Repurposing the liquid-based Pap test for the detection of ovarian cancer protein biomarkers.

Under an IRB approved protocol (1112M07362), we developed “Mock” Pap test kits to mimic the sample collection that would occur in a clinical setting, using a Pap test broom and a SurePath™ vial (BD Biosciences). After obtaining signed consent, “Mock” Pap tests were collected prior to surgery from chemo-naïve women with abdominal masses suspected to have ovarian cancer. High grade serous ovarian carcinomas (HGSOC) are the most common subtype of ovarian cancer and are the focus of this study. As samples were collected at the time of surgery, most HGSOC cases were late stage (III or IV) with only a few early stage samples collected. “Mock” Pap tests were also collected from women with benign ovarian conditions as controls for ovarian cancer. As an additional control, “Mock” Pap tests were collected from healthy women over age 40 in the University of Minnesota Women’s Clinic. Samples were sent to the lab where they were vortexed and the collection broom was removed from the SurePath™ vial before storage at 4 °C. SurePath™ vials were vortexed and 5 mL of the Mock Pap fluid was centrifuged at 400xg to remove cells. Cell-free Pap test fluid samples were precipitated using acetone to remove contaminants and concentrate the proteins, then solubilized in lysis buffer (8 M urea, 75 mM NaCl, 50 mM Tris, pH 8.0, 1 mM EDTA). Protein concentration was determined by BCA assay (Thermo Scientific). The recovered proteins, up to 100 µg, were reduced with 5 mM DDT for 30 min and subsequently alkylated with 10 mM iodoacetamide for 30 min; then diluted 1:3 with 50 mM Tris-HCl (pH 8.0) and subjected to proteolytic digestion with LysC (Wako Chemicals) at 1 mAU:50 µg enzyme-to-substrate ratio for 2 h at room temperature, followed by the addition of sequencing-grade modified trypsin (Promega) at a 1:50 enzyme-to-substrate ratio and overnight incubation at room temperature. The digested proteins were acidified with 50% trifluoroacetic acid (TFA, Sigma) to a pH value of approximately 2.0. Tryptic peptides were desalted on reversed-phase C18 SPE columns (Waters) and dried using a Speed-Vac (Thermo Scientific). Peptide recovery was then determined by BCA assay. Thirty µg of desalted peptides from each sample were labeled with 11-plex TMT reagents according to the manufacturer’s instructions (ThermoFisher Scientific) with minimal modification to accommodate reduced sample mass (ThermoFisher Scientific). Mass tags for 40 “Mock” Pap test samples (Additional Table  1 ) were distributed between the three sample types (20 HGSOC, 10 benign ovarian conditions, and 10 healthy controls) across four TMT 11-plex experiments using a randomized block design. Channel 131 N was used for labeling the internal reference sample (pooled from all samples) throughout the sample analysis. Peptides labeled by different TMT reagents were mixed, dried using a Speed-Vac, reconstituted with 3% acetonitrile, 0.1% formic acid and desalted on tC18 SepPak SPE columns. The pooled sample of TMT labeled peptide samples was fractionated by high-pH RPLC as described [ 20 , 33 , 34 ]. Approximately 500 µg of 11-plex TMT labeled sample was loaded on the reversed phase Agilent Zorbax 300 Extend-C18 column and 96 fractions were collected. The fractions were then concatenated into 24, by combining 4 fractions that are 24 fractions apart, to create 24 total for subsequent analysis by LC-MS/MS analysis. Fractionated samples were separated using a nanoACQUITY UPLC system (Waters) by RPLC [ 20 , 34 , 35 ] coupled to an Orbitrap Fusion Lumos MS (ThermoFisher Scientific). Orbitrap precursor spectra (AGC 4 × 10 5 ) were collected from 350 to 1800 m/z for 110 min at a resolution of 60 K along with data dependent Orbitrap higher-energy collision dissociation (HCD) MS/MS spectra (centroid) at a resolution of 50 K (AGC 1 × 10 5 ). MS data sets were searched against protein sequence database using MS-GF+ [ 36 , 37 ], followed by mzRefinery [ 38 ] to characterize and correct for any instrument systematic mass measurement errors. Results were filtered with a false discovery rate (FDR) of < 1%. The intensities of all TMT reporter ions were extracted using MASIC software [ 39 ]. Protein abundances were summarized from peptide-level reporter ion intensities and normalized to the plex-specific reference channel by dividing each protein intensity by its corresponding reference intensity. The resulting relative protein abundances were log2-transformed, globally median-normalized across samples to account for loading/labeling differences and then corrected for TMT plex-specific batch effects. Differential expression analysis was performed comparing HGSOC against both healthy controls and benign cases using the limma R package [ 40 ] . Significantly differentially expressed proteins were defined as those with p-value < 0.05. Peptides for targeted assays were selected from proteins identified in the TMT data and based on tryptic peptides ranging from 6 to 24 amino acids in length which produce multiply charged ions and uniquely represent the target proteins. Fully tryptic peptides allowing K/R.P [ 41 ] were selected and those peptides with post-translational modifications (PTMs) were excluded [ 42 – 44 ], and in general, methionine-containing peptides were excluded due to its potential for oxidation during storage and sample preparation except when there were no other alternatives. UniProt Knowledgebase was used to check if any PTMs has been reported for a given sequence. The detectability of peptides was determined using online depositories such as PeptideAtlas [ 45 ], and our own data from discovery proteomics studies. For peptides which were not in the database or existing data, computational tools such as PREGO [ 46 ] were used. The peptide sequence uniqueness was checked with protein coverage summarizer [ 47 ], which maps candidate peptides against the whole human proteome and reports the number of proteins matched by each peptide. Only peptides that matched a single protein (uniquely mapping peptides) were selected. Four to five candidate peptides for each target protein were selected and a total of 144 crude heavy peptides were synthesized (Vivitide, Gardner, MA) and used to optimize the LC-SRM-MS assays. For the full list of the 144 peptides used for assay development see Additional Table 2A . A crude heavy peptide mixture at 3 pmol/µL for each peptide was created by mixing stock solutions of all heavy peptides (2 nmol/µL in 15% ACN), and analyzed by LC-MS/MS using HCD on a Lumos MS coupled to nanoACQUITY UPLC system to check the peptide quality and to obtain information on peptide fragmentation (i.e., transitions). Skyline was used to build the peptide library [ 48 ]. The output file from Skyline was loaded onto a Thermo Scientific TSQ Altis triple quadrupole (QQQ) MS for LC-SRM-MS analysis to obtain optimal transitions and collision energy (CE) for each peptide. This was initially performed in a background of diluted Shewanella oneidensis digest, and subsequently in a pooled Pap test protein digest composed of multiple Pap test protein samples from all three types of cases (HGSOC, benign, and healthy) to remove any transitions with endogenous interference. The data quality of endogenous and heavy peptides was monitored based on signal, peak shape and stability in retention time. Seven of the peptides lacked sufficient heavy and endogenous signals and were therefore excluded from the final assay. A minimum of three transitions per peptide was selected. An additional 112 peptides of 56 proteins (6 peptides comprising 5 proteins overlapped with the TMT list) from the CPTAC earlier studies [ 21 , 49 ] were also included in the final assay (Additional Table 2B ). The final SRM-MS assay consisted of 243 peptides comprising 81 proteins (Additional Table 3 ). The final list of 1470 transitions, including both endogenous