Characterization of the Biological Variability of the Angiome Biomarkers over Time in Healthy Participants

Abstract

Background:

Biomarker analyses are an integral part of cancer research. Despite the intense efforts to identify and characterize biomarkers in cancer patients, little is known regarding the natural variation of biomarkers in healthy populations. Here we conducted a clinical study to evaluate the natural variability of biomarkers over time in healthy participants.

Methods:

The angiome multiplex array, a panel of 25 circulating protein biomarkers, was assessed in 28 healthy participants across 8 timepoints over the span of 60 days. We utilized the intraclass correlation coefficient (ICC) to quantify the reliability of the biomarkers. Adjusted ICC values were calculated under the framework of a linear mixed-effects model, taking into consideration age, sex, body mass index (BMI), fasting status, and sampling factors.

Results:

ICC was calculated to determine the reliability of each biomarker. HGF was the most stable marker (ICC=0.973), while PDGF-BB was the most variable marker (ICC=0.167). In total, ICC analyses revealed that 22 out of 25 measured biomarkers display good (≥0.4) to excellent (>0.75) ICC values. Three markers (PDGF-BB, TGF-β1, PDGF-AA) had ICC values <0.4. Greater age was associated with higher IL-6 (p=0.0114). Higher BMI was associated with higher levels of IL-6 (p=0.0003) and VEGF-R3 (p=0.0045).

Conclusions:

Of the 25 protein biomarkers measured over this short time period, 22 markers were found to have good or excellent ICC values, providing additional validation for this biomarker assay.

Impact:

This data further supports the validation of the angiome biomarker assay and its application as an integrated biomarker in clinical trial testing.

Introduction

In the field of oncology, biomarker studies have gained increasing attention due to their ability to provide insights into patient diagnosis, prognosis, and patient selection, furthering our goal of precision medicine(1,2). Compared to tissue-based biomarkers, blood-based biomarkers have many key pragmatic advantages, including the ease of collecting samples before, during, and after therapy. Many proteins of interest in cancer biology can be accurately measured in blood, including factors related to angiogenesis, inflammation, immunity, and tumor growth. Multiplex enzyme-linked Immunosorbent assay (ELISA) technologies allow for the assessment of multiple proteins at a lower cost and require less sample volume than the standard, single marker testing ELISA formats(3). These techniques are readily adapted to clinical diagnostics and markers identified by these approaches can be further explored as therapeutic targets. These features allow the biomarker findings to inform both precision medicine and drug development initiatives related to the use of cancer therapies in patients.

As the importance of biomarkers to guide cancer therapies escalates, the need for proper analytical and clinical validation is paramount. Our laboratory has devoted considerable effort to the development and optimization of the plasma angiome, a multiplexed ELISA panel that quantifiably measures key circulating angiogenic and inflammatory markers(4). Over the past decade, we have employed the angiome assay as a correlative biomarker across dozens of clinical trials, covering numerous cancer types and thousands of patients(5–9). In 2018, NCI Biomarker Review Committee approved the angiome to be used as an integrated biomarker in NCTN trials.

While the technical validation of the angiome has been performed and shown to be robust(8), the biologically variability of these biomarkers in healthy participants is unknown. It has long been recognized that the levels of some proteins are stable over time, while other proteins are more variable. Whether the variability in these biomarkers reflects the true heterogeneity across patients or primarily reflects noise due to a host of critical factors is key to the development of robust biomarkers(10,11). Compared to extensive biomarker assessment in cancer patients, the characterization of biomarker variability in healthy populations is limited.

In order to address this unmet need, we conducted a clinical study to evaluate the natural variability of the angiome protein biomarkers over a short time period of 60 days. We accrued 28 healthy participants and collected plasma samples over the course of two months to examine short-term variability. We controlled for the impact of key covariates, such as age, sex, BMI, and fasting or fed status, and modeled variability using the intraclass correlation coefficient (ICC) as our measure of biomarker variability.

