Discovery and preliminary validation of a new panel of personalized ovarian cancer biomarkers for individualized detection of recurrence

Sample collection

All HGSC serum samples were collected and stored by the UHN Gynecologic Blood Biobank (Toronto, ON, Canada) according to a standardized protocol, to avoid systemic biases. Approval was obtained from the Research Ethics Board of UHN, Toronto, Canada (REB #10-0591).

For the discovery phase, the retrospective cohort consisted of serum samples from 30 HGSC serially collected at four timepoints (120 samples total) as follows: Timepoint 1 - Before primary debulking surgery, Timepoint 2-0.75 months after surgery, Timepoint 3 - A median of 5 months after surgery (ranged from 5 – 8 months), and Timepoint 4 - A median of 11 months (ranged from 8 – 14 months) after surgery. Twenty-five of these patients were diagnosed with recurrence at timepoint 4, according to CA125 elevation, radiological findings, or clinical symptoms. Five patients did not experience recurrence during the study period and were used as non-recurrence controls. All patients received primary debulking surgery followed by six cycles of platinum/taxane-based chemotherapy. We also included serum samples from female, age-matched, non-cancer controls (n=9), taken at three timepoints (27 samples total). This group consists of dementia patients in a 24-month follow-up study (three timepoints at month 0, month 12, and month 24), as described in a recent publication by our group (306). Although these patients were identified from an unrelated study, they were used as biological and technical references in this study for gauging how much the protein levels change between longitudinal measurements due to day-to-day biological variations and technical factors, in a non-cancer population. PEA analysis was performed for the same 1,104 analytes, in the same run, and by the same analytical provider (Olink Proteomics, Boston, MA, USA) as the HGSC study.

For the validation phase, we included retrospective, serum samples collected from 39 HGSC patients, with 4-8 timepoints (median of 6) per patient depending on availability (ranging from three months to seven years of follow-up depending on the time of relapse). Approximately 72% (28/39) of the HGSC patients received primary debulking surgery followed by six cycles of platinum/taxane-based chemotherapy, while 28% (11/39) received neoadjuvant chemotherapy with interval debulking surgery with subsequent platinum/taxane-based chemotherapy (6 cycles total). Longitudinal timepoints were chosen from those collected at pre-primary debulking surgery or pre-neoadjuvant chemotherapy, 0.75 months post-surgery or post neoadjuvant chemotherapy, and after 3rd and 6th cycle of 1st-line chemotherapy or after interval debulking and subsequent chemotherapy. Several collections every three months during remission, and collections at three months before and at imaging confirmed relapse were also included. Patients are categorized as follows: 1) Long-Term Survival/“curative” control (>7 years free of disease from last dose of chemotherapy) (n=5), 2) Platinum sensitive (relapse ~1-3 years from last dose of chemotherapy; most common HGSC case) (n=25), and 3) Platinum resistant (relapse within six months of last dose of chemotherapy) (n=9). We included patients with various timelines of recurrence to ensure our candidates correlate to relapse in a wide range of patients encountered clinically.

Discovery of candidate personalized biomarkers of recurrence

For the discovery phase, PEA analyses were conducted using the twelve available 92-plex panels: cardiometabolic, cell regulation, cardiovascular III, cardiovascular II, development, immune response, inflammation, metabolism, neurology, neuro-exploratory, oncology II, and organ damage. A total of 1,104 analytes were measured. The assay is described in detail in our previous publication.¹⁶

