Machine learning developed a fibroblast-related signature for predicting clinical outcome and drug sensitivity in ovarian cancer

Abstract

Ovarian cancer (OC) is the leading cause of gynecological cancer death. Cancer-associated fibroblasts (CAF) is involved in wound healing and inflammatory processes, tumor occurrence and progression, and chemotherapy resistance in OC. GSE184880 dataset was used to identify CAF-related genes in OC. CAF-related signature (CRS) was constructed using integrative 10 machine learning methods with the datasets from the Cancer Genome Atlas, GSE14764, GSE26193, GSE26712, GSE63885, and GSE140082. The performance of CRS in predicting immunotherapy benefits was verified using 3 immunotherapy datasets (GSE91061, GSE78220, and IMvigor210) and several immune calculating scores. The Lasso + StepCox[forward] method-based predicting model having a highest average C index of 0.69 was referred as the optimal CRS and it had a stable and powerful performance in predicting clinical outcome of OC patients, with the 1-, 3-, and 5-year area under curves were 0.699, 0.708, and 0.767 in the Cancer Genome Atlas cohort. The C index of CRS was higher than that of tumor grade, clinical stage, and many developed signatures. Low CRS score demonstrated lower tumor immune dysfunction and exclusion score, lower immune escape score, higher PD1&CTLA4 immunophenoscore, higher tumor mutation burden score, higher response rate and better prognosis in OC, suggesting a better immunotherapy response. OC patients with low CRS score had a lower half maximal inhibitory concentration value of some drugs (Gemcitabine, Tamoxifen, and Nilotinib, etc) and lower score of some cancer-related hallmarks (Notch signaling, hypoxia, and glycolysis, etc). The current study developed an optimal CRS in OC, which acted as an indicator for the prognosis, stratifying risk and guiding treatment for OC patients.