Materials and Methods

Participants

A total of 31 healthy participants were recruited at Duke University Medical Center from 5/13/19 to 12/30/19. We had established a time frame of approximately six months to complete this study. Flyers were posted in the hospital clinics and participants were a mix of employees, students, or other regular visitors to the hospital. Participants were compensated if they completed all blood donations. Participants gave written informed consent regarding the usage of collected samples.

Each participant self-reported as not having any known underlying disease or condition. All participants were required to make four visits, on days 1, 15, 30, and 60. Three participants did not complete the required visits and were excluded from this analysis. The 28 participants evaluated consisted of 16 females and 12 males with a median age of 64, ranging from 27 to 79. Participants’ heights and weights were measured for BMI at study entry. At each visit, participants were asked if 1) any anti-inflammatory drug was used in the past eight hours, and 2) any food had been eaten in the past two hours. Samples were collected twice on each visit (“AM” and “PM”), at least four hours apart.

The study was reviewed and approved by the Duke Institutional Review Board (IRB) and the studies were conducted in accordance with the Good Clinical Practice (GCP) guidelines and institutional standard operating procedures (SOPs).

Samples

Peripheral venous blood was collected from each participant into EDTA vacutainers, centrifuged twice at 2500 x g for 15 minutes. Double-spun, platelet-poor EDTA-plasma was aliquoted, snap frozen, and stored at −80°C freezer until analysis (12). The plasma samples were frozen for less than six months before testing. Previous testing indicated that the assay provided robust data from samples that had been stored for greater than five years.

Biomarker Assessment

Levels of the 25 angiome biomarkers were measured using the multiplex ELISA platforms from Quanterix (Billerica, MA), Meso Scale Discovery (Rockville, MD) and Protein Simple platforms as previously described(4,8,9,12). The product numbers of the kits from Quanterix were 100–0443, 85–0200, 100–0112, 100–0444, 85–0187, 100–0380, 100–0409, 100–0021, and 100–0105. The catalog number for the VEGF-C kit from Protein Simple was SPCKB-PS-000348. Assays using MSD multi-array 96-well plate (catalog # L15XA-3) to test BMP9(13), CD73(7), and TGFβ-R3(5) were reported previously. All the samples were run in duplicates and the mean value used for statistical analysis.

Statistical Considerations

On the basis of eight longitudinal measurements for each participant and across 28 participants, we estimated intra- and inter-participant variability under the framework of a linear mixed-effects model. Initially, we considered a model including a participant-level random intercept. Later, we extended the error structure of the model, nesting day within each participant. We extended these analyses by assessing potential confounding of age, sex, BMI, fasting or fed status by including these covariates as fixed additive effects. We assumed the random effects (inter-participant variation, across the four visit days for the same participant, AM vs. PM) and measurement errors were homoscedastic normal and assumed independence all around.

All analyses were conducted using the R statistical environment (version 4.3.1). The variance estimation of each model was conducted using restricted maximum likelihood estimation (REML) implemented by the lme4 (version 1.1.34) extension package. To get p-values for the fixed effects, Satterthwaite approximation of the degrees of freedom implemented by the lmerTest (version 3.1.3) extension package was used.

Using these models, for each biomarker we calculated ICC by dividing the inter-participant variance from a linear mixed model by the total variance, using the formula ICC = (inter-participant variance)/[(inter-participant variance) + (intra-participant variance)]. Furthermore, adjusted ICCs were calculated taking into consideration age, sex, BMI, fasting status, and sampling factors. For the model including only a participant-level random effect, the estimate was of the form v[id]/(v[id]+v[e]) where v[id] and v[e] were the estimated variances for the participant-level random effect and measurement error. For the nested model, this estimate was of the form (v[id]+v[id:day])/(v[id]+v[id:day]+v[e]), where v[id:day] was the estimated variance for the day-level random effect nested within participant. Confidence intervals were calculated using bootstrapping with 1000 iterations. P-values were adjusted by controlling the family-wise error rate across the 25 biomarkers using Bonferroni correction.

Spaghetti plots were created for each biomarker to show the trend of each participant’s expression levels over time (Supplemental Figure S1). To visualize intra-patient and inter-patient variations in biomarkers, boxplots of log₂ transformed expression levels were created, with each box representing a given participant’s eight measurements (Supplemental Figure S2). Residual plots and Q-Q plots were used as exploratory visuals to assess deviations from homoscedasticity and normality assumptions (Supplemental Figure S3).