Because we are identifying personalized biomarkers in a surveillance cohort, we are interested in proteins that show significant within-individual increase upon recurrence, in each recurrent patient. Therefore, we first calculated the % rate of change in protein levels between consecutive timepoints to longitudinally assess the protein kinetics within each patient. This is calculated by the % change between timepoints divided by the number of months that elapsed. Mainly, we considered the timepoint before recurrence to the timepoint of clinically confirmed recurrence (i.e., timepoint 3-4 in the 25 recurrent HGSC patients). We calculated the protein-specific reference change value (RCV) for all 1,104 proteins based on the intra-individual biological variation (day-to-day difference) observed in the longitudinal non-cancerous controls and the analytical variation (differences between technical replicates) that was previously calculated (our unpublished data). RCV was calculated with the formula: RCV=2^1/2 * Z * (CV_(A)² + CV_(I)²)^1/2, where Z is 1.96 for 95% probability, CV_(A) is the analytical variation, and CV_(I) is the biological variation. RCV reflects the natural intra-individual changes across serial measurements for a given analyte. A % change in serial biomarker measurement is deemed significant if it is higher than its RCV. Our defined term of “rate of RCV” (the minimal % change per month that denotes significance) was also calculated for the purpose of this study to be comparable to the % rate of change in the proteins. We defined the rate of RCV as the RCV divided by three months since surveillance biomarkers are clinically assessed every ~three months for the first year following first-line treatment. The rate of % change in biomarker value between timepoints 3-4 (upon relapse) in the recurrent patient must exceed the “rate of RCV” of the respective protein, to be considered statistically significant. We outline below our further defined selection criteria. The candidate personalized biomarkers we selected were proteins that sequentially passed the following filtering criteria that we pre-defined and were determined a priori (proteins remaining means proteins that meet the predefined criteria).

Validation of candidate personalized biomarkers of recurrence

PEA analyses were conducted using the seven 92-plex panels. A total of 644 proteins were measured. Unfortunately, we could only perform PEA analysis for seven 92-plex panels due to cost constraints. We selected the panels that covered most of our 23 candidates, which included 21 candidates. Unfortunately, two candidates, DNAJB1 and PSIP1 were not included in the seven panels and were not analyzed further in our validation study.

We were interested in assessing if the 21 out of the 23 candidate proteins that we identified in our discovery phase, in addition to CA125 and HE4, showed significant within-individual increase upon recurrence in the 35 recurrent HGSC patients in an independent validation study. As per the discovery study, we first calculated the % rate of change in the 21 proteins between sequential timepoints, to longitudinally assess the protein kinetics within each patient. Mainly, we considered the timepoint immediately before recurrence to the timepoint of clinically confirmed recurrence for each patient. We used the protein specific RCV and rate of RCV that was previously calculated in the discovery study. The % rate of change in the 21 proteins from timepoint before relapse to relapse in the recurrent patient must exceed the rate of RCV of the respective protein to be considered statistically significant.

Selection of refined panel of personalized markers

To select a refined panel of personalized markers, we performed the same analysis as described for the 21 candidate proteins, applied to all 644 proteins measured in the validation study. In addition, we outline below our further pre-defined selection criteria. The candidate personalized biomarkers we selected were proteins that sequentially passed the following filtering criteria:

Statistical analysis

Statistical analyses were performed using the GraphPad Prism software version 9.0.2 (GraphPad Software, San Diego, CA) and R statistical software version 4.0.4 (www.rproject.org). The Kruskal-Wallis test with Bonferroni’s multiple correction was employed to compare the % rate of change in protein level from timepoint 3 to 4 between three groups: 1) Elevated CA125 at recurrence, 2) Non-elevated CA125 at recurrence, and 3) Non-recurrence. Pearson correlation analysis was performed to examine the relationship between % rate of change in protein level from timepoint 3 to 4 and age at diagnosis. The Kruskal-Wallis test was used to assess the relationship between the % rate of change in protein level from timepoint 3 to 4 between clinical groups (BRCA1/2 mutation status and residual tumor volume). P-values of less than 0.05 were considered to be statistically significant.

Patient population and clinical implications

Demographics and clinicopathological characteristics of the 30 HGSC patients in the discovery study cohort are summarized in Table 1. For one patient (LR45) the CA125 value was not elevated (<35 U/mL) at diagnosis and thus CA125 monitoring would not be recommended for surveillance. In our study, 17/25 (68%) of recurrent patients showed increase in CA125 above the reference range of 35 U/mL, which strongly indicates biochemical relapse, and would prompt referral to imaging in standard clinical practice. It is important to note that a median of 6 months elapsed between timepoint 3-4 (ranges from 3-9 months). Therefore, it is unclear whether the CA125 value doubled in the other patients in the clinical follow-up prior to clinical relapse (3 months period). We calculated the % change in 3 months in clinical CA125 value based on the change from T3-4 and the number of elapsed months. Patients LR22 and LR45 did not show a % increase by ≥100% (CA125 did not double) in 3 months. Patients LR38, LR39, LR42, and LR43 did show % increase by >100% (at least doubling in CA125 value) in 3 months.