and heavy peptides, was assessed using a scheduled LC-SRM-MS method with under 500 concurrent transitions per acquisition cycle, using a 12-minute retention time window. For LC-SRM-MS analyses, a heavy internal standard mixture of about 120 fmol/µL for each peptide was prepared, and 2 µL of the heavy standard mixture was added to 18 µL of each patient’s peptide sample containing 5 µg of peptide digests. Ninety “Mock” Pap test samples were analyzed in the SRM-MS assay: 40 HGSOC, 20 benign ovarian conditions, and 30 healthy controls (Additional Table 4 ). The final concentration of the samples was adjusted to 0.2 µg/µL and analyzed on a nanoACQUITY UPLC system interfaced to an TSQ Altis QQQ MS equipped with a nano-ESI source operated in positive ion-mode with the ESI voltage set to 2,100 V and a capillary temperature of 350 °C [ 50 , 51 ]. The generated raw data was analyzed using Skyline 4 [ 48 ]. All LC-SRM data were imported into Skyline software, where peak boundaries were manually inspected to ensure accurate peak assignment and integration. Undetected peptides and peptides that did not correlate well with other peptides within the proteins were excluded, leaving a total of 110 peptides, corresponding to 66 proteins. The total peak area ratios of endogenous light peptides to their corresponding heavy isotope-labeled internal standards were calculated using Skyline and used for subsequent analysis. The raw intensity values for each peptide were mean-normalized by dividing each peptide intensity by its mean value across samples. In four cases, the heavy peptide standard was not detected, so the peptide could not be quantified. For the small number of peptides with zero values (< 1%: 0.44% at the protein level and 0.64% at the peptide level), values were imputed using one-half of the minimum observed value for the corresponding peptide. Because the normalized peptide intensities exhibited right-skewness, a log2 transformation was applied. To further correct for sample-level variation, median normalization was performed by subtracting the median value of each sample from all peptide measurements within that sample. For protein-level analysis, peptide measurements corresponding to the same protein were averaged after peptide-level mean normalization and prior to log2 transformation and sample-level median normalization. Benign cases were selected as “controls” in this study, based on their close pathology to ovarian cancer. Thus, of the 20 women with benign ovarian conditions, there were 8 cases of ovarian serous cystadenomas, 3 cases of benign ovarian fibromas, 3 cases of ovarian serous cystadenofibroma, 2 cases of fibrothecoma, serous adenofibroma, 2 cases of hemorrhagic cysts, 1 case of benign Brenner tumor with serous cystadenoma, and 1 case of a benign simple cyst. Comparisons between groups were performed using the limma method on the log-transformed intensity values with adjustment for the three known batches. The Benjamini & Hochberg adjusted p-value was used to control the false discovery rate (expected proportion of false discoveries among the rejected hypotheses). The fold change was calculated as the ratio of the geometric means. To plot the individual protein and peptide data by group, the log intensity values were adjusted for the batch effect using removeBatchEffect from the limma package in R [ 52 ]. The multi-protein classifier comparing the data from 40 HGSOC samples versus 30 healthy samples was developed using L1-penalized logistic regression. The tuning parameter for the regression model was selected via 10-fold cross-validation to be the largest value for which the cross-validation error was within 1 standard error of the minimum cross-validation error. To account for the known batch effect, an unpenalized batch indicator was included in the model. The area under the Receiver Operating Characteristic curve (AUC) and sensitivity at 90% specificity were calculated using the predicted probabilities from the held-out cross-validation folds with confidence intervals based on the bias-corrected bootstrap case cross-validation method of Jiang et al. [ 53 ]. Following determination of the proteins included in the multi-protein classifier, stability selection was conducted using 100 bootstrap resamples [ 54 ] as a validation step to assess robustness, with proteins selected in more than 50% of resamples considered stable. In a similar manner, the HGSOC-healthy multi-peptide classifier was developed. The HGSOC versus healthy multi-protein classifier was applied to HGSOC and healthy sample groups to compare predicted HGSOC risk profiles. For healthy and HGSOC samples, predicted probabilities of HGSOC were obtained from the held-out testing folds of the original 10-fold cross-validation procedure, ensuring out-of-sample evaluation. The same classifier was then applied, without retraining, to an independent set of benign samples to generate predicted probabilities of HGSOC. Predicted probabilities were summarized and compared across 40 HGSOC, 30 healthy, and 20 benign groups to assess whether benign samples exhibited cancer-like or healthy-like prediction patterns under the HGSOC-healthy classifier.

Forty “Mock” Pap test samples were run in four TMT 11-plex experiments for biomarker discovery. Twenty 20 HGSOC, 10 benign ovarian, and 10 healthy cases were selected (Additional Table  1 ) with a pooled sample as the reference channel. A total of 3105 proteins were identified in the “Mock” Pap test samples in the four TMT experiments. A total of 700 proteins were identified with complete TMT quantification across all 40 Pap test samples and showed potential differential expression when comparing HGSOC against either the healthy or benign cases. When the expression levels of these 700 proteins were compared by Eisen clustering (cosine metric) maintaining samples within their diagnostic groups, group-specific patterns were observed (Fig.  1 A), with the majority of the 20 HGSOC samples showing patterns that distinguished them from the 10 benign and 10 healthy samples. When unsupervised Eisen clustering was performed using the average linkage method and cosine metric, the majority of the HGSOC samples grouped together in 3 major clusters, with most of the 10 benign ovarian samples clustering nearby (Fig.  1 B). The majority of the 10 healthy samples clustered away from the HGSOC samples. The separation of HGSOC from the healthy samples was further exhibited by three-dimensional principal component analysis (Fig.  1 C), where confidence (95%) ovals highlight the clustering of HGSOC, benign, and healthy cohorts. Thirty proteins that showed differential expression (p-value < 0.05) across HGSOC, benign, and healthy cases were selected as candidate biomarkers for SRM-MS assay development (Table  1 ) encompassing both over- and under-expressed proteins in the HGSOC cases relative to the benign and healthy (Fig.  1 A). The abundance distributions of these 30 selected proteins across the three diagnostic groups are shown as boxplots in Additional Fig.  1 . Fig. 1 Differential expression of proteins based on diagnosis using TMT-labeled quantitative proteomics. A total of 700 proteins that had complete TMT quantification across all 40 Pap test samples [(20 HGSOC (red), 10 benign ovarian conditions (blue), and 10 healthy (green)] and showed potential differential expression (uncorrected p-value < 0.05) when comparing HGSOC against either the healthy or benign cases. (A and B) Heatmap of protein-level clusters (rows) derived via average-linkage Eisen clustering (cosine metric). ( A ) Samples (columns) are maintained in a predefined order to highlight group-specific patterns without additional column clustering; ( B ) Heat-map of unsupervised Eisen clustering of both proteins (rows) and samples (columns). Clustering was performed using the average linkage method and cosine metric; ( C ) Three-dimensional Principal component analysis (3D-PCA). Total explained variances are shown on the axes. Confidence (95%) ovals highlight the clustering of HGSOC, benign, and healthy cohorts Differential expression of proteins based on diagnosis using TMT-labeled quantitative proteomics. A total of 700 proteins that had complete TMT quantification across all 40 Pap test samples [(20 HGSOC (red), 10 benign ovarian conditions (blue), and 10 healthy (green)] and showed potential differential expression (uncorrected p-value < 0.05) when comparing HGSOC against either the healthy or benign cases. (A and B) Heatmap of protein-level clusters (rows) derived via average-linkage Eisen clustering (cosine metric). ( A ) Samples (columns) are maintained in a predefined order to highlight group-specific patterns without additional column clustering; ( B ) Heat-map of unsupervised Eisen clustering of both proteins (rows) and samples (columns). Clustering was performed using the average linkage method and cosine metric; ( C ) Three-dimensional Principal component analysis (3D-PCA). Total explained variances are shown on the axes. Confidence (95%) ovals highlight the clustering of HGSOC, benign, and healthy cohorts Table 1 Proteins from TMT experiments differentially expressed between HGSOC cases and healthy controls and benign conditions UniProtID Entry name Protein name P07741 APT Adenine phosphoribosyltransferase P19652 A1AG2 Alpha-1-acid glycoprotein 2 (Orosomucoid-2) P04003 C4BPA C4b-binding protein alpha chain P08603 CFAH Complement factor H Q16531 DDB1 DNA damage-binding protein 1 Q9C0E8 LNP Endoplasmic reticulum junction formation protein lunapark Q8N9N8 EIF1A Eukaryotic translation initiation factor 1 A domain-containing protein Q96CS3 FAF2 FAS-associated factor 2 P23142 FBLN1 Fibulin-1 P11216 PYGB Glycogen phosphorylase P00738 HPT Haptoglobin (Zonulin) P00739 HPTR Haptoglobin-related protein P07910 HNRPC Heterogeneous nuclear ribonucleoproteins C1/C2 P26583 HMGB2 High mobility group protein B2 P16402 H13 Histone H1.3 Q12906 ILF3 Interleukin enhancer-binding factor 3 Q9Y608 LRRF2 Leucine-rich repeat flightless-interacting protein 2 P55083 MFAP4 Microfibril-associated glycoprotein 4 O00533 NCHL1 Neural cell adhesion molecule L1-like protein Q8NDX1 PSD4 PH and Sect. 7 domain-containing protein 4 Q13492 PICAL Phosphatidylinositol-binding clathrin assembly protein O95865 DDAH2 Putative hydrolase DDAH2 Q9NWH9 SLTM SAFB-like transcription modulator P82979 SARNP SAP domain-containing ribonucleoprotein P50452 SPB8 Serpin B8 Q01082 SPTB2 Spectrin beta chain, non-erythrocytic 1 Q15459 SF3A1 Splicing factor 3 A subunit 1 Q16563 SYPL1 Synaptophysin-like protein 1 P48643 TCPE T-complex protein 1 subunit epsilon O60635 TSN1 Tetraspanin-1 Proteins from TMT experiments differentially expressed between HGSOC cases and healthy controls and benign conditions Table 2 Sixty-one peptides with significantly higher levels of expression when comparing HGSOC versus healthy controls Protein name Peptide sequence Fold change p -value Adjusted p -value 1433B AVTEQGHELSNEER 5.539 < 0.001 < 0.001 A1AG2 EQLGEFYEALDCLCIPR 5.574 < 0.001 < 0.001 A1AG2 EHVAHLLFLR 6.480 < 0.001 < 0.001 A1AG2 NWGLSFYADKPETTK 8.151 < 0.001 < 0.001 A1AG2 SDVMYTDWK 8.904 < 0.001 < 0.001 A1BG ATWSGAVLAGR 3.689 0.002 0.004 A1BG CLAPLEGAR 4.334 0.001 0.003 A2GL TLDLGENQLETLPPDLLR 4.258 < 0.001 < 0.001 A2GL DLLLPQPDLR 4.955 < 0.001 < 0.001 ANXA1 GTDVNVFNTILTTR 3.307 0.011 0.022 APOE AATVGSLAGQPLQER 2.886 0.004 0.009 APT IDYIAGLDSR 1.903 0.018 0.033 APT DALEPGQR 2.673 < 0.001 0.001 APT AAIGLLAR 2.806 0.001 0.003 C4BPA EDVYVVGTVLR 3.899 < 0.001 < 0.001 CD47 STVPTDFSSAK 2.976 0.001 0.002 CERU ALYLQYTDETFR 4.171 < 0.001 0.001 CERU DIASGLIGPLIICK 5.031 < 0.001 < 0.001 CFAH CFEGFGIDGPAIAK 2.252 0.012 0.024 CFAH LSYTCEGGFR 4.620 < 0.001 < 0.001 CFAH WQSIPLCVEK 5.705 < 0.001 < 0.001 CO3 TIYTPGSTVLYR 4.478 0.001 0.003 CO3 VHQYFNVELIQPGAVK 5.446 < 0.001 0.001 CO4B ASSFLGEK 4.829 < 0.001 0.001 CRP GYSIFSYATK 7.861 < 0.001 < 0.001 CRP ESDTSYVSLK 8.818 < 0.001 < 0.001 DAG1 VVENGALLSWK 4.104 0.001 0.002 DDAH2 GAEIVADTFR 1.747 0.024 0.044 DDAH2 GGGDLPNSQEALQK 2.513 0.001 0.003 DDB1 YLAIAPPIIK 1.647 0.018 0.033 DDB1 IVVFQYSDGK 2.037 0.003 0.008 EIF1A TPGNNLHEVETAQGQR 1.853 0.004 0.009 FBLN1 TGYYFDGISR 3.525 < 0.001 0.001 FBLN1 GYHLNEEGTR 3.816 0.001 0.002 FBLN1 EFTRPEEIIFLR 6.012 0.001 0.002 HPT VVLHPNYSQVDIGLIK 8.243 < 0.001 < 0.001 HPT DYAEVGR 11.310 < 0.001 < 0.001 HPT VGYVSGWGR 11.541 < 0.001 < 0.001 HPT VTSIQDWVQK 12.322 < 0.001 < 0.001 IBP3 ALAQCAPPPAVCAELVR 3.012 0.008 0.016 ICAL QAEPELDLR 2.405 0.001 0.002 ICAL AAAPAPVSEAVCR 4.732 0.003 0.007 IL18 SDIIFFQR 4.010 < 0.001 0.001 KNG1 TVGSDTFYSFK 4.795 < 0.001 < 0.001 KNG1 YFIDFVAR 5.505 < 0.001 < 0.001 LEG1 FNAHGDANTIVCNSK 2.392 0.005 0.012 LEG3 IQVLVEPDHFK 2.307 < 0.001 0.002 LEG3 IALDFQR 4.048 0.001 0.002 LG3BP SDLAVPSELALLK 3.394 0.001 0.003 LUM FNALQYLR 3.357 0.001 0.003 MSLN EIDESLIFYK 2.134 0.014 0.027 MUC16 VLQGLLGPIFK 2.610 0.010 0.020 NCHL1 VIAVNEVGR 1.846 0.010 0.020 PIGR ADEGWYWCGVK 6.172 0.003 0.008 PYGB LVTSIGDVVNHDPVVGDR 2.951 < 0.001 0.001 S10A4 ELPSFLGK 1.968 0.018 0.033 S10A4 ALDVMVSTFHK 2.533 0.006 0.013 SYPL1 TVTATFGYPFR 3.200 0.001 0.003 TCPE DVDFELIK 1.562 0.008 0.016 TCPE IADGYEQAAR 5.098 < 0.001 < 0.001 TSN1 VEGCFNQLLYDIR 4.017 < 0.001 0.001 Names of the proteins with peptides having adjusted p-values <0.05. The fold changes are listed along with the p-values and adjusted p-values Sixty-one peptides with significantly higher levels of expression when comparing HGSOC versus healthy controls Names of the proteins with peptides having adjusted p-values <0.05. The fold changes are listed along with the p-values and adjusted p-values We developed a multiplex SRM-MS assay as a high throughput method for targeted validation of the 30 biomarkers identified in our TMT experiments. We selected a total of 144 target peptide sequences from the 30 candidate biomarker proteins to use for assay development (Additional Table 2A ). A minimum of three transitions from 3 peptides per protein were retained for all target proteins. The final list of transitions for both endogenous and heavy peptides was assessed using a scheduled LC-SRM-MS method under 500 transitions per event. The multiplexed LC-SRM-MS assay developed from the 30 TMT protein biomarkers was combined with 112 peptides from 56 ovarian cancer candidate proteins developed by the CPTAC (Additional Table 2B ), based on differential expression between HGSOC and normal fallopian tube epithelium [ 21 ]. Following optimization, the resulting SRM-MS assay was used to quantify a total of 243 peptides from 81 ovarian cancer candidate biomarker proteins (Additional Table 3 ). We used our multiplex SRM-MS assay to quantify the expression of 243 peptides corresponding to 81 proteins in 90 “Mock” Pap test samples (40 HGSOC; 20 benign; 30 healthy) (Additional Table  4 ). The expression levels of 110 peptides corresponding to 66 proteins were quantified in the 90 Pap test samples relative to the expression of the corresponding heavy isotope peptide. The SRM-MS data for the 90 Pap test samples was analyzed by limma to determine which peptides/proteins were differentially expressed in HGSOC versus healthy controls. Sixty-one of the 110 peptides were present at significantly (adjusted p  < 0.05) higher levels in the HGSOC samples compared to the healthy controls (Fig.  2 A; Table  2 ). The levels of 8 peptides (corresponding to proteins A1AG2, CRP, and HPT) were 8–12 times higher in the HGSOC samples compared to the healthy controls. Fig. 2 Quantification of candidate protein biomarkers by SRM-MS . Results of the SRM-MS targeted assay for 110 peptides representing 66 proteins are shown as volcano plots of the adjusted p-values vs. the fold change. Points above the dashed line correspond to adjusted p-values < 0.05 (red circles). ( A ) Sixty-one peptides were significantly upregulated in HGSOC vs. healthy control samples. ( B ) 36 proteins were