Biobehavioral factors predict an exosome biomarker of ovarian carcinoma disease progression

Participants

Women with suspected ovarian cancer were recruited from three University Medical Centers, two in the Midwest and one in Florida, as part of a larger study examining biobehavioral factors and tumor progression in ovarian cancer. Inclusion criteria included primary invasive epithelial ovarian, peritoneal or fallopian tube carcinomas, or carcinosarcoma. Histology was confirmed by pathology review. Patients were excluded who were under 18 years of age, had a history of previous cancer within the last 5 years, a comorbid condition with known immune effects, current pregnancy, regular use of systemic steroid medication in the last month, metastases to the ovaries from other organs, or inability to answer questions; these exclusion criteria were established for the larger study. Patients recruited before 2013 were excluded for a history of cancer; after that date, patients were only excluded for cancer in the last 5 years. Data from a subgroup of 90 women were used in discovery analyses to establish a molecular measure predicting progression-free survival. Of these 90 exosome samples, 78 yielded valid RNA sequencing data (12 had either insufficient RNA or RNA that was too severely degraded to yield useful results) and two had recurrence data that was nonevaluable, yielding a final sample of 76 for the first subgroup. Data from a second subgroup of 31 women were used to test whether this molecular signature showed longitudinal elevation as a function of biobehavioral risk factors. Several individual samples failed to yield valid RNA data, but RNA data were available on at least one of the time points for all 31 individuals. However, some individuals were missing data on other measures that were the targets of analysis (social support, depression) or covariates that reduced the analyzable number of cases to 24 in the case of social support and 22 in the case of depression. Therefore, the final sample was 24 for the second subgroup. For the first subgroup of patients, the earliest date of treatment initiation was in December 2004, and all progression information was censored on December 31, 2016 or last contact before that time, and primary treatment began at least 3 years before the date of censoring. Treatment for the second subgroup was initiated between April 15, 2016 and September 25, 2017. All procedures were approved by institutional review boards at the clinical sites, and all patients provided signed informed consent.

Procedures

Participants were recruited at an initial clinic visit and completed psychosocial assessment and demographic information at home within approximately 2 weeks before surgical resection or initiation of neoadjuvant therapy (NACT). Blood sampling was done the morning of surgery or NACT and during the patient's 6-month and 12-month clinic follow-ups. Follow-up psychosocial surveys were administered in conjunction with the 6- and 12-month visits. Social support was assessed using the attachment subscale of the Social Provisions Scale (SPS-Attachment).¹⁷ This subscale measures perceived emotional connection with others and has been the measure most closely associated with disease-related biomarkers in our previous research. Based on prior studies,^{3, 18-20} we defined poor social support as an SPS-Attachment score of less than 15 (the median value). Depressive symptoms were assessed by the Center for Epidemiological Studies Depression Scale (CES-D), a 20-item self-report scale measuring frequency of depressive symptoms over the last week.²¹ The CES-D is considered a valid and reliable assessment for depressive symptoms in cancer patients.²² Scores 16 or higher are considered to be consistent with clinical depression²³; therefore patients were categorized as having high/low depressive symptoms based on this cutoff.

Clinical disease characteristics (e.g., grade, stage, histology, body mass index [BMI], and date of progression) were extracted from medical records. Demographic and clinical characteristics (age, race, alcohol use, smoking, and BMI) were assessed as potential confounding factors. Alcoholic beverages were coded as ≤2 drinks/day (0) versus ≥3 drinks/day (1); smoking was coded as never (0) versus ever (1). Progression-free survival (PFS) was calculated between date of tumor resection or initiation of neoadjuvant chemotherapy and date of the first recurrence or progression, or last recorded contact before the date of censoring when the patient had not progressed, or date of death.²⁴ Progression was documented by computed tomography, initiation of new therapy, or clinical evidence of progression. The 6-month cutoff was defined as 190 days or less (see Figure S1 for a summary of study time points).

Exosome RNA profiling analyses

Detailed information about exosome data extraction and RNA profiling is provided in the Supporting Methods. For the discovery-based biomarker development analyses for study 1, normalized transcript abundance values were screened to identify all gene transcripts showing r > 0.40 association with disease progression class (1 = PFS ≤6 months; 0 otherwise) using SAS PROC CORR. The resulting 30 transcripts served as predictor features for the elastic net machine learning algorithm implemented in SAS PROC GLMSELECT, using progression class as the criterion with predictor selection by random 5-fold external cross-validation, 40 selection cycles, and L2 tuning by log grid search over the interval 0 to 1. The elastic net is a commonly used method for deriving predictors when the number of candidate variables greatly exceeds the number of cases.²⁵ The resulting composite biomarker (continuous value) was a linear combination of 22 elastic net selected genes; this was analyzed for point biserial correlation with disease progression class using SAS PROC CORR, and dichotomized at a cost-weighted value of 0.25 (i.e., assuming it was three times as important to identify a high-risk case as to misidentify a low-risk case) to quantify association with disease progression class using SAS PROC FREQ. SAS PROC GLM was used for linear model analyses relating biomarker values to disease progression class while controlling for covariates. PROC MIXED was used for mixed effect linear model analyses quantifying longitudinal change in biomarker values from presurgical baseline to 6- and 12-months postsurgery while controlling for nonindependence of repeated measures using a fully saturated (unstructured) variance-covariance matrix on residuals.

