sensitivity and 76% specificity which outperforms previous results of genomic biomarkers based on gynecological liquid biopsy (18)(19)(20). Unlike mutation analysis in UtL samples which looks at a negligible percent of cancer cells, proteomics reflects the complexity of a cancer-associated program that, theoretically, captures expression changes in multiple cell types within the tumor microenvironment, thus can potentially provide a wider array of early-detection biomarkers.
Further improvement of the proteomic signature and its predictive power, requires analysis of more early-stage HGOC UtL samples or STICs, however, these samples are inherently exceedingly rare.
Coupled with the intra-uterine liquid biopsy method, this assay holds promise for clinically significant early detection of HGOC.
The UtL sampling technique that we propose hereby is a simplified version of the originally reported method (18), making it suitable for routine testing of healthy young women at high risk for HGOC, including women who have not undergone vaginal delivery. Fundamental parameters for clinical feasibility, such as patient-reported outcomes, physicians-reported workload and compliance of the target population to undergo routine UtL sampling need to be investigated. Semi-annual monitoring with clinical proteomic assays may be implemented as a measure of reassurance for high-risk populations willing to delay RRBSO until after menopause, and thus become practice changing.
To consolidate the specificity of the signature proteins to HGOC tissues, we examined their expression in independent tissue specimens, comparing FTE and HGOC, using complementary techniques: RT-PCR and IHC. We obtained confirmatory IHC results for five proteins and supportive RT-PCR results for five of the nine genes tested, highlighting the aberrant expression of these proteins in HGOC tissues. These results reinforce the potential of the proteomic signature as a diagnostic test. Of note, discordance between the proteomic predictions and transcriptomic validation results may arise from the differences between mRNA and protein expression patterns, and between the extracellular vesicles and intracellular levels of expression. Alternatively, it is possible that the expression of several proteins is not altered when FTE evolves into HGOC, since they are not directly involved in the cancerous process, but may still be useful biomarkers, like CA125, for aberrant expansion of the cell lineage within HGOC lesions.
Ultimately, the genomic and the proteomic approaches, as well as other possible methodologies of liquid biopsy analysis, may be integrated to yield a multi-modality classifier with an adequate sensitivity and specificity to guarantee early detection of HGOC in both average-and high-risk populations, and potentially enable personalized risk stratification and delay of RRBSO in predisposed women without increasing HGOC incidence.