Conclusions
Despite showing original promise, when this protein signature is applied to a large external cohort, it is unable to accurately predict surgical outcomes. This could be attributed to overfitting of the original model, or differences in surgical practice between cohorts.
Methods
238 high-grade epithelial ovarian cancer samples were collected from patients who participated in a large multicentre trial (ICON5). Samples were collected at the time of initial surgery and before randomisation. Surgical outcome data were collated from prospectively collected study records. Immunohistochemical scores were generated by two independent observers for the three proteins in the original signature (POSTN, CXCL14 and pSmad2/3). Predictive values were generated for individual and combination protein signatures.
Results
When assessed individually, none of the proteins showed any evidence of predictive affinity for suboptimal surgical outcome in our cohort (AUC POSTN 0.55, pSmad 2/3 0.53, CXCL 14 0.62). The combined signature again showed poor predictive ability with an AUC 0.58. Conclusions: Despite showing original promise, when this protein signature is applied to a large external cohort, it is unable to accurately predict surgical outcomes. This could be attributed to overfitting of the original model, or differences in surgical practice between cohorts.