Building the Semantic Foundation of CancerScan
Cancer research increasingly relies on combining information generated by complementary experimental technologies, from molecular profiling to high-resolution tissue imaging. While each modality provides valuable insights into tumour biology, they describe different aspects of the disease and often use distinct formats, terminologies and levels of biological abstraction. Integrating these heterogeneous data into a coherent and biologically meaningful representation therefore requires a shared semantic foundation.
Within CancerScan, this semantic foundation is being built through the careful selection of biomedical ontologies and publicly available reference datasets that complement the experimental data generated by the project partners. These resources provide, respectively, the language and the scientific context needed to construct a knowledge graph capable of representing the complex interactions between tumour cells, the tumour microenvironment, and the molecular and clinical factors that drive disease progression and treatment resistance.
To this end, we have surveyed nearly forty biomedical ontologies covering complementary aspects of cancer research, including genes, genomic variants, biological processes, phenotypes, cell types, anatomy, diseases, drugs, clinical investigations, imaging, and experimental evidence. Particular attention was given to subdomains with increased relevance to the project, such as cellular processes and oncology. Most of these ontologies originate from internationally recognised community initiatives such as the OBO Foundry and BioPortal, ensuring that the knowledge graph is built upon widely adopted, actively maintained, and interoperable standards. This broad semantic coverage allows concepts originating from different experimental modalities to be described using a shared vocabulary.
For example, a gene identified through transcriptomic analysis, the corresponding protein detected by proteomics, and the biological pathway in which it participates can all be represented using common concepts and relationships defined in biomedical ontologies. This facilitates data integration, reproducibility, and future reuse.
Complementing this semantic layer is a broad collection of FAIR biomedical repositories that provide validated reference knowledge across multiple biological scales. More than twenty major data repositories were evaluated, spanning genomics, transcriptomics, proteomics, immunopeptidomics, molecular interactions, biological pathways, and cancer-specific resources. These include widely used resources such as TCGA and cBioPortal for cancer genomics, UniProt for protein knowledge, Reactome and KEGG for biological pathways, STRING for molecular interactions, and ChEMBL for drug and bioactivity information. These repositories represent decades of accumulated biomedical research and provide a robust reference framework against which the diverse experimental data generated by the CancerScan consortium can be interpreted and integrated.
For instance, they may contain previously reported gene expression profiles, protein interaction networks, signalling pathways, or molecular signatures associated with pancreatic cancer and treatment response, providing biological context for newly generated project data. These datasets were selected according to their relevance and complementarity to the data produced within the project, ensuring that public knowledge enriches the project’s own experimental results.
Together, ontologies and reference datasets form the backbone of the CancerScan knowledge graph. In simple terms, ontologies provide the vocabulary and structure needed to describe biomedical entities consistently, whereas reference datasets provide experimentally validated knowledge associated with those entities. Ontologies define what biomedical concepts are and how they relate to one another, for example, that KRAS is a gene, pancreatic ductal adenocarcinoma is a disease, or apoptosis is a biological process. Reference datasets, on the other hand, contribute experimentally validated observations, such as genes differentially expressed in pancreatic cancer, proteins participating in signalling pathways, or biomarkers associated with therapeutic response.
Beyond supporting data integration, this semantic infrastructure establishes the conditions for trustworthy artificial intelligence. Machine learning models perform best when data are consistently described, interoperable, and contextualised. Thus, the combination of standardised ontologies, curated reference datasets, and project-generated experimental data provides a robust and explainable foundation for downstream analyses, predictive modelling, and hypothesis generation.
As new experimental data become available throughout the project, they will not exist in isolation but will immediately benefit from the accumulated knowledge already represented within the knowledge graph, creating a continuously evolving resource that grows in value over time.
Keywords
cancer research, biomedical ontologies, knowledge graph, data integration, FAIR reference datasets, artificial intelligence