Data and code availability

The clinical and angiome biomarker data used in this study are available upon request from the corresponding author. The analyses presented here were carried out with adherence to the principles of reproducible analysis using the knitr package for generation of dynamic reports, tables, and figures, and git source code management. The scripts to reproduce the results presented in this report from source data are available through a public source code repository https://gitlab.oit.duke.edu/dcibioinformatics/pubs/nixon-healthy-volunteers, and the data for this study will be made available through the Open Science Framework (OSF) at https://osf.io/kycqe/.

Results

Participants and biomarker levels

To investigate normal biomarker variation, we recruited 28 healthy participants and collected plasma samples at eight time-points over the span of 60 days. The demographics of all participants are shown in Table 1. Prior to blood draw, participants were asked whether they had eaten in the past two hours (fed/fasting status) and whether any anti-inflammatory drugs had been taken within eight hours prior to visit. Only one participant at one time-point acknowledged taking an anti-inflammatory drug within the past eight hours. Two participants did not provide fasting or fed information at a specific visit, and BMI was not available for eight participants. Median levels and ranges for all markers are shown in Table 2. The assays performed well technically with low coefficients of variation (CV<10%, ranging 1.6 – 7.2%) for all biomarkers tested.

Table 1.

Demographics of healthy participants.

Characteristics	Median (range)
Age	64 (27–79)
Weight (kg)	79.4 (57–153.3)
Height (cm)	169.6 (152.4–185.4)
BMI (kg/m2)	30 (22–66)
Gender (n, %)
Female	16, 57%
Male	12, 43%
Race (n, %)
Caucasian	16, 57%
African American	10, 36%
Other	2, 7%

Table 2. Circulating levels of 25 biomarkers in plasma.

The median and range of each marker are shown. Units vary based on the concentration of the individual biomarker.

Biomarker	Unit	Median (range)
Ang-2	pg/ml	343.8 (13.9 – 1095.7)
BMP-9	pg/ml	66.6 (34.3 – 1377.7)
CD73	ng/ml	2.5 (0.6 – 16.7)
GP130	ng/ml	301.1 (143.7 – 459.3)
HGF	pg/ml	123.3 (13.4 – 290.0)
ICAM1	ng/ml	397.3 (172.4 – 982.8)
IL-6	pg/ml	1.5 (0.2 – 10.7)
IL-6R	ng/ml	36.4 (18.5 – 68.9)
OPN	ng/ml	73.1 (23.4 – 214.9)
PDGF-AA	pg/ml	120.7 (6.7 – 1683.0)
PDGF-BB	pg/ml	850.8 (43.8 – 9692.5)
PlGF	pg/ml	11.3 (3.1 – 30.8)
SDF-1	ng/ml	1.3 (0.07 – 3.7)
TGF-β1	ng/ml	16.1 (2.4 – 129.9)
TGF-β2	pg/ml	26.8 (3.4 – 107.7)
TGFβ-R3	ng/ml	180.2 (54.8 – 274.4)
TIMP-1	ng/ml	57.1 (13.9 – 117.5)
TSP-2	ng/ml	168.0 (55.5 – 434.1)
VCAM-1	μg/ml	1.8 (0.9 – 3.1)
VEGF	pg/ml	39.9 (13.9 – 160.4)
VEGF-C	pg/ml	509.7 (0.0 – 2406.0)
VEGF-D	ng/ml	1.1 (0.7 – 2.3)
VEGF-R1	pg/ml	71.0 (20.7 – 136.3)
VEGF-R2	ng/ml	4.4 (2.0 – 6.9)
VEGF-R3	μg/ml	0.2 (0.08 – 0.5)

Biomarker variation over time

As an initial visual representation of biomarker variability, each biomarker was plotted over time and displayed as spaghetti plots. As shown in Figure 1, HGF was observed to be a highly stable marker. While marked differences were noted across participants, HGF levels remained consistent over time, including daily variation, within an individual. In contrast, PDGF-BB was noted to be the most variable biomarker. Not only did PDGF-BB levels vary greatly over the course of 60 days, but within the same day, the two AM and PM measurements were often markedly different. Detailed plots for the remaining markers are shown in Supplemental Figure S1.

