On the Bayesian Derivation of a Treatment-based Cancer Ontology

Traditional cancer classifications are primarily based on anatomical locations. As knowledge is heavily compartmentalized in the onco-logical specialties, discovering new targets for existing drugs (drug inference) can take years. Furthermore, our lack of understanding of the mechanisms underlying drug efficacy sometimes undercuts the effectiveness of genetic approaches to drug inference. This study tack-les the twin problems of cancer reclassification and drug inference by constructing a global cancer ontology inductively from treatment regimens. A topological abstraction algorithm was performed on the bipartite graph of drugs and cancers to highlight important edges, and a Bayesian algorithm was then applied to determine a new treatment-based classification of cancer, producing 6 highly significant clusters (p < 0.05), confirmed by Fisher’s exact test and enrichment analyses. Edge probabilities derived from its drug inference routine matched real edge frequencies (R2 ≈ 0.96). Drug inference results were reinforced by the identification of relevant published Phase II and III clinical trials, and the drug inference routine differentiated between high- and low-likelihood targets (p < 0.05). This novel treatment-based ontology has the potential to reorganize cancer research and provide powerful tools for drug inference using global patterns of drug efficacy.