Serum immuno-oncology markers carry independent prognostic information in patients with newly diagnosed metastatic breast cancer, from a prospective observational study

Patients with newly diagnosed MBC were enrolled into a prospective observational trial (ClinicalTrials.gov NCT01322893) conducted at Skåne university hospital and Halmstad county hospital, Sweden, between April 2011 and June 2016. Briefly, the inclusion criteria were MBC diagnosis, age ≥ 18 years, Eastern Cooperative Oncology Group (ECOG) performance status score of 0–2, and a predicted life expectancy of more than two months. Exclusion criteria included previous systemic therapy for metastatic disease or other malignant disease diagnosis within the last five years, and inability to understand the study information. The study was approved by the regional research ethics committee at Lund University (Dnr 2010/135) and conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from all patients prior to study inclusion. The cohort constitutes 156 patients with newly diagnosed MBC, planned for first-line systemic treatment and has previously been described in detail [19‐21]. Serum samples were taken at baseline (before start of systemic therapy) and out of the 156 patients, 20 serum samples were lost or excluded (for 14 patients; no serum sample was available, five samples did not pass the quality control and one sample did not provide enough material) (Fig. 1). The number of circulating tumor cells (CTCs) and presence of CTC clusters were evaluated using the CellSearch™ system, and the results have been reported in a previous study [19].

Serum samples were collected at baseline (before start of systemic treatment). The samples were collected in serum tubes, centrifugated, aliquoted and stored at -80 °C until analysis. The proteomic analyses were performed at the Olink facility in Uppsala, Sweden, using the OLINK Proteomics PEA technology [11, 22] (Olink Target 96 Immuno-Oncology panel, analyzing 92 protein biomarkers). Patient information was blinded to personnel at Olink Bioscience and samples were distributed randomly into the analysis plates. Samples were processed at Olink Bioscience according to their manuals, including quality control of data, and normalization of measured results. Intensity normalization was used to adjust for inter-plate variability, with the plate median as the normalization factor. Results were provided as relative values; normalized protein expression (NPX), an arbitrary unit on a log₂ scale. On this scale, a one-unit increase of the NPX value corresponds to a doubling of the protein content, with high NPX values corresponding to a high protein concentration. NPX values can, however, not be converted to absolute protein concentrations. Nor are the NPX values directly comparable between different serum proteins, as they are calculated separately for individual analytes. Data were pre-processed as follows: Proteins with NPX values below the limit of detection (LOD) in more than 15% of the samples and samples with > 15% proteins below LOD were removed from further analysis. For the remaining values, we rescaled the NPX values by subtracting the LOD except for those already below two standard deviations of the LOD, which in turn were assigned as not applicable (NA). All missing values were then imputed based on a nearest neighbor approach (kNN imputation with k = 5).

Random forest (RF) analyses and random survival forest (RSF) analyses were performed to identify predictors of overall survival (OS) or progression-free survival (PFS). OS was calculated from the time of baseline blood sampling to death from any cause. PFS was calculated from the time of baseline blood sampling to progression evaluated using modified RECIST criteria, as described previously [19]. If the outcome was not reached, the time variables were censored at the last follow-up. To remove potential biasing effects that established clinical prognostic factors may have, we regressed the effect of the following factors through multi-linear modeling prior to RF/RSF: age, ECOG, metastasis-free interval, number of metastases, site of metastasis, breast cancer subtype, NHG and number of CTCs (> / ≤ 5 CTCs). RF was initially employed for parameter selection, i.e., to identify serum proteins with little informative value in predicting OS/PFS, from which we used the top 20 proteins as input for further parameter tuning, as indicated for the RSF below. RSF is a random forest method that is used for analysis of right-censored survival data. The method identifies the variables that best predict survival, OS and PFS, respectively. For RSF, we performed the analyses in a three-step process as follows: First, hyperparameter tuning was performed for maximum number of features, number of total estimators, minimum sample number per leaf, and minimum number of samples per split through a randomized grid search process coupled with fivefold cross-validation, repeated 100 times. Second, with the best fit hyperparameters, we then performed tenfold cross-validation for final model training and prediction and repeated this process 100 times. Third, with the best selected parameters and final trained model, we identified feature importance by permutation importance to assess how removal of different features affected the model performance [23]. Recursive feature elimination was then used to examine how many features were needed to maximize the model concordance, Harrell’s C-statistic, for both PFS and OS. We further validated the selected serum proteins using both Cox regression and Penalized Cox regression, both with regressed and un-regressed data. Machine learning technique analyses (RF, RSF, Cox regression and penalized Cox regression) were performed using R 4.1, Rstudio 1.1.456, limma 3.50.0, mixomics 6.16.0, jupyter 1.0.0, scikit-learn 0.24.2, scikit-survival 0.15.0.post0 and python 3.9.9.