significantly upregulated in HGSOC vs. healthy control samples Quantification of candidate protein biomarkers by SRM-MS . Results of the SRM-MS targeted assay for 110 peptides representing 66 proteins are shown as volcano plots of the adjusted p-values vs. the fold change. Points above the dashed line correspond to adjusted p-values < 0.05 (red circles). ( A ) Sixty-one peptides were significantly upregulated in HGSOC vs. healthy control samples. ( B ) 36 proteins were significantly upregulated in HGSOC vs. healthy control samples At the protein level, 36 of the 66 proteins were significantly (adjusted p  < 0.05) higher in HGSOC samples compared to the healthy controls (Fig.  2 B; Table  3 ). The levels of the proteins A1AG2, CRP, and HPT were 6–9 times higher in the HGSOC samples than in the healthy controls, confirming similar behavior at the peptide and protein levels. No peptides or proteins were present at significantly (adjusted p  < 0.05) lower levels in the HGSOC samples compared to the healthy controls (Fig.  2 A and B). Proteins and peptides identified as upregulated in the SRM analysis showed consistent upregulation in the TMT dataset, supporting the robustness of these findings across platforms. Table 3 Thirty-six proteins with significantly higher levels of expression between HGSOC versus healthy controls Protein name Fold change p -value Adjusted p -value A1AG2 6.261 < 0.001 < 0.001 A1BG 3.891 0.001 0.004 A2GL 4.503 < 0.001 < 0.001 APOE 2.886 0.004 0.009 APT 2.304 0.002 0.004 C4BPA 3.899 < 0.001 < 0.001 CD47 2.976 0.001 0.002 CERU 4.418 < 0.001 < 0.001 CFAH 3.684 < 0.001 0.001 CO3 4.883 < 0.001 0.001 CO4B 4.829 < 0.001 0.001 CRP 7.876 < 0.001 < 0.001 DAG1 4.104 0.001 0.002 DDAH2 2.044 0.002 0.005 DDB1 1.774 0.008 0.018 EIF1A 1.429 0.027 0.049 FBLN1 3.994 < 0.001 0.001 HPT 9.285 < 0.001 < 0.001 IBP3 3.012 0.007 0.018 ICAL 3.354 < 0.001 < 0.001 IL18 4.010 < 0.001 0.001 KNG1 4.820 < 0.001 < 0.001 LEG1 2.114 0.014 0.027 LEG3 2.963 < 0.001 0.001 LG3BP 3.394 0.001 0.004 LUM 3.357 0.001 0.003 MSLN 2.134 0.013 0.027 MUC16 2.610 0.010 0.020 NCHL1 1.846 0.009 0.020 PIGR 3.696 0.020 0.038 PYGB 1.874 0.002 0.005 S10A4 2.185 0.009 0.019 SYPL1 3.200 0.001 0.004 TCPE 2.141 < 0.001 0.001 TSN1* 4.017 < 0.001 0.001 X1433B 5.539 < 0.001 0.001 Proteins with adjusted p-values <0.05.  Shown are fold changes, p-values and adjusted p-values. Proteins shown in bold are the six proteins in the multi-protein classifier *Also significantly different when comparing HGSOC versus benign cases Thirty-six proteins with significantly higher levels of expression between HGSOC versus healthy controls Proteins with adjusted p-values <0.05.  Shown are fold changes, p-values and adjusted p-values. Proteins shown in bold are the six proteins in the multi-protein classifier *Also significantly different when comparing HGSOC versus benign cases In contrast, when the peptide levels in the HGSOC samples were compared to the samples from benign cases, none of the 110 peptides in the SRM-MS assay were present at significantly higher levels (adjusted p  < 0.05) in the HGSOC cases. Although the levels of several of the peptides measured in the HGSOC Pap tests were up to 5 times higher than in the benign controls, the adjusted p-values were not significant, potentially due to a smaller number of benign samples. Only one of the 66 proteins (TSN1) was present at significantly (adjusted p  = 0.040) higher levels in the HGSOC samples compared to the benign controls (Additional Fig.  2 ). The levels of several other proteins present in the HGSOC Pap tests were over 3 times higher than in the benign controls (CFAH, IBP3, FBLN1) and one protein (ILF3) was 2.3 times lower in HGSOC compared to benign Pap tests although none were statistically significant; again this is potentially due to a smaller number of benign samples. The expression levels of each of the peptides and proteins were graphed as boxplots to visualize their expression in HGSOC, benign conditions, and healthy controls. An example of the expression levels of one of the proteins (FBLN1) and the two peptides that were used in the SRM-MS assay is shown in Fig.  3 . In this case, the level of expression of the peptides (Fig.  3 A and B) and protein (Fig.  3 C) showed an upward trend from healthy controls to benign ovarian conditions to HGSOC. Fig. 3 Quantification of the candidate ovarian cancer biomarker FBLN1 by SRM-MS in Pap tests . Normalized values from each sample in the SRM-MS data were plotted with the median and 25th and 75th percentiles for the healthy controls (green), the benign ovarian conditions (blue) and the HGSOC patients (red). The fold change (FC) for comparison between HGSOC and healthy or benign samples is shown below each category, stars correspond to statistical significance (***: adjusted p < 0.001, **: adjusted p < 0.01). Plots for the protein FBLN1: ( A ) peptide GYHLNEEGTR, ( B ) peptide TGYYFDGISR, and ( C ) protein FBLN1. The expression values for these peptides and protein FBLN1 showed a trend for low levels in healthy controls, with higher levels in benign, and significantly higher levels in HGSOC compared to healthy controls Quantification of the candidate ovarian cancer biomarker FBLN1 by SRM-MS in Pap tests . Normalized values from each sample in the SRM-MS data were plotted with the median and 25th and 75th percentiles for the healthy controls (green), the benign ovarian conditions (blue) and the HGSOC patients (red). The fold change (FC) for comparison between HGSOC and healthy or benign samples is shown below each category, stars correspond to statistical significance (***: adjusted p < 0.001, **: adjusted p < 0.01). Plots for the protein FBLN1: ( A ) peptide GYHLNEEGTR, ( B ) peptide TGYYFDGISR, and ( C ) protein FBLN1. The expression values for these peptides and protein FBLN1 showed a trend for low levels in healthy controls, with higher levels in benign, and significantly higher levels in HGSOC compared to healthy controls However, the boxplots for the majority of the 61 peptides and 36 proteins showed a trend whereby the level of expression in the HGSOC and benign samples were elevated compared to the healthy cases, as shown for the protein A1AG2 and its peptides (Fig.  4 ). Fig. 4 Quantification of A1AG2 , a candidate ovarian cancer biomarker , levels by SRM-MS in “Mock” Pap tests . Normalized values from each sample in the SRM-MS data were plotted with the median and 25th and 75th percentiles for the healthy controls (green), the benign ovarian conditions (blue) and the HGSOC patients (red). The fold change (FC) for comparison between HGSOC and healthy or benign samples is shown below each category, stars correspond to statistical significance (***: adjusted p < 0.001, **: adjusted p < 0.01). Plots for: ( A ) peptide EHVAHLLFLR, ( B ) peptide EQLGEFYEALDCLCIPR, ( C ) peptide NWGLSFYADKPETTK ( D ) peptide SDVMYTDWK and ( E ) protein A1AG2. The expression values for these peptides and protein A1AG2 were significantly higher in HGSOC compared to the healthy controls Quantification of A1AG2 , a candidate ovarian cancer biomarker , levels by SRM-MS in “Mock” Pap tests . Normalized values from each sample in the SRM-MS data were plotted with the median and 25th and 75th percentiles for the healthy controls (green), the benign ovarian conditions (blue) and the HGSOC patients (red). The fold change (FC) for comparison between HGSOC and healthy or benign samples is shown below each category, stars correspond to statistical significance (***: adjusted p < 0.001, **: adjusted p < 0.01). Plots for: ( A ) peptide EHVAHLLFLR, ( B ) peptide EQLGEFYEALDCLCIPR, ( C ) peptide NWGLSFYADKPETTK ( D ) peptide SDVMYTDWK and ( E ) protein A1AG2. The expression values for these peptides and protein A1AG2 were significantly higher in HGSOC compared to the healthy controls The classification performance distinguishing between healthy controls and HGSOC for each protein individually was