On the translational potential of atlases in precision oncology

The proliferation of publicly available imaging datasets, combined with widespread access to computational power, has boosted research in neuroscience, neurobiology, and systems biology while inspiring projects centered on building atlases. Atlases are methodological tools that provide global overviews of detailed thematic information, are usually organized in grids with assigned resolution to facilitate inference, and offer a reliable knowledge base from multimodal evidence data. Two critical goals have been generally addressed through atlases: (I) warehousing baseline information of normal anatomical structures under physiological (i.e., non-pathological) conditions; and (II) establishing a reference for typical anatomy across demographic groups by aggregating high-resolution imaging data that cover extensively diverse populations. Compared to more traditional atlases often referring to homogeneous groups of populations, the recent atlas developments have leveraged data multimodality and utilized machine learning (ML) and artificial intelligence (AI) tools for inference purposes. Together with the possibility of representing normal variation within specific demographic cohorts and gaining usability and reliability in clinical applications, data multimodality is particularly impactful in precision oncology and personalized therapy. This review discusses the translational potential of atlases in cancer studies through their property of integrating multiple types of cancer data and inspiring predictive learning algorithms that account for the correlations between anatomical and imaging features with genetic and omics markers.