Figure 1. Biomarker variability over time for HGF and PDGF-BB.

Each line represents an individual participant with biomarkers measured twice a day at days 0, 15, 30, and 60.

Inter- and intra-patient variation

To capture a more simplified view of inter- and intra-patient variability, boxplots are used to display the average expression levels for each participant across all timepoints. HGF and PDGF-BB are illustrated again as two representative examples of low- and high-variability biomarkers (Figure 2). HGF exhibited little fluctuation within each participant over time, yielding small boxes with tight ranges. However, variation across participants was substantial, with diverse HGF expression across the healthy participants. In contrast, PDGF-BB varied markedly within each participant, reflected by the large boxes exhibiting of a wide range of biomarker expression over time. When inter-participant variation of PDGF-BB was evaluated, the high degree of intra-participant variation resulted in overlapping expression patterns across participants. Detailed plots for the remaining markers are shown in Supplemental Figure S2.

Figure 2. Intra- and inter-participants variability for HGF and PDGF-BB.

The box represents the 1^st and 3^rd quartile, the middle bar represents the median, and the whiskers extend from the quartiles to the most extreme values no further than 1.5 times the inter-quartile range for each participant.

Intraclass correlation (ICC) model

In order to model the variability, we opted to use ICC as our composite measure. ICC values for all 25 biomarkers were shown in Table 3. We categorized the inferred reliability of each marker as excellent, good, or poor based on the cutoffs of greater than 0.75, between 0.4 and 0.75, or below 0.4 respectively, as proposed in the criteria by Rosner(14). Based on this rubric, of the 25 biomarkers tested, 22 markers were considered to have either excellent (ICC≥0.75, n=17) or good (0.4≤ICC<0.75, n=5) reliability. The median ICC for these 22 reliable biomarkers was 0.858, ranging from 0.417 to 0.973. Only three markers (PDGF-BB, TGF-β1, PDGF-AA) had ICC values lower than 0.4, indicating poor reliability (Table 3).

Table 3: Intraclass correlation coefficient (ICC) for all markers.

Adjusted ICC estimates were generated taking into consideration age, sex, BMI, fasting or fed status, and random effects including day and AM/PM collection within participant. % Change is calculated as (ICC – Adjusted ICC)/ Adjusted ICC.

Biomarker	ICC	Adjusted ICC	Δ(ICC-adj ICC)	% Change
HGF	0.973 (0.948, 0.984)	0.951 (0.894, 0.975)	0.022	2.3
BMP9	0.963 (0.934, 0.978)	0.963 (0.912, 0.98)	0	0
Ang2	0.956 (0.923, 0.974)	0.955 (0.904, 0.975)	0.001	0.1
SDF1	0.938 (0.89, 0.96)	0.948 (0.883, 0.971)	−0.01	−1.1
VEGFR2	0.908 (0.841, 0.945)	0.902 (0.81, 0.947)	0.006	0.7
ICAM1	0.893 (0.814, 0.935)	0.855 (0.702, 0.916)	0.038	4.4
IL6R	0.889 (0.802, 0.93)	0.818 (0.632, 0.894)	0.071	8.7
VCAM1	0.879 (0.783, 0.922)	0.85 (0.699, 0.916)	0.029	3.4
PlGF	0.87 (0.769, 0.92)	0.815 (0.653, 0.892)	0.055	6.7
OPN	0.87 (0.778, 0.921)	0.875 (0.731, 0.932)	−0.005	−0.6
CD73	0.861 (0.754, 0.91)	0.885 (0.761, 0.936)	−0.024	−2.7
GP130	0.855 (0.755, 0.907)	0.788 (0.598, 0.879)	0.067	8.5
VEGFR3	0.848 (0.745, 0.905)	0.743 (0.538, 0.846)	0.105	14.1
VEGFD	0.834 (0.708, 0.895)	0.84 (0.695, 0.908)	−0.006	−0.7
TSP2	0.817 (0.708, 0.879)	0.796 (0.62, 0.882)	0.021	2.6
IL6	0.768 (0.631, 0.851)	0.383 (0.151, 0.557)	0.385	100.5
TGFbR3	0.751 (0.618, 0.837)	0.719 (0.503, 0.835)	0.032	4.5
VEGFC	0.66 (0.498, 0.775)	0.726 (0.529, 0.837)	−0.066	−9.1
TGFb2	0.631 (0.456, 0.748)	0.622 (0.382, 0.755)	0.009	1.4
TIMP1	0.619 (0.438, 0.729)	0.64 (0.411, 0.786)	−0.021	−3.3
VEGFR1	0.566 (0.377, 0.697)	0.508 (0.262, 0.668)	0.058	11.4
VEGF	0.417 (0.237, 0.574)	0.408 (0.173, 0.585)	0.009	2.2
PDGFAA	0.257 (0.107, 0.396)	0.191 (0.036, 0.352)	0.066	34.6
TGFb1	0.238 (0.091, 0.381)	0.169 (0.013, 0.345)	0.069	40.8
PDGFBB	0.167 (0.043, 0.29)	0.165 (0.017, 0.325)	0.002	1.2

