This study was conducted in the Mammary Carcinoma Risk Factor Investigation (MARIE) breast cancer patient cohort, which has been described in detail previously [20]. In brief, 3813 women with a breast cancer diagnosis were initially enrolled between August 2002 and September 2005 in two regions (Rhine-Neckar-Karlsruhe and Hamburg) in Germany. Participants were aged 50–74 years at baseline and had a histologically confirmed primarily invasive or in situ carcinoma diagnosis. At recruitment, participants completed an in-person interview and provided data on lifestyle, health, and anthropometric characteristics and were requested to provide a blood sample. Follow-up interviews were conducted in 2009 (first follow-up) and in 2015 (second follow-up). In the present study, participants with in situ breast cancers, metastasis at diagnosis/stage IV breast cancer, previous tumors other than breast cancer, unknown stage, or missing hormone receptor status were excluded. A total of 2282 participants with available blood sample were included in this study [18]. Main study characteristics such as age and BMI were largely similar between cases of the overall cohort [20] and the study population investigated here; stage at diagnosis differed by design excluding women with metastasis at diagnosis or in situ breast cancer who were part of the original study sample. Eight participants with blood collection before breast cancer diagnosis due to original recruitment as “control” were then re-classification as “case” (median time between blood collection and diagnosis: 5.3 months (range 1.6 months to 14.7 months)); exclusion of these participants (0.4% of the study sample) did not impact the observed associations and therefore these cases were retained in the study. The median time between diagnosis and blood collection for the full study sample was 3.7 months (range − 14.7 months to 57.6 months). Participants of the MARIE study were asked about their last natural menstrual period at baseline and were defined as postmenopausal if the reference date was at least one year after their last natural menstrual period or if they had bilateral oophorectomy or had cessation of menstrual period because of radiation or chemotherapy for a disease other than breast cancer. Participants older than 55 years with unclear menopausal status because of hysterectomy or menopausal hormone therapy were also defined as postmenopausal in the MARIE study [20]. The use of exogenous selective estrogen receptor modulators (SERMs) such as tamoxifen and aromatase inhibitor (AI) was assessed via questionnaires at both follow-up times for the preceding time interval. If self-reported information on endocrine therapy was not available, information derived from medical records was used. A description of the sample selection of this study as well as detailed information on endocrine therapy use has been reported previously [18].

Follow-up information was obtained from participants by telephone interview at each follow-up, and recurrence information was confirmed through medical records or contact with treating physicians. Vital status was obtained from population registries of the study regions up to the end of the second follow-up period in 2015, and copies of the death certificates were obtained from local health offices. Causes of death were coded according to the 10th revision of the International Classification of Diseases (ICD-10-WHO).

The outcomes for this study included all-cause mortality, cause-specific mortality (breast cancer death, other cancer death, cardiovascular disease death, other cause of death), and risk of breast cancer recurrence. “All-cause mortality” was attributed to death by any cause, “breast cancer (BC)-specific mortality” was attributed to breast cancer deaths (ICD-10-C50), and “recurrence” was defined using the definition for recurrence-free interval as described in the Standardized Definitions for Efficacy End Points (STEEP) criteria [21] and includes local/regional invasive breast cancer recurrence, metastasis, contralateral disease of the breast, and deaths due to breast cancer. This study includes 438 all-cause deaths, 237 BC-specific deaths, and 376 recurrences. The outcome “other cancer death” was assigned to all cancer deaths (ICD-10-C) other than breast cancer including cancer of digestive organs (n = 30), respiratory organs (n = 23), genital organs (n = 9), and other cancers (n = 26, n ≤ 5 for any other individual cancer); “cardiovascular deaths” was assigned to all deaths of the cardiovascular system (ICD-10-I) including deaths due to acute myocardial infarction (n=13), heart failure (n=10), chronic ischemic heart disease (n = 8), pulmonary embolism (n = 6), and other (n = 24; n≤6 for any other individual cardiovascular death); “other cause of death” includes all of the deaths not attributed to any of the other outcomes including death due to the respiratory system (ICD-10-J, n = 16), digestive system (ICD-10-K, n = 9), infectious and parasitic diseases (ICD-10-A/B, n = 8), nervous system (ICD-10-G, n = 6), and other (n = 13; n ≤ 6 for any other individual cause of death).

