Biomarkers for early detection of ovarian tumors are urgently needed. Tumors of the ovary grow within cysts and most are benign. Surgical sampling is the only way to ensure accurate diagnosis, but often leads to morbidity and loss of female hormones. The present study explored the deep proteome in well-defined sets of ovarian tumors, FIGO stage I, Type 1 (low-grade serous, mucinous, endometrioid; n = 9), Type 2 (high-grade serous; n = 9), and benign serous (n = 9) using TMT–LC–MS/MS. Data are available via ProteomeXchange with identifier PXD010939. We evaluated new bioinformatics tools in the discovery phase. This innovative selection process involved different normalizations, a combination of univariate statistics, and logistic model tree and naive Bayes tree classifiers. We identified 142 proteins by this combined approach. One biomarker panel and nine individual proteins were verified in cyst fluid and serum: transaldolase-1, fructose-bisphosphate aldolase A (ALDOA), transketolase, ceruloplasmin, mesothelin, clusterin, tenascin-XB, laminin subunit gamma-1, and mucin-16. Six of the proteins were found significant (p < .05) in cyst fluid while ALDOA was the only protein significant in serum. The biomarker panel achieved ROC AUC 0.96 and 0.57 respectively. We conclude that classification algorithms complement traditional statistical methods by selecting combinations that may be missed by standard univariate tests. SignificanceIn the discovery phase, we performed deep proteome analyses of well-defined histology subgroups of ovarian tumor cyst fluids, highly specified for stage and type (histology and grade). We present an original approach to selecting candidate biomarkers combining several normalization strategies, univariate statistics, and machine learning algorithms. The results from validation of selected proteins strengthen our prior proteomic and genomic data suggesting that cyst fluids are better than sera in early stage ovarian cancer diagnostics.
Read full abstract