summarized using the AUC and sensitivity at 90% specificity with bootstrap confidence intervals. The sensitivity was calculated at 90% specificity vs. a higher specificity due to the limited number of healthy control samples. Sixteen proteins had an estimated AUC of > 0.7 (Table  4 ). We also considered whether the classification performance could be improved by integrating results from multiple proteins into a single classifier, which we refer to as a multi-protein classifier. Table 4 AUC values for the 16 proteins with AUC > 0.7 Protein name AUC (95% CI) AUC rank Sens. at 90% spec. (95% CI) Sens. at 90% spec. rank Proportion of selection (%) A1AG2_HUMAN 0.836 (0.737, 0.920) 1 0.600 (0.225, 0.800) 2 77 CRP_HUMAN 0.821 (0.718, 0.912) 2 0.525 (0.275, 0.800) 4 69 HPT_HUMAN 0.800 (0.690, 0.894) 3 0.500 (0.275, 0.800) 5 84 X1433B_HUMAN 0.776 (0.660, 0.879) 4 0.625 (0.200, 0.800) 1 22 TCPE_HUMAN 0.761 (0.635, 0.875) 5 0.275 (0.025, 0.800) 20 27 LEG3_HUMAN 0.760 (0.637, 0.873) 6 0.300 (0.100, 0.800) 15 6 A2GL_HUMAN 0.753 (0.632, 0.863) 7 0.375 (0.075, 0.650) 11 27 SYPL1_HUMAN 0.745 (0.616, 0.865) 8 0.250 (0.000, 0.700) 22 9 ICAL_HUMAN 0.740 (0.617, 0.848) 9 0.550 (0.100, 0.700) 3 59 APT_HUMAN 0.736 (0.610, 0.850) 10 0.450 (0.075, 0.700) 6 2 IL18_HUMAN 0.733 (0.609, 0.845) 11 0.400 (0.025, 0.675) 9 17 KNG1_HUMAN 0.731 (0.612, 0.840) 12 0.400 (0.225, 0.600) 9 0 CERU_HUMAN 0.728 (0.604, 0.838) 13 0.450 (0.200, 0.600) 6 1 PYGB_HUMAN 0.727 (0.605, 0.840) 14 0.425 (0.169, 0.675) 8 55 C4BPA_HUMAN 0.716 (0.591, 0.828) 15 0.325 (0.200, 0.625) 13 77 DDB1_HUMAN 0.713 (0.581, 0.836) 16 0.200 (0.000, 0.525) 31 5 The six proteins in the multi-protein classifier from Fig.  5 A are in bold AUC values for the 16 proteins with AUC > 0.7 The six proteins in the multi-protein classifier from Fig.  5 A are in bold There were six proteins included in the multi-protein classifier (A1AG2, CRP, HPT, ICAL, PYGB, and C4BPA) with all proteins having positive weights, indicating that higher levels are associated with a higher likelihood of the sample being ovarian cancer (Fig.  5 A). The predicted HGSOC risk score is equal to: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&\:expit(2.644\:+\:0.301\:\times\:\:ICAL\:+\:0.287\:\\&\times\:\:A1AG2\:+\:0.265\:\times\:\:C4BPA\:\\&+\:0.243\:\times\:\:PYGB+\:0.185\:\times\:\:HPT\\&+\:0.113\:\times\:\:CRP\:)\:\end{aligned}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:expit\left(x\right)\:=\:{e}^{x}/(1\:+\:{e}^{x}\:)$$\end{document} . While this risk score would typically equal the estimated probability of HGSOC, the intercept estimate is biased given the case-control study design and, thus, we call it more generally a risk score. The AUC for the multi-protein classifier was 0.880 (95% CI: 0.738–0.989) and the sensitivity at 90% specificity was 0.800 (95% CI: 0.470-1.00), higher than any of the individual proteins (Table  4 ; six proteins in bold). Fig. 5 Multi-protein and multi-peptide classifiers for HGSOC vs. healthy Pap tests . ( A ) The ROC curves for the HGSOC-healthy multi-protein classifier and each of the six individual proteins included in the multi-protein classifier (A1AG2, CRP, HPT, ICAL, PYGB, and C4BPA) as described in Table  4 . ( B ) Predicted probability of HGSOC from the HGSOC-healthy multi-protein classifier across true sample conditions. Boxplots summarize cross-validated predictions for HGSOC and healthy samples and projected predictions for independent benign samples. The dashed line indicates the 0.5 classification threshold. ( C ) The ROC curves for the multi-peptide classifier and each of the four individual peptides included in the multi-peptide classifier as described in Additional Table  5 . Peptide NWGLSFYADKPETTK (protein A1AG2), ESDTSYVSLK (protein CRP), VVLHPNYSQVDIGLIK (protein HPT), and IADGYEQAAR (protein TCPE) Multi-protein and multi-peptide classifiers for HGSOC vs. healthy Pap tests . ( A ) The ROC curves for the HGSOC-healthy multi-protein classifier and each of the six individual proteins included in the multi-protein classifier (A1AG2, CRP, HPT, ICAL, PYGB, and C4BPA) as described in Table  4 . ( B ) Predicted probability of HGSOC from the HGSOC-healthy multi-protein classifier across true sample conditions. Boxplots summarize cross-validated predictions for HGSOC and healthy samples and projected predictions for independent benign samples. The dashed line indicates the 0.5 classification threshold. ( C ) The ROC curves for the multi-peptide classifier and each of the four individual peptides included in the multi-peptide classifier as described in Additional Table  5 . Peptide NWGLSFYADKPETTK (protein A1AG2), ESDTSYVSLK (protein CRP), VVLHPNYSQVDIGLIK (protein HPT), and IADGYEQAAR (protein TCPE) We evaluated the robustness of protein selection using a bootstrap-based stability analysis, repeating the L1-penalized logistic regression 100 times on resampled datasets. Among the top-performing individual proteins, only the six proteins included in the multi-protein classifier were selected in more than 50% of bootstrap iterations, indicating that the multi-protein classifier preferentially incorporated proteins with greater selection stability. The AUC values for the 15 peptides that were used in 6 proteins that comprise the multi-protein classifier are shown in Table  5 . Table 5 AUC values for the 15 peptides that comprised the six proteins in the multi-protein classifier Protein name Protein AUC rank Protein proportion of selection (%) Peptide sequence Peptide AUC (95% CI) Peptide AUC rank Peptide proportion of selection (%) A1AG2_HUMAN 1 77 NWGLSFYADKPETTK 0.865 (0.774, 0.939) 1 77 (AUC = 0.836) EQLGEFYEALDCLCIPR 0.856 (0.761, 0.937) 2 38 SDVMYTDWK 0.815 (0.708, 0.906) 4 4 EHVAHLLFLR 0.744 (0.618, 0.857) 17 0 CRP_HUMAN 2 69 ESDTSYVSLK 0.829 (0.726, 0.918) 3 66 (AUC = 0.821) GYSIFSYATK 0.761 (0.645, 0.869) 13 9 HPT_HUMAN 3 84 VVLHPNYSQVDIGLIK 0.808 (0.698, 0.899) 5 68 (AUC = 0.800) VTSIQDWVQK 0.802 (0.692, 0.897) 6 15 DYAEVGR 0.788 (0.674, 0.887) 7 2 VGYVSGWGR 0.774 (0.658, 0.876) 9 4 ICAL_HUMAN 9 59 AAAPAPVSEAVCR 0.704 (0.577, 0.822) 29 32 (AUC = 0.740) QAEPELDLR 0.683 (0.554, 0.803) 43 18 PYGB_HUMAN 14 55 LVTSIGDVVNHDPVVGDR 0.738 (0.620, 0.848) 20 34 (AUC = 0.727) DFYELEPEK 0.626 (0.494, 0.749) 61 31 C4BPA_HUMAN 15 77 EDVYVVGTVLR 0.716 (0.591, 0.828) 27 54 (AUC = 0.716) Boxplots showing the level of expression of A1AG2 and its peptides are shown in Fig.  4 AUC values for the 15 peptides that comprised the six proteins in the multi-protein classifier Boxplots showing the level of expression of A1AG2 and its peptides are shown in Fig.  4 Predicted probabilities from the HGSOC-healthy multi-protein classifier demonstrated clear separation between HGSOC and healthy samples (Fig.  5 B). HGSOC samples exhibited consistently high predicted probabilities, with most values well above the 0.5 decision threshold, whereas healthy samples showed substantially lower probabilities concentrated below this threshold. When the classifier was applied to benign samples, 19 of 20 benign samples were classified as HGSOC based on the 0.5 probability threshold, while only one benign sample was classified as healthy (Fig.  5 B). Compared with healthy controls, benign samples displayed higher median predicted probabilities and greater overlap with HGSOC, indicating cancer-like prediction profiles under the HGSOC-healthy classifier. In bottom-up proteomics approaches, peptide sequences resulting from trypsin digestion are detected and quantified by the mass spectrometer. These abundances are then used to infer the abundance of the protein using the canonical sequences. Therefore, we hypothesized that individual peptide sequences could provide additional insight. Consequently, we also performed peptide-level classification analysis and determined the AUC values for each of the 110 peptides quantified by SRM based on the 40 HGSOC cases and the 30 healthy cases. In total, 29 peptides (corresponding