A new unsupervised clustering algorithm applied to genome-wide profiles of breast cancers in The Cancer Genome Atlas proper subsets triple-negative samples.

Abstract

e23195 Background: Clustering algorithms have identified subtypes of major cancers from analysis of genome-wide gene expression (GE) and somatic mutation (SM) profiles. These algorithms almost never discover a proper subset cluster, a recovered cluster that includes all the samples of a specific subtype. For breast cancer (BC), clustering of genome-wide profiles has been unable to proper subset triple negatives (TNs), TN subtypes, or other major subtypes. Methods: To search for a proper subset cluster for TNs, we applied a new clustering algorithm to the public domain GE and SM data of BC samples in The Cancer Genome Atlas (TCGA). A module of Medidata’s Clinical Trial Genomics (CTG) platform for automated clinical and genomic data integration and analysis, it uses a hierarchical component with tree learned cut points applied to a principal component dimension reduced similarity matrix calculated from a genome-wide data profile. Results: Our analysis of 540 TCGA BC samples run without human supervision produced a proper subset cluster that included all 55 TN samples and only 74 non-TN samples. GE data have previously indicated TN status, but this is the first demonstration that these TCGA BC data contain enough information to proper subset TNs, implying that this broad BC subtype has a strong, quantifiable impact on GE. We show that the genome-wide SMs of TCGA BC samples can be used to proper subset 4 novel subtypes distinguished as classes “TP53 mutated”, “PIK3CA mutated”, “both TP53 and PIK3CA mutated”, and “neither mutated”, signifying an important role for these known driver mutations in producing the subtypes’ genome-wide mutation profiles. We find that most ( > 80%) TN BCs are in “TP53 mutated” but only 1 TN sample ( < 2%) is in “PIK3CA mutated”, indicating distinct biology for these TNs with potential implications for TN therapy. Conclusions: CTG clustering achieves proper subset cancer subtype clustering of TCGA BC samples. These results illustrate the therapeutic discovery potential possible from genomic data of the high quality present in TCGA if combined with detailed clinical data with the Medidata CTG integration and annotation platform.