Cox regression analyses were performed for estimation of hazard ratios (HR) with 95% confidence interval (CI) for both PFS and OS according to the continuous serum protein NPX values, both univariable analyses and multivariable analyses adjusted for the prognostic factors mentioned above. The association between serum protein levels and different patient and tumor characteristics was analyzed using Fisher’s exact test or logistic regression where appropriate, with continuous serum protein NPX values dichotomized using the median value for each protein as the cut-off point. The study was performed according to the REMARK criteria [24].

All P values presented are two-sided and should in general be regarded as continuous measures of evidence, but following Benjamin et al., two thresholds are used throughout this paper: suggestive evidence for P values between 0.05 and 0.005 and significant evidence for P < 0.005 [25]. In addition, false discovery rate (FDR) was used to adjust Cox regression P values for multiple testing. P values for all 92 proteins studied were used to calculate so-called q values separately for uni- and multivariable analysis and for each endpoint. SPSS Statistics version 27.0.1.0 (IBM, Armonk, NY, USA) was used for correlation and Cox regression analyses. FDR analyses, as described in Benjamini and Hoffberg [26], were carried out using the user contributed program qqvalue.ado in Stata 17.0, (StataCorp LLC, College Station, TX, USA).

A total of 156 patients with newly diagnosed metastatic breast cancer (MBC) were enrolled in the study and serum samples were collected at baseline, before start of systemic therapy. Among the 156 patients, serum samples were not available for 20 patients (Fig. 1). Information on the whole patient cohort has been published before [19]. Patient and tumor characteristics for the 136 patients included in this study, summarized in Table 1, are representative for the whole cohort. Twenty-eight patients were diagnosed with de novo MBC and 108 patients were diagnosed with distant recurrence. The median follow-up time from baseline was 20 months (range 0–66) for patients alive at the last medical visit. Median age of the patients at time of MBC diagnosis was 65 years (range 40–90 years). Breast cancer subtype was determined in metastases first-hand and primary tumors second-hand, with 94 patients (69%) having estrogen receptor-positive (ER +) and human epidermal growth factor receptor 2 negative (HER2-) tumors, 15 patients (11%) had HER2 + (ER ±) tumors, and 24 patients (18%) had triple-negative breast cancer (TNBC), with the subtype missing for two patients. Eighty patients (59%) had visceral metastases (defined as lung, liver, brain, peritoneal, and/or pleural involvement).

In order to identify novel immuno-oncology markers that predict survival in MBC patients, we analyzed the levels of 92 serum proteins from 136 patients with newly diagnosed MBC using multiplex proximity extension assay (PEA; the Olink Target 96 Immuno-Oncology panel). Due to the large number of serum proteins measured for each MBC patient, the likelihood for false positive findings can be inflated. To identify the proteins with highest importance for survival analyses, we first used random forest (RF) analyses for an unbiased selection of variables of importance for PFS and OS and eliminated the 25% of proteins with the lowest median importance for PFS and OS. Additional File 1: Fig. S1 shows feature importance after the RF analyses, sorted according to OS feature importance (top) and PFS feature importance (bottom).

Using the top 20 proteins for prediction of survival, identified using RF (Additional File 1: Fig. S1), we next used random survival forest (RSF) (Fig. 2) to further select the proteins of highest relevance. Throughout, we employed Harrell's concordance index (c-index) to quantify RSF model concordance with PFS or OS and considering the right-censoring of the data. For OS, the median training and test C-indices were 0.689 and 0.630, respectively, with Mucin-16 (MUC16) showing a higher overall importance than other proteins, followed by natural killer cell receptor 2B4 (CD244), interleukin (IL) 8, Fas antigen ligand (FasL), and carbonic anhydrase IX (CAIX). The 20 serum proteins with highest overall importance can be seen in Fig. 2 (left). For PFS, the median training and test C-indices were 0.727 and 0.600, respectively, with the top five variables being similar to the top five variables for OS. IL-8 shows the highest overall importance, followed by IL-10, CD244, Adenosine Deaminase (ADA), and MUC16 (Fig. 2 (right)).

To examine how many features were needed to maximize the model concordance, we utilized recursive feature elimination by considering the median feature importance from the RSF, removing the features one by one and parameterizing the models, using cross-validation, and computing the median score for all models. This identified the number of features that lead to the highest median score to be considered. For OS, the top nine variables of highest predictive importance were: MUC16, CD244, IL-8, FasL, CAIX, tumor necrosis factor receptor superfamily member 4 (TNFRSF4), IL-10, IL-6, and macrophage colony-stimulating factor 1 (CSF-1) (Fig. 2, left). The top ten variables for predicting PFS were: IL-8, IL-10, CD244, ADA, MUC16, IL-6, Caspase-8 (CASP8), CSF-1, T-cell surface glycoprotein CD8 alpha chain (CD8A) and monocyte chemotactic protein 2 (MCP2) (Fig. 2, right). Five of the top nine proteins for predicting OS were also found within the top ten proteins for predicting PFS, comprising in total 13 proteins identified as being most relevant for prediction of outcome.

