Abstract 1979: Algebraic factorization of gene expression profiles reveals subtype-specific drug sensitivities among 54 breast cancer cell lines

Breast cancer is a heterogeneous disease, with reproducible and prognostically important subclasses. Breast cancer cell lines are widely used to study preclinical investigational agents, but the relationships, if any, between breast cancer subclass and drug response are not well understood. To help bridge this gap, we have profiled drug responses across a large panel of well-annotated breast cancer cell lines. In order to translate in vitro drug responses into clinically useful predictions, it is important that the cell line panel be organized into subtypes representative of the cancer subtype diversity found in the clinic. Recent studies have demonstrated that an algebraic clustering method known as “non-negative matrix factorization” (NMF) can be applied to gene expression profiles to resolve clinically meaningful cancer subtypes in greater detail than achievable using other clustering methods. To test whether NMF improves our ability to resolve drug sensitivities in breast cancer cell lines, we clustered pretreatment gene expression profiles of 54 breast cancer cell lines using two different methods: (1) NMF-based consensus clustering and (2) hierarchical consensus clustering. Using NMF-based consensus, we identified five robust subtypes (two Basal-A, one Basal-B and two Luminal classes). In contrast, using hierarchical consensus clustering, we could identify only three robust subtypes (one Basal-A, one Basal-B and one Luminal class). The drug response profiles were then segregated by subtype to determine whether either of the two clustering methods improves our ability to resolve drug sensitivities. Of the 67 drug compounds included in our study, the 5-subtype NMF-based classification scheme revealed three compounds (AG1024, CPT-11 and topotecan) exhibiting subtype-specific drug effects (p