Demographics and clinicopathological characteristics of the 39 HGSC patients in the validation cohort are summarized in Table 2. CA125 was non-elevated at diagnosis in 13% (5/39) of the HGSC cases, representing patients who would not be recommended for CA125 monitoring by US and international guidelines and have no conventional biomarkers for surveillance.¹²

Discovery of candidate personalized biomarkers

We employed a patient-centric analysis to identify the top proteins that sensitively reflect changes in tumor burden upon relapse in individual patients. We identified a panel of the top 23 candidate personalized markers, in addition to CA125 and HE4. The names and biological functions of the 23 candidate proteins are listed in Table 3. The 23 candidate personalized markers each showed a sensitivity of 16-52% at 100% specificity for detecting relapse (Figure 1). The % change in 3 months (for clinical relevance, as biomarkers are usually assessed every ~3 months) for each personalized marker and patient are shown in Figure 1. The marker is considered to signify significant risk of relapse (shown in orange) for the patient, if it showed a % change greater than the RCV and the maximum % change seen in all non-recurrent patients was less than the RCV (100% specificity). In addition to this criterion, markers that showed a % change greater than two-times the RCV, are considered to denote a highly significant risk of relapse in that patient (shown in red).

Figure 1. Heatmap of % change^† for top 23 candidate personalized biomarkers, plus CA125 and HE4.

To complement CA125, the panel of 23 personalized markers, plus HE4 (24 proteins in total), was used to select a custom marker combination for predicting relapse in each patient (combinations comprised of unique 1-19 informative markers per patient). These patient-centric marker combinations predicted relapse in 92% of patients. Nine of the 25 recurrent patients had at least one personalized marker (ranging from 1-11 markers) that showed larger % increase upon relapse compared to CA125, as measured by PEA.

PSIP1 and STK4 showed weak but significant correlation between % change upon relapse and age at diagnosis (r = 0.57 and 0.49, respectively). None of the 23 candidate personalized markers showed significant association between % change upon relapse with BRCA1/2 mutation status and residual tumor volume after primary or interval debulking surgery.

Validation of candidate personalized biomarkers

We examined the sensitivity of the 21 candidate proteins in this independent HGSC cohort. In Figure 2 we show a heatmap of their % change in 3 months per patient.

Figure 2. Heatmap of % change^† for 21 candidate proteins identified from our discovery study, plus CA125 and HE4.

The 21 candidates we previously identified in the discovery study each showed a sensitivity of 3-35% (at 100% specificity) for detecting relapse. To complement CA125, our previously identified panel of the 21 candidate personalized markers, plus HE4, was used to select a custom marker combination for predicting relapse in each patient (combinations comprised of unique 1-13 informative markers per patient). These patient-centric marker combinations predicted relapse in 71% (24/34) of the recurrent patients. Twelve of the 34 recurrent patients had at least one personalized marker (ranging from 1-7 markers) that showed larger % increase upon relapse compared to CA125 as measured by PEA.

Refined panel of personalized markers

Using patient-centric selection criteria, we identified a panel of the top 33 candidate personalized markers, in addition to CA125 and HE4. This panel includes 18 candidates that were identified in our discovery cohort. The names and biological functions of the 15 additional candidate proteins are listed in Table 4. The 33 candidate personalized markers each showed a sensitivity of 6-41% at 100% specificity for detecting relapse (Figure 3). To complement CA125, the panel of 33 personalized markers, plus HE4 (34 proteins total), was used to select a custom marker combination for predicting relapse in each patient (combinations comprised of unique 1-22 informative markers per patient). These patient-centric marker combinations predicted relapse in 91% of patients. Eighteen of the 34 recurrent patients had at least one personalized marker (ranging from 1-11 markers) that showed larger % increase upon relapse compared to CA125 as measured by PEA.

