An approach for deciphering patient-specific variations with application to breast cancer molecular expression profiles.

Several studies have successfully used molecular expression profiling in conjunction with classification techniques for discerning distinct disease groups. However, a majority of these studies do not provide sufficient insights into potential patient-specific variations within the disease groups. Such variations are ubiquitous and manifests across multiple scales with varying resolution. There is an urgent need for novel approaches that falls within the objective of precision medicine and provide novel insights into patient-specific variations and sub-populations within disease groups while discerning the disease groups of interest so as to enable timely and targeted intervention of select subjects. This study presents a selective-voting ensemble classification approach (SVA) for discerning good and poor-prognosis breast cancer samples from their 70-gene molecular expression profile revealing patient-specific variations within the poor-prognosis group. In contrast to traditional classification, SVA adapts the feature sets in a sample-specific manner capturing the proclivity of the samples to each of the disease groups. Correlation between normalized vote counts from SVA and clinical outcomes of the subjects is elucidated. Performance of Support Vector Machine and Naïve Bayes classifier is investigated within the SVA framework and compared to established clinical criteria (Nottingham Prognostic Index, Adjuvant Online, St. Gallen) and Mammaprint approach. Weighted undirected graph abstractions of the ensemble sets of the poor-prognosis test samples is also shown to exhibit markedly different topologies with varying proclivities. These patient-specific networks may reflect inherent variations in underlying signaling mechanisms in the poor-prognosis subjects and reveal potential targets for personalized therapeutic intervention.