The following oxysterols were considered for inclusion in this study: 22R-hydroxycholesterol (22R-HC), 24S-hydroxycholesterol (24S-HC), 5α,6α-epoxycholesterol (5a,6a-EC), 5β,6β-epoxycholesterol (5b,6b-EC), 7α-hydroxycholesterol (7a-HC), 7β-hydroxycholesterol (7b-HC), 7-dehydrocholesterol (7-DC), 7-ketocholesterol (7-KC), desmosterol (desmos), lanosterol (lan), 24,25-dihydrolanosterol (24-DHLan), 24,25-epoxycholesterol (24,25-EC), 5α,6β-dihydroxycholestanol (THC). Oxysterols 27-hydroxycholesterol (27-HC; systematic name (25R),26-hydroxycholesterol) and 25-hydroxycholesterol (25-HC), and estradiol have been investigated in a previous study [18] and were evaluated as potential covariates in the current study. Oxysterol levels were measured by biocrates life sciences ag (Innsbruck, Austria) using UHPLC-MS/MS with multiple reaction monitoring (MRM) in positive mode using a mass spectrometer with electrospray ionization (ESI). Inter-assay coefficients of variation (CV) were determined by including 16 blinded replicate quality controls. Analyte concentrations and coefficients of variation (CV) are displayed in the additional file (Additional file 1, Table S1). In brief, mean intra-assay CVs were below 20% (Additional file 1, Table S1), and mean inter-assay CVs were below 25%, with the exception of 5a6a-EC (31.8%), 24-DHLan (33.7%), 7-KC (42.4%), 7b-HC (74.9%). 7b-HC was excluded from the survival analyses due to the high CV (74.9%). Of note, CVs were highest for oxysterols with low concentrations (median concentrations of oxysterols with CV > 30% are between 4.8 and 34.8 nM, whereas median concentrations of oxysterols with CV ≤ 30% are between 47.2 nM and 2705.8 nM), and median concentrations of quality controls with CV > 30% were lower in the quality control samples than median concentrations measured in the study samples (e.g., median, 7b-HC, QC samples: 5.8 nM, participant samples: 202.3 nM; 7-KC, QC samples: 34.8 nM, participant samples: 184.9 nM). Estradiol concentrations were measured using an ELISA in the Division of Cancer Epidemiology at the German Cancer Research Center (DKFZ) (inter-batch CV of 16.2%).

In the following, the term “oxysterol” is attributed to cholesterol metabolites (24S-HC, 7a-HC, 7b-HC, 7-KC, THC, 5a6a-EC, 5b6b-EC, 22R-HC, 27-HC, 25-HC) as well as precursors of cholesterol (lanosterol, 24-DHLan, desmosterol, 7-DC, 24,25-EC).

We applied a log2-transformation to all biomarker concentrations to obtain approximately normal biomarker distributions and to estimate the effect of a doubling in biomarker concentrations.

Values below the limit of detection (LOD) were detected for the following biomarkers: 24S-HC, n = 7; 7b-HC, n = 14; desmosterol, n = 11; 7-DC, n = 150; THC, n = 492; 25-HC, n = 89; estradiol, n=12 (Additional file 1, Table S1). We imputed these values with the midpoint between 0 and the lowest detectable value stratified by the study region. 22R-HC and 24,25-EC were excluded from all analyses due to the high number of values below LOD (> 88%). Values exceeding the calibration range were detected for the following biomarkers: 24S-HC, n = 2; 5a6a-EC, n = 1; 5b6b-EC, n = 4; 7a-HC, n = 11; 7b-HC, n = 10; desmosterol, n = 1; 7-DC, n = 61 (Additional file 1, Table S1); these values were excluded from all analyses. After imputing values below LOD and excluding values exceeding the calibration range, we evaluated outlying values using the Generalized ESD Many-Outlier Procedure [22]. Overall, ≤ 30 outliers were detected for any biomarker, and analyses were conducted with and without inclusion of the outliers.