To further confirm the proteins of highest importance, we used penalized Cox regression by employing elastic nets, where variable weight on the model is identified for individual variables. Penalized Cox regression (Additional File 2: Fig. S2) indicated that the top variables for both OS and PFS are similar to those observed using RSF, further supporting the relevance of these proteins. Similarly, the results of penalized Cox regression on un-regressed data were in accordance with the results from the RSF analyses (Additional File 3: Fig. S3 and data not shown). Additional File 4: Table S1 and Additional File 5: Table S2 show how the top 13 proteins selected were ranked for prediction of OS and PFS, respectively, in each model used (RF, Cox regression, RSF, penalized Cox regression and penalized Cox regression on un-regressed data). These results indicate a high degree of concordance between the models used in identifying the serum proteins of highest importance for predicting survival.

To further determine the potential of the identified 13 serum proteins to predict OS or PFS, we next quantified the association (HR, Hazard Ratio) and the evidence for an association (P value) between the serum levels (NPX values) of the proteins of highest importance, as selected using RSF, and survival using unadjusted and adjusted Cox regression analysis with either OS or PFS as end point (Table 2). For OS, all the nine evaluated serum proteins showed suggestive or significant evidence for correlation with survival in both uni- and multivariable analysis, with the exception of CD244 that showed a weak correlation with survival in multivariable analysis (Table 2). A twofold increase in the serum level of FasL was significantly associated with improved OS (UV: HR = 0.58, 95%CI: (0.40–0.83), P = 0.0031; MV: HR = 0.47, 95%CI: (0.30–0.75), P = 0.0013), while a doubling of the levels of CD244 showed evidence of being associated with improved OS (UV: HR = 0.57, 95%CI: (0.35–0.95), P = 0.030). Elevated levels of CAIX, IL-8 and IL-10 all showed significant association with worse OS in both uni- and multivariable analysis (Table 2). Interestingly, CSF-1 displayed a high HR (UV: HR = 11.16, 95% CI: (4.12–30.20); MV: HR = 6.05, 95% CI: (1.59–23.00)) for a twofold increase in uni- and multivariable analysis, respectively, compared to the other serum proteins (HR around 1.0).

For PFS, eight out of ten analyzed top serum proteins showed suggestive or significant evidence for association with PFS in both uni- and multivariable analyses (Table 2). Again, elevated levels of CD244 showed evidence for association with improved PFS (UV: HR = 0.50, 95%CI: (0.31–0.79), P = 0.0028; MV: HR = 0.50, 95%CI (0.27–0.91), P = 0.022). Higher serum levels of ADA, CASP8, IL-8 and IL-10 all showed significant association with worse PFS in both uni- and multivariable analysis (Table 2). In addition, FDR analysis was used to adjust Cox regression P values for multiple testing. P values for all 92 proteins studied were used to calculate so-called q values separately for UV and MV analyses for OS as well as PFS. Corresponding q values for all P values of the top 13 proteins are shown in Table 2, illustrating that IL-8 shows the strongest evidence for correlation to OS and PFS in both UV and MV analyses after FDR correction (OS: UV, q < 0.0001; MV, q = 0.0002; PFS: UV, q = 0.0001; MV, q = 0.0310). Regarding the other identified top proteins CAIX (q = 0.0017), CSF-1 (q = 0.0002), IL-6 (q < 0.0001), MUC16 (q = 0.0001), IL-10, (q = 0.0058) and TFNSFR4 (q = 0.0307) all showed strong evidence for correlation to OS in UV analyses after FDR correction, whereas the evidence for correlation in adjusted MV analysis was weaker (CAIX, q = 0.1399; CSF-1, q = 0.5976; IL-6, q = 0.4126; MUC16 q = 0.9818; IL-10, q = 0.1492; TFNSFR4, q = 0.4062) (Table 2). For correlations to PFS, ADA (q = 0.0067), CASP8 (q = 0.0099), CSF-1 (q = 0.0059) and IL-6 (q = 0.0245) all showed strong evidence for correlation in UV analyses after FDR correction, whereas the evidence for correlation in adjusted MV analyses were weaker (ADA, q = 0.1703; CASP8, q = 0.0786; CSF-1, q = 0.4584; IL-6, q = 0.5677 and IL-10, q = 0.2114).

To investigate the potential correlation of serum proteins with clinicopathological features, the continuous protein levels were dichotomized into high or low levels using the respective median as the cut-off point (Table 3). Eight out of 13 serum proteins showed evidence for correlation with ECOG performance status. High levels of CD244 and FasL showed evidence for correlation with better ECOG performance status whereas high CSF-1, IL-8 and MUC16 all correlated significantly with worse ECOG performance status. Nine out of 13 proteins showed suggestive or significant evidence for correlation with increased number of metastatic sites, with higher levels of IL-6 and MUC16 correlating significantly with > 3 metastatic sites (Table 3). Interestingly, higher levels of ADA, IL-8, IL-10 and MUC16 all correlated significantly with the presence of CTC clusters, with ADA and MUC16 further correlating with high numbers of CTCs (≥ 5), and IL-8 showing evidence for correlating with high levels of CTCs. All correlations to clinicopathological features and serum proteins are summarized in Table 3.