Breast cancer gene expression datasets do not reflect the disease at the population level

Publicly available tumor gene expression datasets are widely reanalyzed, but it is unclear how representative they are of clinical populations. Estimations of molecular subtype classification and prognostic gene signatures were calculated for 16,130 patients from 70 breast cancer datasets. Collated patient demographics and clinical characteristics were sparse for many studies. Considerable variations were observed in dataset size, patient/tumor characteristics, and molecular composition. Results were compared with Surveillance, Epidemiology, and End Results Program (SEER) figures. The proportion of basal subtype tumors ranged from 4 to 59%. Date of diagnosis ranged from 1977 to 2013, originating from 20 countries across five continents although European ancestry dominated. Publicly available breast cancer gene expression datasets are a great resource, but caution is required as they tend to be enriched for high grade, ER-negative tumors from European-ancestry patients. These results emphasize the need to derive more representative and annotated molecular datasets from diverse populations.