to 18 proteins) had an estimated AUC of > 0.7; the 15 peptides with the highest AUCs are shown in Additional Table  5 . There were four peptides included in a multi- peptide classifier, with all peptides having positive weights, indicating that higher levels are associated with a higher likelihood that the sample is ovarian cancer (Fig.  5 C). In a similar manner, the predicted HGSOC risk score is equal to: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}\:&expit(1.640+0.378\times NWGLSFYADKPETTK[A1AG2]+0.070\\&\times ESDTSYVSLK[CRP]+0.070\\&\times IADGYEQAAR[TCPE] + 0.024\\& \times VVLHPNYSQVDIGLIK[HPT\left]\right)\end{aligned}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:expit\left(x\right)\:=\:{e}^{x}/(1\:+\:{e}^{x}\:)$$\end{document} . The AUC for the multi-peptide classifier was 0.825 (95% CI: 0.724–0.913), and the sensitivity at 90% specificity was 0.625 (95% CI: 0.400–0.850). We evaluated the robustness of peptide selection using a bootstrap-based stability analysis, repeating the L1-penalized logistic regression 100 times on resampled datasets. Among the top-performing individual peptides, three of the four peptides included in the multi-peptide classifier were selected in more than 50% of bootstrap iterations. These four peptides were derived from the proteins A1AG2, CRP, HPT, and TCPE; only TCPE was not one of the proteins in the multi-protein classifier above. All four peptides were confidently identified in the SRM assay based on co-elution of endogenous and heavy isotope-labeled internal standards and consistent transition ion-ratio patterns, as shown in Additional Fig.  3 .

Earlier detection is critical to improved prognosis for women with ovarian cancer, which is the 5th leading cause of cancer deaths among women in the U.S [ 1 ]. Given the lack of specific symptoms, most patients are diagnosed at later stages of disease, when the long-term survival rate is < 20% [ 2 ]. Serum CA-125 levels and transvaginal ultrasound, even when used in combination, are neither sensitive nor specific enough for general population screening given disease prevalence [ 3 , 4 ]; highlighting the need for novel “early detection” strategies. In this study, we set out to develop a paradigm shift that will change the way that women are screened for ovarian cancer by leveraging the standard Papanicolaou (Pap) test. Screening for cervical cancer by Pap tests has been routinely performed for over 50 years [ 5 ] and > 30 million are performed annually in the U.S [ 6 ]. The liquid-based Pap test consists of collecting cervical cells from the external cervical opening and placing them into a vial containing an alcohol-based fixative [ 7 , 8 ]. After processing the cells from the vials and staining them, they are examined by a pathologist to identify the presence of premalignant and malignant cells. The fact that ovarian cancer cells have been identified in liquid-based Pap tests [ 9 – 13 ] suggests that ovarian cancer proteins and peptides can also traverse the fallopian tube to the cervical opening. The liquid fixative solution in which the cells are collected for Pap tests is typically discarded after examination of the cells. To our knowledge, this study is the first to analyze cervical effluent in the Pap test fixative by mass spectrometry (MS)-based proteomic techniques to search for ovarian cancer peptide/protein biomarkers. Cervical effluent present in the residual Pap test fixative provides several important advantages for biomarker development. First, proteins present in Pap test fixative are collected from a site anatomically near the ovarian cancer (i.e. proteins may be secreted or shed from the tumor and flow through the fallopian tube into the uterus and out the cervical opening). Tumor-specific biomarkers have been shown to be present at higher levels in biospecimens proximate to the tumor, such as tumor interstitial fluid, compared to blood [ 14 ]. Our group and others showed that ascites fluid contains higher levels of ovarian cancer biomarkers than serum from the same patient by immunoassay [ 15 , 16 ] and quantitative proteomic analysis [ 17 , 18 ]. In addition, emerging paradigms have focused on the fallopian tube epithelium as the initial site of at least some high grade serous ovarian cancer (HGSOC) [ 19 ], further emphasizing the potential of cervical effluent as a source of ovarian cancer biomarkers. These studies, therefore, would suggest that cervical effluent samples obtained during a routine Pap test may contain higher levels of ovarian cancer proteins than blood. Second, comprehensive quantitative proteomic analyses of HGSOC tumor tissue by investigators for the Clinical Proteomic Tumor Analysis Consortium (CPTAC) has revealed a proteomic signature associated with chromosomal instability and homologous recombination defects [ 20 ]. These genomic alterations are hallmarks of HGSOC and could provide insight into relevant biomarkers discovered in Pap test samples. More recently, CPTAC reported a comparison of proteins from ovarian cancer tumors to proteins expressed in the fallopian tube [ 21 ]. These results strengthen the idea that cervical samples are a rich source of cancer biomarkers. These data identified 509 proteins that were differentially expressed between tumor tissue and normal fallopian tube, which could be used to discern biomarkers with the potential for early detection in Pap test samples. In an earlier feasibility study, we optimized sample preparation and handling procedures for SurePath™ Pap test fixative from 120 post-menopausal women with normal cervical cytology [ 22 ]. Briefly, proteomic techniques were used to separate the proteolytic peptides in the samples by reversed-phase liquid chromatography (RPLC), then detected by MS. More than 700 proteins were identified from a pool of 40 healthy individuals; an average of 427 proteins were identified in Pap tests from each of 5 healthy individuals. We coined the term, “Normal Pap Test Core Proteome” to represent the 153 proteins that were present in at least 4 out of 5 individuals, and quantified the level of protein by calculating a normalized spectral abundance factor for each protein. When we compared our “Normal Pap Test Core Proteome” to proteins identified by others who had directly sampled cervical vaginal fluid [ 23 – 26 ], we found many similarities [ 27 ]. We concluded that the Pap test fixative may serve as a proxy for samples derived from the genital tract. Other studies have examined the Pap test for somatic mutations in DNA for the identification of gynecological cancers [ 28 ]. Similar to our study, Kinde et al. used the Pap test sample; however, they examined DNA isolated from the cellular component of the Pap test for mutations that were known to be present in tumors from the same patient. Subsequently, this group used DNA isolated from Pap brush samples to detect mutations in 18 genes commonly mutated in ovarian cancer, combined with a test for aneuploidy [ 29 ], however the sensitivity of ovarian cancer detection was low (33%). More recently, two groups in Europe used droplet digital PCR technology to detect TP53 variants identified in tumors in Pap test samples from ovarian cancer patients [ 30 , 31 ]. Interestingly, both groups were able to detect the TP53 variants in archival Pap test samples ranging from 20 months [ 30 ] to 6 years [ 31 ] prior to diagnosis. However, another recent analysis of TP53 mutations in Pap test samples from women with and without HGSOC demonstrated that using a highly sensitive method of sequencing, they were able to detect “abundant, low frequency” TP53 mutations in women without ovarian cancer [ 32 ]. HGSOC are the most common subtype of ovarian cancer, and they are typically diagnosed at a late stage when treatments are less effective [ 5 ]. Thus, HGSOC was the focus of this study for the discovery and validation phases of cancer biomarkers. In this study, we have used Tandem Mass Tags™ (TMT) isobaric labeling and 2-dimensional liquid chromatography coupled to tandem mass spectrometry (2D LC-MS/MS) to compare the protein profiles of Pap tests collected from women prior to gynecologic oncology surgery versus Pap tests collected from healthy women. Selected proteins that we found to be differentially expressed between these groups were then examined by selected reaction monitoring (SRM) targeted MS on a larger cohort of Pap tests to develop a multi-protein classifier. Our rationale for generating a multi-protein classifier that could distinguish HGSOC from healthy controls is based on our vision of developing an at-home self-sampling test to be used for population-based screening; women with positive results would then require additional tests in the clinic. By incorporating the multi-protein classifier into a routine Pap test, it may be possible for women to be screened simultaneously for cervical and ovarian cancer.