The association between covariates and biomarker levels

To refine the ICC model, we tested if factors such as age, sex, BMI, and fasting/fed status were associated with biomarker levels. Greater age was significantly associated with higher IL-6 levels (Effect size 3.42, p=0.0005, adjusted p=0.0114, Supplemental Table S1). While male versus female was associated with lower TGF-β1 (Effect size −0.705, p=0.0111, adjusted p=0.2676), PDGF-AA (Effect size −1.128, p=0.0261, adjusted p=0.6257), and IL-6 (Effect size −0.583, p=0.0279, adjusted p=0.67), all associations lost statistical significance after multiple testing correction (Supplemental Table S2).

We further explored the association between BMI and biomarker levels and found higher BMI was associated with higher IL-6 and VEGF-R3. For example, a one unit increase of BMI was associated with an increase of IL-6 levels by a factor of 2.8 (2^1.507). This association remained statistically significant after p-value correction (adjusted p=0.0003). A complete list of the associations of BMI with each biomarker is shown in Supplemental Table S3.

We last investigated whether fasting or fed status was associated with these biomarker levels. Based on whether a participant had eaten within two hours of blood collection, the samples were divided into those from participants who had not eaten (n=121) and those from participants who had (n=99). After adjustment for multiple testing, no evidence was found that the biomarkers tested were associated with fasting or fed status. A complete list of associations of fasting or fed status with each of the markers is shown in Supplemental Table S4.

Adjusted ICC Analyses

Given the potentially significant impact of the above-described factors and sampling effects on biomarker expression, we adjusted the ICC model, accounting for age, sex, BMI, fasting/fed status, and random sampling effects. Sampling consisted of three variables: inter-participant variation, variation across the four sampling days for the same participant, and AM/PM variation for the same participant on the same day (Supplemental Table S5). The impact of adjusting for these variables is shown in Table 3, where differences between ICC and adjusted ICC values are listed. While few markers were greatly impacted, several biomarkers had adjusted ICC value changes of greater than 0.07 units. The ICC values most impacted by adjustment were those of IL-6, VEGF-R3, and IL6R (Table 3).

Discussion

The angiome has gone through rigorous analytic validation as quality metrics have been optimized and all assays have been locked down and perform well. While ensuring robust technical performance, we wanted to further investigate the role of natural variability in these biomarker levels and understand both inter-participant and intra-participant variability. While different ICC approaches with varying cut points have been reported(15), we chose to use the cut-point proposed by Rosner to distinguish good vs. poor reliability on our standard ICC calculations (16).