Study 2 analyses (biobehavioral prediction analyses) quantified associations between longitudinal variation in progression biomarker values and biobehavioral risk factors in a separate data set. Gene expression data were derived as described above for study 1, with the multi-gene progression biomarker scored using SAS PROC GLMSELECT to apply the algorithm developed in study 1. SAS PROC MIXED was then used for mixed effect linear model analyses examining the association between biobehavioral risk factors (social isolation and depressive symptoms) over time and longitudinal variation in biomarker values from presurgical baseline to 6- and 12-months post-surgery while controlling for clinical characteristics (stage, grade, and histological subtype) and patient demographics (age, BMI, race, and smoking history) and accounting for nonindependence of repeated measures using a fully saturated (unstructured) variance–covariance matrix on residuals.

Study 1: discovery analysis identifying an exosome RNA biomarker of ovarian cancer progression

To define a “liquid biopsy” measure of disease progression propensity in ovarian cancer, we conducted machine learning analyses of genome-wide messenger RNA (mRNA) profiles in plasma exosome samples collected pretreatment from 76 women with known disease progression times. Cases were classified as short PFS ≤6 months; n = 13, 17%) versus longer PFS (>6 months; n = 63, 83%) and the elastic net machine learning algorithm was applied to identify a composite biomarker based on pre-surgical exosome RNA profiles that optimally predicted subsequent postsurgical progression class (i.e., short PFS vs. longer PFS). Results of this analysis identified a 22-gene predictive biomarker that was strongly correlated with progression class (r = 0.95; p < .0001) and correctly classified 98.6% of cases (75 of 76) with good sensitivity (92%) and specificity (100%). In multi-variable linear model analyses, the biomarker's predictive accuracy emerged above and beyond the effects of disease grade and histological subtype (serous vs. non-serous) as well as age, BMI, race, and smoking history (adjusted r = 0.95, p < .0001). The 22-gene composite biomarker was approximately 2.4-fold greater in cases with a short PFS compared to those with a longer PFS (difference: 1.25 log2 mRNA abundance units ± SE 0.07, t(65) = 17.95, p < .0001) (Figure 1). As would be expected for a valid measure of disease progression, additional plasma samples collected longitudinally from the same patients at 6 and 12 months postsurgery showed elevated biomarker levels compared to pretreatment baseline (F [2122] = 132.65, p < .0001) with a modest increase over the first 6 months (difference: 0.07 ± 0.06 log2 mRNA abundance, t(122) = 1.19, p = .2346) and a more substantial increase by 12 months (difference: 1.37 ± 0.09, t(122) = 15.08, p < .0001). The overall association of progression biomarker levels with time since surgery was strong, r = 0.70 (p < .0001).

Study 2: biobehavioral influences on exosome progression biomarker

After deriving a “liquid biopsy” measure of disease progression biology in study 1, we next sought to use that measure in longitudinal analyses assessing how biobehavioral risk factors (social isolation and depressive symptoms) might affect disease progression biology in the interval between surgery/neoadjuvant treatment and subsequent clinical progression. To ensure that these analyses did not capitalize on chance by using the same data set to derive the biomarker outcome and to test its association with other potential predictors, these analyses were conducted in a second subgroup of 31 cases that were collected from the same clinical context as in study 1 but for whom clinical disease and survival end points were not yet available. In these analyses, mixed effect linear models analyzed changes in exosome RNA biomarker activity from pre-surgical baseline to 6- and 12-months post-surgery, and additionally tested whether social isolation or depressive symptoms might be associated with higher levels of the biomarker while controlling for potentially confounding effects of disease stage, tumor grade, tumor histological subtype, age, BMI, race, and smoking history. In analyses of 62 longitudinal observations from 24 individuals for whom social isolation data were available, results showed significantly elevated levels of the exosome RNA progression biomarker over time in those who were socially isolated (b = 0.174, t(16) = 4.26, p = .0006) (Figure 2A). In analyses of 58 longitudinal observations from 22 individuals for whom depressive symptom data were available, results also showed significantly elevated levels of the exosomal RNA progression biomarker over time in those with high levels of depressive symptoms (b = 0.145, t(14) = 2.83, p = .0134) (Figure 2B). In multi-variable analyses examining both depression and social isolation simultaneously, social isolation continued to be associated with significantly elevated progression biomarker values while controlling for depressive symptoms (b = 0.233, t(13) = 4.66, p = .0004), whereas depressive symptoms were not significantly associated with the biomarker values after controlling for social isolation (b = 0.092, t(13) = 1.81, p = .0941).

Limitations

It is possible that some of the observed transcriptional activity may have been from exosomes derived from normal cells. Previous research has indicated that in cancer patients, tumor is the primary source of circulating exosomes,⁴⁸ which have similar molecular composition as the cells from which they were derived.⁴⁹ As additional techniques are developed to isolate tumor-specific exosomes, these methods can be used to further validate these findings. In addition, exosomes originating from nontumor cells may well contribute to disease progression by several possible mechanisms, including changes in vascular biology, stromal cells, inflammation, and other changes in nontumor tissues that affect the development of metastatic niches and other aspects of disease progression.¹⁰ The present findings have focused on potential biobehavioral modulation of exosomal content and future research will be needed to assess other processes such as exosome production, release, and concentration. Although depressive symptoms and social support independently predicted the propensity biomarker, it is also possible that the reverse is true, and this is a potential limitation of the data. Additionally, the sample size was limited and additional validation in larger, more diverse independent samples is needed.

Clinical significance

The present study identifies an exosome-based biomarker of ovarian cancer progression and finds that biobehavioral factors of social isolation and depressive mood predict its variation over time. This biomarker may have additional utility as an easily accessible measure of clinical disease progression risk that can be used for intervention targeting and clinical monitoring.