Affinity proteomic profiling of plasma for proteins associated to area-based mammographic breast density

This study included samples collected from participants of the KARMA cohort [14]. KARMA is a population-based cohort initiated in January 2011, which comprises 70,877 women attending routine mammography screening or clinical mammography at four hospitals in Sweden [14, 15]. The overarching goal of KARMA is to reduce the incidence and mortality of breast cancer by focusing on individualised prevention and screening.

Raw (unprocessed) digital mammograms for each study participant were collected at KARMA study enrolment [14, 15]. Mammograms were taken from cranial-caudal and mediolateral oblique views by full-field digital mammography. Mammographic density was measured two-dimensionally as an absolute dense area (AD) (cm²) using the newly developed in-house STRATUS program as previously described [15, 16] and three-dimensionally as an absolute dense volume (VD) (cm³) using the automated Volpara system. STRATUS analyses both raw and processed mammograms and estimates the breast and dense area based on mammographic textures. Each pattern segment is analysed for several statistical features including pattern area, circumference, intensity, positioning, relation to other areas and shape. This quantified texture structure of the breast is compared to a reference library of matching breast texture-density-level pairs. The reference library was created using the penalised lasso regression machine-learning method.

The total mammographic dense area and percent mammographic density did not differ significantly between the two sample sets (p = 0.80 and p = 0.90, respectively). AD and VD measures from the right breast were considered for statistical analysis.

KARMA participants were included in the study based on measured VD and selected from the total KARMA study population (N = 70,773). For practical reasons, the study was conducted in two phases resulting in two sample sets; sample set 1 included 729 women from three sample groups and sample set 2 included 600 women from two sample groups (Table 1 and Fig. 1). No participant had a prior history of breast cancer or other malignant cancer at the time of sampling. One individual developed breast cancer 2 years after blood draw.

For each sample set, women were allocated into two subsets of VD (high and low). The high-density sample groups (sample set 1, median VD = 104.9 cm³; sample set 2, VD = 100.2 cm³) were women from the highest quintile of absolute volumetric density in the KARMA cohort. The low-density sample groups (sample set 1, median VD = 33.5 cm³; sample set 2, median VD = 33.5 cm³) were women from the lowest quintile of absolute volumetric density in KARMA. The sample groups (high and low VD) were matched on age and BMI (Fig. 2). An additional 139 samples from another KARMA study (denoted “Karma Normal”) were selected in the same way based on the highest and lowest quintiles of absolute volumetric density (median VD = 68.3 cm³) and included in sample set 1 (Fig. 2). Karma Normal is a nested study within KARMA with the objective to study normal breast physiology and only includes samples from healthy participants in KARMA, without any history of breast cancer or other cancers. Karma Normal has been described in detail elsewhere [17]. Participants in both sample sets were matched to the Information Network for Cancer treatment (INCA) to ensure disease-free status at the time of sample collection. BMI was calculated at the time of the mammogram and was based on self-reported height and weight. Distributions of sample characteristics and breast cancer risk factors were similar between the two study sample sets (Table 1) and between each study sample set and the total KARMA cohort. Each study participant signed an informed consent form before joining the KARMA project. The Stockholm ethical review board approved the study (2010/958-31/1).

Non-fasting EDTA plasma samples of peripheral blood were collected from the KARMA study participants at enrolment [14, 15]. All blood samples were handled in accordance to a strict 30-h cold-chain protocol and were processed in the Karolinska Institutet high-throughput biobank. The majority (97.5%) of blood samples were taken on the same day as the mammogram. Mean time between the mammogram and blood sample collection was 4.8 h (SD 57.6 h). The time interval from the mammogram to blood collection did not differ significantly between sample set 1 and sample set 2 (p = 0.80).

For multiplexed protein profiling, sets of 382 and 393 antibodies derived from the Human Protein Atlas [13] were used. These targeted a total 445 unique protein-encoding genes, and a complete list of all antibodies included in the study is provided in Additional file 1: Tables S1-S2. The 382 antibodies included in the first suspension bead array (SBA1) were selected based on a possible relationship with mammographic breast density, cancer development and/or progression or tissue composition and/or remodelling. The 393 antibodies included in the second bead array (SBA2) targeted proteins annotated to extracellular matrix (Uniprot.​org; N = 156) [18] and proteins enriched in breast tissue according to RNA sequencing (RNAseq) data [13]. The list also included antibodies selected from immunohistochemistry (IHC) primary data [13]. Further details about antibody generation and selection can be found in Additional file 2.