Figure 3. Heatmap of % change^† for refined panel of 33 candidate personalized markers, plus CA125 and HE4.

Three of the 33 candidate personalized markers (TBCB, PMVK, FKBP5) showed significant association between the % change upon relapse and BRCA1/2 mutation status. KLK6 and PTX3 showed significant but weak correlation of % change upon relapse with age at diagnosis (r = 0.34 and 0.35, respectively). The % change upon relapse was not associated with residual tumor volume for any of the markers.

Putative role of candidate personalized biomarkers in OvCa

Some of our 23 candidate proteins are well-studied in OvCa. Cisplatin stimulation during first-line chemotherapy has been shown to trigger an increase in CCL20 production by ovarian tumor-associated macrophages.^31–35 CCL20 activates C-C chemokine receptor 6 (CCR6) on ovarian tumor cells and, in turn, promotes epithelial-to-mesenchymal transition and cancer cell migration.^31–35 High expression of CHI3L1 is associated with poor prognosis and chemoresistance in OvCa.^36–39 CHI3L1 induces AKT and ERK signaling pathways to promote OvCa stem-like cells.^36–39 TGM2 is a player in the chromatin-binding protein family, high-mobility-group-box (HMGB), interactome in OvCa. TGM2 is correlated to poor outcome and promotes epithelial-to-mesenchymal transition.^40–43 STK4 plays a key role in the kinase cascade elicited in the Hippo signaling pathway, which controls organ growth.^44–47 Deregulation of the Hippo signaling pathway is implicated in ovarian carcinogenesis.^44–47 SDC1 serves as a receptor for collagen alpha-1(I) chain (COL1A1) proteins secreted by fibroblasts. Ligand-receptor binding of SDC1 and COL1A1 activates AKT signaling pathways to promote OvCa cell motility.^48–50

Validation of candidate personalized biomarkers

Our goal was to evaluate the ability of the top 21 of the 23 personalized marker candidates to detect relapse in a larger, independent, longitudinal HGSC cohort. The 21 candidates demonstrated a significant increase in biomarker value upon relapse in at least 1 recurrent HGS patient in the validation cohort (Figure 2). However, the 21 candidates showed lower sensitivity (3-35%, at 100% specificity for recurrence) for detecting relapse in the validation cohort compared to the discovery findings (16–52%). Personalized marker combinations detected relapse in 71% of patients in the validation cohort, compared to 92% in the discovery study. The discrepancy between the two findings is likely due to the limited sample sizes. As personalized markers are defined by being expressed in a small subset (5-30%) of tumors and are not informative for most patients of a cancer type, a large sample size is needed to accurately encompass the heterogeneity that exists in ovarian cancer. Our discovery and validation studies provide the necessary data for calculating an accurate sample size that would provide statistical power in future large-scale prospective studies. Studies have shown differential molecular profiles for platinum resistant (relapse in <6 months) and platinum sensitive (relapsed in > 6 months) cases.^51,52 The validation data may better reflect the biomarkers’ performance for surveillance in clinical settings, where there is currently no clinically used predictor of platinum sensitivity/resistance.⁵³ Finally, the validation cohort may represent a different subset of tumor composition, response to chemotherapy, and molecular drivers for relapse.⁵⁴

Refined panel of personalized markers

Our patient-centric analysis selected 18 of the top 21 candidates that we identified in our discovery phase, which we included in our refined panel of the top 33 personalized tumor markers (Figure 3). The 15 new candidate personalized markers each showed a sensitivity of 15-41% at 100% specificity for tracking recurrence. Further validation studies are needed for these 15 candidates. The early detection of recurrence, even before recurrent tumors can be detected through imaging modalities, may provide benefit on survival outcome, as there are increasing selections of maintenance treatment and experimental targeted trials for patients to enroll in.^8,55–58

Limitations and possible mitigations

We recognize that our study has important limitations, as follows: