Improving the Uncertainty Estimation in PAM50: Impact on Subtype Assignment and ROR.

CRISTOBAL FRESNO; GERMAN GONZALEZ; GABRIELA MERINO; ANA GEORGINA FLESIA; OSVALDO PODHAJCER; ANDREA LLERA; ELMER FERNÁNDEZ

Conferencia; 4th International Society for Computational Biology-Latin America Bioinformatics Conference (ISCB-LA); 2016

Background: PAM50 is currently the benchmark for the molecular classification of breast cancer. The algorithm consists of a five centroid-based Spearman?s correlation classifier, which analyzes the expression levels of 50 genes. Patients are assigned to the highest correlation subtype irrespectively of the obtained correlation. Then, PAM50 classification is used to build the risk of recurrence (ROR), currently used in therapeutic decisions. The aim of this work was to develop a population-independent uncertainty control methodology to enhance the PAM50 subtype assignment.Methods: Thirty-three public datasets (n=5,228) were analyzed using PAM50. The subject uncertainty assessment was performed by an estimation of gene-based permutations over the five subtype correlations. According to the simulated p-values, the subjects were then labeled as ?Assigned? (A), ?Ambiguous? (Amb) or ?Not Assigned? (NA). We compared our assignments to three alternative PAM50 implementations that use different uncertainty measurements. Survival discrimination and ROR assignment among the Aand NA subjects were assessed.Results: Only 61% of the subjects could be reliably assigned to a PAM50 subtype using the proposed strategy. In contrast, 33% of individuals were NA due to an observed median correlation