Antibody bead arrays were generated using carboxylated magnetic beads of up to 393 unique bead identities (MagPlex-C, Luminex Corp.) as previously described [12]. All plasma samples within each study set were retrieved from the biobank and analysed at the same time points. Plasma samples stored at − 80 °C were thawed at 4 °C and transferred to 96-well microtitre plates in a semi-randomised plate layout, where samples from different sampling locations were balanced across the different plates and each matched pair of the two sample groups (high and low VD) were placed within the same plate. The randomised plate layouts resulted in an even distribution of AD across all 96-well plates (Kruskal-Wallis p values 0.94 and 0.57 for sample sets 1 and 2, respectively). All plates included four aliquot replicates from a crude plasma pool from all individuals included in that study set. Samples were biotinylated, diluted, heat-treated at 56 °C and combined with the bead array on two separate 384-well assay plates in accordance with previously described protocols [19]. Further details can be found in Additional file 2.

Different types of assays were used to validate the antibodies. Detailed descriptions about epitope mapping by high-density peptide arrays, western blot and immuno-capture mass spectrometry analysis of plasma samples can be found in Additional files 1 and 2.

Data from SBA assays were processed separately according to the following procedure: blank (sample-free with buffer) wells were excluded from analysis. In sample set 2, the replicated data from one 96-well plate was used only for quality control; meaning only one of each sample in a duplicated pair was included for statistical analysis. Outlying samples, detected by robust principal component analysis (PCA) [20], were replaced by missing values (N/A) using the “rrcov” R package. Probabilistic quotient normalisation [21] was then applied for all data points originating from each 96-well plate, followed by between-plate normalisation using a multidimensional normalisation method [22]. Prior to statistical analyses, antibody profiles were annotated based on assay performance. The annotations were based on four different criteria, including median signal intensities above that of the negative control bead identity (rIgG). A more detailed description is provided in Additional file 2. Filtering of antibody profiles based on such technical quality assessment resulted in a refined list of 245 (SBA1) and 244 (SBA2) antibodies against a total number of 357 proteins that were targeted within each study set. PCA was applied for quality control and to detect potential sampling location effects. Prior to PCA, data were log-transformed, centred and subjected to unit variance scaling, and missing data points were replaced by the median of the complete data set.

The initial aim of the study was to contrast plasma protein profiles of women with high and low VD. However, during the proceeding time, a study by Nguyen et al. showed that breast cancer risk is more strongly associated with the denser part of the breast [23]. AD is thus likely a better representation of the true dense tissue in the breast. We therefore updated our strategy and performed our density-protein association analyses using absolute area-based density measures, which targets the most radio-dense tissue in the breast. Identification of protein profiles in relation to AD while controlling for age and BMI is thus relevant for providing new biological insights into the mechanisms of mammographic breast density. Accordingly, and prior to data analyses, we decided to use AD as our primary endpoint. We also report results according to our original design, using VD as endpoints and samples matched for age and BMI (Fig. 4 and Additional file 2).

For contrasting high and low VD, the paired Wilcoxon signed-rank test of normalised and log transformed data was used. In order to keep the matching of age and BMI between paired sample IDs, sample IDs matched to those classified as outliers, based on signal intensities in robust PCA in the experimental data, were removed prior to analysis. This resulted in 5 matched pairs (10 samples) within sample set 1 and 6 matched pairs (12 samples) within sample set 2 being excluded from two-group comparisons.

To assess the association between antibody profiles and AD, normalised, log-transformed data scaled to unit variance were used for statistical analysis. Eleven sample outliers, as identified by robust PCA, were excluded from the statistical analysis. Quantile regression models were computed using the “quantreg” package in R. In both sample sets, the correlation between AD and BMI differed between the matched groups of high and low VD; there was stronger negative correlation between AD and BMI in the low VD group (rho = −0.51 and −0.53, respectively) than in the high VD group (rho = −0.19 and −0.12, respectively) (see Additional file 2: Figure S1). Consequently, the effect of AD was adjusted for age, BMI, sample group (high or low VD) and the interaction between BMI and VD. When stated, p values from each study set (set 1 and set 2) were combined by Fisher’s method and adjusted for multiple testing using the Benjamini-Hochberg method (referred to as “adj. p”). Data analysis and statistical analysis were performed in R.

First, we investigated the associations between the plasma profiles and age and BMI. Both variables influence AD but were not associated with one another in the studied sample sets (p > 0.1). Age was associated with 11 plasma profiles at p < 10⁻¹⁰, with concordant trends in both sample sets (Additional file 2: Table S3). Among these, the profiles for AMBN, TMEM86A, MLH1, PTGR1 and SPNS1 were less strongly associated with BMI (p > 0.001), and all but SPNS1 decreased with age. For association with BMI, the overall significance levels were lower compared to those for age, and there were concordant trends for 10 profiles in both sample sets (p < 10⁻⁵; Additional file 2: Table S4). Among the profiles associated with BMI, only TPP1 and ENG profiles were less strongly associated with age (p > 0.001). Interestingly, the trends of the slopes for BMI and age only differed for TPP1.

We subsequently analysed the linear relation between protein profiles and AD. The data were adjusted for age, BMI and the interaction between BMI and VD group. The distributions of AD and VD within the three sample groups are illustrated in Fig. 2. Using quantile regression models, we identified 20 candidate profiles that were significantly associated with AD (p < 0.05) in both sample sets. All proteins remained significant (adj. p < 0.05) after combining the p values from both sample sets and adjusting for multiple testing. In total, 11 of the 20 proteins (55%) were negatively associated with AD (Tables 2 and 3). Among these were ACOX2, ITGB6 and SHC1, which had been observed as proteins strongly associated with age and BMI. Next, we investigated the candidates for expression in breast tissue. Annotations of gene expressions were obtained from publically available RNAseq or immunohistochemistry data (Additional file 2: Table S5) [13, 24]. Table 2 lists those 11 candidates for which gene or protein expression has been detected in human breast tissue. Figure 3 demonstrates linear associations with the eleven candidates and shows that plasma levels of ABCC11, TNFRSF10D, F11R and ERRF were positively associated with AD, while SHC1, CFLAR, ACOX2, ITGB6, RASSF1, FANCD2 and IRX5 were negatively associated with AD. The additional nine candidates lacking RNA expression in breast tissue are shown in Table 3.

The aforementioned analysis revealed 20 proteins with concordant associations with AD in the two study sets. We also analysed the data in relation to VD in accordance with the initial study design. When comparing women with different VD, we identified significantly elevated levels of forkhead box P3 (FOXP3) using HPA045943 in the high VD group compared to the low VD group (sample set 1, p = 0.004; sample set 2, p = 0.01; Fig. 4). The anti-FOXP3 antibody, however, was not among the 20 antibodies that were the linearly associated with AD.

We conducted several experiments to support the indications obtained for the high throughput immunoassays. Acknowledging the challenge to validate antibodies due to their context-dependent and assay-dependent functionality and methods of different sensitivity, we first investigated if paired antibodies raised towards a common protein target would reveal concordant information in the high throughout assay. The data for HPA054101, raised against an internal region of FANCD2 were indeed supported by a second anti-FANCD2 antibody (HPA063742), which was generated against the N-terminal part of the protein. HPA063742 was associated with AD within sample set 1 (p = 0.0005) but was not statistically significant in sample set 2 (p = 0.09). The two antibodies were correlated, with rho = 0.46 (in sample set 1) and rho = 0.62 (in sample set 2). Similarly, the association between increasing AD and higher levels of ABCC11 that was observed for HPA031981 was supported by two additional anti- ABCC11 antibodies in sample set 2 (HPA031979; p = 0.0004 and HPA031982; p = 0.02, respectively). The three anti-ABCC11 antibodies were generated against separate regions of ABCC11.

Next, epitope mapping was conducted on high-density peptide arrays (Additional file 2: Figure S2), which revealed overlapping epitope regions for HPA054101 and four distinct epitope regions for HPA063742. None of these epitopes was found to be homologous with abundant plasma proteins, hence supporting on-target recognition of FANCD2 (Additional file 2: Table S6). For HPA064845, raised against ACOX2, four distinct epitope regions were identified. Furthermore, western blot analysis (Additional file 2: Figure S3) revealed a single band within +/− 20% of the expected weight range of ACOX2 (75 kDa). Also, for HPA065387 (TNFRSF10D), which was positively associated with AD in our data, a single band at the expected molecular weight ± 42 kDa was identified. Last, immuno-capture mass spectrometry (IC-MS) was used to assess the selectivity of the highlighted proteins in plasma (Additional file 2: Figure S4). On-target binding was confirmed for six targets, namely ERRF (HPA026676; z score = 8.4), RASSF1 (HPA040735; z score = 9.1), IL4 (HPA042270; z score = 7.5), and ITGB6 (HPA023626; z score = 8.1), F11R (HPA061700; z score = 3.3) and ABCC11 (HPA031981; z score = 11.4). For the latter three, additional proteins were co-enriched, suggesting either multiple off-targets or a complex formation. We suggest a complex has been formed between ITGB6 and LDHA, because of minimal overlap between the HPA023626 antigen region and LDHA (residues CxIxxL). In IC-MS, off-target enrichment was observed for HPA001577 (anti-SHC1; off-target THBS4: z score = 4.1) and HPA049377 (anti-PSMA8; off-target LGALS3BP: z score = 8.4). Antibodies for FANCD2 (HPA063742) and LIN28B (HPA061745) did not reveal a specific enrichment over the population of commonly identified peptides (z score <3). Results from validation experiments are summarised and annotated in Additional file 1: Table S7.