We calculated Spearman partial correlations for circulating oxysterols adjusted for age at diagnosis and study region. To account for the left-truncated survival data, delayed-entry Cox proportional hazards regression was used to estimate hazard ratios (HR) and 95% confidence intervals (CI) for risk of all-cause mortality. HRs were calculated using continuous log2-transformed oxysterol values. We assessed associations between each individual oxysterol and cause-specific mortality (BC-specific death, other cancer death, cardiovascular death, and death due to other causes) and risk of recurrence (recurrence vs. non-breast cancer cause of death) using the competing risks approach described by Lunn and McNeil [23].

Non-parametric restricted cubic splines were used to examine possible non-linearity, comparing models with linear and cubic terms to models with only the linear term [24]. We did not observe significant deviation from linearity. The start date for follow-up time (time-to-event) was the date of diagnosis and the start date for participants at risk (time-at-risk) was the date of blood collection. End of follow-up time was the date of the defined outcome, the date of last contact, or the end of follow-up 2 (June 2015), whichever came first. Associations were evaluated for all participants and by hormone receptor status (ER/progesterone receptor (PR)-positive and ER/PR-negative). In the ER/PR-negative subset, only the outcomes all-cause death, BC-specific death, and recurrences were evaluated due to limited sample size (n = 320).

Covariates for all of the models were selected a priori and included: age at diagnosis, body mass index (BMI), smoking status (never, former, current smoker), alcohol consumption (g/day), Charlson Comorbidity Index (CCI), and tumor size (≤ 2 cm, 2–5 cm, > 5 cm), nodal status (0, 1–3, >3), and tumor grade (low, moderate, high). All models were stratified by ER/PR-status (ER+/PR+, ER+/PR− or ER−/PR+, ER−/PR−), and study region (Hamburg, Rhine-Neckar-Karlsruhe). The biomarkers 27-HC, 25-HC, and estradiol, as well as physical activity, parity, and educational status were considered as further potential covariates; however, adding these covariates to the model minimally changed the results (< 10 %) and they were excluded from the final models. Of the evaluated oxysterols, 7-DC was strongly associated with endocrine therapy (endocrine therapy vs. no endocrine therapy, 16% difference) and THC was strongly associated with endocrine therapy and chemotherapy (endocrine therapy vs. no endocrine therapy, -18% difference; chemotherapy vs. no chemotherapy, 29% difference). In further analyses, we adjusted associations with these oxysterols additionally for chemotherapy or endocrine therapy. Controlling for smoking status using a continuous variable (duration of smoking in years or average pack years of smoking) resulted in similar results as use of categorical variables smoking status (never, former, current), which we used due to completeness of data. Study participants with missing information for any of the covariates were excluded (n participants = 17). The proportional hazards assumption was tested using Schoenfeld residuals; we did not observe violation of the proportional hazards assumption.

In an exploratory analysis, we applied a regularized Cox proportional hazard regression model using the elastic net penalty, a combination between LASSO and Ridge regression [25], controlling for the covariates listed above with the aim to select variables strongest associated with the outcomes. Hyper-parameter tuning was performed to identify the best α (elastic net mixing parameter) and λ (shrinkage parameter) applying 5-fold cross-validation and optimizing the C-index. In a sensitivity analysis, we conducted analyses including biomarkers with low CV (< 30%) only. Associations between the selected oxysterols and the outcomes were evaluated in competing risks models. In further sensitivity analyses, we conducted stratified 5-year survival analyses. Finally, we evaluated the effect of adjusting for multiple comparisons (60 comparisons in our main analysis for 6 outcomes and 10 oxysterols).

All statistical tests were two-tailed and considered significant at p < 0.05. Statistical analyses were conducted using SAS 9.4 (SAS Institute Inc., Cary, NC, USA) and R (version 4.2.1).