Abstract
BACKGROUND: Cancer screening tests are considered pivotal for early diagnosis and survival. However, the efficacy of these tests for improving survival has recently been questioned. This study aims to test if cancer screening could be improved by biomarkers in peripheral blood based on multi-omics data. METHODS: We utilize multi-omics data from 500,000 participants in the UK Biobank. Machine learning is applied to search for proteins, metabolites, genetic variants, or clinical variables to diagnose cancers collectively and individually. RESULTS: Here we show that the overall performance of the potential blood biomarkers do not outperform clinical variables for collective diagnosis. However, we observe promising results for individual cancers in close proximity to peripheral blood, with an Area Under the Curve (AUC) greater than 0.8. CONCLUSIONS: Our findings suggest that the identification of blood biomarkers for cancer might be complicated by variable overlap between molecular changes in tumor tissues and peripheral blood. This explanation is supported by local proteomics analyses of different tumors, which all show high AUCs, greater than 0.9. Thus, multi-omics biomarkers for the diagnosis of individual cancers may potentially be effective, but not for groups of cancers.