Breast cancer molecular subtypes with uncertainty assestment ina pan cancer context

GONZALEZ MONTORO, ALDANA; PRATO, LAURA; ROCHA, DARIO; FERNÁNDEZ, ELMER ANDRÉS

Mendoza

Congreso; A2B2C 10th Meeting; 2019

A2B2C

Although molecular classification of cancer has traditionally been developed and applied in a tissue-based manner, a pan-cancer approach that defines distinct pathological states (subtypes) irrespectiveof their tissue of origin will improve our ability to understand and treat cancer. To search for pan-cancer subtypes, cancer-specific molecular signatures may be useful as long as classification errors arecontrolled. Recent studies have used PAM50 (a breast cancer-specific molecular signature) to identifybreast cancer subtypes on different cancer types, although without considering potential classificationerrors. Our objective was to evaluate the effect of uncertainty assessment on the use of PAM50 inmultiple cancer types, regarding classification errors.The PAM50 classifier was used on TCGA transcriptomic data to classify primary tumor samples from33 cancer types into breast cancer subtypes: Basal, Her2, Luminal A, Luminal B and Normal-like.Uncertainty assessment categorized samples as ?assigned? or ?unassigned? according to classificationconfidence. The relevance of PAM50 subtyping was assessed by estimating classification errors andcharacterizing PAM50 subtypes across all cancer types.RESULTSThe percentage of classification errors was lower in the assigned group (median 10%) compared tothe unassigned one (median 43%). PAM50 subtypes showed clear molecular differences in theassigned group, but not in the unassigned group. Known poor prognosis markers like TIMELESS, PCNAand MCM family members were overexpressed in Basal subtype across cancer types. PGAP3, STARD3and other genes belonging to ERBB2 amplicon were overexpressed in Her2 PAM50 subtype in multiplecancer types. In the assigned group, Basal subtype prognosis was poor compared to the othersubtypes (Hazard ratios = 0.66, 0.5, 0.54 and 0.58 for Her2, Luminal A, Luminal B and Normal-like,respectively).CONCLUSIONUncertainty assessment revealed that a high proportion of non-breast cancer samples may bemisclassified by PAM50. Analyzing only confidently classified samples revealed relevant molecularpatterns in multiple cancer types. This approach opens the door to study the biological and clinicalrelevance of any cancer-specific molecular signature across cancer types. Moreover, it suggests thattherapeutic approaches previously reserved for particular cancer types could be extended to othercarcinomas with the help of uncertainty assessment.