We used SRM-MS, a targeted quantification approach, to quantify the expression of specific proteins in the cervical effluent from Pap tests. By comparing the AUC values of the proteins for their ability to distinguish HGSOC from healthy controls, we developed a multi-protein classifier comprised of six proteins (A1AG2, CRP, HPT, ICAL, PYGB, and C4BPA) with an AUC of 0.880; higher than any of the individual proteins. These data demonstrate proof-of-principle that the multi-protein classifier can differentiate HGSOC from healthy controls. Pap tests are a routine, readily accepted method for cervical cancer screening. Thus, by combining this testing into routine Pap tests, it could prove to be a cost-effective strategy that would be agreeable to patients and clinicians, and thereby may increase uptake. Three of the six proteins that comprised the multi-protein classifier [A1AG2 (orosomucoid2), CRP (C-reactive protein), and HPT (haptoglobin)] had the highest individual AUC values and the highest fold change between HGSOC and healthy controls. Although their levels of expression were significantly different between HGSOC and healthy control, they were not significantly different between HGSOC and benign ovarian conditions. All three proteins are acute phase reactants that are components of blood and considered biomarkers of inflammation. The elevated levels of these proteins in Pap test samples likely reflects the presence of serum-derived proteins in cervical mucus, or possibly they were derived from the ascites fluid, which has a similar composition to serum [ 17 ]. We previously identified HPT and A1AG2 as serum biomarkers for ovarian cancer by two other MS-based proteomic techniques, iTRAQ and DIGE [ 55 , 56 ]. Similarly, serum CRP levels are also known to be elevated in ovarian cancer; high levels of CRP are associated with tumor size and stage [ 57 ] and a poor prognosis [reviewed [ 58 ] ]. In addition, in pre-diagnostic serum, women with elevated CRP levels are at risk for developing breast and ovarian cancer [ 59 – 61 ]. Furthermore, CRP is an approved biomarker for inflammation and cardiovascular disease [ 62 ]. Although CRP and other biomarkers might not be ovarian cancer-specific, they may still have a clinical application whereby women would be triaged for further gynecologic work-ups, including potentially CA125 blood tests and transvaginal ultrasounds. The fourth protein in the multi-protein classifier, C4BPA (Complement component 4 binding protein alpha), is also secreted into the blood. It may play a role in the immune response by negative regulation of complement C4 [ 63 ]. In an analysis of data from The Cancer Genome Atlas (TCGA), C4BPA expression in breast cancer was positively correlated with expression of several immune cell marker genes (B-cells, neutrophils, CD8 T-cells) and was a favorable prognostic marker [ 64 ]. Similarly, using RNA-seq data from the ovarian cancer TCGA, C4BPA was identified as part of a 7 gene signature to predict prognosis [ 65 ]; high C4BPA expression was associated with a favorable prognosis. The over-expression of C4BPA protein in ovarian cancer tissues compared to normal ovary was also validated by immunohistochemistry (IHC) [ 65 ]. In pancreatic adenocarcinoma (PDA), C4BPA was identified as a serum biomarker using TMT-labeled serum from matched pre- and post- surgery specimens. Subsequent biomarker validation of C4BPA by ELISA showed a significant increase in C4BPA levels in PDA compared to both pancreatitis and healthy controls [ 66 ]. In another study, fully sialylated C4BPA was identified as a serum biomarker in ovarian cancer using glyco-peptide screening [ 67 ]. In both studies [ 66 , 67 ], C4BPA was elevated in early-stage disease, however, the ovarian cancer study focused on ovarian clear cell carcinoma, although they found elevated sialylated C4BPA in serum from HGSOC patients as well [ 67 ]. The fifth protein in the multi-protein classifier, PYGB (Glycogen phosphorylase, brain isoform) plays an important role in glycogen metabolism and ATP production. It is over-expressed in the tissues of multiple cancer types (gastric, hepatocellular carcinoma, non-small cell lung, and prostate) including ovarian cancer, and is associated with advanced stage disease and poor prognosis [ 68 ] (reviewed by [ 69 ]). In colorectal cancer, PYGB expression correlated with increased dysplasia, and was found prior to p53 overexpression, suggesting it may be an early marker of carcinogenesis [ 70 ]. The final protein in the multi-protein classifier, ICAL [Calpastatin, Calpain-10 (also known as Sperm BS-17 component)], is a specific inhibitor of the proteinase calpain. The calpain-calpastatin system plays a role in cancer progression by regulation of cell adhesion, migration, proliferation and apoptosis [reviewed by [ 71 ] ]. According to the Human Protein Atlas, calpastatin is a “Cancer Related Protein”, although it has not been recognized as a “Candidate Cancer Biomarker” [ 72 ]. ICAL is highly expressed in over half of ovarian cancer tumors, and high expression was associated with the HGSOC subtype and better overall survival [ 73 ]. A recent proteomic analysis of saliva from patients with newly diagnosed or recurrent glioblastoma found ten proteins, including ICAL, that were expressed only in saliva prior to surgery and not in post-treatment saliva samples [ 74 ] suggesting that salivary calpastatin levels could be a biomarker for disease recurrence. Moreover, in glioblastoma, high levels of ICAL is associated with a poor prognosis [ 72 , 75 ]. In our SRM analysis, we found many more proteins with differential expression when comparing the HGSOC samples to healthy controls than when compared to the benign samples. This cannot be attributed to the way that the candidate biomarkers were selected, since we selected proteins from the TMT data that were differentially expressed between HGSOC and healthy and benign controls. Many serum biomarkers for ovarian cancer may also be elevated in subjects with benign ovarian conditions [ 76 ]. In our SRM analysis only one protein, TSN1 (tetraspanin 1) was significantly elevated in the HGSOC samples compared to benign. Tetraspanins are components of the membrane that form microdomains involved in receptor clustering and signal transduction and participate in cell adhesion, migration, proliferation and extracellular vesicle formation [reviewed by [ 77 ]]. TSN1 is expressed in ovarian cancer tissues [ 78 ] and overexpression of TSN1 in benign endometriosis promoted cell proliferation and invasion, potentially leading to ovarian clear cell carcinoma [ 79 ]. In a study to identify plasma protein biomarkers for ovarian cancer detection, TSN1 was part of a 7-protein signature that could distinguish early-stage ovarian cancer from benign ovarian disease [ 80 ], supporting our finding that TSN1 is a biomarker of ovarian cancer. Many of the proteins that were identified as candidate protein biomarkers for ovarian cancer