Identifying a protein biomarker in blood applying an unbiased data-driven approach

The search for protein biomarkers in blood often involves a pre-selection of potential candidates based on literature, followed by an experimental validation of these proteins in a given experimental or clinical context. The success rate of this approach is often low because a) the upfront choice of candidates is biased towards previously studied proteins, b) these proteins are often not specific enough, and c) the number of screened proteins is limited. By contrast, a huge number of gene expression experiments have been published, providing a valuable information source for a huge variety of experimental contexts. However, it is crucial to curate and structure the data prior to integrating it into a searchable platform and thus being able to efficiently use it. The GENEVESTIGATOR platform turns out to be an extremely powerful tools to support our search for proteins strongly over-expressed in cancers and secreted into the bloodstream. We first identified genes strongly and specifically expressed in chosen cancer types (as compared to normal tissues and over 1,000 other cancer types). The search was performed across 40'000 curated Affymetrix expression arrays covering a wide variety of well-described experiments. It resulted in sets of up to 20 marker genes being highly specific for a particular cancer type. The corresponding proteins were then filtered by molecular properties, resulting in a few individual testable candidates expected to be abundant, cancer specific, and secreted to the blood. The method was successfully applied to a chosen cancer type as demonstrated by experimental validation. In fact, from four candidate proteins tested by ELISA assays in protein extracts from blood samples of cancer patients, one protein was well detected and showed significant discriminative power. The method could similarly also be used to identify drug targets after applying protein filters for druggability, cellular localization or molecular function. The power of this approach is the very large number of experimental conditions that can be simultaneously screened to find genes with an extremely specific profile, combined with protein property data. As opposed to literature search, it offers an unbiased, data-driven and global context-based approach for selecting testable candidates.