In this analysis, 22 out of 25 markers (88%) had ICC values >0.4, and 17 markers (68%) had ICC >0.75, indicating high reliability of the angiome panel overall. In our previous analyses, several potential predictive biomarkers have already been identified. We have shown predictive value for IL-6 in renal cell cancer(9) and ovarian cancer(6) patients treated with bevacizumab. Additional studies in GI-related malignancies have revealed that circulating levels of VEGF-D, an alternate ligand for VEGFR2, were predictive of anti-angiogenic therapy in GI diseases, including pancreatic cancer(5) and colorectal cancer(8). The high ICC values for both IL-6 (0.768) and VEGF-D (0.834) further supports the potential utility of these biomarkers clinically.

For the 25 markers tested in this study, only three markers had ICC estimates less than 0.4 (TGF-β1, PDGF-AA, and PDGF-BB). These are functionally related proteins and it is well established that TGF-β1 can directly regulate the expression of PDGF-AA and PDGF-BB(17). With respect to the use of anti-angiogenic therapies, PDGF signaling is crucial in recruiting pericytes to endothelial cells during blood vessel maturation(18,19). Notably, these three biomarkers have rarely been identified to be prognostic or predictive in previous angiome analyses(20,21). This does not imply these three markers have no value. In fact, TGF-β1 has been an extensively studied marker having wide ranging applications(22). Validating the correlation between ICC and the prognostic/predictive value of a marker, or defining precise cutoffs for ICC, is beyond the scope of this study. Yet our observations suggests that ICC may be useful in determining the robustness of biomarkers, especially when being considered for clinical decision-making.

In evaluating the biological variability of these biomarkers, we considered additional covariates, including age, sex, BMI and fasting/fed status. Aging is an extensively studied area and multiple biomarkers have been shown to correlate with chronological age, including soluble biomarkers of inflammation such as IL-6(23). Our data also demonstrated a significant association between age and IL-6 levels. We observed that sex was weakly associated with TGF-β1, PDGF-AA, and IL-6 levels, although these associations lost significance after multiple testing adjustment.

BMI was observed to be associated with IL-6 and VEGF-R3 levels (adjusted p < 0.05, Supplemental Table S3). Obesity is a known inducer of chronic inflammation, and the correlation between obesity and the inflammatory cytokine, IL-6, has been widely reported(24). Obese participants have been shown to have elevated IL-6 levels(25), while significant decreases in IL-6 levels have been observed after bariatric surgery or due to cachexia(26). The impact of BMI on VEGF-R3 is less clear; however, a recent review including 1449 participants revealed that dietary-based weight loss interventions led to significantly decreased level of VEGF(27). Additionally, Karaman et al. studied the role of the lymphangiogenic factors (VEGF-C and -D) in the mediation of metabolic syndrome-associated adipose tissue inflammation and proposed the use of soluble VEGF-R3 for the prevention of obesity-associated insulin resistance(28). Significant crosstalk exists among these angiogenic ligands and receptors.

Results from the linear mixed effects model are sensitive to deviations from distributional assumptions including those of homoscedascity and normality of the random effects and measurement errors. In view of that and the small cohort size, the statistical analysis should be considered exploratory. Another limitation of this study is that the cutoffs used for the ICC categorizations are somewhat arbitrary and vary across studies. These caveats notwithstanding, we still observed that most angiome biomarkers were stable and resistant to external influences.

In summary, we tested the biological variability of the circulating biomarkers in the angioma multiplex ELISA panel. Of the 25 biomarkers, 22 biomarkers demonstrated good to excellent ICC estimates, i.e., low intra-participant variation over time despite high inter-participant variation. In general, biomarkers with higher ICC values were not observed to be associated with factors such as age, sex, BMI, or fasting versus fed status. As a result, these markers may be more reliably correlated with disease status, drug effects, and clinical outcomes. However, for IL-6, consideration should be given to the impact of BMI on this biomarker. Finally, three biomarkers (TGF-β1, PDGF-AA, and PDGF-BB) had very low ICC values and may not be appropriate choices for further testing in this setting.

This data satisfied an unmet need to show the level of natural variation of 25 biomarkers in a healthy population, with the potential to serve as reference points for biomarker analyses in cancer patients. The significance of this data goes beyond the specific angiome panel used here and results should be generalizable to other validated platforms measuring these biomarkers.