Search

Background Despite recent progress in liquid biopsy technologies, early blood-based detection of breast cancer is still a challenge. Methods We analyzed secretion of the protein cellular communication network factor 1 (CCN1, formerly cysteine-rich angiogenic inducer 61) in breast cancer cell lines by an enzyme-linked immunosorbent assay (ELISA). Soluble CCN1 in the plasma (2.5 µL) of 544 patients with breast cancer and 427 healthy controls was analyzed by ELISA. The breast cancer samples were acquired at the time of primary diagnosis prior to neoadjuvant therapy or surgery. A classifier was established on a training cohort of patients with breast cancer and age-adapted healthy controls and further validated on an independent cohort comprising breast cancer patients and healthy controls. Samples from patients with benign breast diseases were investigated as additional controls. Samples from patients with acute heart diseases (n = 127) were investigated as noncancer controls. The diagnostic accuracy was determined by receiver operating characteristic using the parameters area under the curve, sensitivity, and specificity. Results CCN1 was frequently secreted by breast cancer cell lines into the extracellular space. Subsequent analysis of clinical blood samples from patients with breast cancer and age-adjusted healthy controls revealed an overall specificity of 99.0% and sensitivity of 80.0% for cancer detection. Remarkably, 81.5% of small T1 cancers were already CCN1-positive, while CCN1 concentrations in patients with benign breast lesions were below the threshold for breast cancer detection. Conclusions Circulating CCN1 is a potentially novel blood biomarker for the detection of breast cancer at the earliest invasive stage.

Investigating the proteome can add a significant layer of information to manifold existing methylation, mutation, and transcriptome data on brain tumors as proteins represent the pharmacologically addressable phenotype of a disease. Small cohorts limit the usability and validity of statistical methods, and variable technical setups and high numbers of missing values make data integration from public sources challenging. Using a newly developed framework being able to reduce batch effects without the need for data reduction or missing value imputation, we show –based on in-house and publicly available datasets- successful integration of proteomic data across different tissue types, quantification platforms, and technical setups. Exemplarily, data of a Sonic hedgehog (Shh) medulloblastoma mouse model were analyzed, showing efficient data integration independent of tissue preservation strategy or batch. We further integrated batches of publicly available data of human brain tumors, confirming proposed proteomic cancer subtypes correlating with clinical features. We show that, missing value tolerant reduction of technical variances may be helpful to identify biomarkers, proteomic signatures, and altered pathways characteristic for molecular brain cancer subtypes.

Dataset integration is common practice to overcome limitations, e.g., in statistically underpowered omics datasets. This is of particular importance when analyzing rare tumor entities. However, combining datasets leads to the introduction of biases, so called 'batch effects', which are due to differences in quantification techniques, laboratory equipment or used tissue type. A common problem is the missing quantification for features like gene transcripts or proteins within a dataset. These missing values can appear at random in a given dataset and also get introduced by combination of multiple datasets. Currently, strategies beyond common normalization for batch effect reduction are either missing entirely or are unable to handle absence of data points and therefore rely on error-prone data imputation. We introduce a framework that enables batch effect adjustments for combined datasets while avoiding data loss by appropriately handling missing values without imputation. The underlying idea is based on a matrix dissection approach, adjusting common information from the integrated dataset under guarantee of sufficient data presence. The strategy is implemented within the R environment and linked with popular software stacks that are built on top of R. Successful data adjustment is exemplarily shown for proteomic data generated by different quantification approaches and LC-MS/MS instrumentation setups.