in the discovery phase of this study (i.e. the TMT experiments), could be classified as highly abundant serum proteins. However, those proteins still exhibited differential levels of expression in the SRM-MS assay when comparing HGSOC to healthy samples. Some of these proteins have been suggested to serve as possible ovarian cancer biomarkers in the literature, e.g. HPT and LRG [ 55 , 56 , 81 ]. Of interest, many of the “classical” ovarian cancer biomarkers [e.g. CA125 (MUC16) and HE4] that had been discovered in tissue or serum were also included in our SRM-MS assay as heavy-labeled peptides from earlier CPTAC studies [ 21 , 49 ]. However, these proteins were not sufficiently robust to be included in our multi-protein classifier. Other groups have developed a technique to rinse the inside of the uterus, then analyze the DNA of the cells from the washings for mutations [ 82 ]. Recently, a comprehensive proteomic analysis of microvesicles isolated from uterine lavage samples was used to create and validate a multi-protein classifier for ovarian cancer detection with 70% sensitivity and 76.2% specificity [ 83 ]. This methodology may prove to be too cumbersome to perform on an annual basis for ovarian cancer screening. Given the relative ease of collecting a Pap test or vaginal swab in comparison to a uterine lavage, the Pap test may be accepted more readily in clinics to screen for both ovarian and cervical cancers. Because of the inherent variability in the collection of Pap test samples, the amount of cervical effluent present in the SurePath™ vials was determined by measuring the protein concentration following acetone precipitation. On average, each Pap test SurePath™ fixative sample yielded ~ 140 µg of protein, excluding 5 Pap tests which contained > 500 µg of protein, most likely due to blood contamination which was visible in the fixative. To control for variability in sample collection, we used the same amount of protein (50 µg) from each sample when performing the TMT experiments, and the same amount of protein (5 µg) from each sample when performing the SRM-MS assay. In future applications, it may be necessary to include a panel of “housekeeping” proteins that will control for variability in sample collection. A limitation of this preliminary study was the inability of our multi-protein classifier to distinguish HGSOC from benign ovarian conditions. The two major drivers of this limitation are the size and selection of the clinical cohort. First, our selection of “Mock” Pap tests from 20 women with benign ovarian serous pathology were very similar to cases of HGSOC, in that they had all progressed to a palpable abdominal mass that required surgery. Second, practical constraints limited the total number of patients included in our TMT discovery study to 40. The limited number of samples that were used in this pilot study, and the limited clinical information available about the patients, precluded us from studying the effect of possible confounders that may have affected the outcome of our results, e.g. parity, smoking, BRCA, BMI, and menopausal status. Future studies could include these factors and others that were not captured in this study. In future studies, we intend to use a larger cohort of women for the discovery of additional candidate protein biomarkers and include more of these candidate proteins into our SRM-MS assay. This pilot study was limited to only 30 proteins developed for the targeted SRM-MS assay from among the 700 potential protein candidates. We will also use a larger cohort of women for the targeted SRM-MS assay and the development of a multi-protein classifier, including samples from women with early stage HGSOC. This preliminary study was limited to the use of Pap tests primarily from women with late stage HGSOC; women with other gynecologic malignancies, including endometrial cancer and cervical cancer, were not studied. This, too, will be the focus of future studies to independently confirm the cancer of origin. The U.S. Preventive Services Task Force (USPSTF) cervical cancer screening guidelines recommend that clinicians stop testing for cervical cancer by use of the Pap test when women reach 65 years of age [ 84 ]. In addition, the guidelines recommend the frequency of Pap testing decrease from annually to every 3 to 5 years, depending on the patient’s age and other circumstances [ 84 ]. Although these current recommendations may initially be seen as a limitation to the utility of the Pap test for ovarian cancer, we envision that once further studies are completed, the Pap test recommendations may be updated to increase the age at which women are tested, since ovarian cancer predominantly is detected in post-menopausal women, with a median age of 63 years [ 85 ]. We also envision that this testing will be converted to a self-sampling at-home test using a cervical swab, so that ovarian cancer detection can be monitored in women of all ages on a frequent basis. Since Pap tests are routinely performed in the clinic and relatively inexpensive, we envision that the multi-protein classifier could be used as a screening tool first for women at high risk of ovarian cancer (e.g. BRCA mutation carriers) and later in a normal risk population. To do so, it may be necessary to perform longitudinal sampling to determine whether levels of some of these proteins increase over time. Women with asymptomatic benign lesions who have a risk of ovarian cancer based on the multi-protein classifier would benefit from being sent to a gynecologic oncologist for a more extensive work-up, which may involve blood tests for CA125/HE4 levels and a transvaginal ultrasound. In the future, state-of-the-art technology for DNA mutations coupled with MS technology for proteins may be used in the clinic for cancer detection. While our results clearly demonstrate the future promise of this approach, finding a clinically acceptable screening strategy will require additional experimentation.

This study demonstrates that proteins are shed/secreted by ovarian cancer cells into the cervical opening and can be captured during a routine Pap test. These promising results provide proof-of-principle that it is possible to quantify ovarian cancer biomarkers from the cervical effluent in the cell-free Pap test fixative. These results provide a new paradigm in which the routine Pap test may have a new application: to detect both cervical cancer and ovarian cancer. Future studies will be needed for validation of sensitivity and specificity using significantly larger patient cohorts, ideally in a prospective cohort of women prior to diagnosis. The assay could initially be used to test women in high-risk groups, and eventually for screening the general population. Perhaps longitudinal tracking of changes in protein abundance will emerge as the best method to distinguish between benign and malignant lesions in an early detection context. Such studies have the potential to improve the detection of early stages of ovarian cancer, when the disease is treatable, thereby increasing survival rates of women. In summary, we have developed a novel method for population-level screening of normal risk women that piggybacks on Pap tests, an existing approved public health screening protocol. Our results have been validated in an independent and clinically relevant population, demonstrating robust discrimination between healthy patients and those with tumors or benign lesions. Although our multi-protein classifier cannot distinguish between tumor and benign, it may still prove useful as a referral tool for a more extensive gynecological exam, including a transvaginal ultrasound, which is currently the best tool we have for early detection of ovarian cancer.

Supplementary Material 1: Additional Figures 1–3. Supplementary Material 1: Additional Figures 1–3. Supplementary Material 2: Additional Tables 1–5. Supplementary Material 2: Additional Tables 1–5.